AIDD

Teaching AI to Speak Chemistry, See the Big Picture, and Play by the Rules

title: “Teaching AI to Speak Chemistry, See the Big Picture, and Play by the Rules” description: | The FARM model forces AI to learn the fundamental language of…
Jixing Liu, Ph.D.

AI in Drug Discovery: Three Trends in Synthetic Data, Reinforcement Learning, and Automated Peer Review

Explore the latest advances in AI for drug development. This article breaks down how affinity models trained on rigorously screened AI-predicted structures can perform as…
Jixing Liu, Ph.D.

AI in Drug Discovery: From SciGPT Beating GPT-4 to Trustworthy Drug Repurposing

This article explores three AI technologies for drug discovery and computational chemistry. SciGPT is a language model designed for scientific literature that outperforms…
Jixing Liu, Ph.D.

Machine Learning Potential Energy Surfaces: From Alchemy to Trustworthy Molecular Dynamics

Here’s the rewritten text.
Jixing Liu, Ph.D.
This is a line graph comparing the information retention of protein language models. The x-axis represents protein length, and the y-axis represents information retention. Among the several curves, one labeled 'SCG-Lang' remains nearly a horizontal line as protein length increases, showing extremely high information retention. In contrast, another curve labeled 'Fine-grained' drops sharply as protein length increases. The graph visually demonstrates the superiority of the SCG coarse-graining method discussed in the article: it effectively solves the 'amnesia' problem where traditional models lose information on long sequences, ensuring a complete understanding of the entire protein sequence.

Can We Cure AI’s ‘Amnesia’ with Large Proteins?

Protein Language Models
Coarse-Grained Modeling
Secondary Structure
Large Proteins
Computational Drug Discovery

Standard language models often develop “amnesia” when processing extra-large proteins like Titin, losing critical information because the sequence is too long. The SCG framework offers a new solution. Instead of reading amino acid “letters” one by one, it models a protein’s secondary structures (like alpha-helices) as “words.” This structure-aware, coarse-graining method solves the semantic truncation problem, allowing AI to understand and analyze the function of massive proteins in their entirety. The technique opens up new ways to study the giant molecular targets involved in complex diseases and to design drugs for them.

Tuesday, the 13th of January, 2026
Jixing Liu, Ph.D.
A conceptual diagram illustrating protein design with the SeedProteo model. At the center is a clear, complex 3D model of a polypeptide chain shown in blue and purple ribbons. Surrounding it are several translucent, blurry outlines of protein structures. This represents the process of a diffusion model generating a final, precise structure from random noise. The image visualizes the core concept of SeedProteo: creating specific, functional molecules from a vast space of possibilities.

SeedProteo: Reverse-Engineering AlphaFold3 to Design New Proteins

Protein Design
Diffusion Models
SeedProteo
AlphaFold3
Drug Discovery

The SeedProteo model applies AlphaFold3’s predictive power in reverse to achieve high-precision, all-atom protein design. Using a diffusion model, it starts from random atomic noise and gradually generates physically sound, novel protein structures. With innovative self-conditioning guidance and Markov Random Field decoding, SeedProteo ensures the generated 3D structures and amino acid sequences are highly consistent, overcoming a common weakness in traditional design. It performs exceptionally well in designing protein binders for drug discovery, providing a powerful new tool for creating functional proteins on demand.

Monday, the 12th of January, 2026
Jixing Liu, Ph.D.
A comparative diagram explaining the T-cell receptor (TCR) mispairing problem and the solution offered by computational design. The top half, 'Mispairing,' shows that when two different α and β chains are introduced, they combine randomly, forming the correct pairs along with two ineffective or harmful mispaired TCRs. The bottom half, 'Orthogonal,' shows TCR chains engineered with computational design. Their binding interfaces are shaped to be uniquely complementary, forcing the α1 chain to bind only with the β1 chain, and the α2 chain only with the β2 chain, thus eliminating the mispairing problem. The diagram visually demonstrates how orthogonal TCR design provides a foundation for developing safer multi-specific T-cell therapies.

Biggest Hurdle for Dual TCR Therapy Solved? Computational Design Creates ‘Orthogonal’ T-Cell Receptors

T-cell therapy
computational design
protein engineering
immunotherapy
TCR

In dual-target T-cell therapies, the random mispairing of TCR chains is a key obstacle, leading to reduced efficacy and a serious risk of autoimmunity. This article covers a breakthrough study that uses computational design (Rosetta) to precisely re-engineer the binding interface of the TCR constant region, creating “orthogonal” TCR pairs. This design forces correct α/β chain pairing with up to 95% accuracy, paving the way for safer and more effective multi-specific T-cell therapies and solving a major challenge in immunotherapy.

Sunday, the 11th of January, 2026
Jixing Liu, Ph.D.
This is a 3D molecular structure diagram showing a protein binding with a small molecule ligand. The protein's structure is displayed as a complex, colorful surface and ribbon, forming a recessed 'pocket'. In the center of this pocket, a small molecule ligand is clearly shown as a ball-and-stick model. The image illustrates the core goal of the PocketX method: using reinforcement learning to design a protein pocket from scratch that can precisely recognize and bind a specific small molecule. It makes the abstract concept of 'protein pocket design' tangible, showcasing an ideal, biologically functional structure generated after AI model optimization.

PocketX: How Reinforcement Learning Makes Protein Pocket Design More Than Just a Blueprint

Reinforcement Learning
Protein Pocket Design
Generative AI
Drug Discovery
GRPO

PocketX addresses a common problem in AI protein design: structures that look good on screen but don’t work in the lab. It first uses a generative model to quickly sketch a protein pocket. Then, it refines the design using a reinforcement learning framework called Group Relative Policy Optimization (GRPO). By providing ranked feedback, this method efficiently guides the model to optimize its designs. The result is not only strong ligand binding but also an amino acid sequence that is more biologically and evolutionarily plausible. This strategy is a key step in moving AI-designed proteins from theoretical perfection to practical lab use.

Saturday, the 10th of January, 2026
Jixing Liu, Ph.D.
A scientific illustration showing the core concept of molecular docking. On the left is a complex 3D protein structure with a distinct groove on its surface, the 'binding pocket.' A small molecule ligand is poised to bind with this pocket. Terms like 'Docking score' are labeled, visually explaining how computational tools evaluate the quality of the fit. This image illustrates the task that the BENTO benchmark assesses—predicting the binding mode of a molecule to a protein—and highlights the generalization challenge that existing AI tools face when the protein pocket is a novel structure.

A New Benchmark for AI Drug Discovery: Why Even AlphaFold3 Struggles with New Targets

BENTO
Molecular Docking
Generalization
AI Drug Discovery
Benchmarking

The latest BENTO benchmark reveals a key industry pain point: from classical methods to advanced AI models like AlphaFold3, all tools face a serious generalization challenge when it comes to brand-new protein targets. The research finds that the root cause is data bias in past benchmarks, which led to models that looked good on paper but were actually overfitted. By using a stricter data-splitting method, BENTO provides a realistic assessment of the limitations of various tools in solving core drug discovery problems, offering practitioners a sober perspective and a practical guide for tool selection.

Friday, the 9th of January, 2026
Jixing Liu, Ph.D.
A diagram showing the core principle of how the CurvePotGCN model predicts protein-protein interactions. In the center are the surfaces of two opposing protein molecules, differing in shape and color. The protein on the left has a concave surface with blue areas, representing positive electrostatic potential; the protein on the right has a convex surface with red areas, representing negative electrostatic potential. This complementary 'lock-and-key' shape and 'opposites-attract' charge visually explains the physical basis for how the two proteins can bind. The image illustrates how the model uses these two fundamental features—surface curvature and electrostatic potential—to determine if an interaction exists, highlighting the method's simple and effective approach.

A New Way to Predict Protein Interactions: CurvePotGCN Goes Back to Basics

Protein-Protein Interaction
CurvePotGCN
Graph Convolutional Network
Computational Biology

The CurvePotGCN model offers a new, back-to-basics paradigm for predicting protein-protein interactions (PPI). The method ignores complex evolutionary information and instead focuses on two core physical properties of the protein surface: curvature and electrostatic potential. By “coarse-graining” the protein surface into graph nodes and using a Graph Convolutional Network (GCN) to learn their combined patterns, CurvePotGCN can predict interactions efficiently and accurately. A major advantage of this model is its ability to directly use monomer structures predicted by AlphaFold2, greatly speeding up the process of drug target screening and off-target effect analysis.

Thursday, the 8th of January, 2026
Jixing Liu, Ph.D.
A concept diagram showing how the SynCraft model improves synthesizability through molecular editing. On the left, a complex molecule with a 'synthetic liability' is shown, with a difficult-to-construct chemical group highlighted. An arrow or icon representing a large language model points to a new, edited molecular structure on the right. The new structure retains the core scaffold and pharmacophore of the original molecule but has the synthetic bottleneck removed. The diagram illustrates SynCraft's core idea of crossing the 'synthesis cliff' by making precise, atom-level modifications rather than completely regenerating the molecule to improve the real-world synthesizability of AI-designed molecules.

AI Molecular Editing: Crossing the Synthesis Cliff

Large Language Model
Molecular Editing
Synthesizability Optimization
Drug Discovery
SynCraft

In drug discovery, AI-designed molecules often fall into a “synthesis cliff” because they are too difficult to make. SynCraft offers a new solution. It uses a Large Language Model (LLM) to perform precise, atom-level “edits” on molecules, much like a senior chemist would, rather than generating entirely new ones. The model produces an executable sequence of edits to improve a molecule’s synthesizability while preserving its key pharmacophore, providing a new tool to bridge the gap between AI design and real-world synthesis.

Wednesday, the 7th of January, 2026
Jixing Liu, Ph.D.
This is an architecture diagram of the StructBioReasoner system, illustrating how multi-agent AI is used to design drugs for intrinsically disordered proteins. At the center of the diagram might be a coordinating platform named 'Academy,' surrounded by several AI agent modules representing different functions, such as 'Protein Structure Prediction,' 'Molecular Dynamics Simulation,' and 'Binder Design.' Arrows clearly show the collaboration flow and data exchange between these agents as they work together to complete the complex design task, from target analysis to the final candidate drug molecule. The diagram visually represents the system's automated and scalable scientific reasoning framework.

AI Agents Design Drugs for IDPs, Tackling ‘Undruggable’ Targets

AI Agents
Intrinsically Disordered Proteins
Drug Design
Computational Biology
Supercomputing

A multi-agent AI system called StructBioReasoner designs biologics for intrinsically disordered proteins (IDPs), which are difficult for traditional methods to handle. By simulating a team of experts, the system’s AI agents perform specialized tasks like structure prediction and molecular dynamics simulations. It uses a tournament-style reasoning framework to efficiently screen vast numbers of candidate solutions. This drug design method has already outperformed human experts in computational tests, providing a powerful automated tool for scientific discovery against “undruggable” targets.

Tuesday, the 6th of January, 2026
Jixing Liu, Ph.D.
This diagram shows the workflow of the ReACT-Drug framework, an AI method for designing synthesizable new drugs. It likely includes several key parts: on the left, an initial molecular fragment and a library of chemical reaction templates; in the middle, a reinforcement learning agent that selects a reaction to expand the molecule; on the right, the newly generated molecule is docked to a target protein, and a reward score is calculated based on binding affinity. The entire process emphasizes building molecules by simulating real chemical reactions, which ensures the final product's synthesizability and solves the core problem of traditional AI drug design where generated molecules are often hard to make.

ReACT-Drug: Teaching AI to Design Synthesizable Drugs with Reaction Templates

AI Drug Discovery
Reinforcement Learning
De Novo Drug Design
Synthesizability

How can an AI design molecules that are not just potent but also synthesizable? The ReACT-Drug framework tackles this by giving a reinforcement learning agent a set of real chemical reaction templates. Instead of just sticking atoms together, the AI mimics a chemist, building molecules by executing real reactions like Suzuki coupling. This method ensures the resulting molecules are synthesizable from the ground up, solving a major pain point in AI drug discovery and creating a flexible and powerful platform for automated drug design.

Monday, the 5th of January, 2026
Jixing Liu, Ph.D.
A scientific diagram showing the workflow of the Latent-X2 model. The figure has multiple panels. The left side depicts the model encoding an antibody structure 'X' into a latent space 'Z', and decoding from the latent space to generate new antibodies. The right panels display the excellent properties of the generated results, including high affinity, low immunogenicity, and good developability metrics like stability and solubility. The diagram visually explains how Latent-X2 achieves one-shot generation of high-quality antibody drug candidates in AI drug discovery, addressing several key bottlenecks of traditional R&D.

AI Designs Drug-like Antibodies in a Single Step, Solving Immunogenicity

Latent-X2
Antibody Drugs
Immunogenicity
Developability
AI Drug Discovery

Developing antibody drugs the traditional way is a long, difficult process. It involves screening, maturing affinity, and reducing immunogenicity. The Latent-X2 model changes this. It generates antibody candidates in a single step with high affinity, low immunogenicity, and excellent drug-like properties. Molecules designed by the model were tested with human T cells and showed very low immune risk, solving a key bottleneck in AI drug discovery. This technology shifts drug discovery from tedious screening and optimization to a new era of precise, efficient AI design.

Sunday, the 4th of January, 2026
Jixing Liu, Ph.D.
A scientific illustration showing complex biomolecular structures, used to explain how the CONFIDE model solves the 'hallucination' problem in AI protein structure prediction. The image might contrast a physically plausible protein fold with an incorrect structure generated by AI that has a 'topological conflict.' The latter could show a peptide chain unnaturally weaving through or passing through itself, forming an unbreakable 'knot.' This image visually demonstrates the limitations of traditional pLDDT metrics and highlights the importance of the CONFIDE framework in assessing global physical plausibility, helping researchers identify and filter out unreliable AI predictions.

CONFIDE: A New ‘Quality Inspector’ for AI Drug Discovery, Ending Protein Structure Hallucinations

CONFIDE
Protein Structure Prediction
AI Hallucination
Topological Conflict
Computational Drug Discovery

AI protein structure prediction often generates “hallucinations”—structures that look plausible but are physically impossible. The traditional pLDDT metric struggles to identify these global topological conflicts. CONFIDE is a new framework designed to solve this. It introduces the CODE metric, which assesses the global physical realism of a structure by analyzing the model’s internal data, effectively catching “AI hallucinations.” By combining CODE’s global check with pLDDT’s local check, CONFIDE provides a reliable quality control service for AI-predicted biomolecular structures, significantly improving their reliability in key applications like molecular glue design and drug resistance prediction.

Friday, the 2nd of January, 2026
Jixing Liu, Ph.D.
A diagram of the Synthelite framework's user interface, showing its core function of planning chemical synthesis routes through natural language. On the left is a dialog box where a user inputs a command asking the AI to synthesize a target molecule with a specific strategy. On the right, the AI-generated detailed retrosynthesis route is displayed, including multiple chemical reaction steps and molecular structure diagrams. This image illustrates how Synthelite combines the strategic intent of human chemists with the computational power of AI to design synthesis plans that better meet practical needs.

Synthelite: An AI Chemist That Designs Synthesis Routes Using Plain Language

Retrosynthesis
Large Language Model
Cheminformatics
Synthesis Planning
Human-AI Collaboration

Traditional retrosynthesis software is often criticized for being out of touch with reality. The Synthelite framework addresses this by combining Large Language Models (LLMs) with Monte Carlo Tree Search. It lets chemists give high-level strategic commands in natural language. The AI then finds specific reaction pathways, making the designed routes more practical for the lab. This human-AI collaboration turns AI into a true assistant for chemists, helping them explore the best synthesis strategies together.

Thursday, the 1st of January, 2026
Jixing Liu, Ph.D.
An illustration of the SculptDrug model using Bayesian Flow Networks to generate a drug ligand. The image shows a molecular structure inside a protein's binding pocket, guided by the boundary awareness module as it converges from a disordered noise distribution into a geometrically sound drug ligand. The diagram emphasizes the model's effective avoidance of steric clashes and its precise adaptation to the protein surface features.

SculptDrug: Sculpting the Perfect Ligand with Bayesian Flow

SculptDrug
Bayesian Flow Networks
SBDD
AIDD
Generative Models

SculptDrug uses Bayesian Flow Networks (BFN) to advance structure-based drug design. This article explains how the model’s “Boundary Awareness Block” acts like a parking sensor to avoid steric clashes. It also uses a hierarchical encoder to accurately capture the geometry and chemistry of protein pockets, successfully generating high-affinity, drug-like ligands. This work addresses the challenge of traditional generative models struggling to adapt to complex protein environments.

Wednesday, the 31st of December, 2025
Jixing Liu, Ph.D.
A diagram illustrating the core idea of the HD-Prot model. On the left is a colorful 3D protein structure model, representing continuous physical coordinates. An arrow points from it to the right, which shows two data streams: discrete tokens representing the amino acid sequence above, and continuous tokens representing the 3D structure below. The image visually explains how HD-Prot processes and models both the sequence information and the continuous structural information without discretization. It highlights the model's breakthrough in achieving high-fidelity joint sequence-structure modeling, which is the core technical innovation discussed in this chapter.

HD-Prot: Continuous Tokens Are a New Paradigm for Protein Modeling

Protein Language Model
Continuous Tokens
Structure Modeling
Diffusion Model
HD-Prot

Traditional protein language models discretize continuous 3D structures into tokens, causing significant information loss. The HD-Prot model solves this problem with an innovative continuous token paradigm. It uses a non-quantizing autoencoder and a unified absorbing diffusion process to achieve high-fidelity joint modeling of protein sequence and structure. This breakthrough not only excels in tasks like motif scaffolding but also greatly reduces computational costs, opening up a more efficient and precise path for protein engineering and drug discovery.

Wednesday, the 31st of December, 2025
Jixing Liu, Ph.D.
A schematic diagram showing how the PXDesign framework works. At the center is a large 3D structure of a target protein with a specific binding site on its surface. It is surrounded by several candidate protein binders generated by AI, each with different shapes and structures. Arrows indicate the process from the AI model to the generation of candidate molecules, ultimately filtering down to one binder that fits the target protein perfectly. This image visually explains how PXDesign uses diffusion models and a multi-predictor screening strategy to design high-affinity protein drugs from scratch, solving the low success rate problem of traditional methods.

ByteDance PXDesign: AI-Designed Proteins Hit an 82% Success Rate

Protein Design
Generative AI
Diffusion Models
AlphaFold
Drug Discovery

ByteDance’s PXDesign framework has made a remarkable breakthrough in AI protein design, achieving an experimental success rate as high as 82%. The method combines two generative strategies: PXDesign-d, based on a diffusion model, explores novel structures, while PXDesign-h, based on a hallucination model, performs precise optimization. The key to its success is an innovative multi-predictor screening strategy that uses AlphaFold and other models to collectively ‘review’ and select the most reliable candidates from a vast number of designs. This research not only significantly boosts the efficiency of de novo protein binder design but also provides a powerful open-source tool for drug discovery.

Wednesday, the 31st of December, 2025
Jixing Liu, Ph.D.
An illustration showing the dynamic trajectory of a protein. At the center is a clear 3D structure of a protein, represented by a colored ribbon model with red alpha-helices and yellow beta-sheets. Surrounding this central structure are several semi-transparent copies of the same protein in different poses. This multiple-exposure effect vividly depicts the protein's movement and conformational changes over time. The image visually explains the core goal of the '4D Diffusion Model' mentioned in the article: to use AI to generate a 'short movie' of protein motion, replacing traditional, time-consuming molecular dynamics simulations for studying protein dynamics.

AI Makes Proteins Move: Introducing the 4D Diffusion Model

4D Diffusion Model
Protein Dynamics
Molecular Dynamics
Generative AI
Drug Discovery

Traditional molecular dynamics (MD) simulations are powerful but take too long to capture the dynamic changes in proteins. The 4D Diffusion Model described here offers a new solution. This model can directly generate a “short movie” of a protein’s movement. It uses a reference network and a motion alignment module to make sure the dynamic trajectory is physically plausible and continuous over time. This AI method provides a new way to replace time-consuming MD simulations. It could reveal a protein’s flexible regions and conformational changes in minutes, which would greatly speed up target validation and screening in early drug discovery.

Tuesday, the 30th of December, 2025
Jixing Liu, Ph.D.
This image illustrates the core idea of the RARB framework: using retrieval to improve retrosynthesis accuracy. On the left is a complex target product molecule that a chemist needs to synthesize. In the middle, an icon of a database or library symbolizes the first step of the RARB system: retrieving structurally similar 'relatives' from a vast collection of known reactions. On the right, the AI model, guided by the retrieved information, successfully predicts the reactant molecules. The diagram clearly conveys the point made in the third section of the article: giving an AI a 'literature search system' allows it to 'look things up' instead of guessing, which significantly improves the accuracy and generalization of its predictions.

Giving AI a Search Library Boosts Retrosynthesis Accuracy by 15%

Retrosynthesis
Retrieval-Augmented Generation
RARB
Diffusion Model
Computational Chemistry

Traditional retrosynthesis AI often gets stuck when it encounters new molecules. The RARB framework solves this by giving the AI its own “literature search system,” mimicking how human chemists work. The method first searches a database for known reactions involving molecules similar to the target. It then extracts key information into a “prompt” that guides a diffusion model to make a more accurate prediction. This retrieval-augmented strategy increases the Top-1 accuracy of retrosynthesis predictions by nearly 15% and significantly improves the model’s ability to handle unknown molecules. It shifts the AI from rote memorization to evidence-based reasoning.

Monday, the 29th of December, 2025
Jixing Liu, Ph.D.
A comparison image showing the denoising effect of the CryoDDM model on a Cryo-EM image. The left panel, titled 'Noisy Cryo-EM Image', shows a blurry micrograph of protein particles full of noise, where details are heavily obscured. The right panel, titled 'Denoised by CryoDDM', shows the same image after being processed by the model. The protein's three-dimensional structure is exceptionally clear, with rich detail and various conformational features clearly visible. The image directly illustrates how CryoDDM technology recovers high-quality biomolecular structural information from noisy raw data, helping scientists see the dynamic details of proteins.

CryoDDM: AI Helps Cryo-EM See the Dynamic Details of Proteins

Cryo-EM
Denoising Diffusion Model
Structural Biology
Protein Conformation
Drug Discovery

CryoDDM is a denoising diffusion model designed for Cryo-Electron Microscopy (Cryo-EM) images, built to tackle noise that traditional methods struggle with. By learning the unique noise distribution of Cryo-EM data and using an efficient two-stage diffusion process, the model accurately removes noise while preserving critical structural details. This allows researchers to resolve complex conformational heterogeneity in proteins with new clarity, capturing dynamic changes once lost in the noise. For drug discovery, this technology can reveal new targets like transient allosteric pockets, providing a powerful tool for finding new medicines.

Sunday, the 28th of December, 2025
Jixing Liu, Ph.D.
A chemical structure diagram illustrating the explainability of an AI model. The image shows the 2D structure of a complex organic molecule. Parts of the molecule, such as a nitroaromatic ring, are highlighted with a red heatmap. This visualizes the 'sparse mask' technique mentioned in the article, where the AI marks the chemical groups it considers most relevant to a specific toxicity. This transforms the model from a black box into a tool that provides medicinal chemists with concrete, visual evidence for molecular optimization by pointing out potential 'toxicophores'.

AI Predicts Molecular Toxicity: More Accurate, and Now Explainable

Molecular Toxicity Prediction
Multi-task Learning
Explainable AI
Chemical Language Models

Traditional AI models for toxicity prediction are like black boxes—they give a conclusion without explaining why. This article introduces an innovative multi-task learning framework that uses a special “sparse mask” (L1 sparse penalty) to force the model to focus on the key chemical groups that actually cause toxicity. This method not only significantly improves prediction accuracy but also makes the model explainable, allowing the AI to pinpoint the “culprit” in a molecule, just like a chemist would. This technology transforms AI from a simple classification tool into an intelligent partner that can guide the optimization of drug molecules.

Saturday, the 27th of December, 2025
Jixing Liu, Ph.D.
A complex concept map illustrating how to dissect the internal workings of a molecular transformer model. The diagram might include modules representing the transformer architecture, like attention heads and neural network layers, with arrows pointing to SMILES strings or molecular structures. It visually demonstrates how researchers use tools like sparse autoencoders to extract interpretable chemical features from the model's complex internal activations, such as specific neuron activation patterns corresponding to a benzene ring or a carbonyl group. The core purpose of the image is to visualize the process of an AI model understanding chemical grammar and concepts, revealing its 'thought circuits'.

Dissecting the AI Chemist: The Thought Circuits of a Molecular Transformer

Molecular Transformer
Explainable AI
Sparse Autoencoders
Molecule Generation
Computational Chemistry

Deep learning models are often seen as ‘black boxes,’ but a new study reveals the ‘thought circuits’ inside a molecular transformer. Using tools like sparse autoencoders, researchers found that the model not only learns the grammatical rules of chemical language but also forms internal representations that correspond to chemical concepts like valence and functional groups. This advance improves the use of explainable AI in science. More importantly, it provides a knob to directly control the model’s ‘chemical thinking,’ allowing us to guide the AI to generate molecules with specific chemical features. This opens up a new, interactive design model for drug discovery.

Friday, the 26th of December, 2025
Jixing Liu, Ph.D.
This is a conceptual diagram explaining how Sparse Autoencoders (SAEs) improve the interpretability of chemistry language models. On the left, a complex neural network represents the 'black box' model. Data flows from its hidden layers to the central component, the SAE, depicted as a structure that separates and purifies mixed information. Its outputs are clear, distinct chemical concepts or features, like icons of a benzene ring or a sulfonamide group. The diagram visually shows how SAEs 'dissect' the model's fuzzy internal representations into specific chemical knowledge that humans can understand and control, turning the 'black box' into a 'white box'.

Decoding the AI Brain: SAEs Make Chemistry Language Models Interpretable

Sparse Autoencoders
AI Interpretability
Chemistry Language Models
Drug Discovery

Chemistry language models have long been ‘black boxes,’ limiting their use in drug development. This article introduces a powerful tool called Sparse Autoencoders (SAEs). Think of them as a ‘scalpel’ that can go inside a model and break down jumbled knowledge into understandable, controllable chemical concepts, like specific functional groups. This breakthrough in AI interpretability not only lets us ‘read’ an AI’s chemical thinking for the first time but also allows us to precisely control molecule generation. It paves the way for building more trustworthy and efficient AI-assisted drug discovery tools.

Thursday, the 25th of December, 2025
Jixing Liu, Ph.D.
An infographic comparing two drug discovery techniques: AI co-folding and traditional molecular docking. The left side shows the 'co-folding' method, highlighting its reliance on a large protein structure database (PDB), implying its predictions are highly correlated with training data similarity. The right side illustrates 'docking,' depicting a small molecule finding its optimal binding pose in a protein's active site based on physicochemical principles. The graphic visually explains the article's core point in section 5: for truly novel drug targets, physics-based docking may be more reliable than AI models that depend on training data.

AI Drug Discovery: Co-folding vs. Docking—Which is More Reliable?

AI Drug Discovery
Co-folding
Molecular Docking
Computational Chemistry
Benchmarking

In the AI drug discovery boom, which is more reliable: emerging co-folding models or classic molecular docking methods? This study’s rigorous benchmarking reveals that AI co-folding models perform poorly on new targets not seen in their training data, as their success heavily depends on data similarity. In contrast, physics-based molecular docking is more robust when exploring unknown chemical space. This serves as a warning for practitioners in drug discovery: for new targets and new molecules, traditional docking might be the more trustworthy choice, while the future likely lies in combining the strengths of both.

Wednesday, the 24th of December, 2025
Jixing Liu, Ph.D.
A scientific illustration showing two different conformations of the metamorphic protein Mad2. On the left, the 'Open Mad2' conformation is shown as a ribbon diagram in blue and green. On the right, the 'Closed Mad2' conformation is shown in red and orange. An arrow between them indicates that the protein structure can convert from one to the other. This image visually explains the core concept of the article: some proteins are like 'transformers' with multiple stable structures. AlphaFold2 can be used to predict and identify these multi-state proteins, which is vital for understanding their biological functions and for drug development.

A New Trick for AlphaFold2: Predicting ‘Transformer’ Proteins

AlphaFold2
Metamorphic Proteins
Conformational Ensemble
Statistical Learning
Protein Structure Prediction

Traditional methods struggle to find “metamorphic proteins,” which can switch between multiple structures. A new study cleverly uses AlphaFold2 by manipulating its input, coaxing it to generate multiple possible conformations for a protein. Researchers then combined this with statistical learning to train a classifier that can accurately identify these “transformer” proteins from the conformational ensembles. This efficient computational tool makes it possible to screen for metamorphic proteins on a large scale. It also opens up new ideas for developing drugs that target specific protein shapes, potentially changing how we design medicines.

Tuesday, the 23rd of December, 2025
Jixing Liu, Ph.D.
A conceptual diagram of the CI-LLM framework. It shows several chemical polymer structures connected by graphics and lines, symbolizing the HAPPY (Hierarchically Abstracted Repeat Units) representation method. This method decomposes complex long-chain polymers into chemically meaningful substructures like repeat units and end groups. The diagram illustrates how CI-LLM teaches a large language model to understand molecular composition hierarchically, from macro to micro, like a chemist. This enables more precise property prediction and 'inverse design.' It reflects the article's core idea: integrating chemical knowledge into language models to solve challenges in polymer informatics.

CI-LLM: Teaching Large Models to Think About Polymers Like a Chemist

Large Language Models
Polymer Informatics
Inverse Design
Explainable AI

Traditional methods struggle to help AI effectively understand complex polymers. The CI-LLM framework uses a hierarchically abstracted representation to break down polymers into structural units familiar to chemists, teaching the large language model to think like an expert. The model not only accurately predicts molecular properties but also achieves ‘inverse design’—designing molecules from performance requirements. Its interpretability also provides key insights for rational materials development, marking a significant breakthrough in polymer informatics.

Monday, the 22nd of December, 2025
Jixing Liu, Ph.D.
This image illustrates the core concept of the FoldBench benchmark, designed to evaluate the structure prediction capabilities of AI models in real-world drug discovery scenarios. It might include visualizations of various biomolecular complexes, such as the binding interface of an antibody and antigen, a small-molecule drug targeting RNA, and a protein undergoing an allosteric effect. Together, these diverse structures form a comprehensive set of 'test questions' to examine the accuracy and reliability of models like AlphaFold 3 in predicting complex interactions beyond single proteins, revealing the practical limitations of current AI technology in frontier areas like antibody design and nucleic acid targeting.

The FoldBench Benchmark: A High-Stakes Exam for AI Structure Prediction

FoldBench
AlphaFold 3
Structure Prediction
Drug Discovery
Benchmarking

The FoldBench benchmark sets up a high-stakes exam for AI structure prediction models like AlphaFold 3, tailored for drug discovery. It covers nine real-world R&D scenarios, including antibody design, allosteric effects, and nucleic acid targeting, to fully assess a model’s practical use. The results show that while AI has progressed significantly, major challenges remain in predicting CDR loop conformations, small-molecule-induced conformational changes, and RNA structures. The benchmark gives researchers a calibration tool, clarifying the capabilities and limits of current AI structure prediction technology in modern drug development.

Sunday, the 21st of December, 2025
Jixing Liu, Ph.D.
A diagram showing the workflow of the GFFUTILSAI tool. On the left, a user's natural language question points to a central 'GFFUTILSAI Agent' module. This agent then breaks down the task, pointing to a 'Toolkit' and a 'GFF Database' on the right. The flow clearly illustrates how the AI agent translates human language commands into executable code to query genomic data, solving the efficiency and accuracy problems of large language models directly handling GFF files. The image visualizes the core concept from the fourth section of the article: using an AI agent framework to enable interactive conversation with genomic data.

A New AI Agent Tool: Talk Directly to Your Genome GFF Files

AI Agent
GFF File
Bioinformatics
Natural Language Processing
Genomics

Tired of writing complex scripts to query GFF files? GFFUTILSAI introduces an AI agent framework that lets you talk to your genomic data using natural language. The tool uses a Large Language Model (LLM) as a “translator” to break down user commands into precise tasks that local, specialized tools can execute. This effectively solves the “hallucination” and context window limitations of LLMs in bioinformatics. This “zero-to-hero” model allows non-experts to explore GFF files efficiently and accurately, significantly lowering the barrier to genomic data analysis.

Saturday, the 20th of December, 2025
Jixing Liu, Ph.D.
A flowchart showing how the GRASP framework operates. In the center, a hexagonal module labeled 'GRASP' represents the system. Arrows indicate the flow of information: A 'Human Expert' provides instructions in natural language; GRASP parses MATLAB code into a 'Knowledge Graph'; a 'Reasoning' module then operates on the graph; finally, a 'Code Generation' module produces new MATLAB code based on the updated graph. The process forms a closed loop, illustrating how GRASP acts as an intelligent translator and executor between human experts and complex pharmacology model code, enabling model construction and modification through conversational interaction.

GRASP: Building Complex Pharmacology Models with Conversation

Quantitative Systems Pharmacology
GRASP
AI Agent
Graph Reasoning
Knowledge Graph

Building Quantitative Systems Pharmacology (QSP) models usually requires deep programming knowledge, which limits their use. The GRASP framework changes this by combining multi-agent and graph reasoning technologies. It turns complex code into a knowledge graph, letting pharmacologists build and modify models directly with natural language. GRASP acts like a virtual assistant that understands both biology and programming. It handles complex coding tasks automatically, freeing scientists from coding to speed up drug development.

Friday, the 19th of December, 2025
Jixing Liu, Ph.D.
A scientific concept diagram illustrating the core principle of the CryoBoltz method. In the center is a clear, colorful 3D model of a protein structure, shown with ribbons representing helices and folds. This model is surrounded by a translucent, glowing blue cloud, which represents the heterogeneous density map data from a cryo-EM experiment, containing mixed information from multiple dynamic protein conformations. Several glowing arrows point from the blurry cloud to the clear protein model, visually explaining how CryoBoltz uses the blurry experimental data as a 'guide' to steer an AI model toward generating a precise protein structure that fits within these constraints. The image shows how 'blurry' experimental snapshots can be turned into clear insights into a protein's dynamic processes, which is critical for discovering drug targets associated with specific functional states.

CryoBoltz: Using AI to See Protein Dynamics and Unlock New Drug Targets

CryoBoltz
Cryo-EM
Protein Dynamics
Diffusion Model
Drug Discovery

Traditional protein structure prediction is like taking a static ‘ID photo,’ but drug development needs to see a ‘video’ of the protein in action. The CryoBoltz method uses blurry data from cryo-electron microscopy (cryo-EM) to guide an AI model in resolving multiple dynamic protein conformations. Through a ‘multi-scale guidance’ mechanism, it can reconstruct a series of clear structures from ‘ghostly’ data containing many mixed states. This helps unlock new drug targets by revealing specific conformations for drug design, making R&D more efficient.

Thursday, the 18th of December, 2025
Jixing Liu, Ph.D.
A flowchart showing how VCWorld works. On the left, the input is labeled 'Biological Knowledge' with a network diagram. In the center is the core component, labeled 'Biological World Model (VCWorld)'. On the right, the output is split into two parts: 'Prediction,' pointing to a bar chart of gene expression, and 'Explanation,' pointing to a clear signaling pathway diagram. The image illustrates how VCWorld integrates biological principles to not only make accurate predictions but also provide interpretable mechanistic paths, highlighting its core advantage as a 'white-box' model in drug development.

VCWorld: Building an Explainable ‘White-Box’ Virtual Cell with a Large Model

Computational Biology
Large Language Models
Explainable AI
Virtual Cell
Knowledge Graph

Traditional AI models in drug discovery are often like “black boxes”—hard to understand. VCWorld introduces an innovative “white-box” virtual cell simulator. It deeply integrates biological knowledge graphs with large language models. This not only allows it to accurately predict how cells respond to drugs but also to provide verifiable, mechanism-level explanations. This explainable white-box model solves the prediction challenge when data is scarce, offering a powerful tool for computational biology and drug discovery to move from “predicting” to “understanding.”

Wednesday, the 17th of December, 2025
Jixing Liu, Ph.D.
A comparison image. On the left is a 3D diagram of a protein structure. On the right is a per-residue energy map generated by the PRIMRose model. The map is a heatmap with the protein's amino acid sequence on the y-axis and an energy metric on the x-axis. Colors ranging from blue to red indicate energy changes from stable to unstable. The image visually demonstrates how an InDel mutation specifically affects the energy of each amino acid, revealing structural details that traditional methods cannot provide. This is a core feature of PRIMRose for high-resolution analysis of protein stability.

New AI Tool PRIMRose Decodes InDel Mutations, Amino Acid by Amino Acid

PRIMRose
InDel mutations
protein engineering
deep learning
Rosetta

The PRIMRose model uses deep learning to provide a per-amino-acid energy analysis for protein InDel mutations. It’s much faster than traditional Rosetta simulations and generates high-resolution energy maps to pinpoint structural weaknesses caused by a mutation. This makes it an efficient and precise screening tool for protein engineering and drug discovery, capable of quickly evaluating thousands of mutation designs.

Tuesday, the 16th of December, 2025
Jixing Liu, Ph.D.
A concept diagram showing how the TissueNarrator framework works. On the left is a circular tissue slice made of multi-colored dots representing cells. An arrow points from the tissue map to a sequence of text on the right, labeled a 'spatial sentence.' This sentence clearly contains cell coordinate information, like '(X, Y)', and related gene expression data. The image illustrates how the model transforms complex biological spatial information into a linear sequence that a large language model can understand and process, which is the key step connecting spatial transcriptomics and natural language processing.

TissueNarrator: Using Large Language Models to Read the Cellular Landscape

Large Language Model
Spatial Transcriptomics
Computational Biology
Gene Perturbation
Generative AI

The TissueNarrator framework turns complex tissue slice data into “spatial sentences” that a Large Language Model (LLM) can understand. By putting the physical coordinates of cells and their gene expression data into a sequence, the model can simulate complex cell-to-cell interactions and run virtual gene perturbation analyses. This saves a lot of the cost of traditional experiments. The method brings spatial transcriptomics data analysis into a conversational format, letting researchers explore biology with natural language and offering a powerful new model for computational drug discovery.

Monday, the 15th of December, 2025
Jixing Liu, Ph.D.
A concept diagram illustrating how the CanBART model works. The image might depict the process of feeding genetic mutation data from a large number of real cancer patients into a 'black box' AI model. After processing this information, the model outputs new, structurally complete virtual patient genetic profiles. The diagram visually represents how CanBART learns the patterns of cancer's 'genetic language,' similar to how a language model learns 'vocabulary' and 'grammar,' to generate synthetic data for research. It emphasizes the technology's central role in solving the problem of data scarcity for rare cancers.

CanBART: Using a Language Model to Generate Synthetic Cancer Patient Data

CanBART
Generative AI
Synthetic Data
Cancer Genomics
Rare Cancers

Cancer drug development is often stalled by a lack of samples for rare diseases. The CanBART model offers a solution. It learns the “genetic language” of cancer, much like a large language model learns human language. By analyzing massive amounts of genetic mutation data, CanBART can complete incomplete genetic profiles and generate biologically plausible synthetic patient data. This technology addresses the data bottleneck in rare cancer research. Using this high-quality generative AI data improves disease classification accuracy and opens a new path to accelerate drug development.

Sunday, the 14th of December, 2025
Jixing Liu, Ph.D.
A 3D visualization of a molecule's electron density map. Instead of the traditional ball-and-stick model of atoms, the image shows a continuous, cloud-like surface rendered in various colors, representing the probability of finding electrons in space. This image illustrates the article's core idea: using these information-rich electron density maps as direct input for AI models can predict protein-ligand interactions and quantum properties more accurately than atom-based representations, transforming AI drug discovery.

Beyond Atoms: How Electron Density Maps Are Changing AI Drug Discovery

AI Drug Discovery
Electron Density Maps
3D Molecular Learning
Binding Affinity Prediction
Structure-Based Drug Design

Traditional AI models for drug discovery treat molecules as collections of atoms, but chemistry is really about the interaction of electron clouds. This study explores a new approach: feeding electron density maps directly into a 3D convolutional neural network. The results show that this physics-based representation performs better at predicting protein-ligand binding affinity and quantum properties, especially when data is limited. This opens up a new path for structure-based drug design and shows why returning to physical principles is so important for AI in drug discovery.

Saturday, the 13th of December, 2025
Jixing Liu, Ph.D.
A concept image illustrating a protein design framework based on 'swarm intelligence.' In the center is a complex sphere made of lines, representing the protein backbone being designed. It is surrounded by dozens of small, glowing spheres connected to each other and to the central structure by lines of light. Each small sphere symbolizes an independent large language model agent responsible for a specific amino acid residue in the protein sequence. The image vividly shows the framework's decentralized idea: without central coordination, a large number of simple agents explore and construct a stable and functionally new protein through local interaction and collaboration, just as the article discusses this new method for designing proteins without training.

Swarms of Large Language Models: Designing New Proteins Without Training

Protein Design
Large Language Models
Swarm Intelligence
Agents

Explore a new paradigm for protein design. Inspired by “swarm intelligence,” this study proposes a decentralized framework where each amino acid residue is assigned an independent large language model agent. Through local collaboration, these agents autonomously explore and design new protein sequences that are experimentally verified to fold stably—all without extra training and at very low computational cost. This approach opens a new path for discovering functional proteins that don’t exist in nature, greatly improving the efficiency and flexibility of computer-aided drug discovery.

Friday, the 12th of December, 2025
Jixing Liu, Ph.D.
An abstract digital art illustration shows a glowing sphere in the center, symbolizing a molecule. Multiple blue and purple data streams flow into it, forming a complex network. The image visualizes the core idea of the Rep3Net model: fusing multimodal information to predict drug activity. The central molecule represents a PARP-1 inhibitor, while the different data streams represent molecular descriptors, ChemBERTa embeddings, and graph features. The overall network structure illustrates how the model integrates this heterogeneous data to achieve accurate predictions of molecular bioactivity.

Rep3Net: Using Multimodal Information to Accurately Predict PARP-1 Inhibitor Activity

Multimodal Learning
Drug Activity Prediction
PARP-1
Graph Neural Network
ChemBERTa

How does the Rep3Net model fuse multimodal information to accurately predict the activity of PARP-1 inhibitors? This framework combines molecular descriptors, ChemBERTa language model embeddings, and graph neural network features to overcome the limitations of single molecular representations. Benchmarked on the cancer target PARP-1, Rep3Net shows superior performance over traditional QSAR models, providing a more reliable tool for virtual screening in early-stage drug discovery.

Thursday, the 11th of December, 2025
Jixing Liu, Ph.D.
A diagram illustrating the FoldSAE concept. On the left is a complex 'dense representation' black box, representing the hard-to-understand data inside the RFdiffusion model. In the middle is the 'Sparse Autoencoder (SAE)' module, which decodes the dense data into a series of sparse and interpretable features. On the right, these features are shown acting like control knobs to precisely manipulate the generation of the protein backbone, ultimately outputting a protein with a specific secondary structure (like α-helices and β-sheets). The entire diagram visually explains how FoldSAE transforms a random protein generation process into a controlled engineering design.

FoldSAE: Precisely Controlling Protein Folding with a Sparse Autoencoder

FoldSAE
Sparse Autoencoder
Protein Design
RFdiffusion
Secondary Structure

RFdiffusion is a powerful protein design tool, but its internal mechanics are a black box. The FoldSAE framework uses an innovative Sparse Autoencoder (SAE) to successfully unpack its dense internal representations, breaking down complex vectors into interpretable, sparse features. The research found that these features act like “knobs” for controlling protein secondary structure, allowing for precise increases in helices or strands through simple adjustments. This breakthrough transforms the random protein generation process into a controllable engineering task, opening new paths for future computer-aided drug design.

Wednesday, the 10th of December, 2025
Jixing Liu, Ph.D.
A 3D diagram shows a protein's folding 'energy funnel.' A funnel-shaped surface narrows from a wide top to a deep blue valley, representing the energy landscape of protein conformations. At the bottom of the funnel sits a model of a protein's molecular structure, symbolizing its most stable, lowest-energy native state. This image is used in the article to visually explain the core idea: by learning from vast amounts of data, the internal mathematical mechanism of an AI diffusion model has 'reverse-engineered' the laws of thermodynamics. The energy scores it generates can accurately map the physical process of finding the lowest-energy stable state, just like the traditional physics engine Rosetta.

Diffusion Models Learn Thermodynamics: Extracting Physical Energy from AI

Diffusion Models
Protein Structure Prediction
Computational Chemistry
Thermodynamics
Generative AI

When a diffusion model predicts a protein’s structure, is it just memorizing shapes or does it understand the laws of physics? New research shows its denoising process closely matches the Boltzmann distribution from thermodynamics, and its negative log-likelihood can be treated as physical energy. This means the AI builds an energy landscape similar to Rosetta’s, allowing it to map out an “energy funnel.” This opens a new path for drug discovery by merging AI with physics-based simulations.

Tuesday, the 9th of December, 2025
Jixing Liu, Ph.D.
This image shows the core principle of the Apo2Mol model in drug design. In the center is a complex 3D protein structure with a recessed pocket, which is the binding site for drug molecules. Inside this pocket, a small molecule is being generated and optimized. The key is that the protein pocket's conformation is not static; it adjusts dynamically to fit the molecule being generated. The image illustrates how Apo2Mol uses a diffusion model to simultaneously generate a ligand and adjust the pocket conformation, solving the prediction inaccuracies caused by ignoring protein flexibility in traditional methods. It reflects the biological process of 'induced fit' and is a key technology for overcoming the challenges of drug design based on Apo structures.

Apo2Mol: A New Path for AI Drug Design that Tackles Protein Flexibility

Apo2Mol
Diffusion Model
Protein Flexibility
Drug Design
Induced Fit

Traditional drug design often gets stuck when dealing with protein flexibility, especially when only an unbound (Apo) structure is available. The Apo2Mol model offers a solution. It uses a dynamic pocket-aware diffusion model to generate a drug molecule while simultaneously adjusting the protein pocket’s conformation in real time. This co-optimization method simulates the real “induced fit” effect, accurately predicting the stable state after a ligand binds to a protein. The technique overcomes the core challenge of protein flexibility, opening a new path for AI drug design against highly flexible or novel targets.

Monday, the 8th of December, 2025
Jixing Liu, Ph.D.
A scientific illustration showing the 3D molecular structure of a protein complex. A larger protein molecule, shown with a light blue surface and internal ribbon structures, is tightly bound to a smaller antibody fragment. The antibody fragment, composed of orange and green ribbon structures, fits precisely into the binding site of the larger molecule. This image visualizes the core concept discussed in the article: using a modified AlphaFold3 model for the co-design of antibody sequence and structure. It represents an antibody-antigen complex successfully generated by the AI model, where the precise structure and sequence together determine its effective target-binding ability, validating this new AI design paradigm.

A New AI Paradigm: Generating Antibody Sequence and Structure in One Step with an AlphaFold3 Model

AlphaFold3
Antibody Design
Sequence-Structure Co-design
Protein Structure Prediction
Drug Discovery

This is a new way to discover antibody drugs. Researchers modified the AlphaFold3 model, adding a sequence diffusion head to generate antibody sequences and structures at the same time. The method solves the problem of scarce antibody-antigen complex data, boosting sequence recovery in the critical CDR-H3 region by nearly 87% while maintaining the same structural prediction accuracy as the original model. This sequence-structure co-design technique opens a new path for developing effective therapeutic antibodies.

Sunday, the 7th of December, 2025
Jixing Liu, Ph.D.
An illustration of the core idea behind the FuncBind framework. In the center is an abstract shape representing a unified generative model. It is surrounded by three different types of molecules: a small molecule, a macrocyclic peptide, and an antibody CDR loop. All molecules are shown as similar clouds of atomic density, symbolizing how neural fields represent them uniformly. The image shows how FuncBind uses a single model to cross the boundaries between different drug modalities, enabling unified molecule generation and providing a general theoretical framework for AI drug design.

Neural Fields for Unified Molecule Generation: A Unified Theory for AI Drug Design?

Neural Fields
FuncBind
Drug Design
Molecule Generation
Multimodal

Traditional drug discovery uses specialized models for different molecule types, like small molecules or antibodies. The FuncBind framework offers a different approach. It uses Neural Fields to represent all molecules as continuous atomic densities, allowing a single model to generate multiple drug modalities. In simulations, the method performs as well as specialized models, and experiments have confirmed its ability to design functional molecules. This unified perspective opens a new path for cross-modality drug design and could become a foundational theory for AI in drug discovery.

Saturday, the 6th of December, 2025
Jixing Liu, Ph.D.
A diagram showing the core idea of the GraphCliff model. It depicts the graph structure of a molecule, with circles as atoms and lines as chemical bonds. Some nodes are highlighted in orange, representing the key substructures that cause an 'activity cliff.' The entire graph is set against a dark grid background, visualizing how a graph neural network processes molecular information. The image illustrates how GraphCliff's gating mechanism identifies and focuses on the small but critical structural features that lead to drastic changes in biological activity, allowing it to accurately predict the activity cliff phenomenon.

GraphCliff: A New Graph Neural Network for Predicting Activity Cliffs

Activity Cliffs
Graph Neural Network
Drug Discovery
Gating Mechanism
QSAR

“Activity cliffs” are a big challenge in drug discovery. Tiny changes in a molecule’s structure can cause huge differences in its biological activity. The GraphCliff model was created to tackle this. It’s a graph neural network (GNN) with a unique gating mechanism at its core. This gate helps it capture both local details and global structural information in a molecule, avoiding the “over-smoothing” problem common in other GNNs. By balancing short- and long-range information, GraphCliff can more accurately predict the sharp changes in activity caused by subtle structural tweaks, offering a new way to think about drug development.

Friday, the 5th of December, 2025
Jixing Liu, Ph.D.
A conceptual diagram of a Graph Neural Network (GNN). The image shows multiple interconnected nodes, with information passing between them along edges, forming a complex network. This visually represents how a GNN works by aggregating information from neighboring nodes to update the state of a central node. In the 'Atomic Hybridization Empowers GNNs' section of the article, this image illustrates how the model abstracts molecular structure into graph data. The GNN architecture then learns local chemical features, and its accuracy in predicting drug targets is enhanced by incorporating atomic hybridization states.

Atomic Hybridization Empowers GNNs for More Accurate Drug Target Prediction

Graph Neural Network
Atomic Hybridization
Drug Target Prediction
Molecular Descriptors
Computer-Aided Drug Discovery

Standard Graph Neural Networks (GNNs) often overlook deep chemical properties like atomic hybridization when processing molecules. This study gives GNNs a chemist’s perspective by identifying atomic hybridization states, allowing the model to understand molecular structure at an electronic level. By combining the local insights of GNNs with the global physicochemical properties from traditional molecular descriptors, this hybrid framework significantly improves the accuracy of drug target prediction. It offers a more chemically intuitive tool for computer-aided drug discovery.

Thursday, the 4th of December, 2025
Jixing Liu, Ph.D.
A flowchart explaining how the MedRule-KG framework works. On the left, an icon represents a Large Language Model (LLM) generating content. In the middle, a knowledge graph with nodes and connections symbolizes verified scientific rules. A module named 'Soft Constraint Controller' connects the LLM and the knowledge graph, validating the generated content in real-time. On the right, the reliable, logically corrected output is shown. The diagram illustrates how MedRule-KG acts like 'logic guardrails,' embedding the hard logic of a knowledge graph into the AI's generation process to prevent basic scientific errors and improve model reliability in fields like drug discovery.

MedRule-KG: Adding ‘Logic Guardrails’ to AI Drug Discovery

MedRule-KG
Knowledge Graph
Large Language Model
Neuro-symbolic AI
Drug Discovery

Large Language Models (LLMs) often “hallucinate” and produce unscientific results in drug discovery. The MedRule-KG framework installs “logic guardrails” for the LLM’s generation process by introducing a lightweight knowledge graph and a soft constraint controller. This method corrects outputs that violate scientific rules in real-time, without retraining the model, and reduces rule violations by 83.2%. This neuro-symbolic AI approach effectively improves the reliability and accuracy of AI for critical tasks like drug design.

Wednesday, the 3rd of December, 2025
Jixing Liu, Ph.D.
A scientific illustration showing a complex macromolecular structure, likely a protein or ribosomal complex. The structure is composed of ribbon and surface models in various colors, including blue, purple, green, and orange, depicting its different subunits or conformational states. A sense of motion in the image suggests molecular movement and change. This figure visualizes the core result of the CryoDyna technology: reconstructing thousands of static 2D cryo-EM images into a continuous, near-atomic resolution dynamic model of a protein. It shows how the model captures multi-scale structural changes, from global to local, revealing the dynamic processes of proteins as they perform their biological functions, which is important for drug development and structural biology research.

CryoDyna: Turning Cryo-EM Stills into a Dynamic Movie of Proteins

CryoDyna
Cryo-EM
Protein Dynamics
Deep Learning
Structural Biology

Cryo-electron microscopy (Cryo-EM) captures beautiful static structures of proteins but can’t reveal their dynamic processes. The CryoDyna tool uses an end-to-end deep learning framework to directly reconstruct tens of thousands of 2D particle images into a continuous dynamic model of a protein. It combines cross-view attention and multi-scale deformation prediction, and it also incorporates constraints from physical laws to generate a dynamic “movie” at near-atomic resolution. This technology fills the gap between static structures and molecular dynamics simulations, offering a new way to understand target function and drug interactions.

Tuesday, the 2nd of December, 2025
Jixing Liu, Ph.D.
A flowchart shows how the 'Atom-Anchoring' Large Language Model works for retrosynthesis. On the left is a target molecule with numbered atoms. The middle section shows the LLM reasoning to identify a suitable bond to break. On the right are the two starting reactants predicted by the model based on that break. The diagram illustrates how the model simulates a chemist's thinking process, using atom-level reasoning to plan a synthesis route rather than just matching a database.

A New Way to Use Large Models: Atom-Level Reasoning for Retrosynthesis

Retrosynthesis
Large Language Models
Atom-Anchoring
Chemical Reasoning

Traditional retrosynthesis methods rely on huge amounts of data, like making a model memorize recipes. A new study proposes an “Atom-Anchoring” method that teaches a Large Language Model (LLM) to reason at the atomic level, just like a chemist, by giving each atom in a molecule a unique number. The method doesn’t require training on special reaction data. Using only zero-shot and few-shot prompts, it enables an LLM to accurately find breaking points in a molecule and predict the reactants. This shifts the AI from a search tool to a reasoning assistant with chemical intuition.

Monday, the 1st of December, 2025
Jixing Liu, Ph.D.
This cover image for the MultimerMapper section visualizes the evolution from predicting single, static protein structures to mapping complex, dynamic assembly processes. The image reflects the article's core idea: using AI analysis to infer unknown stoichiometry and context dependency is like stitching isolated photos into a coherent movie. This visual emphasizes the tool's technical breakthrough in decoding protein interactions and dynamic behavior, showcasing its ability to work accurately without a predefined recipe.

AI Protein Complex Prediction: From Still Snapshots to a Dynamic Movie

MultimerMapper
AlphaFold
Protein Structure Prediction
Stoichiometry Inference
Bioinformatics

MultimerMapper is a major step forward for AI protein structure prediction. It solves a big problem with tools like AlphaFold, which need you to know the exact recipe of a protein complex beforehand. MultimerMapper figures out the recipe by analyzing how the complex assembles over time. This lets researchers go from looking at static snapshots to understanding the whole movie of how protein complexes behave depending on their environment. The result is a more biologically useful view of potential drug targets.

Sunday, the 30th of November, 2025
Jixing Liu, Ph.D.
This image shows the technical architecture of the KSMoFinder model, detailing how it uses a knowledge graph to connect proteins, kinases, and motifs into a complex biological network. The diagram illustrates the complete workflow, from building network associations and generating graph embedding vectors for nodes to feeding them into a deep learning classifier for prediction. It visually represents how the model improves prediction accuracy by capturing the social network context of proteins.

KSMoFinder: Knowledge Graphs Predict Kinases Better Than Large Protein Models

Knowledge Graph
Kinase Prediction
Deep Learning
Drug Development
Bioinformatics

The KSMoFinder model builds a “social network” for proteins using a knowledge graph. It integrates sequence data with biological context through graph embeddings. In predicting kinase substrates, this approach significantly outperforms major protein language models. It offers a more precise and efficient computational tool for analyzing off-target effects and discovering new targets in drug development.

Saturday, the 29th of November, 2025
Jixing Liu, Ph.D.
This image shows the core workflow of the EVOSYNTH framework. The left side depicts the first stage, latent space evolutionary optimization, where an algorithm iteratively screens for molecular structures with high binding potential to multiple targets (like JNK3 and GSK3β). The right side shows the second stage, the synthesis-aware prioritization module, which simulates a chemist's process of retrosynthetic analysis and cost evaluation for the selected molecules. The diagram clearly illustrates how the model balances biological activity with chemical synthesizability at the very beginning of the drug design process.

A New AI Framework for Multi-Target Drug Discovery: Balancing High Activity and Synthesizability

AI Drug Discovery
Multi-Target Drug Design
Evolutionary Algorithms
EVOSYNTH
Synthesizability

This article explains EVOSYNTH, an AI framework designed to solve the conflict between biological activity and manufacturing viability in multi-target drug discovery. It details how the model uses an evolutionary algorithm for multi-target optimization in a latent space and incorporates a “synthesis-aware prioritization” module to filter out molecules that are difficult to produce. The framework shows significant potential for designing drugs for complex diseases like Alzheimer’s, offering a practical path for moving AI-designed drugs from theory to the clinic.

Friday, the 28th of November, 2025
Jixing Liu, Ph.D.
An illustration of the SIGMADOCK model's core mechanism for molecular docking. The image shows a complex, flexible ligand molecule being broken down into several rigid fragments with fixed internal structures, rather than being treated as a single entity. Using an SE(3) Riemannian diffusion model, these fragments are then guided, assembled, and oriented into their optimal positions within the binding pocket of a protein target. This visualization highlights how the model uses geometric priors and diffusion algorithms to reduce the complexity of calculating degrees of freedom, leading to more accurate and efficient predictions of molecular conformations than traditional methods.

SIGMADOCK: Rethinking Molecular Docking with a Fragment Diffusion Model

Molecular Docking
Diffusion Model
Deep Learning
Drug Discovery
SE(3) Equivariance

Molecular docking is a key challenge in drug discovery, especially when dealing with highly flexible ligand molecules. SIGMADOCK introduces a new approach by breaking the ligand into rigid fragments and combining this with an SE(3) Riemannian diffusion model. This simplifies the complex conformational search space. The method achieved a 79.9% Top-1 success rate on the PoseBusters benchmark, outperforming existing deep learning models and traditional physics-based methods. It offers a powerful new tool for high-precision drug screening.

Thursday, the 27th of November, 2025
Jixing Liu, Ph.D.
This image shows the workflow of the CHEFNMR model. On the left are the inputs: a 1D NMR spectrum and a molecular formula. The middle section depicts the AI model's atomic diffusion process. On the right is the 3D molecular structure generated by the model. The core of this model is its use of an atomic diffusion model to learn the precise 3D positions of atoms from spectral data.

A New Way for AI to Read NMR: The CHEFNMR Atomic Diffusion Model

AI
Drug Development
NMR
CHEFNMR
Atomic Diffusion Model

CHEFNMR is an AI tool based on an atomic diffusion model. It takes a 1D NMR spectrum and a compound’s molecular formula to directly generate its 3D structure, offering a new way to solve molecular structures in drug development. The model is over 65% accurate on complex natural products and shows strong zero-shot generalization.

Wednesday, the 26th of November, 2025
Jixing Liu, Ph.D.
A diagram showing the core architecture or performance of the DMol model. The image illustrates the graph diffusion process for molecule generation, highlighting how motif compression treats key chemical structures like benzene rings as supernodes. This visual helps explain how the model achieves generation speeds 10 times faster than mainstream methods while maintaining the chemical accuracy of the molecules.

DMol: Generate Molecules 10x Faster While Keeping Key Chemical Motifs

DMol
Graph Diffusion Model
Motif Compression
Drug Discovery
Generative AI

This article explains the DMol platform, a model that speeds up molecule generation by 10x using graph diffusion models and motif compression. We look at how DMol uses hybrid noise and a supernode strategy to significantly reduce diffusion steps, ensuring the generated molecules are chemically sound and drug-like. This provides an efficient tool for drug discovery that balances speed and quality.

Tuesday, the 25th of November, 2025
Jixing Liu, Ph.D.
An illustration of how the Peptide2Mol model works. It shows how AI translates the binding features of a peptide and its target into a small-molecule design. The diagram follows the process from the peptide's structure to the generated small molecule, showing how the diffusion model uses 3D information to build a 'new key' that precisely fits a protein target. The image highlights the technology's innovative use in medicinal chemistry.

Peptide2Mol: AI Directly Generates Small Molecules that Mimic Peptides for Targeted Protein Binding

Peptide2Mol
Diffusion Model
Drug Discovery
Small Molecule Generation
AI for Science

Peptide2Mol uses an advanced diffusion model to directly generate small molecules that can mimic natural peptides and bind to their targets. This offers a new way to solve the drug-like properties problem of peptide drugs. The model incorporates 3D spatial symmetry, suggests effective chemical group substitutions, and turns the abstract problem of mimicry into a concrete chemical design, which greatly speeds up drug development.

Monday, the 24th of November, 2025
Jixing Liu, Ph.D.
A workflow diagram of the SPECTRA framework. The image shows how molecular structures are mapped to a spectral domain. The Gromov-Wasserstein coupling algorithm aligns two real molecules, and linear interpolation between their eigenvalues generates new spectral features. Finally, these features are used to reconstruct chemically valid ghost molecules to augment a sparse dataset.

AI Drug Discovery: Using ‘Ghost Molecules’ to Fill in Sparse Data

SPECTRA
Drug Discovery
Data Augmentation
Molecular Property Prediction
Generative Models

The SPECTRA framework offers a new way to handle imbalanced data in drug discovery. It generates chemically valid “ghost molecules” in the spectral domain to fill sparse data regions. Using Gromov-Wasserstein couplings, it improves model accuracy in key property ranges. This interpretable data augmentation strategy solves the problem of scarce “golden molecule” data without adding computational overhead.

Sunday, the 23rd of November, 2025
Jixing Liu, Ph.D.
An illustration of the CryoSPARC software's automated workflow. The background shows a complex 3D rendering of a protein structure in blue and purple. In the foreground, the title 'Automated Workflows' sits above four icons representing key steps in the process: 'Automated Preprocessing', 'Automated Particle Picking', 'Automated 2D Classification', and 'Automated 3D Reconstruction'. The image clearly shows how the technology transforms the labor-intensive steps of cryo-EM data processing into a standardized, automated operation, accelerating the drug discovery pipeline.

Cryo-EM Automation: CryoSPARC Makes High-Throughput Drug Discovery a Reality

Cryo-EM
CryoSPARC
Structural Biology
Drug Discovery
Automated Workflow

For structural biologists, cryo-electron microscopy (cryo-EM) data processing was once a time-consuming bottleneck. The CryoSPARC team at Structura Biotechnology has developed a fully automated, end-to-end data processing workflow that handles everything from data screening and particle picking to 3D structure reconstruction. When applied to challenging GPCR targets, the results surpassed the quality and resolution of expert manual processing, providing a powerful automation tool for high-throughput and structure-based drug discovery (SBDD).

Saturday, the 22nd of November, 2025
Jixing Liu, Ph.D.
A flowchart titled 'StoPred' with the subtitle 'Accurate Stoichiometry Prediction for Protein Complexes Using Protein Language Models and Graph Attention'. The diagram shows how the StoPred model works. On the left, protein subunit sequences or 3D structures are input and processed by a protein language model (pLM) or a graph neural network (GNN) for feature extraction. These features are then fed into a graph attention network (GAT), which simulates interactions between subunits to predict the most stable stoichiometric combination. Finally, the model outputs the most likely stoichiometry, such as A2B2. The chart clearly illustrates how StoPred uses AI to solve the core problem of determining the recipe for protein complexes.

StoPred: AI Accurately Predicts the ‘Recipe’ for Protein Complexes

Stoichiometry Prediction
Protein Complex
Protein Language Model
Graph Attention Network
AI-aided Drug Discovery

In drug discovery, knowing a protein complex’s stoichiometry is the first step to understanding its function. The StoPred tool offers a new solution. It combines a protein language model (pLM) with a graph attention network (GAT) to predict the composition of protein complexes directly from their sequence or structure, without needing a template. For heteromeric complexes, its prediction accuracy is 41% higher than the previous best method. This helps avoid wasting computing power on incorrect models and significantly boosts research efficiency.

Friday, the 21st of November, 2025
Jixing Liu, Ph.D.
An illustration of the 'Speak to a Protein' AI system interface. On the left side of the screen is a dialog box where a user has typed a natural language command. On the right is a 3D visualization window displaying the complex three-dimensional structure of a protein molecule, with specific regions highlighted. This image clearly shows the system's core function: the user interacts with the AI through natural language, and the AI translates these commands into visual manipulations and analyses of the protein structure, demonstrating the integration of natural language processing with structural biology analysis.

New AI Tool: Analyze Proteins Like a Conversation with Real-Time 3D Interaction

Protein Structure Analysis
Large Language Model (LLM)
Natural Language Processing
Retrieval-Augmented Generation (RAG)
3D Visualization

Structural biology analysis is often tedious and requires specialized knowledge. The “Speak to a Protein” system changes this by merging natural language processing, code execution, and 3D visualization. It turns complex protein analysis into a simple, conversational process. Users just state their needs in plain language. The system then uses a large language model and retrieval-augmented generation (RAG) to automatically handle tasks like data retrieval, structural alignment, and visualization. Results appear in real-time in an interactive 3D window, greatly speeding up the process of testing scientific hypotheses.

Thursday, the 20th of November, 2025
Jixing Liu, Ph.D.
A flowchart showing how the 'Rank-Conditioned Committee' (RCC) framework handles the conformational uncertainty of proteins. On the left, a single protein structure is used to generate multiple conformations through molecular dynamics simulations. These conformations are sorted by energy level and grouped into different ranks, such as 'Rank 1' and 'Rank 2'. On the right, each rank of conformations is fed into a specially trained deep neural network 'committee' for evaluation and scoring. The flowchart clearly illustrates how the framework breaks down the problem by modeling different conformational states separately, allowing it to more accurately distinguish between model prediction errors and the protein's own flexibility in AI-assisted drug discovery.

A New Way for AI to Discover Antibodies: Solving the Protein ‘Shape-Shifting’ Problem

Antibody Discovery
Conformational Uncertainty
Active Learning
Deep Neural Networks
Drug Design

In computational antibody discovery, the changing shapes of proteins create a big challenge. When an AI model gives an antibody a low score, it’s hard to know if the molecule is actually bad, or if the model just “wrongly rejected” a promising molecule caught in a bad pose. The “Rank-Conditioned Committee” (RCC) framework, introduced in this paper, solves this by modeling different energy levels of protein shapes separately. It successfully tells the difference between the model’s own uncertainty and the protein’s natural shape-shifting. This new approach allows AI to evaluate antibody potential more intelligently, preventing good drug candidates from being missed due to their flexibility. It offers a more reliable and efficient computational strategy for the active learning loop in drug discovery.

Wednesday, the 19th of November, 2025
Jixing Liu, Ph.D.
A concept diagram shows how the ConforMix algorithm guides an AI diffusion model to explore different protein conformations. On the left is a stable and common protein shape, labeled as the 'default' or 'low-energy' state. A series of arrows points from left to right, symbolizing the algorithm's guidance process. On the right are several different, even rare, protein conformations. One of these clearly shows an open 'allosteric pocket,' a key target for drug design. The diagram uses simple graphics and smooth lines to illustrate how ConforMix unlocks and discovers these functionally vital hidden molecular states at inference time, without retraining the model.

ConforMix: Unlocking Hidden Biomolecular Shapes at AI Inference Time

ConforMix
Diffusion Models
Molecular Conformation
Allosteric Pockets
Computer-Aided Drug Discovery

Diffusion models are good at predicting protein structures, but they usually only generate the most common, stable shapes. They often miss the rare but functionally critical ones. The ConforMix algorithm, developed at Stanford, uses an innovative “classifier guidance” technique. This method guides the AI to efficiently explore a biomolecule’s hidden conformational states during the inference phase, without needing to retrain the model. This approach can find “allosteric pockets” needed for drug targeting and estimate the free energy differences between conformations. It provides a powerful “plug-and-play” tool for computer-aided drug discovery (CADD) and for studying dynamic molecular behavior.

Tuesday, the 18th of November, 2025
Jixing Liu, Ph.D.
A bioinformatics illustration showing a protein structure. On the left is a complex 3D model of a protein, representing the HIV gp120 target, with a multi-colored and intricately folded surface. A smaller, ring-shaped molecule, the cyclic peptide, is precisely docked into a pocket on the target protein's surface. The entire image is set against a blue and purple gradient background, highlighting the details of the protein and peptide molecules. It visually explains the core concept of using AlphaFold for precise cyclic peptide design to target specific protein binding sites.

Tuning AlphaFold: Precisely Designing Cyclic Peptides to Hit HIV Targets

Cyclic Peptide Design
AlphaFold
Drug Discovery
Protein-Protein Interaction
HIV

By adding custom loss functions and spatial constraints, researchers have successfully “tuned” AlphaFold to design cyclic peptides that target HIV with high precision. This work transforms AlphaFold from a prediction tool into a powerful generative one by defining a bullseye, adding a centroid distance penalty, and optimizing binding affinity. It opens up new paths for computational drug design and the development of cyclic peptide therapeutics.

Monday, the 17th of November, 2025
Jixing Liu, Ph.D.
This image illustrates the workflow of the AIMS-PAX framework. It depicts multiple parallel exploration processes, symbolizing the parallel active learning mechanism, as they explore a vast molecular conformational space. These exploration paths converge on a central region where high-precision Density Functional Theory (DFT) calculations are performed. This shows how the framework efficiently identifies and calculates only the most valuable conformations, reducing the data required for training a machine learning force field by three orders of magnitude. The overall image aims to show how AIMS-PAX significantly accelerates the construction of high-precision force fields through parallelization and active learning.

AIMS-PAX: A 1000x Faster Way to Build Force Fields, and a New Paradigm for AI Simulation

Machine Learning Force Fields
Molecular Dynamics
Active Learning
Density Functional Theory
Computational Chemistry

Molecular dynamics simulations have long struggled with a trade-off between accuracy and speed. Training high-precision machine learning force fields (MLFFs) is itself a computational bottleneck. The AIMS-PAX framework uses a parallel active learning strategy to autonomously identify the model’s knowledge gaps, performing expensive DFT calculations only on the most valuable molecular conformations. This approach reduces the data needed to build a force field by up to three orders of magnitude, greatly speeding up the training process. It makes it possible to quickly customize high-precision force fields for complex drug and material systems, like PROTAC molecules or perovskite materials, offering a powerful new tool for molecular simulation and drug design.

Saturday, the 15th of November, 2025
Jixing Liu, Ph.D.
A conceptual diagram of the Omni-DNA model. On the left is a stylized DNA double helix with the letters 'A', 'T', 'C', 'G' representing the base sequence. In the middle, a brain-shaped icon with circuit patterns symbolizes the AI model processing information. Arrows from the brain point to a book and a text file icon on the right, representing the model's ability to 'translate' DNA sequences into readable functional text annotations. The core idea of the image is the transformation of complex genomic sequences into human-understandable language through an AI model, revealing their biological functions.

Omni-DNA: AI Reads DNA, Unlocking the Long Narrative of the Genome

Omni-DNA
Genomics
Natural Language Processing
DNA Sequence Analysis
Multi-task Learning

The Omni-DNA model uses an innovative SEQPACK compression method, enabling AI to efficiently process and understand ultra-long DNA sequences and capture interactions between distant gene-regulatory elements. By deeply integrating genomics with Natural Language Processing (NLP), the model not only classifies DNA sequences but also generates text annotations describing their functions. Through a multi-task learning strategy, Omni-DNA shows excellent performance and generalization across multiple biological tasks, offering a new paradigm for functional genomics research.

Friday, the 14th of November, 2025
Jixing Liu, Ph.D.
A diagram illustrating the FolDE framework's workflow, divided into three main stages. The left section shows the initial screening using a protein language model, emphasizing a 'naturalness' score. The middle section demonstrates how the first batch of experimental activity data is combined with the 'naturalness' score to train a predictive model, showing the transition from general rules to specific patterns. The right section explains its batch selection strategy through a counter-intuitive 'pessimistic' assumption, aimed at increasing the diversity of candidate mutants to avoid getting stuck in local optima. The diagram clearly communicates how FolDE efficiently guides protein optimization experiments when data is scarce.

FolDE: Efficiently Optimizing Protein Activity with Limited Data

Active Learning
Protein Engineering
Directed Evolution
Language Models
Cold Start Problem

FolDE is an open-source active learning framework designed to solve the “cold start” problem in protein directed evolution, where early experimental data is scarce. The framework combines the “naturalness” predictions of protein language models with a small amount of real activity data. It uses a “pessimistic” strategy to ensure diversity in mutant selection. For researchers in protein engineering, FolDE offers a ready-to-use tool that significantly improves the efficiency of discovering highly active mutants, speeding up research and reducing experimental costs.

Thursday, the 13th of November, 2025
Jixing Liu, Ph.D.
A diagram showing the workflow of the CIDD framework. On the left is the 3D structure of a protein target's binding pocket, displaying its complex surface and internal cavity. To the right of center is an icon of a brain containing gears and chemical structures, symbolizing a Large Language Model (LLM) thinking like a chemist. Four boxes labeled 'Analyze', 'Design', 'Reflect', and 'Decide' extend from the brain icon. Together, these steps form a cyclical 'chain of thought' process. The overall color scheme is a technological blue, illustrating how the CIDD framework blends artificial intelligence with the workflow of a medicinal chemist to generate more chemically reasonable drug molecules.

The CIDD Framework: Using LLMs to Design Molecules Like a Veteran Medicinal Chemist

Drug Design
Large Language Models
CIDD
Chain-of-Thought
Molecule Generation

Molecules generated by traditional computer-aided drug design often lack chemical common sense. The CIDD framework teaches Large Language Models (LLMs) to design molecules step-by-step, mimicking the “chain of thought” of a medicinal chemist: first analyze target interactions, then design a scaffold, reflect and evaluate, and finally, make a decision. This approach significantly improves the drug-likeness and synthesizability of generated molecules, transforming AI from a “molecule generator” into a smarter “collaborative partner” and offering a new path to discover drug candidates with greater potential.

Wednesday, the 12th of November, 2025
Jixing Liu, Ph.D.
This image shows a workflow diagram for the RETRO-R1 model, titled 'RETRO-R1: LLM-based Agentic Retrosynthesis'. At the center is a brain icon, representing the core Large Language Model (LLM), which interacts via 'Policy' with the 'Environment' (chemical environment) on the right. The environment includes tools like 'Single-step Retrosynthesis' and 'Building Block Inventory'. Below the brain, the 'Reward' and 'Observation' mechanisms of reinforcement learning are shown, forming a closed feedback loop. The diagram clearly explains how RETRO-R1 simulates a chemist, using reinforcement learning to interact with chemical tools to autonomously learn and optimize the design of retrosynthetic routes.

RETRO-R1: Optimizing Retrosynthesis with Large Language Models and Reinforcement Learning

AI Agents
Retrosynthesis
Reinforcement Learning
Large Language Models
Drug Development

Chemists spend a great deal of time and rely heavily on experience when designing retrosynthetic routes, a critical first step in drug development. To make this process more efficient, researchers have developed an AI agent called RETRO-R1. This model uses a large language model (LLM) at its core and is trained with reinforcement learning. This combination allows it to think like a chemist, interacting with a chemical tool environment to continuously try, get feedback, and learn. The result is the ability to design molecular synthesis routes with a higher success rate and shorter paths. This method significantly outperforms existing technologies and offers a new way to solve complex scientific planning problems.

Tuesday, the 11th of November, 2025
Jixing Liu, Ph.D.
A scientific illustration showing the 3D structure of a protein, featuring a complex molecular model composed of helical ribbons and folded sheets in various colors, predominantly blue and purple. The background is a soft, light gradient. This image visually represents the 3D conformation of a VH domain optimized through computational methods. In the article, it illustrates how researchers enhanced the stability of the VH domain by introducing disulfide bonds or point mutations, a foundational step for developing new multispecific antibody drugs.

Computationally Designing Stable VH Domains to Develop Conditionally Active Multispecific Drugs

VH Domain
Antibody Engineering
Computational Design
Rosetta
Multispecific Biologics

The VH domain is a core functional part of antibody drugs, but its inherent instability limits its use.This study uses the Rosetta software to computationally analyze VH domains and design two strategies to stabilize them: introducing disulfide bonds and making point mutations. After modification, the expression of VH domains increased by over 600 times and their thermal stability improved by more than 20°C. This work provides a solid foundation for developing conditionally active multispecific biologics based on VH domains, such as Tentacles™ molecules, which hold promise for precise targeting and reduced side effects in cancer immunotherapy.

Monday, the 10th of November, 2025
Jixing Liu, Ph.D.
A scientific illustration shows the process of a ligand docking with an RNA molecule. On the left is a complex 3D RNA structure, depicted as twisted ribbons and chains. On the right, a small molecule ligand is about to bind with it. The image uses blue and purple gradients to highlight the molecular interactions and conformational matching. This visual explains the core application of the IRIS tool: using machine learning to accurately predict and rank the binding poses of ligands to RNA targets in drug discovery, thereby improving docking accuracy.

Machine Learning Masters RNA Docking: IRIS Tool Boosts Pose Prediction Accuracy

Machine Learning
RNA-Ligand Docking
Pose Re-ranking
Drug Discovery
Computational Chemistry

IRIS is a machine learning tool for re-ranking molecular poses. By learning from a large dataset of real nucleic acid-ligand complexes, it significantly improves the accuracy of RNA-ligand docking predictions. The tool addresses the poor performance of traditional scoring functions on flexible targets. Its model, which combines multiple features, more reliably selects near-native structures from a large pool of poses, providing efficient and practical support for RNA-targeted drug discovery.

Sunday, the 9th of November, 2025
Jixing Liu, Ph.D.
A flowchart illustrating how the DiSE model works. On the left, various NMR spectral data inputs are shown, including MS, 1H NMR, 13C NMR, HSQC, and COSY. Arrows point from these inputs to the central component, a diffusion model labeled 'DiSE,' which is iteratively converting random noise into a clear 2D molecular structure through a denoising process. On the right, the final, precise molecular structure is displayed. The entire workflow demonstrates the automated AI-driven process of resolving complex spectral data into an accurate chemical structure.

DiSE: AI Decodes Molecular Structures by Thinking Like a Chemist

DiSE Model
Diffusion Model
Structure Elucidation
NMR Spectroscopy
AI Chemistry

The DiSE model mimics a chemist’s step-by-step reasoning process by integrating multiple types of NMR spectral data to achieve automated, precise elucidation of organic compound structures. At its core, it uses a Diffusion Model that progressively builds a molecular structure from random noise and spectral data, effectively avoiding the problem of generating invalid SMILES strings common in traditional methods. This approach not only improves prediction accuracy but also enhances the interpretability of the AI’s decision-making process, providing a powerful AI tool for automated laboratories and new drug development.

Saturday, the 8th of November, 2025
Jixing Liu, Ph.D.
A schematic showing the mechanism of protein interaction. The image depicts two protein molecular models touching and binding at specific amino acid residues. These key contact points are highlighted or connected by lines, illustrating how the PLMDA-PPI model's dual-attention mechanism first predicts residue contacts and then determines the overall interaction. The image visually represents the model's core idea of learning the physical basis of protein interaction from sequence information, which is closely related to the content of the first section, 'AI Predicts Protein Interactions: A Dual-Attention Mechanism Marks a New Step Forward'.

AI Predicts Protein Interactions: A Dual-Attention Mechanism Marks a New Step Forward

Protein-Protein Interaction
Dual-Attention Mechanism
PLMDA-PPI
Drug Discovery
Residue Contact Prediction

The PLMDA-PPI model, using an innovative dual-attention mechanism, not only predicts if proteins interact but also precisely pinpoints the key amino acid contact residues. This model doesn’t rely on sequence alignment. By learning the physical basis of interactions, it shows exceptional generalization ability when handling entirely new proteins. Its performance even surpasses the resource-intensive AlphaFold2-Complex, offering a highly efficient and accurate new tool for drug target screening and interaction interface design.

Friday, the 7th of November, 2025
Jixing Liu, Ph.D.
An instructional diagram showing the core idea behind the DL4Proteins tutorial. On the left, a student sits at a computer displaying a 3D protein structure and related code, symbolizing learning through an interactive Jupyter Notebook. On the right is an icon of a brain with a complex neural network inside, representing artificial intelligence and deep learning models like AlphaFold2. An arrow connects the two sides, signifying how the tutorial makes advanced AI technology easy to learn and use, lowering the barrier for students and researchers to enter the field of AI protein design. The entire image clearly communicates the tutorial's main theme of using AI tools for protein structure prediction and design.

AI for Protein Design: Meet DL4Proteins, a Free Interactive Tutorial

DL4Proteins
Jupyter Notebook
Google Colab
Protein Design
AI Tutorial

Researchers at Johns Hopkins University have developed DL4Proteins, a free online tutorial, to address the high computational resource barrier of AI protein design tools. Based on Jupyter Notebooks on the Google Colab platform, the tutorial allows students and researchers to learn and practice with advanced AI technologies like AlphaFold2 and diffusion models without needing a local GPU. Through interactive case studies and a structured curriculum, DL4Proteins aims to make AI more accessible in protein structure prediction and design, advancing bioinformatics education.

Thursday, the 6th of November, 2025
Jixing Liu, Ph.D.
An illustration showing the Pearl model's role in high-precision protein-ligand docking. On the left is a complex 3D protein structure in gray, with many folds and cavities. On the right, a small molecule ligand is precisely fitted into the protein's binding pocket, with glowing lines connecting them to symbolize exact interaction forces. The deep blue background highlights the molecular structures. The image effectively illustrates how the Pearl model can navigate and place every atom of a drug molecule in the correct position on the protein target, achieving sub-angstrom accuracy, which is crucial for drug design.

Pearl AI: Placing Every Atom with Precision, Surpassing AlphaFold 3

Molecular Docking
Drug Discovery
AI-Aided Drug Design
Protein Structure Prediction
Diffusion Models

The Pearl model from Genesis Molecular AI tackles the protein-small molecule docking challenge in drug discovery. Trained on massive synthetic datasets and built with a diffusion architecture that incorporates physical symmetries, it achieves outstanding performance on multiple benchmarks. Pearl particularly excels in sub-angstrom accuracy (RMSD < 1Å), an area critical for medicinal chemists, significantly outperforming top models like AlphaFold 3 and providing a powerful new tool for high-precision drug design.

Wednesday, the 5th of November, 2025
Jixing Liu, Ph.D.
A flowchart titled &#039;MetaPepticon' illustrates an automated bioinformatics workflow for discovering anticancer peptides from microbial genomic data. The diagram is divided into several main modules from left to right. On the left are the input data types, including genomes, metagenomes, and transcriptomes. In the middle are the core processing steps, such as data preprocessing, consensus prediction using multiple tools (AntiCP 2.0, ACPred-BMF, ConACP), and toxicity prediction. On the right is the final output: filtered, non-toxic candidate anticancer peptides. The chart clearly shows how MetaPepticon integrates a complex screening process into an automated pipeline, highlighting its efficiency and reliability.

Mining the Microbiome: MetaPepticon Automates Anticancer Peptide Discovery

Anticancer Peptides
Bioinformatics
Microbiome
Drug Discovery
Automated Pipeline

Screening for anticancer peptides (ACPs) in vast microbial genomic data is like finding a needle in a haystack. MetaPepticon is an innovative, fully automated bioinformatics pipeline that integrates multiple prediction algorithms and a toxicity screening mechanism. It uses a consensus strategy to significantly improve the efficiency and reliability of discovery. Researchers can use this open-source tool to quickly mine high-confidence drug candidates from massive sequencing data at scale, effectively shortening the early drug discovery cycle.

Tuesday, the 4th of November, 2025
Jixing Liu, Ph.D.
A flowchart titled &#039;CombiMOTS' depicts a computational method for generating dual-target molecules. On the left, inputs are shown, including two target proteins and a chemical fragment library like Enamine REAL. At the center is the core algorithm, 'Pareto Monte Carlo Tree Search (PMCTS)', which builds molecules through a series of decision steps (select, expand, simulate, backpropagate). The right side displays the output: a set of molecules on the Pareto front, representing the best balance across multiple objectives like affinity for both targets, drug-likeness, and synthesizability. The diagram visually explains how CombiMOTS integrates multi-objective optimization with synthesizability from the very beginning of molecule design to generate drug candidates with practical value.

AI Generates Dual-Target Molecules and Guarantees They Can Be Made

dual-target drugs
multi-objective optimization
CombiMOTS
drug design
synthesizability

CombiMOTS combines multi-objective optimization with fragment-based drug design to solve a core conflict in dual-target drug development: activity versus synthesizability. The method uses a Pareto Monte Carlo Tree Search (PMCTS) framework that builds synthesizability in from the start. By selecting chemical fragments from the Enamine REAL Space database, it ensures that every generated molecule has a clear synthetic route. This allows AI-assisted drug design to produce dual-target molecules that are both highly active and realistically synthesizable, effectively bridging the gap between computational design and lab synthesis.

Monday, the 3rd of November, 2025
Jixing Liu, Ph.D.
An explanatory diagram of how the DeepVul model works. The chart shows how a multi-task Transformer architecture integrates three different types of data: gene expression profiles, gene perturbation effects (like CRISPR knockouts), and drug perturbation effects. These three information streams are fed into a shared encoder and mapped to a unified latent space. This process forces the model to learn the common underlying biological principles they share. The diagram illustrates how DeepVul uses gene essentiality data to more accurately predict a cell&#039;s response to specific drugs, offering a new approach to precision medicine that goes beyond genetic mutations.

DeepVul: Predicting Drug Response with a Transformer, a New Path for Precision Medicine

Precision Medicine
Transformer
Drug Response Prediction
Transcriptome
Gene Essentiality

The DeepVul model overcomes the limitation of precision medicine’s reliance on specific gene mutations. It uses a Transformer architecture to analyze a cell’s transcriptome data, finding treatment options for a broader group of patients. By jointly learning gene expression profiles, gene perturbations, and drug perturbations, the model uncovers the core biological logic of cell survival. A key finding is that patterns learned from gene essentiality data are more universal than those from drug response data. This provides a more robust foundation for predicting drug sensitivity and offers good interpretability.

Sunday, the 2nd of November, 2025
Jixing Liu, Ph.D.
A concept diagram showing how the Thoth framework generates high-quality biological experiment protocols using a structured reward mechanism. On the left, a brain icon representing a Large Language Model (LLM) processes the input experimental goal. The middle section illustrates the &#039;Sketch-and-Fill' design process with a flowchart, including Analysis, Structuring, and Expression phases. On the right, an evaluation icon with a checkmark and a score represents the SCORE reward mechanism, used to validate the accuracy and executability of the generated protocol. The entire image clearly communicates the core idea of the Thoth framework: breaking down complex scientific tasks and using strict evaluation criteria to guide the AI in producing reliable results.

Thoth Framework Generates Biology Experiment Protocols with AI

Thoth Framework
AI Generation
Biology Experiment Protocols
Large Language Models
Structured Rewards

Introducing Thoth, an AI framework that designs and generates precise, executable biology experiment protocols, much like a scientist would. Thoth uses a unique “Sketch-and-Fill” paradigm and a structured reward mechanism (SCORE) to evaluate quality, ensuring the generated protocols are logical and actionable. Tests show Thoth outperforms general Large Language Models (LLMs) on multiple fronts, offering a new approach for complex scientific reasoning tasks.

Saturday, the 1st of November, 2025
Jixing Liu, Ph.D.
A concept diagram showing the application of artificial intelligence in the drug discovery process. The image includes icons of a DNA double helix, molecular structures, cells, and a brain, symbolizing the entire process from biological data to AI models to intelligent design.

From Designing Anti-Cancer Molecules to Speeding Up Molecular Docking

AI
Drug Discovery
Generative AI
Molecular Docking
Bioinformatics

This article covers three recent breakthroughs using AI to advance drug discovery. The first study combines predictive and generative AI to design small molecules that precisely target specific cancer cells. The second introduces the MATCHA model, which balances speed and accuracy in molecular docking using multi-stage flow matching. The third is an AI tool that automatically links amino acid residues mentioned in scientific papers to their 3D structures, greatly improving research efficiency.

Friday, the 31st of October, 2025
Jixing Liu, Ph.D.
A flowchart showing how the DeepChem-DEL framework processes DNA-Encoded Library (DEL) data. The diagram illustrates the complete workflow, from raw sequencing data input, through denoising steps like signal amplification and unification, to molecular representation using disynthons and trisynthons, and finally to the output of model predictions.

AI in Life Sciences: A Look at GQVis, DeepChem-DEL, and Odyssey

Artificial Intelligence
Bioinformatics
Genomics
Protein Design
Drug Discovery

This article explores three advances in AI for the life sciences. GQVis is a dataset that uses natural language to create professional genomics visualizations. DeepChem-DEL is an open-source framework that standardizes data analysis for DNA-Encoded Libraries (DELs). The Odyssey model uses an innovative “consensus” mechanism instead of attention to build a 100-billion-parameter language model that understands protein sequence, structure, and function. These tools and models are accelerating drug discovery and genomics research.

Thursday, the 30th of October, 2025
Jixing Liu, Ph.D.
A conceptual diagram showing how digital twin technology is used to predict drug side effects. On the left is an icon representing a drug molecule, and on the right is a circular icon representing a cell. An arrow connects them, symbolizing the interaction between the drug and the cell. The background is a complex network of molecular structures, highlighting the method&#039;s explainability based on biological mechanisms.

AI in Drug Discovery: From a Metabolomics Benchmark to New Parasite Drug Targets

Drug Discovery
Artificial Intelligence
Bioinformatics
Large Language Models
Computational Chemistry

This article brings together five recent advances in AI for the life sciences. It covers MetaBench, a benchmark for evaluating Large Language Models in metabolomics; a digital twin model for predicting and explaining drug side effects; the Slogen model, which unifies molecule generation and screening; Kinome-AI for accurately identifying activating kinase mutations; and a deep learning approach that decodes the “dark proteome” of a parasite to find new drug targets. These studies show how AI is accelerating drug discovery, improving precision medicine, and solving complex biological problems.

Wednesday, the 29th of October, 2025
Jixing Liu, Ph.D.
A conceptual image showing the fusion of artificial intelligence and healthcare. On the left is a stylized human brain filled with binary code and circuit patterns, representing AI and machine learning. On the right is a red DNA double helix, symbolizing genetics and biomedical data. Glowing lines connect the two, illustrating the use of AI technology to analyze genetic data to improve healthcare.

AI Accurately Predicts Breast Cancer Treatment Response, Optimizing Personalized Therapy

Machine Learning
Breast Cancer
XGBoost
Personalized Medicine
Gene Expression

This article explores how the XGBoost machine learning algorithm analyzes gene expression data from the I-SPY2 clinical trial to predict breast cancer patients’ response to neoadjuvant therapy (NAT). The model effectively identifies which patients are likely to benefit from specific drugs, achieving an AUC of up to 0.814, and reveals biological clues from key genes like ABDH1 and DENND1C. The research shows AI’s potential to advance personalized cancer treatment, optimize drug selection, and increase pathological complete response (pCR) rates.

Tuesday, the 28th of October, 2025
Jixing Liu, Ph.D.
A flowchart shows three levels for solving the few-shot molecular property prediction problem. The top level is &#039;Learning paradigm,' the middle is 'Model-level,' and the bottom is 'Data-level.' Each level lists specific technical methods, such as 'Data augmentation' at the data level, 'Pre-training' at the model level, and 'Meta-learning' and 'Multi-task learning' at the learning paradigm level.

AI in Drug Discovery: From Virtual Cells and Few-Shot Learning to Adaptive Virtual Screening

AI in Drug Discovery
Large Language Models
Few-Shot Learning
Virtual Screening
Meta-Learning

This article explores three frontier topics in AI for drug discovery. First, it explains how Large Language Models (LLMs) are speeding up the creation of “virtual cells” by acting as “oracles” and “agents.” Second, it reviews few-shot learning methods that address data scarcity in early-stage drug discovery, providing a roadmap across data, models, and learning strategies. Finally, it details the Drug-TTA paradigm, a new approach that uses test-time adaptation and meta-learning to significantly improve the accuracy and generalization of virtual screening.

Monday, the 27th of October, 2025
Jixing Liu, Ph.D.
A flowchart showing how the LLaDR framework works. On the left, an icon represents a Large Language Model (LLM), which generates text descriptions for entities in a knowledge graph. An arrow in the middle points to the right, where a network-like biomedical knowledge graph (KG) is shown. The nodes in the graph are enhanced and fine-tuned with the text generated by the LLM, ultimately used for downstream prediction tasks like drug repurposing.

New AI Frontiers: A Look at CGBENCH, DemoDiff, and LLaDR

Large Language Model
Biomedicine
Knowledge Graph
Drug Discovery
Molecular Design

This article explores three AI technologies in biomedicine and chemistry. The CGBENCH benchmark reveals that current Large Language Models are seriously flawed in interpreting clinical genetics literature, especially in judging evidence strength and hallucinating facts. The DemoDiff model achieves efficient few-shot molecular design by understanding molecular “functional groups,” outperforming much larger general-purpose models. The LLaDR framework combines the semantic understanding of Large Language Models with biomedical knowledge graphs, significantly improving the accuracy and robustness of drug repurposing predictions.

Sunday, the 26th of October, 2025
Jixing Liu, Ph.D.
A flowchart showing how the K-DREAM framework integrates a biomedical knowledge graph into a molecule generation model. On the left is the knowledge graph, containing relationships between genes, proteins, and diseases. In the middle is a Context Regressor Network (CRN) that connects knowledge to molecular structures. On the right is a diffusion model that generates molecules with specific biological functions, guided by the knowledge graph.

K-DREAM: Teaching AI Biology to Design Molecules

AI Drug Discovery
Generative AI
Protein Design
Diffusion Models
Knowledge Graphs

This article explores three AI technologies for drug discovery. K-DREAM integrates biomedical knowledge graphs into diffusion models for targeted, multi-target drug design. The DiffDock+UniDock pipeline combines generative AI with rapid scoring for large-scale, unbiased blind docking virtual screening. And the ADFLIP model uses all-atom discrete flow matching to design functional proteins that interact precisely with specific ligands or dynamic structures.

Saturday, the 25th of October, 2025
Jixing Liu, Ph.D.
A diagram showing the 3D structure of a protein. The image depicts a complex protein molecule chain in various colors (blue, green, red), folded and coiled in space to form unique alpha-helices and beta-sheets. The background is a dark grid that highlights the protein&#039;s 3D form, with abstract symbols representing data and graphs nearby, symbolizing the use of AI techniques like graph learning for protein analysis.

AI in Drug Discovery: A Look at Three New Techniques

Drug Discovery
AI for Science
Graph Learning
Diffusion Models
Multi-Reward Optimization

This article explores three techniques that use artificial intelligence to speed up drug discovery. The first one gives graph learning models a better understanding of biology by using functional community invariance (FCI) and 3D protein structure invariance (3-PSI) to create more accurate protein representations. The second uses a multi-reward optimization framework to balance binding affinity with drug-like properties, designing molecules that are closer to real drugs. The third uses a diffusion model that incorporates domain knowledge to generate high-quality data for pharmacokinetics (PK) and drug-target interactions (DTI), tackling the problem of data scarcity.

Friday, the 24th of October, 2025
Jixing Liu, Ph.D.
A diagram showing the workflow of the GraphBFN model. It depicts how a complex molecular structure is simplified into several core &#039;supernodes' using a 'graph coarsening' technique. The model then starts from this coarse framework and progressively refines it to generate the complete molecular graph.

From Molecule Generation to Target Prediction

AI Drug Discovery
Molecule Generation
Graph Neural Networks
Protein Binding
Computer-Aided Drug Design

This article covers three recent advances in AI drug discovery. The GraphBFN framework enables faster, more accurate molecular graph generation through hierarchical modeling and Bayesian Flow Networks. A large-scale experiment reveals that the difficulty of protein binding is primarily determined by the target’s intrinsic properties, offering a new perspective for drug design. And the MIPT framework uses lightweight, multi-level prompts to efficiently adapt pre-trained graph neural networks (GNNs) for specific molecular property prediction tasks, significantly improving model flexibility and accuracy.

Thursday, the 23rd of October, 2025
Jixing Liu, Ph.D.
A workflow diagram of the AF-CALVADOS model. It shows a multi-domain protein where stably folded domains are marked as rigid bodies, and low-confidence intrinsically disordered regions (IDRs) are marked as flexible linkers. The diagram illustrates how the model uses AlphaFold&#039;s predicted confidence scores to automatically distinguish between the rigid and flexible parts of a protein, enabling large-scale dynamic conformation simulations.

New AI Frontiers in Protein Research: From Disordered Protein Prediction to Dynamic Conformation Simulation

Protein Structure Prediction
AI-Aided Drug Design
AlphaFold
Bioinformatics
Machine Learning

This article introduces five new AI technologies and their latest applications in biopharma. It covers the PepTron model, which for the first time accurately predicts intrinsically disordered proteins (IDPs), opening new paths for drug development. We also look at the BiomarkerML tool, which lets biologists without coding skills use machine learning to find disease biomarkers. Then there’s the MECo framework, which uses code-based editing to let AI optimize molecules with the precision of a chemist. The AF-CALVADOS model combines AlphaFold with simulations to study the dynamic conformations of tens of thousands of human proteins at scale. And finally, ConforFold actively controls secondary structures to predict multiple protein conformations, providing a new tool for understanding dynamic drug targets.

Wednesday, the 22nd of October, 2025
Jixing Liu, Ph.D.
A conceptual image showing an AI analyzing a protein amino acid sequence. The AI&#039;s 'eyes' are focused on specific residues in the sequence, which are highlighted to represent the High Attention sites the model has identified as key regions determining the protein's function.

A New Way to Use AI: Decoding Protein Function with Attention

Protein Language Model
Attention Mechanism
Drug Design
AI Interpretability
ESM

How do protein language models (like ESM) “think” like experts? This article explores how we can analyze their “attention” mechanism to identify the key amino acid sites (HA sites) that determine a protein’s family and function. This method not only reveals the deep biological rules learned by AI models but also offers new ideas for annotating proteins of unknown function, designing drugs, and studying drug resistance mutations. It’s helping turn AI from a prediction tool into an interpretable “research partner.”

Tuesday, the 21st of October, 2025
Jixing Liu, Ph.D.
An infographic showing the concept of the SAM-DTI model. On the left is a 3D structure representing a drug molecule, and on the right is a 3D structure of a protein target. Arrows and lines connect them, highlighting key interaction areas, symbolizing how the model uses a spatial attention mechanism to predict drug-target binding.

Five Innovations in AI Drug Discovery: From Target Prediction to Chemical Space Exploration

AI in Pharma
Drug Discovery
Machine Learning
Computational Chemistry
AlphaFold

This article explores five recent research advances in AI-driven drug discovery. Topics include the SAM-DTI model, which uses a spatial attention mechanism to innovate drug-target prediction; a new strategy that improves machine learning efficiency with molecular docking scores; the Ensemblify tool, which converts AlphaFold’s uncertainty into dynamic protein conformational ensembles; the FusionCLM framework, which combines the strengths of multiple chemical language models; and a new unbiased random sampling method for chemical space that addresses data bias in AI model training.

Monday, the 20th of October, 2025
Jixing Liu, Ph.D.
A chart showing the performance of different fine-tuning methods on various tasks in the RoFt-Mol benchmark. The background is a dark blue grid, featuring multiple scatter plots and bar charts that use different colors (orange, blue, green) to distinguish fine-tuning strategies. The axes are clearly labeled with evaluation metrics, providing a direct visual comparison of model performance under specific conditions.

RoFt-Mol Benchmark: What’s the Best Way to Fine-Tune AI Models for Drug Discovery?

Drug Discovery
AI for Science
Molecular Models
Model Fine-tuning
Computational Chemistry

This article explores the challenges and strategies of fine-tuning AI molecular models. The RoFt-Mol benchmark reveals how task type (classification/regression) and data volume determine the optimal fine-tuning method, and shows that self-supervised pretrained models excel when paired with advanced methods like DWiSE-FT. We also cover how the QW-MTL framework uses quantum chemical descriptors to improve ADMET prediction, how SurGBSA’s AI surrogate model speeds up binding energy calculations by 6,500x, and how AutoPK and SESAME models offer innovative ways to automate data extraction and predict protein conformational changes.

Sunday, the 19th of October, 2025
Jixing Liu, Ph.D.
A flowchart showing the &#039;Hierarchical Solution Refinement' process of the Eigen-1 framework. It shows multiple initial solutions, with one selected as the 'anchor.' Other solutions act as 'peers' to review and revise it, ultimately producing a more refined solution through multiple iterations.

AI in Drug Discovery: 5 New Models from Molecule Generation to Scientific Reasoning

Drug Discovery
AI Pharma
Large Language Models
Molecule Generation
Reinforcement Learning

This article explores five AI models that are speeding up drug discovery and chemical research. We’ll look at how CAMT5 generates molecules by thinking like a chemist, how FragAtlas-62M is building an open-source fragment library, and how combining a large language model’s knowledge with molecular structure improves property prediction. We’ll also cover how the Eigen-1 multi-agent framework refines scientific reasoning and how ExMolRL blends phenotypic screening with target-based design to create new molecules that are both bioactive and drug-like.

Saturday, the 18th of October, 2025
Jixing Liu, Ph.D.
A scientific chart showing how AlphaTMB predicts survival benefits from immune checkpoint inhibitor therapy more effectively than traditional TMB. The survival curves in the chart clearly indicate that patients in the high-AlphaTMB group have a significantly higher survival rate than those in the low-AlphaTMB group, demonstrating its superiority as a biomarker.

AI in Precision Medicine: From Predicting Protein Expression and Immunotherapy Response to Interpreting Genetic Variants

Precision Medicine
Proteomics
Artificial Intelligence
Genomics
Bioinformatics

This article explores three forward-looking applications of artificial intelligence in precision medicine. First, fine-tuning a deep learning model with large-scale proteomics data can more accurately predict how genetic variants affect protein expression. Second, by incorporating AlphaMissense’s pathogenicity scores, a new biomarker called AlphaTMB can better predict a cancer patient’s response to immunotherapy. Finally, a benchmark study shows that VarChat, a specialized model, performs best at summarizing literature on genetic variants, but all AI tools still require final verification by human experts.

Friday, the 17th of October, 2025
Jixing Liu, Ph.D.
A flowchart shows how the ToolUniverse ecosystem works. On the left, a user inputs instructions in natural language. In the center is the ToolUniverse core, with &#039;Find Tool' and 'Call Tool' functions connected to a large library of tools (machine learning models, databases, etc.). On the right, the system automates the entire process from tool discovery and code generation to workflow execution, aiming to accelerate scientific research.

AI in Biomedical Research: Advances in AlphaFold, ToolUniverse, and Knowledge Graphs

AI Drug Discovery
AlphaFold
Knowledge Graph
Large Language Models
Bioinformatics

This article covers five recent studies in biomedical research. We’ll explore how AlphaFold-Multimer is upending our understanding of receptor structure and how ToolUniverse builds personalized AI scientists to speed up drug discovery. We’ll also look at the MOLERR2FIX benchmark, which reveals the limits of large language models in chemistry. Then, we’ll see how BindPred can predict protein binding affinity from sequence alone, and finally, examine a large-scale pharmacogenomic knowledge graph that provides a reliable foundation for AI to discover new links between drugs and diseases.

Thursday, the 16th of October, 2025
Jixing Liu, Ph.D.
A diagram showing how the DlRNA-BERTa model works. On the left is an RNA sequence, on the right is a small molecule structure. They are connected in the middle by a cross-attention mechanism, which outputs a predicted value for binding affinity. The entire process is end-to-end and does not require 3D structural information.

AI in Drug Discovery: From Sequence Prediction to VR Modeling

AI Drug Discovery
Computational Biology
Graph Attention Networks
Molecular Modeling
Virtual Reality

This article explores three recent breakthroughs in AI for drug discovery. The DlRNA-BERTa model accurately predicts RNA-drug binding affinity without 3D structures. The MolecularWeb ecosystem uses AR/VR to turn molecular modeling into an intuitive, collaborative experience. And a graph attention network, integrated with biological pathway knowledge, not only predicts cell responses but also infers gene interactions, opening new paths for AI-driven drug development.

Wednesday, the 15th of October, 2025
Jixing Liu, Ph.D.
A scientific illustration of the human hemoglobin molecule. It shows the complex quaternary structure, composed of four subunits (two alpha and two beta chains). Each subunit contains a red heme group at its center, which binds and transports oxygen. The background is blue, highlighting the protein&#039;s intricate folding.

AI in Drug Discovery: From Predicting Mutations and Quantifying Solubility to Synthesizable Molecular Design

Drug Discovery
AI
Machine Learning
Computational Chemistry
Explainable AI

This article covers three recent studies combining AI with biomedical research. We’ll look at using AI and protein structure knowledge to accurately predict the pathogenicity of hemoglobin mutations; an explainable AI model that quantifies key factors affecting drug solubility, like LogP and TPSA; and an innovative genetic algorithm, SynGA, that ensures AI-designed molecules are synthesizable by directly manipulating their synthetic routes, speeding up the drug discovery process.

Tuesday, the 14th of October, 2025
Jixing Liu, Ph.D.
A scientific diagram showing the docking process between two protein molecules, one large and one small. Arrows and lines depict the movement, rotation, and internal flexible changes of the molecules, illustrating the complex dynamics of flexible protein docking.

AI in Scientific Discovery: A Look at Three New Models for Drug Resistance, Drug Optimization, and Protein Docking

Artificial Intelligence
Bioinformatics
Drug Discovery
Machine Learning Models
Protein Docking

This article breaks down three recent studies using AI to tackle tough scientific problems. We’ll cover: 1) How a hybrid CNN and XGBoost model accurately predicts bacterial drug resistance from gene sequences; 2) How a new framework called AIM resolves conflicting goals in drug development by intelligently mediating gradients; and 3) How a hierarchical, adaptive diffusion model overcomes the computational hurdles of flexible protein docking. The goal is to show what AI can do in bioinformatics and drug discovery.

Monday, the 13th of October, 2025
Jixing Liu, Ph.D.
A diagram showing the BindFlow workflow, with elements like protein structures, small molecule ligands, a DNA double helix, and charts. It illustrates how the tool automates the process of calculating binding free energy.

AI in Drug Discovery: Three Advances in Gene Perturbation, Binding Energy, and Resistance Prediction

Drug Discovery
Computational Chemistry
AI
Free Energy Perturbation
Genetic Perturbation

This article explores three breakthroughs in drug discovery. The ALIGNED framework unifies genetic perturbation data and biological knowledge using neuro-symbolic learning for more accurate predictions. The BindFlow tool automates and provides free access to complex absolute binding free energy calculations, speeding up hit screening. Finally, new research shows how Free Energy Perturbation (FEP) can predict drug resistance mutations in flexible proteins (IDRs), offering a new way to design more durable drugs.

Sunday, the 12th of October, 2025
Jixing Liu, Ph.D.
A concept diagram showing the workflow of the DARKIN benchmark. On the left, a brain icon represents a protein language model (pLM) receiving kinase and substrate protein sequences as input. An arrow points to a matching process in the middle, which outputs a prediction score on the right. This score determines if the kinase and substrate are a match, thereby revealing the function of the &#039;dark kinase.'

AI Shines a Light on ‘Dark Kinases’: Using Protein Language Models to Find New Drug Targets

Dark Kinases
Protein Language Models
Zero-Shot Learning
New Drug Targets
Computational Biology

This article introduces the DARKIN benchmark, which uses protein language models (pLMs) for zero-shot learning to predict the phosphorylation substrates of “dark kinases” with unknown functions. This computational approach allows researchers to generate high-confidence lists of potential substrates for hundreds of dark kinases, offering an efficient computational roadmap to systematically explore their functions and discover new drug targets, overcoming the high costs of traditional high-throughput screening.

Saturday, the 11th of October, 2025
Jixing Liu, Ph.D.
A concept image showing the process of building a precise digital twin of a biological system from vast, messy biological data (like genomics and proteomics) through complex mathematical models and algorithms.

Frontiers in AI Drug Discovery: Digital Twins, Generative Models, and Large Language Models

AI Drug Discovery
Digital Twins
Molecular Dynamics
Generative Models
Large Language Models

This article explores three key directions in AI for drug discovery. First, we look at the challenges and solutions for building digital twins of biological systems using symbolic regression, sparse regression, and hybrid frameworks. Next, we cover how the BioMD generative model uses a hierarchical strategy to efficiently simulate all-atom molecular dynamics and predict ligand unbinding pathways. Finally, we show how instruction-tuning allows Large Language Models (LLMs) to accurately extract virus-host interaction data from full scientific papers using just a few samples, speeding up knowledge graph construction.

Friday, the 10th of October, 2025
Jixing Liu, Ph.D.
A diagram showing the design framework of the PAFlow model, titled &#039;Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number.' The image contains complex flowcharts and molecular structure diagrams, depicting the process from a target protein pocket to the generation of a molecule with high binding affinity.

AI-Driven Drug and Chemical Design: From Predicting Drug Resistance to Optimizing Chemical Plants

Molecular Machine Learning
Drug Design
Protein Language Models
Resistance Prediction
Chemical Process Design

This article explores three pieces of research showing how AI is changing drug discovery and chemical design. AMRscope fine-tunes a protein language model to predict bacterial resistance mutations, providing an early warning. PAFlow uses a flow matching framework to generate high-affinity drug molecules efficiently, setting a new standard for structure-based drug design. Finally, the article looks at the future of molecular machine learning, where AI will not just design single molecules but will integrate physical knowledge to design and optimize entire chemical processes, moving from analysis to system design.

Thursday, the 9th of October, 2025
Jixing Liu, Ph.D.
A concept diagram showing the workflow of the EffiChem framework. On the left is an icon representing a large chemical language model (like Molformer). An arrow in the middle points to a LoRA adapter module. On the right is a flask icon, symbolizing the final molecular property prediction task. The entire image highlights the core idea of parameter-efficient fine-tuning.

AI in Drug Discovery: Sequence-Based Prediction, Efficient Fine-Tuning, and Practical Strategies

Drug Discovery
AI Pharma
Computational Chemistry
Machine Learning
Large Language Models

This article explores three recent advances in AI for drug discovery. The LINKER model predicts protein-small molecule interactions using only sequence data, solving a common problem when 3D structures are unavailable. The EffiChem framework uses LoRA to efficiently fine-tune large chemical language models and combines them with classical descriptors for low-cost, high-performance molecular property prediction. Finally, a case study from an antiviral drug prediction competition reveals the best use cases for different strategies, like XGBoost and transfer learning, in predicting ADME and drug potency.

Wednesday, the 8th of October, 2025
Jixing Liu, Ph.D.
A diagram comparing two active learning strategies. On the left, traditional &#039;pool-based BALD' shows a massive pool of candidate molecules, where the model must evaluate every single one, a tedious process. On the right, the innovative 'BALD-GFlowNet' depicts a GFlowNet model directly generating new molecules with high uncertainty, bypassing the need to screen the entire pool, which appears much more efficient and direct.

The Frontier of AI in Pharma: From Generative Molecule Design to Predicting Clinical Efficacy

AI in Pharma
GFlowNet
Active Learning
Drug Discovery
Machine Learning

This article explores three forward-looking areas in AI-driven drug discovery. First, how the BALD-GFlowNet framework boosts virtual screening efficiency 2.5-fold by “generating” molecules instead of “screening” them. Second, how various strategies are tackling the challenge of synthesizing AI-designed molecules, bridging the gap from computation to reality. And third, how machine learning can combine molecular structure with real-world data to build models that predict a new drug’s potential efficacy before clinical trials even begin.

Tuesday, the 7th of October, 2025
Jixing Liu, Ph.D.
A flowchart showing how the SAGERank model works. On the left is the 3D structure of an input antibody-antigen complex. An arrow points to the middle, indicating a graph construction process that converts it into an atom-level graph representation. On the right is the graph neural network model, which outputs a predicted score to evaluate the accuracy of the docking pose.

AI in Drug Discovery: Three Advances in Antibody Docking, Cyclic Peptide Permeability, and More

Drug Discovery
Graph Neural Networks
Contrastive Learning
Computational Biology
Cyclic Peptides

This article explores three recent advances in AI for drug discovery. SAGERank uses graph neural networks to innovate antibody-antigen docking scores, solving the challenge of generalizing from small datasets. VECTOR+ provides a practical framework for generating novel molecules in low-data scenarios through property-enhanced contrastive learning. For cyclic peptide membrane permeability, a systematic benchmark shows that DMPNN performs best, but future breakthroughs will depend on integrating the dynamic 3D structures of molecules.

Monday, the 6th of October, 2025
Jixing Liu, Ph.D.
A flowchart showing a parallel workflow using High-Performance Computing (HPC) to accelerate Alzheimer&#039;s drug discovery. The diagram includes steps like molecular docking, molecular dynamics simulation, and energy minimization, using arrows and icons to illustrate how parallel processing shortens computation time.

AI in Drug Discovery: Three New Tools to Speed Up Target Screening and Drug Simulation

Drug Discovery
Computational Biology
Virtual Screening
HPC
Bioinformatics

This article introduces three computational tools designed to accelerate early-stage drug discovery. ProteinWeaver helps researchers quickly understand target function by visualizing protein interaction networks. A High-Performance Computing (HPC) workflow significantly shortens molecular docking and dynamics simulation times for Alzheimer’s drugs using parallel computing. And PlasmoDocking simplifies the complex virtual screening process for anti-malarial drugs into an easy-to-use, open-source web tool, making it much more accessible.

Saturday, the 4th of October, 2025
Jixing Liu, Ph.D.
A diagram illustrating the MolSculptor workflow. It depicts a molecular structure being &#039;sculpted' and optimized, surrounded by icons representing concepts like diffusion models, evolutionary algorithms, and active learning. This visualizes how the framework iteratively generates, evaluates, and selects molecules to meet complex design constraints.

AI in Drug Discovery: A Closer Look at MolSculptor, OTMol, and PertFormer

AI Drug Discovery
Computational Biology
Diffusion Models
Optimal Transport
Foundation Models

This article explains three AI tools for computational biology. MolSculptor combines diffusion models and evolutionary algorithms to design multi-target, highly selective drugs from scratch. OTMol uses optimal transport theory to solve long-standing chemical plausibility problems in molecular alignment. PertFormer is a multiomics foundation model that can predict the cellular outcomes of gene perturbations with zero-shot capability. Together, these tools are accelerating drug target discovery and development from different angles.

Friday, the 3rd of October, 2025
Jixing Liu, Ph.D.
An infographic showing the workflow of the Polypharmacology Browser PPB3 tool. On the left is a chemical molecule structure, with an arrow pointing to a schematic of a deep neural network model in the middle. This leads to a list of predicted targets on the right, which includes target names and prediction scores.

AI in Biology: Insights on Protein Evolution, Glycan Modeling, and Drug Target Prediction

Protein Mutation
AlphaFold 3
Drug Target Prediction
Computational Biology
Deep Learning

This article explores three key topics in computational biology. First, it reveals that natural proteins have evolved a strong robustness to mutations, likely to counteract errors in transcription and translation. Second, it assesses AlphaFold 3’s potential and current limits in predicting complex glycan structures, highlighting the need for result validation. Finally, it introduces PPB3, a free drug target prediction tool based on deep learning and the ChEMBL database, which helps identify a molecule’s polypharmacological effects to aid early-stage drug discovery.

Thursday, the 2nd of October, 2025
Jixing Liu, Ph.D.
An image of an automated chemistry lab. Several glass reaction vessels are connected by tubes and equipment. A robotic arm is manipulating one of the vessels. In the background, complex instruments and blue lighting suggest a process of precise, efficient chemical reaction optimization driven by AI.

AI in Chemistry: From Faster Molecular Simulations to Automated Reaction Optimization

Artificial Intelligence
Computational Chemistry
Molecular Dynamics
Large Language Models
Knowledge Graph

This article explores three applications of AI in chemistry. By reframing a machine learning force field as a Deep Equilibrium Model (DEQ), researchers boosted the speed and accuracy of molecular dynamics simulations. A standardized Chemotion knowledge graph is turning messy experimental data into structured, AI-readable information, paving the way for automated chemistry. Finally, the ChemBOMAS framework uses a large language model to guide Bayesian optimization, achieving a 96% yield in a wet-lab experiment—far surpassing human experts.

Wednesday, the 1st of October, 2025
Jixing Liu, Ph.D.
This study shows that a simple AI model, forced to learn real molecular interactions through a &#039;molecular dropout' training method, can outperform more complex models in predicting binding affinity.

AI Predicts Binding Affinity: Training Matters More Than the Model

AI Drug Discovery
Computational Chemistry
Molecular Generation
Diffusion Models
Machine Learning

Explore recent advances in AI drug discovery. ChemAI incorporates physical laws into its model, allowing it to predict reactions from chemical equations alone. An innovative “molecular dropout” training method shows that simple models can outperform complex ones by learning true interactions. And the CSGD diffusion model enables precise, multi-conditional control over molecular generation, like freely combining recipe ingredients.

Tuesday, the 30th of September, 2025
Jixing Liu, Ph.D.
DeepSEA is an explainable deep learning model that bypasses the limitations of sequence alignment to more accurately identify novel antibiotic resistance proteins and reveal their mechanisms.

DeepSEA: Using AI to Find Hidden Antibiotic Resistance Proteins

AI Drug Discovery
Deep Learning
Bioinformatics
Molecule Generation
Antibiotic Resistance

Explore recent progress in AI drug discovery. The CBYG framework uses Bayesian Flow Networks for more controllable 3D molecule generation. MolEncoder reveals a “sweet spot” for molecular language models, proving that better pre-training strategies are more effective than just scaling up. And DeepSEA, an explainable deep learning model, accurately identifies new antibiotic resistance proteins, overcoming the limits of traditional sequence alignment.

Monday, the 29th of September, 2025
Jixing Liu, Ph.D.
Massive sampling can improve AlphaFold&#039;s predictions for complex protein interfaces, but picking the best model from a huge pool remains a key challenge.

AlphaFold by Brute Force? How Massive Sampling Tackles Protein Complexes

AI Drug Discovery
AlphaFold
Large Language Models
Molecular Docking
Computational Chemistry

Exploring the frontier of AI in drug discovery. This article discusses how massive sampling boosts AlphaFold’s accuracy on complex protein interfaces, introduces LightChem, an AI assistant combining large language models with specialized chemistry tools, and explains a new method that uses water molecules as guides to solve the tricky problem of flexible glycan docking.

Sunday, the 28th of September, 2025
Jixing Liu, Ph.D.
The drGT model not only predicts a drug&#039;s effectiveness on cancer cells but also, like a seasoned researcher, points out the key genes influencing that effectiveness, thus revealing the AI's decision-making process.

drGT: An AI That Not Only Predicts Drug Efficacy but Also Explains Why

AI Drug Discovery
PROTAC
Protein Language Models
Graph Attention Networks
Explainable AI

Explore the frontier of AI in drug discovery. Learn how reinforcement learning is optimizing PROTAC molecular design, uncover the inner workings of how protein language models predict structure, and see how a graph attention network (drGT) not only predicts drug efficacy but also explains the key genetic factors behind it.

Saturday, the 27th of September, 2025
Jixing Liu, Ph.D.
A team of AI agents, a &#039;Discoverer' and a 'Judge,' turns competitive landscape analysis for drug assets from a multi-day manual task into an automated process that takes just a few hours.

AI Due Diligence: 2.5 Days of Work Done in 3 Hours

AI Drug Discovery
Bioinformatics
Large Language Models
Knowledge Graphs
AI Agents

This article explores five new applications of AI in the biomedical field. It covers TEMPL, a benchmark for evaluating molecular docking models; Bio-KGvec2go, a service that automates updates for biological knowledge graphs; Tensor-DTI, a drug-target prediction model focused on binding pockets; a system using large language models to interpret single-cell data; and an agent system that automates drug asset due diligence. These tools and methods are solving real challenges, from basic research to business analysis.

Friday, the 26th of September, 2025
Jixing Liu, Ph.D.
BioLab integrates multiple specialized foundation models and agents to achieve the first end-to-end, fully autonomous antibody design, from target discovery to wet lab validation, successfully optimizing Pembrolizumab.

BioLab: An AI Agent System for End-to-End Drug Discovery

AI Drug Discovery
Graph Neural Networks
AI Agents
Antibody Design

AI is reshaping drug discovery. From the bioSBM model predicting 3D genome structures to the BioLab agent system designing novel antibodies that outperform existing drugs, and Graph Neural Networks (GNNs) being applied across molecular property prediction and target discovery, artificial intelligence is accelerating the process at an unprecedented depth and scale. However, challenges in data, models, and interpretability remain.

Thursday, the 25th of September, 2025
Jixing Liu, Ph.D.
This isn&#039;t just another Q&A bot. STELLA is an AI agent that can evolve like a research team and even 'invent' new tools on demand.

STELLA: The Self-Evolving AI Research Agent That Builds Its Own Tools

AI Agents
Large Language Models
Drug Discovery
AI for Science
Benchmarking

This article explores recent progress in AI for scientific research. It introduces an AI system that programs automatically and outperforms experts, a research agent named STELLA that self-evolves and creates its own tools, and a new benchmark, BioML-bench, that reveals the engineering shortcomings of current AI agents. We’ll look at how AI is evolving from a tool into a partner and the core challenges it faces.

Wednesday, the 24th of September, 2025
Jixing Liu, Ph.D.
ADGSyn achieves a new level of performance and a leap in efficiency for predicting anticancer drug synergy, thanks to its innovative dual-stream attention GNN and AMP training.

ADGSyn is Here! Drug Synergy Prediction is Now 3x Faster, More Accurate, and More Efficient

AI Drug Discovery
Drug Synergy
Graph Neural Network
AlphaFold
Single-Cell Analysis

Exploring the frontier of AI in drug discovery. This article looks at three breakthroughs: using AlphaFold to map the structures of hydrophobic proteins on a large scale and reveal evolutionary relationships; the OmniPert model, which unifies the prediction of gene and drug perturbation responses at the single-cell level; and ADGSyn, which uses an innovative graph neural network to boost the prediction efficiency of anticancer drug synergy by three times while maintaining high accuracy.

Tuesday, the 23rd of September, 2025
Jixing Liu, Ph.D.
MolReasoner uses a two-step &#039;lecture and tutor' method to force AI to stop memorizing and start thinking like a chemist. This makes its answers not only more accurate but, for the first time, explains 'why.'

AI Stops Memorizing Answers: MolReasoner Teaches It to Think Like a Chemist

AI Drug Discovery
Deep Learning
Computational Biology
Protein Interactions

Explore the latest AI advances in drug discovery, from predicting the physical details of protein binding to solving the cold-start problem in personalized medicine. This article covers several new models. The AlphaPPIMI framework, for instance, improves the accuracy of predicting PPI modulators by fusing multimodal pretrained models. Meanwhile, MolReasoner uses a two-step training method to teach AI chemical reasoning, making its predictions not only accurate but also explainable.

Monday, the 22nd of September, 2025
Jixing Liu, Ph.D.
GPT-5 shows impressive performance on biomedical Q&A, but it still has limitations in tasks requiring high precision. This suggests a future where general large models and domain-specific models are used together.

GPT-5 in Biomedicine: A New Turning Point for AI Drug Discovery?

AI Drug Discovery
Large Language Models
Bioinformatics
Causal Inference
Scientific Discovery

Artificial intelligence is evolving from a pattern-recognition tool into a true scientific partner. Through explainable causal inference models, independent hypothesis generation, and precise prediction of molecular activity, AI is not only accelerating biomedical R&D but also beginning to reason like a scientist. This signals a fundamental change in how we discover science.

Sunday, the 21st of September, 2025
Jixing Liu, Ph.D.
The Boltz-ABFE framework is the first to successfully use AI-predicted protein structures for high-accuracy Free Energy Perturbation (FEP) calculations.

AI Finds a Path for FEP, Freeing It from Crystal Structure Dependency

AI Drug Discovery
Molecular Modeling
Protein Design
Large Language Models
Computational Chemistry

This article explores five new developments in AI for drug discovery. We cover a new diffusion model (SLM) that generates biological sequences, a way to get high-accuracy free energy calculations from AI-predicted structures (Boltz-ABFE), and a method for predicting protein-peptide binding without known structures. We also look at a framework that combines a large language model with Bayesian optimization for lead optimization (AutoLead), and a practical trick to improve the rotational invariance of 3D molecular models.

Saturday, the 20th of September, 2025
Jixing Liu, Ph.D.
PETIMOT is an AI framework based on a graph neural network. It can rapidly and accurately infer complex protein dynamics directly from sparse, static 3D structures, providing a transformative tool for conformational analysis in drug discovery.

PETIMOT: An AI that Predicts Protein Dynamics, Incredibly Fast

AI Drug Discovery
Graph Neural Network
Interpretable AI
Molecular Generation
Protein Dynamics

This article covers three recent AI developments in drug discovery. PETIMOT uses an equivariant graph neural network to quickly predict protein dynamics from static 3D structures, offering an efficient tool for finding cryptic pockets and allosteric sites. Another study shows how an interpretable machine learning model can accurately predict inhibitor selectivity and provide chemical design ideas. Lastly, SiMGen achieves zero-shot molecular generation through a local similarity principle, challenging the reliance on big data.

Friday, the 19th of September, 2025
Jixing Liu, Ph.D.
PROTAC-Splitter uses a dual-model hybrid strategy to reliably and automatically deconstruct PROTAC molecules, paving the way for large-scale data analysis and molecular design.

AI Precisely Deconstructs PROTACs, Freeing Up Medicinal Chemists

AI Drug Discovery
Computational Chemistry
PROTAC
Explainable AI
Machine Learning

AI is evolving from a prediction tool into an intelligent partner for chemists. A new method makes AI explanations useful by suggesting chemically sound replacements, while automated tools like PROTAC-Splitter use a hybrid model strategy to handle tedious tasks like molecular deconstruction. This shift, from manually engineering features to letting AI learn on its own, is profoundly changing the future of chemical research.

Thursday, the 18th of September, 2025
Jixing Liu, Ph.D.
DiffGui builds chemical bonds while generating atoms, guided by binding affinity and drug-likeness. This makes the AI-generated molecules chemically valid structures, not just clouds of atoms.

DiffGui: Building Molecules Like an Architect, Drawing the Skeleton Before Adding the Flesh

AI Drug Discovery
Bioinformatics
Large Language Models
Generative AI
Multimodal

Let’s explore five recent advances in AI for biomedicine. Find out how AIDO.ModelGenerator simplifies multimodal model development with YAML, how ChatMDV turns natural language into visualization code, and how DiffGui generates chemically sound molecules. We also look at DGDM, a self-evolving drug discovery framework, and analyze the true capabilities of general-purpose large models on classic bioinformatics problems.

Wednesday, the 17th of September, 2025
Jixing Liu, Ph.D.
OwkinZero used reinforcement learning on a high-quality, verifiable Q&A dataset to prove that small, specialized AI models can beat larger, more general commercial models at biological reasoning.

AI for Biological Reasoning: Small Models Beat Big Models

AI Drug Discovery
Deep Learning
Large Language Models
Molecular Dynamics
ADME Prediction

OwkinZero shows that small models with focused, high-quality data can outperform large, general-purpose models in biological reasoning. At the same time, a generative AI model called HemePLM-Diffuse speeds up molecular dynamics simulations by 100x, directly generating dynamic processes. Another study finds that deep learning wins in complex ADME prediction, but classical methods are just as effective for potency prediction. Data quality, it turns out, is more critical than model choice.

Tuesday, the 16th of September, 2025
Jixing Liu, Ph.D.
We finally have a mathematical tool that speaks chemistry&#039;s native language, allowing AI to see molecules not as a collection of atoms, but as a network of connections.

Graph Neural Networks: The Universal Language of Chemistry

AI Drug Discovery
Graph Neural Networks
Molecule Generation
AI Agents
Physics-Based Energy Models

AI is changing drug discovery. An autonomous AI agent has designed a dual-target antibiotic from scratch. Graph neural networks are teaching AI the language of chemical bonds, moving beyond simple atom counts. And the physics-based IDFlow model guides AI to generate stable, realistic 3D molecular structures, bridging the gap between digital design and physical reality.

Monday, the 15th of September, 2025
Jixing Liu, Ph.D.
Boltzina uses a fast but rough tool for initial work, followed by a slow but precise AI for the final evaluation, balancing speed and accuracy in virtual drug screening.

Boltzina: Making AI Drug Screening Both Fast and Accurate

AI Drug Discovery
Virtual Screening
Protein Function Prediction
Oncology
Multimodal AI

This article explores recent AI advances in drug discovery. PARSE precisely predicts protein function by analyzing local microenvironments; the Honeybee framework reveals that clinical text can be more predictive than genomic data in cancer research; and Boltzina combines fast and accurate AI models for efficient virtual drug screening.

Sunday, the 14th of September, 2025
Jixing Liu, Ph.D.
Chemsmart is an open-source &#039;Swiss Army knife' toolkit designed to automate the tedious manual processes in computational chemistry, letting scientists focus on scientific thinking instead of file format conversions.

Chemsmart: Saying Goodbye to the ‘Manual Workshop’ of Computational Chemistry

AI Drug Discovery
Computational Chemistry
Generative Models
Reinforcement Learning
Molecular Docking

This article explores five new developments in AI drug discovery. They cover automating computational chemistry, generating molecules with physics constraints, predicting drug approval with explainable AI, improving docking accuracy with experimental data, and a new way to generate molecules by learning their complete distribution. These techniques are moving AI from simply imitating to truly understanding, making it a powerful engine for drug discovery.

Saturday, the 13th of September, 2025
Jixing Liu, Ph.D.
AlphaFold3 has evolved, now including non-protein molecules in its predictions for the first time. But official CASP16 results show it isn&#039;t invincible, especially with complexes, RNA, and hard targets. Older methods paired with clever strategies can still compete.

AlphaFold3 at CASP16: Evolved, but Not Yet a God

AI Drug Discovery
Protein Structure Prediction
AlphaFold3
Molecular Generation
Diffusion Models

AI is speeding up drug discovery. A new model, UTGDiff, can design new molecules directly from a chemist’s text instructions. AlphaFold3 is more versatile but, as CASP16 showed, it has weaknesses in predicting complexes and RNA. Meanwhile, ESMDynamic can efficiently predict the dynamic behavior of proteins, offering a new way to understand how these molecular machines work.

Friday, the 12th of September, 2025
Jixing Liu, Ph.D.
Using reinforcement learning, Retro-Expert trains a large model in a &#039;sandbox' of expert models to think like a chemist. It not only provides a retrosynthesis route but also explains its reasoning.

A New Way for AI Chemists: Retro-Expert Makes Retrosynthesis Explainable

AI Drug Discovery
Computational Chemistry
Retrosynthesis
VAE
Drug Design

Explore how AI is reshaping drug discovery and chemical synthesis. This article covers three new directions: making retrosynthesis explainable with reinforcement learning, a practical guide to using VAEs for multi-omics data integration, and a fresh look at how “promiscuous” molecular fragments can help find new drug targets.

Thursday, the 11th of September, 2025
Jixing Liu, Ph.D.
Researchers are building a virtual human driven by a team of AI experts to simulate a drug&#039;s complete journey, from molecule to whole-body effects, tackling the biggest gap in drug development.

The AI Virtual Human: Drug Discovery’s Ultimate Flight Simulator

AI in Pharma
Drug Discovery
Machine Learning
Computational Chemistry
Large Language Models

This article covers three new advances in AI-powered drug discovery. They include using machine learning to quickly predict hydrogen bond strength, building an AI agent framework for a “virtual human” that simulates drug effects across the entire body, and using “knowledge prompts” to give AI a deeper understanding of molecular structures to speed up the discovery of new medicines.

Thursday, the 11th of September, 2025
Jixing Liu, Ph.D.
Researchers trained a model to do two things at once: reconstruct molecules and predict their properties. From this, they created an &#039;unfamiliarity' score that can both identify truly novel molecules and tell us when to be skeptical of an AI's predictions.

When AI Learns to Say ‘I Don’t Know’

AI Drug Discovery
Generative AI
Large Language Models
Protein Structure
Bioinformatics

Let’s explore the latest advances in AI for life sciences. This post covers five key studies: BioEmu uses generative AI to simulate protein dynamics, speeding up drug discovery; a new model learns to spot its own knowledge gaps, making its predictions more reliable; Mol-R1 teaches AI to reason like a chemist; the PEPBI database gives AI high-quality protein-peptide binding data; and rbiol pioneers training AI with “virtual cells.”

Wednesday, the 10th of September, 2025
Jixing Liu, Ph.D.
MD-LLM-1 isn&#039;t a camcorder for traditional molecular dynamics. It's more like a director that can imagine a protein's unseen conformations, leaping over huge energy barriers to see the landscape on the other side.

AI Tackles Molecular Dynamics: Predicting the Other Side of Proteins

AI Drug Discovery
Machine Learning
Molecular Dynamics
Binding Free Energy
Molecular Glues

This article explores three forward-looking applications of AI in drug discovery. We’ll see how MD-LLM-1 uses a Large Language Model to predict rare protein conformations, how machine learning speeds up binding free energy calculations for metal-based drugs, and how MolGlueDB creates the first dedicated database for molecular glue degraders, providing powerful tools and data for new drug discovery.

Tuesday, the 9th of September, 2025
Jixing Liu, Ph.D.
By learning a Chain-of-Thought, PepThink-R1 not only provides results when optimizing cyclic peptides but also explains why it made those changes, making the AI&#039;s decision process transparent.

PepThink-R1: A Cyclic Peptide Optimization AI That Can ‘Think’

AI Drug Discovery
Generative Models
Large Language Models
Graph Neural Networks
Bioinformatics

Explore recent advances in AI drug discovery, including MLFGNN for molecular property prediction, MolSnap as a fast generator challenging diffusion models, and PepThink-R1, a cyclic peptide optimization AI that can explain its design logic. These models are changing how we approach molecular discovery and bioinformatics analysis.

Monday, the 8th of September, 2025
Jixing Liu, Ph.D.
Instead of treating proteins like rigid locks, Q-MOL models them as flexible energy landscapes. This allows it to find hidden allosteric valleys on seemingly flat surfaces, achieving an impressive 36% hit rate in real-world screening.

Q-MOL: Mapping Protein ‘Terrains’ to Find Drugs

AI Drug Discovery
Protein Structure
Computational Drug Design
Bioinformatics

This article covers three recent AI developments in drug discovery. The CAML model accurately predicts protein function by analyzing a gene’s “neighbors.” Q-MOL models proteins as flexible “energy landscapes,” discovering hidden drug-binding sites with a high hit rate. The article also discusses how future drug combination predictions must shift from 2D pathway maps to 3D protein structures, highlighting AI’s central role in understanding molecular interactions.

Sunday, the 7th of September, 2025
Jixing Liu, Ph.D.
By modeling molecular docking as a game where the protein and ligand adapt to each other, a new algorithm improves both speed and accuracy in the difficult problem of flexible docking.

A New Game in Molecular Docking: Let the Protein and Ligand Play Against Each Other

AI Drug Discovery
Molecular Docking
Medical NLP
Chemical Reaction Prediction
AI for Science

This article explores the latest advances in AI for drug discovery. It introduces MedTE, a medical NLP model that understands complex medical records; a new algorithm that treats molecular docking as a game; and Uni-Mol3, a foundation model that predicts 3D chemical reactions. Together, they show how AI is accelerating drug discovery from text comprehension and molecular docking to reaction prediction.

Saturday, the 6th of September, 2025
Jixing Liu, Ph.D.
AgentD tries to package the complex computational work of early drug discovery into an LLM-driven automated workflow. It&#039;s a cool idea, but the real test is how well each module performs on its own and how much chemical intuition they have when combined.

AgentD: An AI Drug Discovery Pipeline. Can It Really Do It All with One Click?

AI Drug Discovery
Molecule Generation
Large Language Models
Binding Site Prediction
Computational Chemistry

This is a quick look at three recent papers in AI drug discovery. FRATTVAE generates complex molecules using fragments and tree structures, just like a chemist. AgentD uses a Large Language Model to build an automated drug discovery pipeline. And VN-EGNNrank accurately pinpoints protein binding sites using ligand information. Together, they show AI’s huge potential in molecule generation, workflow automation, and target discovery.

Thursday, the 4th of September, 2025
Jixing Liu, Ph.D.
By borrowing models trained on big data, researchers have efficiently screened potential new antibiotics for specific bacterial strains where data is scarce, offering a new approach to the difficult problem of drug resistance.

The Frontier of AI Drug Discovery: Uncovering Protein Dynamics, Fighting Superbugs, and Solving Activity Cliffs

AI Drug Discovery
AlphaFold2
Transfer Learning
Hyperbolic Space
Protein Conformation

Explore the latest advances in AI for drug discovery. The CF-random method reveals proteins’ hidden multiple conformations by reverse-engineering AlphaFold2. Transfer learning enables efficient screening of new antibiotics for data-scarce bacteria. The HypSeek model uses hyperbolic geometry to accurately predict and solve the “activity cliff” problem in drug discovery.

Wednesday, the 3rd of September, 2025
Jixing Liu, Ph.D.
NovoMolGen, the largest study on molecular language models to date, found little correlation between pre-training metrics and downstream task performance. A small model with only 32 million parameters performed just as well as much larger ones.

AI in Drug Discovery: A Deep Dive into Five Recent Papers

AI Drug Discovery
Molecular Language Model
Multimodal Learning
Property Prediction
Knowledge Graph

This article breaks down five recent advances in AI for drug discovery. Topics include building biomedical knowledge graphs with multimodal fusion and contrastive learning, tackling data scarcity in ADMET prediction with multi-task learning, and improving molecular property prediction using vision-language models. We also explore the scaling laws of molecular language models and introduce a new framework for fusing diverse molecular data.

Tuesday, the 2nd of September, 2025
Jixing Liu, Ph.D.
Computational methods and AI are bringing new weapons to bear on undruggable targets, from allosteric modulation to protein degradation. The rules of the game are changing.

AI in Drug Discovery: From Repurposing Old Drugs to Tackling ‘Undruggable’ Targets

AI Drug Discovery
Drug Repurposing
Graph Neural Network
Contrastive Learning
Undruggable Targets

Let’s look at the latest in AI drug discovery. DREAM-GNN is using a dual-channel graph neural network to revolutionize drug repurposing. DecoyDB is significantly improving binding affinity prediction with a high-quality dataset of “wrong answers” for contrastive learning. And computational methods and AI are making “undruggable” targets druggable through new strategies like allosteric modulation and targeted protein degradation.

Monday, the 1st of September, 2025
Jixing Liu, Ph.D.
The Prompt-to-Pill framework uses a centrally coordinated group of AI agents to try to integrate the long, fragmented process from molecular design to virtual clinical trials into a single automated pipeline.

From Automated Literature Extraction and Full-Process Agents to Real-World Physical Validation

AI Drug Discovery
Large Language Models
AI Agents
Computational Chemistry

The ReactionSeek framework uses Large Language Models (LLMs) and clever prompt engineering to automatically extract information from chemistry literature with high accuracy. The “Prompt-to-Pill” framework builds a team of agents coordinated by a central AI, trying to merge the long process from molecular design to virtual clinical trials into a single automated pipeline. Finally, a nested learning framework successfully enabled a generative AI to design novel drug molecules that were synthesizable and effective in wet-lab experiments by introducing a “reality check” from physics.

Friday, the 29th of August, 2025
Jixing Liu, Ph.D.
MEVO cleverly combines molecule generation with evolutionary strategies, training its model on vast amounts of &#039;unbound' ligand data to open a new path for structure-based drug design.

Frontiers in AIDD: From Water Molecules in Docking to the Language of Chemists

AI Drug Discovery
Molecular Docking
Computational Chemistry
Explainable AI

Schrödinger’s new Glide WS significantly improves molecular docking accuracy by explicitly handling water molecules in the binding site. A new method uses neural networks to iteratively find reaction pathways, while MEVO combines molecule generation with evolutionary strategies. ChemDFM-R helps AI understand functional groups, and another study shows Large Language Models can speed up synthetic health data creation, but expert human oversight remains essential.

Wednesday, the 27th of August, 2025
Jixing Liu, Ph.D.
When predicting a protein designed to fold into two different conformations, AlphaFold3 often tries to cram both structures together, generating a physically impossible stitched-together monster and revealing its fundamental limitation in handling sequence ambiguity.

From Molecular Language and Virtual Labs to AlphaFold3’s Stumble

AI Drug Discovery
Large Language Models
Protein Engineering
Research Tools

SmilesT5 shows that smart training beats brute-force data. A Nature paper presents a “virtual lab” run by AI agents that successfully designed a COVID nanobody. At the same time, research reveals AlphaFold3’s fundamental limits in handling conformational ambiguity, while SKiM-GPT and sequence-based prediction models offer powerful new tools for validating literature hypotheses and discovering PPI drugs.

Wednesday, the 27th of August, 2025
Jixing Liu, Ph.D.
UniDock-Pro integrates structure-based, ligand-based, and an innovative hybrid virtual screening method on a single, GPU-accelerated platform. It can screen over a million molecules per day on a single GPU, significantly improving enrichment efficiency.

GPU-Accelerated Virtual Screening, AlphaFold3 for Preclinical Research, and an AI Force Field for Classical Simulations

AI Drug Discovery
Virtual Screening
Protein Structure Prediction
Molecular Dynamics

UniDock-Pro uses GPUs to combine structure-based, ligand-based, and a new hybrid virtual screening method on a single platform, achieving unprecedented screening efficiency. XenoSignal uses AlphaFold3 to predict human-mouse cell interactions in xenograft models with atomic-level accuracy, which will deeply impact how we interpret preclinical research. Finally, TorchANI-Amber uses a clever interface to let the classical molecular dynamics software Amber seamlessly use AI neural network force fields, enabling large-scale biomolecular simulations with quantum-level accuracy.

Wednesday, the 27th of August, 2025
Jixing Liu, Ph.D.
The Seekrflow platform seamlessly integrates machine learning force fields with enhanced sampling methods, enabling end-to-end, automated, high-throughput prediction of drug-target binding kinetics and thermodynamics.

From Topological Vision and Dynamics Prediction to a Corporate AI Architecture

AI Drug Discovery
Computational Chemistry
Graph Neural Networks
Molecular Dynamics

PACTNET gives AI models the ability to see the overall “topological shape” of a molecule, which significantly improves the prediction accuracy of graph neural networks. The Seekrflow platform seamlessly integrates machine learning force fields with enhanced sampling methods, enabling automated, high-throughput prediction of drug-target binding kinetics. Finally, the H-MoE framework creates a “corporate” AI architecture that intelligently assigns different molecules to “expert models,” solving the challenge of representing vast chemical space in a way that’s closer to a human chemist’s intuition.

Wednesday, the 27th of August, 2025
Jixing Liu, Ph.D.
Using de novo computational design, researchers successfully engineered the difficult membrane protein CCR8 into a water-soluble form while preserving its key antibody-binding site, removing a major obstacle in GPCR drug development.

The Frontier of AI Drug Discovery: Soluble GPCRs, Explainable GNNs, and Precise Molecular Generation

AI Drug Discovery
Protein Engineering
Explainable AI
Molecular Generation

Researchers used de novo computational design to successfully convert the membrane protein CCR8, a popular target, into a water-soluble protein, clearing a major hurdle for GPCR drug development. The SEAL model addresses the “black box” problem of Graph Neural Networks (GNNs) in molecular property prediction with an innovative fragment-wise attribution method, making its results intuitive and trustworthy. Finally, the CMCM-DLM framework uses plug-and-play control modules to achieve precise dual control over a molecule’s structure and properties without retraining the model.

Monday, the 25th of August, 2025
Jixing Liu, Ph.D.
By combining graph wavelet transforms and contrastive learning, GHCDTI not only improves the accuracy of drug-target interaction prediction but is also the first model to account for the dynamic flexibility of proteins.

Three Papers in AI Drug Discovery: Liver Injury, Target Prediction, and ADME

AI Drug Discovery
Graph Neural Networks
Computational Chemistry
ADME Prediction

DILIGeNN improves drug-induced liver injury (DILI) prediction by adding detailed 3D chemical features to graph neural networks. GHCDTI is the first to account for protein flexibility in drug-target interaction prediction using graph wavelets and contrastive learning. And a Merck study shows that combining public and internal data with multi-task learning can effectively improve the generalization of ADME models.

Monday, the 25th of August, 2025
Jixing Liu, Ph.D.
LABind&#039;s new approach to predicting protein binding pockets considers ligand information, moving beyond blind guessing to find the best-fitting seat for each unique ligand.

Three New AI Tools for Medicinal Chemistry: GP-MOBO, LABind, and B3clf

Multi-Objective Bayesian Optimization
Tanimoto Kernel
Cross-Attention
Resampling Techniques
Blood-Brain Barrier Prediction

Three new algorithms are being applied to medicinal chemistry. GP-MOBO uses a Tanimoto kernel for efficient multi-objective molecular optimization. LABind uses a cross-attention mechanism for ligand-aware protein binding pocket prediction. And B3clf improves blood-brain barrier prediction accuracy with smart resampling techniques. These tools help chemists screen high-quality candidate molecules faster, speeding up drug development.

Monday, the 25th of August, 2025
Jixing Liu, Ph.D.
NetMD combines graph theory and time warping to give us a powerful new tool that can automatically align and compare multiple molecular dynamics trajectories without presets.

Automated agents, multi-ancestry target discovery, and a new way to look at molecular dynamics

AI Drug Discovery
AI Agents
Computational Biology
Molecular Dynamics

The FROGENT AI agent tries to automate the entire workflow from target discovery to molecule design. The BLISS framework cleverly uses public pQTL summary data to build powerful multi-ancestry protein prediction models, unlocking new drug targets. And NetMD offers a new unsupervised solution for comparing and analyzing complex molecular dynamics trajectories using graph embedding and time warping.

Sunday, the 24th of August, 2025
Jixing Liu, Ph.D.
A concept diagram of the ESM-LoRA-Gly method, showing how tiny LoRA patch matrices are added to and trained on the ESM2 protein language model. This allows for efficient fine-tuning of the model for glycosylation site prediction, improving accuracy while reducing computational resource requirements.

Large Language Models’ Blind Spots, Lightweight Fine-Tuning, and Designing Antibiotics from Scratch

AI Drug Discovery
Large Language Models
Generative AI
Protein Engineering

An evaluation of nine major LLMs reveals they still struggle to understand complex biochemical reaction networks. Meanwhile, ESM-LoRA-Gly uses LoRA to efficiently fine-tune a large protein model, significantly improving glycosylation site prediction while cutting computational costs. Finally, a study successfully used generative AI to design entirely new antibiotics that proved effective in animal models, showing AI’s potential to tackle the tough problem of drug resistance.

Saturday, the 23rd of August, 2025
Jixing Liu, Ph.D.
The Model Quality Assessment (MQA) competition at CASP16 shows that after AI predicts massive numbers of high-quality structures, the real challenge has shifted from predicting to selecting—how to pick the conformation closest to reality from thousands of similar models.

From Structure Selection and Virtual Docking to Small-Data Molecule Generation

AI Drug Discovery
Protein Structure Prediction
Virtual Screening
Molecule Generation

The CASP16 assessment revealed a new post-AlphaFold challenge: shifting from structure prediction to selecting the best conformation from countless high-quality models. HelixVS improves virtual screening efficiency and accuracy by combining deep learning rescoring with traditional molecular docking. Finally, the IBEX framework tackles high-quality molecule generation from extremely sparse data by merging information theory with physics-based optimization.

Saturday, the 23rd of August, 2025
Jixing Liu, Ph.D.
A concept image of an automated chemistry lab controlled by a Large Language Model (LLM). A central brain, representing the AI, analyzes data and directs robotic arms to perform precise chemical reactions in a fume hood, symbolizing AI&#039;s move from theoretical prediction to a closed-loop system of autonomous synthesis.

Autonomous Synthesis, Open-Source Structure Prediction, and a 69-Billion-Molecule Screen

AI Drug Discovery
Autonomous Synthesis
Protein Structure Prediction
Virtual Screening

Large Language Models (LLMs) are evolving from theoretical predictors into “field commanders” directing robotic autonomous chemical synthesis. The Baker lab’s open-source AtomWorks framework and RF3 model are closing the gap with AlphaFold3 while providing modular tools to accelerate the entire field. Lastly, the open-source AdaptiveFlow platform efficiently screened a 69-billion-molecule library to discover a nanomolar inhibitor for the difficult target FSP1 and solve its first-ever co-crystal structure.

Wednesday, the 20th of August, 2025
Jixing Liu, Ph.D.
A concept diagram of the Chem3DLLM model, showing how reversible compression encoding translates a complex 3D molecular structure into a text sequence that a Large Language Model (LLM) can understand and process, enabling end-to-end structure-based drug design.

Interactive Genomics, 3D Molecular LLMs, and High-Efficiency Virtual Screening

AI Drug Discovery
Large Language Models
Virtual Screening
Computational Chemistry

ChemGenXplore packages complex chemogenomic analysis into an interactive platform friendly to bench scientists. Chem3DLLM, through an innovative reversible compression encoding, allows a Large Language Model (LLM) to directly understand and generate 3D molecular structures for the first time. Finally, AuroBind aligns structural and bioactivity data to achieve an astonishingly high hit rate in virtual screening, successfully discovering picomolar-level active molecules for orphan targets.

Wednesday, the 20th of August, 2025
Jixing Liu, Ph.D.
A concept diagram of the CROP architecture, showing how a Large Language Model (LLM), through an efficient prefix mechanism, can process a one-dimensional SMILES text sequence while also receiving and integrating a molecule&#039;s two-dimensional topological graph and three-dimensional spatial conformation. This gives it a more complete and deeper understanding of the chemical world.

3D Generation, Data Ratios, and a Reality Check for AlphaFold 3

AI Drug Discovery
Computational Chemistry
Large Language Models
Protein Structure Prediction

Research shows that while unconditional 3D molecular generation has improved, it still lacks physical realism. For handling imbalanced data, a 1:10 ratio of active to inactive compounds is key for boosting model performance. AI is also evolving from a single black box into a mature toolbox; the CROP architecture gives Large Language Models the ability to see 2D and 3D molecular structures, while a reality check on AlphaFold 3 highlights the importance of providing accurate biological context.

Monday, the 18th of August, 2025
Jixing Liu, Ph.D.
A conceptual image showing a protein&#039;s binding pocket overlaid with a 3D target preference map. In the image, different colored regions clearly indicate the site's preference for various chemical groups (such as hydrogen bond acceptors, hydrophobic groups, charged groups), providing intuitive design guidance for medicinal chemists.

Automated Virtual Cells, Target Preference Maps, and Modeling Solvent Effects

AI Drug Discovery
Computational Biology
Drug Design
Molecular Docking

The CellForge framework uses multi-agent collaboration to automate the pipeline from biological questions to efficient computational models, freeing up scientists’ modeling work. A new model can create “target preference maps” for protein pockets, visually guiding chemists in molecular design. Finally, a new framework forces AI to confront solvent effects by introducing “solvent-aware” data augmentation, significantly improving the accuracy of molecular docking and virtual screening.

Sunday, the 17th of August, 2025
Jixing Liu, Ph.D.
A concept image for the FGBench benchmark, showing a large language model (LLM) looking confused and struggling with a complex test paper on functional group chemical reasoning. This symbolizes the significant challenge current AI faces in understanding fundamental chemical principles.

Genomic Embeddings, Functional Group Reasoning, and an All-in-One Drug Discovery Platform

AI Drug Discovery
Genomics
Cheminformatics
Explainable AI

One paper shows that for genomic prediction, simply extracting features from a large model is far more efficient than expensive fine-tuning. A new benchmark, FGBench, reveals that top AI models are failing to understand chemistry’s basic language—functional groups. Finally, the BSL AI platform successfully discovered new active molecules for a difficult target, validated in real lab experiments, closing the loop from algorithm to reality.

Sunday, the 17th of August, 2025
Jixing Liu, Ph.D.
A concept diagram of the ActivityDiff model, showing a diffusion model during molecular generation being influenced by both positive guidance (enhancing activity against a target) and negative guidance (reducing activity against off-targets) to design more precise and safer drug molecules.

Precise Molecular Control, Epilepsy Drug Prediction, and the Real-World Challenges of Deep Learning

AI Drug Discovery
Molecular Generation
Precision Medicine
Cheminformatics

The ActivityDiff model uses positive and negative guidance to precisely control the activity of drug molecules, balancing efficacy and toxicity like a mixologist. One study uses a vision-language model and an expert knowledge tree to predict the effectiveness of anti-seizure medications from MRI scans, even for drugs the model has never seen. Finally, a large-scale benchmark reveals that for molecular representation tasks, traditional chemical fingerprints (ECFP) still outperform most new deep learning models.

Thursday, the 14th of August, 2025
Jixing Liu, Ph.D.
A concept diagram of a transfer learning framework. It shows an AI model first undergoing &#039;general education' pre-training on a massive dataset of general chemical knowledge. It is then fine-tuned with a small amount of specialized antibacterial data for 'specialized training,' ultimately leading to the efficient discovery of new antibiotics.

Explaining Activity Cliffs, Discovering Antibiotics with Transfer Learning, and Tackling Data Imbalance

AI Drug Discovery
Explainable AI
Transfer Learning
Cheminformatics

An AI model called ACES-GNN learns from human chemists’ explanations to understand and interpret the tricky “activity cliff” phenomenon for the first time. A transfer learning framework, trained first on general chemical knowledge and then on specialized data, successfully screened a library of over one billion molecules to find new sub-micromolar antibacterial drugs. Finally, a new method uses GFlowNets to generate rare drug-drug interaction samples, effectively addressing the data imbalance problem in DDI prediction.

Thursday, the 14th of August, 2025
Jixing Liu, Ph.D.
A concept diagram of a multi-agent AI drug design platform, showing multiple AI agents with different roles (like Principal Researcher, Medicinal Chemist, AI Expert) working together like a committee to optimize molecules, with all decision-making processes recorded for auditability.

Decoding GPCR Mutations, Explainable Predictions, and the Challenges of AI Teamwork

AI Drug Discovery
GPCR
Explainable AI
Multi-Agent Systems

The BOLD-GPCRs model not only predicts GPCR ligand activity but also assesses the impact of protein mutations, offering a new tool for precision medicine. AttriLens-Mol uses reinforcement learning to force AI to “show its reasoning” before predicting molecular properties, solving the black box problem. Finally, a multi-agent AI platform reveals the profound challenge of how an “expert team” can sacrifice overall drug-likeness when optimizing for a single objective.

Tuesday, the 12th of August, 2025
Jixing Liu, Ph.D.
A concept diagram of a new method combining quantum chemistry and machine learning, showing a large protein-ligand complex being broken down into smaller fragments for calculation. A machine learning model then corrects key physical forces, ultimately yielding highly accurate binding affinity rankings in a very short time.

Speeding Up Molecular Simulations, Cracking Solubility, and Ranking Binding Affinity with Precision

Computational Chemistry
Molecular Dynamics
Machine Learning
Drug Screening

af2rave uses AlphaFold2’s uncertainty to generate diverse starting points for molecular dynamics simulations, efficiently capturing protein’s dynamic conformations. A new method breaks down solubility into basic physical processes and tackles each with machine learning for more accurate predictions. Lastly, a new technique combining quantum chemistry and machine learning can rank binding affinities with high accuracy in under 5 minutes, accelerating drug screening.

Monday, the 11th of August, 2025
Jixing Liu, Ph.D.
General-purpose large models are all show in specialized fields. To make AI truly useful in pharmacy, you have to feed it high-quality, specialized databases using a RAG architecture.

From Protein Puzzles to Topological Networks and RAG-Powered Pharmacy

AI Drug Discovery
Computational Chemistry
Large Language Models
Cheminformatics

A clever “protein sequence puzzle” pre-training method significantly improves an AI’s ability to make zero-shot predictions on new targets and drugs. Another model, GLDPI, finds rare “active” signals in a sea of “inactive” data by forcing the AI to respect the chemical “social network” of molecules. Finally, research shows that general-purpose large models are all show and no substance in pharmacy until they are “fed” high-quality, specialized databases using a RAG architecture.

Sunday, the 10th of August, 2025
Jixing Liu, Ph.D.
The Tripleknock AI model bypasses the speed bottlenecks of traditional metabolic models by brute-forcing the lethal effects of three-gene combinations, opening a new fast lane for finding multi-target antibiotics.

From 3D Molecular Sculpting to a Virtual Human Simulator

AI Drug Discovery
Diffusion Models
Computational Biology
Large Language Models

Diffusion models are taking AI molecular generation from 2D drawings to 3D sculpting. A new benchmark shows that Large Language Models can now write bioinformatics code as well as top human experts. The Tripleknock model brute-forces three-gene combinations to open new paths for multi-target antibiotics. Researchers have also proposed a grand vision for a “programmable virtual human,” while PairReg effectively tackles the oversmoothing problem in EGNN models.

Friday, the 8th of August, 2025
Jixing Liu, Ph.D.
Using nanoGPT as a benchmark for AI agent code optimization is a great idea, but the results pour cold water on the AI self-improvement hype: AI still has a long way to go.

Frontiers in AI Research: From Code Optimization and Literature Reviews to Molecular Dynamics Data Compression

AI in Science
Large Language Models
Computational Chemistry
Research Tools

A benchmark based on nanoGPT reveals that AI agents are still far from human-level performance in code optimization, pouring cold water on the hype around “AI self-improvement.” A dual-LLM system significantly improves the efficiency and quality of scientific literature reviews by combining a paper’s “impact” and “relevance.” Finally, the MDZip framework achieves over 95% compression of molecular dynamics data using a neural network, offering a new solution for large-scale data archiving and sharing.

Thursday, the 7th of August, 2025
Jixing Liu, Ph.D.
DeepADR uses the new KAN architecture to integrate drug structure, targets, and side effect text information, aiming to flag potential problems early in drug development.

Frontiers in AIDD: From KANs for Side Effect Prediction to Automated Gene Analysis

AI Drug Discovery
Computational Chemistry
Explainable AI
Multi-omics

The DeepADR model uses the new KAN architecture, combining multimodal information to predict adverse drug reactions (ADRs). A separate study reveals that SHAP, a deep learning explanation tool, can produce highly unstable results in multi-omics analysis, urging caution. Lastly, the GenoMAS framework automates complex gene expression analysis by assembling a diverse team of AI agents, showing AI’s potential as a collaborator.

Wednesday, the 6th of August, 2025
Jixing Liu, Ph.D.
The CellAtria AI framework tries to turn the tedious process of single-cell data analysis into a simple conversation with a chatbot.

Three New AI Tools: Precise Metabolism Prediction, Multiscale Bio-Networks, and a Single-Cell Analysis Agent

AI Drug Discovery
Computational Chemistry
Bioinformatics
Single-Cell Analysis

BioTransformer 4.0 uses smart filtering and a modular design to boost the signal-to-noise ratio in drug metabolite prediction, turning it into a practical map for drug design. The BIND framework leverages knowledge graphs and two-stage training to tackle the multiscale challenge of predicting biological networks from molecules to phenotypes. Finally, an AI agent framework called CellAtria aims to transform the tedious process of single-cell data analysis into a simple conversation with a chatbot, democratizing research.

Tuesday, the 5th of August, 2025
Jixing Liu, Ph.D.
No matching items
Back to top