Computational Drug Discovery Methods

Benchmarking Large Language Models for Assay-Aware Drug-Target Binding Affinity Prediction

This paper benchmarks large language models for drug-target binding affinity prediction, showing that assay descriptions provide predictive value beyond molecular and protein data.

RePurpose: An LLM-Assisted, Source-Traceable Drug Repurposing Tool with Retrospective Validation on Clemastine in Multiple Sclerosis

Large-language-model-assisted repurposing web tool recovered clemastine for multiple sclerosis remyelination only through mechanism-worded, model-dependent queries; its rankings missed later evidence of harm in progressive disease.

NoiTox: A Machine Learning Strategy for Peptide Toxicity Prediction Reveals Label Noise in Benchmark Datasets

NoiTox presents a multi-model approach for peptide toxicity prediction, revealing label noise in datasets and showing that physicochemical descriptors are informative.

NCI-Profiler: automated protein-ligand interaction detection and visualization

NCI-Profiler detects and visualizes nine types of protein-ligand non-covalent interactions with high accuracy and speed, outperforming existing tools in several categories.

OmniScore: Universal Scoring of Diverse Biomolecular Complexes via Equivariant Geometry-Aware Discrete Representation Learning

OmniScore learns shared geometric representations of diverse biomolecular complexes and improves quality assessment and interaction prediction across molecular classes.

A Structure-Based hERG Pose Selection Workflow Integrating Induced-Fit Docking, Shape Similarity, Interaction Fingerprints, and Molecular Dynamics

The paper evaluates a pose-selection framework using hERG structures to prioritize ligand poses, finding shape similarity and interaction fingerprints are effective for different ligand groups.

Toxicity by Consortium Revisited: Structural Neighbors and Cross-Family Reach

The paper revisits a toxicity prediction method using structural similarity and reverse screening, showing improved performance on public data and demonstrating the value of target breadth in predicting CYP3A4 inhibition.

In Vivo Transcriptomics-Based Framework for Systematic Indication Expansion

This paper presents a drug repurposing framework using in vivo transcriptomics to expand indications by matching drug signatures with disease signatures, successfully recovering established uses of semaglutide and proposing new ones.

Prospective discovery of LasB inhibitors by integrating generative design, free energy ranking, and experimental validation

This paper discovers three new nanomolar inhibitors of LasB by combining generative design, free energy calculations, and experimental validation.

Pocket-Conditioned Generation and Down-Selection of Non-Oxime Reactivators

This paper compares generative models for designing non-oxime reactivators of paraoxon-inhibited acetylcholinesterase, finding that down-selection methods outperform DrugDiff in practical drug-likeness and synthesis feasibility.

Do LLM Proposers Improve Budgeted Search for Ionizable Lipids?

This study evaluates if language models improve combinatorial molecular search for ionizable lipids compared to classical methods under budget constraints, finding no significant improvement.

CLARA: Compound Library Active Learning with Reasoning Agents

CLARA uses a language model to make strategy decisions in compound library active learning, outperforming random sampling across multiple protein targets.

Probing isn't Causality in Molecular Models

The paper evaluates whether molecular models can causally represent chemical variables by comparing intervention effects with random controls and pretraining advantages.

Large scale prospective evaluation of co-folding across hundreds of Mac1-ligand complexes and three virtual screens

On 489 evaluable new Mac1 complexes, AlphaFold3, Chai-1 and Boltz-2 placed over half of ligands within 2 Å, but on docking hit-lists AF3 and Boltz-2 added little over docking.

MtbAIM: AI/ML-guided endpoint-specific prediction of small-molecule activity across drug-susceptible and drug-resistant Mycobacterium tuberculosis

Tree models on RDKit/Morgan features predict M.tb activity and MIC endpoints for drug-susceptible and resistant isolates, slightly beating pretrained embeddings; released as a Colab tool.

Structure-Based Drug Design of Novel Inhibitors Targeting the Thiamine Biosynthesis Pathway in Bacteria

The paper designs novel inhibitors targeting thiamine biosynthesis in bacteria using structure-based drug design, resulting in a potent orally bioavailable broad-spectrum bactericidal molecule.

Tensorizing Pairwise Structural Analysis: GPU-Accelerated RMSD and FCC Clustering in HADDOCK3 Without Structural Regression.

The paper accelerates structural analysis in HADDOCK3 by reformulating RMSD and FCC clustering as GPU-based linear algebra operations, achieving significant speedups without compromising accuracy.

Human Intestinal Absorption Revisited: The 1998 Descriptors Survive, and the Method Extends to Permeability

The paper re-evaluates a 1998 method for predicting human intestinal absorption using molecular descriptors and shows improved performance with repeated extraction and consensus descriptor sets.

Perturb-Seqr: Signature Search Engine Integrating Connectivity Mapping Resources

Perturb-Seqr integrates multiple Connectivity Mapping datasets to enable drug and target discovery by searching for gene expression signatures of drug and gene perturbations.

Protein Language Model-Conditioned Graph Neural Networks for Multitask Ligand Activity Prediction Across Human GPCRs

The paper develops a graph neural network that combines ligand molecular graphs with protein language model embeddings to predict GPCR ligand activity, showing improved performance over baseline models.

Synthesis-aware generative design in trillion-scale chemical spaces for automated drug discovery

This paper introduces a generative AI framework for drug discovery that efficiently navigates large chemical spaces while ensuring synthetic feasibility and functional success.

DEPICT enables sample-centric therapeutic prioritization across evolving and spatially heterogeneous tumor states

DEPICT prioritizes therapies within tumor contexts using deep learning, showing effectiveness across different tumor states and spatial heterogeneity.

Frustration Quenching and Network Topology of the Energy Landscape as Primary Determinants of Protein-Ligand Binding Pose Prediction by Deep Learning Models

This paper shows that deep learning models' performance in predicting protein-ligand binding poses depends on the energy landscape's frustration quenching and network topology, not algorithmic flaws.

Anti-target chemical-space coverage of Indian medicinal plant phytochemicals

The paper evaluates the chemical-space coverage of Indian medicinal plant phytochemicals against experimentally tested compounds for anti-target endpoints, finding most lack close structural representation.

Comparative Analysis of Relative Ligand Binding Free Energy Simulation Methods: Amber-TI, GROMACS-NETI, OpenMM-FEP, and BLaDE-MSLD

This study compares four ligand binding free energy methods using the same force field and finds they show comparable accuracy with no significant performance differences.

SIMORGH: Ensemble-Aware Geometric Deep Learning for Apo-State and Cryptic Ligand Binding Site Prediction

SIMORGH predicts ligand binding sites from apo proteins by integrating conformational ensembles, outperforming existing methods in ligand-free settings and identifying cryptic pockets.

FragDockRL: A Reinforcement Learning Method for Fragment-Based Ligand Design via Building Block Assembly and Tethered Docking

FragDockRL uses reinforcement learning to design ligands by assembling building blocks and tethered docking, outperforming random search on three protein targets.

Evaluating Autonomous LLM Agents Across Molecular Prediction and Optimization Benchmarks

This paper evaluates autonomous LLM agents in molecular prediction and optimization benchmarks, showing they can develop methods that match or exceed human-developed approaches in several tasks.

Two Comparators May Be All We Need

Random forests on sequence and 2D chemistry pick the favored protein for a molecule, or the favored molecule for a protein, at 0.71 to 0.78 accuracy on unseen compounds.

Efficient detection of cryptic pockets on challenging drug targets using MD-free co-vibrational modes

The paper identifies cryptic pockets on challenging drug targets using a fast MD-free method and validates them experimentally.