Computational Drug Discovery Methods

The paper develops a meta-model that combines multiple machine learning models to predict ADMET properties, outperforming existing methods on a benchmark dataset.

The paper identifies FDA-approved molecules as potential inhibitors of H1N1 viral surface proteins through virtual screening and molecular dynamics simulations.

This study applies offline reinforcement learning to generate protein binders with improved docking and chemical properties by addressing the computational cost of protein-ligand docking.

This study evaluates GPT-4 and GPT-3.5 in molecular tasks, showing their strengths in molecular optimization but challenges in generating complex molecules.

ChemXTree improves small-molecule drug discovery by integrating tree-based algorithms with a graph model, showing superior performance in feature optimization on benchmark datasets.

This paper introduces Ligand-Transformer, a deep learning framework for predicting protein-ligand interactions, and demonstrates its application in identifying inhibitors for specific targets and studying disordered proteins.

RNAmigos2 improves RNA virtual screening by combining semi-supervised graph learning and large-scale docking data, achieving high accuracy and speed.

This paper introduces a generative diffusion model for molecular docking trained on artificial binding pockets, showing improved performance over existing methods.

This paper introduces a distance-aware Transformer model that enhances ligand-based virtual screening by capturing 3D shape similarity, enabling efficient screening of large compound databases.

The paper develops a system for predicting protein-ligand binding sites from Cryo-EM maps and generates drug-like ligands using a combination of PLBP methods and a de novo structure modeling tool.

PanScreen assesses off-target liabilities of small molecules using structure-based modeling and deep learning, offering accurate predictions and insights into potential modes of action.

This study designs new BACE-1 inhibitors using computational modeling, showing improved docking scores compared to Verubecestat for Alzheimer's disease treatment.

This paper evaluates dataset quality and splitting techniques for predicting protein-ligand binding affinities, introducing a tool (po-sco) and showing improved model performance with better data practices.

TranSiGen denoises and reconstructs chemical-induced transcriptional profiles using deep generative modeling, effectively predicting differential expression genes and serving as alternative phenotype information for drug discovery.

The paper trains a chemical property prediction model using English and SELFIES, achieving better performance than message passing neural networks in some tasks and enabling natural language drug searches.

This study identifies E96 as a potential inhibitor of MTHFD2 through consensus docking and molecular dynamics, showing it outperforms other compounds in binding stability and free energy.

This study evaluates if pre-training chemical language models on large compound datasets improves compound representation, finding that standard masked language modeling is insufficient.

DeepVASP-S predicts amino acid contributions to binding specificity using CNNs and explains steric mechanisms in protein-ligand interactions.

This paper introduces Umol, an AI system that predicts protein-ligand complex structures from sequence and ligand SMILES, outperforming existing methods at high accuracy.

Bio-Mol improves small molecule property prediction by integrating missing bioactivity profiles through pretraining and finetuning, achieving a 5.2% higher AUROC than previous models.

ActFound is a foundation model for bioactivity prediction trained on 2.3 million compounds and 50,869 assays, showing strong generalization and performance comparable to FEP+(OPLS4).

This paper introduces two Transformer-based methods for Drug-Target-Interaction prediction, emphasizing interpretability and performance through contrastive learning and cross-attention mechanisms.

The paper investigates potential inhibitors of ROCO4 kinase for Parkinson's disease treatment, identifying Mongolicain-A as a promising candidate with high specificity for the G2019S mutation.

This paper introduces EADTN, an ensemble model for drug-target interactions that improves predictive performance and identifies false-negative samples through adaptive feature weighting and clustering fine-tuning.

Herb-CMap identifies bioactive ingredients in herbal medicine by integrating gene perturbation data with a random walk algorithm, revealing molecular mechanisms of action for treating cough variant asthma.

This study develops public classifiers for predicting drug-induced cardiotoxicity using FDA DICTrank data, identifying key chemical and biological features that distinguish toxic from non-toxic compounds.

The paper uses large-scale perturbation profiles to analyze drug mechanisms of action and polypharmacology across multiple tumor subtypes and tissues.

This paper uses metadynamics to predict small molecule-RNA binding modes, showing agreement with experimental structures and providing binding pathways for RNA aptamers.

The paper introduces the EFFECT benchmark suite to evaluate cancer sensitivity prediction models by addressing data biases and improving benchmarking accuracy.

MMD-DTA predicts drug-target binding affinity and regions using a multi-modal deep learning framework, outperforming existing models and showing strong generalization.