Computational Drug Discovery Methods

The paper introduces a method to visualize chemical space by projecting molecular data onto a lower-dimensional space using linear methods, outperforming non-linear techniques in efficiency and neighborhood preservation.

The paper introduces methods to identify and avoid activity cliffs in drug design using the BitBIRCH clustering algorithm.

The paper presents a framework that predicts endocrine disruptors using a multimodal cross-attentive graph model, achieving high accuracy and providing mechanistic explanations.

MOAST provides a rapid, accessible tool for predicting mechanism of action of compounds using KDE and CatBoost, achieving better accuracy than existing methods.

The paper validates a generative AI model using experimental assays to design small molecule drugs targeting specific cancer-related phenotypes, showing 88% morphological impact and 37% transcriptomic pathway effects.

The paper uses structure-based virtual screening to identify novel TMPRSS2 inhibitors from a large compound library, selecting 10 hit candidates through a reproducible filtering workflow.

This paper introduces AiPP, an AI-based platform for predicting ligandable sites in the human proteome, and demonstrates its application in discovering covalent inhibitors for PTPN6.

This study compares five PK software tools for predicting fraction absorbed and unbound in humans, finding ANDROMEDA to be the most accurate with better quantitative and qualitative performance.

DLRNA-BERTa predicts RNA-drug binding affinities using a transformer model, achieving high Pearson correlations across RNA classes and identifying 2,859 compounds with predicted affinities.

The paper identifies rosmarinic acid as a potential neuroprotective compound for FTD by analyzing transcriptome data and molecular interactions with HDAC1.

This paper uses in silico methods to identify terpenoids from Cannabis sativa that may interact with cannabinoid receptors to relieve pain.

This paper introduces FORGE, a framework that jointly models drug response and gene essentiality to identify treatment groups and predict efficacy using basal gene expression.

This paper introduces VVS, a machine learning framework for efficiently searching ultra-large chemical libraries by decomposing molecules into building blocks and improving retrieval precision.

This paper presents an AI framework for multimodal drug discovery that generates natural language hypotheses using structured data, knowledge graphs, and LLMs, demonstrating potential for restructuring fragmented biomedical knowledge.

PepCCD generates target-specific peptides using a contrastive conditioned diffusion framework, achieving strong binding affinity and improved diversity compared to existing methods.

The paper evaluates if evolutionary proximity correlates with chemical similarity in natural products across multiple taxa using deep learning embeddings.

HyperLab accelerates drug discovery by providing an AI-driven platform with tools for ligand-protein binding prediction, molecular optimization, and compound screening, achieving high accuracy and speed in benchmark tests.

This paper demonstrates a prompt-driven approach to identify and contextualize biological targets, applying it to PARP1 to uncover its roles in DNA repair and disease.

This paper introduces SAM-DTI, a spatial attention model for predicting drug-target interactions using SMILES strings and protein sequences, achieving competitive performance against existing methods.

This paper develops GCN-based QSAR models to predict toxicity of SARMs, finding mixed success with transfer learning and moderate accuracy in predicting DILI, DIRI, and DICT.

This study uses an agentic LLM to design and evaluate 1,002 broad-spectrum antibiotic candidates targeting Enterobacteriaceae, with 774 passing ADMET benchmarks and 391 showing moderate binding to dual targets.

The paper introduces DGDM, a self-improving AI framework for drug discovery that enhances molecular design and pipeline optimization, achieving improved binding affinity while maintaining drug-likeness and novelty.

PathPCNet predicts drug sensitivity using multi-omics data and biological pathways, achieving high correlation and uncovering key gene-drug interactions.

AMODO-EO discovers and integrates new optimization criteria during drug discovery, improving multi-objective optimization by revealing hidden molecular relationships.

This study uses in silico methods to identify potential inhibitors for NDM-1 metallo-β-lactamase in K. pneumoniae, finding two sulfonamide compounds as lead candidates.

The paper uses a conditional diffusion model to design a small molecule inhibitor of secondary nucleation in alpha-synuclein aggregation, showing it effectively inhibits fibril proliferation.

The paper uses machine learning to predict PPAR-γ agonist bioactivity, achieving an R2 score of 0.83 using molecular descriptors from ChEMBL data.

This paper introduces H-MoE, a hierarchical MoE model for molecular representation learning that improves generalization by incorporating scaffold-level routing and curriculum learning.