#7 · ML / AI methods
Assay-aware and protocol-matched ML models for prioritizing small molecules
The dominant thread is machine learning used to prioritize small molecules by activity, selectivity, and ADMET properties. D2ML builds assay-aware models that predict activity and selectivity from crude and purified direct-to-biology screening data, guiding ROS1 selectivity. Several ADMET studies apply controlled comparisons: one benchmarks seven active-learning acquisition strategies and two uncertainty methods across eight regression endpoints under a strict protocol, while another compares GNNs, frozen pretrained models (ChemBERTa, MolCLR), and fingerprints on six TDC endpoints under scaffold splits. A hERG study systematically compares imbalance-handling strategies for QSAR regression with region-aware evaluation. RangeFinder predicts ranges of nine ligand properties from protein pockets, and PSICHIC+ predicts affinity distributions from sequence and SMILES, trained on about 82M records. What is new is assay-aware or protocol-matched training and evaluation, alongside pocket-conditioned property prediction and large pretrained bioactivity models.
Preprints
D2ML: Models for Direct-to-Biology Small Molecule Screening
Assay-aware ML models for direct-to-biology small molecule screening predict activity and selectivity from crude and purified data, guiding ROS1 selectivity…
RangeFinder: Predicting Ligand Property Ranges from Protein Pockets
RangeFinder, an SE(3)-equivariant transformer, predicts ranges of nine ligand physicochemical properties from protein pockets, reducing ligand search space…
Interpreting Graph Neural Networks for Antibacterial Drug Repurposing against Acinetobacter baumannii
Graph neural network ensemble for A. baumannii antibacterial prediction, using atom features only; explainability highlights piperazine and prospective Drug…
Bioactivity foundation model for hit identification and hit-to-lead optimization from sequence data
PSICHIC+, a bioactivity foundation model trained on ~82M protein-ligand records, predicts binding affinity distributions from sequence and SMILES for virtual…
Parameter-Efficient Adaptation of Protein Language Models for Dual-Target Antimicrobial Peptide Discovery Under Strict Homology Constraints
Multi-task LoRA fine-tuning of ESM-2 (150M) jointly predicts AMP antibacterial potency and hemolytic toxicity under 60% homology splits, screening 749 designed…
XAMPGen: An Interpretable SI Pipeline For Novel Antimicrobial Peptide Discovery
Fine-tunes ProGen on known antimicrobial peptides to generate candidates, then uses probing classifiers and self-attention analysis to interpret layers and…
Topic history
| Date | Rank |
|---|---|
| #7 | |
| #25 | |
| #22 | |
| #36 |