#3 · ML / AI methods
Protocol-matched benchmarks test whether learned molecular models beat conventional…
The dominant thread treats evaluation protocol, not architecture, as the experimental variable in molecular property prediction. A multi-endpoint study compares seven active-learning acquisition strategies and two uncertainty estimators across eight ADMET regression endpoints under a strict protocol, asking when acquisition actually pays. A matched-budget, scaffold-split comparison of GNNs, frozen ChemBERTa/MolCLR embeddings and fingerprints across six TDC endpoints finds no model family dominant. Ark-ADMET pushes the same logic to tuning, showing endpoint-specific choice of fingerprints, descriptors, GIN/MiniMol embeddings and CatBoost hyperparameters matches or beats leaderboard scores on 7 of 21 endpoints. PeptSol extends benchmark construction to protected-peptide solubility with 1388 curated measurements and chemistry-aware splits over frozen encoders. Adjacent work applies the same auditing stance elsewhere: single-cell foundation models benchmarked against simple baselines, degree-matched evaluation separating discrimination from retrieval in synthetic lethality, and antibody immunogenicity labels where a format lookup beats sequence predictors.
Published papers
Advancing morphological profiling with self-supervised learning through comparative insights into representation learning of cellular perturbations.
Large-scale benchmark of self-supervised foundation models (DINO, DINOv2, MAE, iTPN) for Cell Painting morphological profiling on JUMP-CP, outperforming…
Preprints
PeptSol: A Translational Benchmark for Protected-Peptide Solubility in Drug Manufacturing
PeptSol: an open benchmark of 1388 curated protected-peptide solubility measurements with chemistry-aware splits, evaluating seven frozen molecular encoders…
Attribution-based diagnostics for single-cell foundation models
Systematic post-hoc attribution analysis of single-cell RNA foundation models (scGPT, Cell2Sentence, STATE) for cell annotation and perturbation prediction,…
Separating benchmark discrimination from experimental retrieval in synthetic-lethality prediction
Degree-matched AUROC and screen retrieval dissociate in synthetic-lethality prediction; near-chance models still gave 2.62-fold enrichment.
When Does Active Learning Improve ADMET Prediction? A Multi-Endpoint Benchmark
Protocol-matched benchmark of seven active-learning acquisition strategies and two uncertainty methods across eight ADMET regression endpoints, using a strict…
Task-Dependent Performance of GNNs and Pretrained Models in ADMET Prediction
Matched-budget, scaffold-split benchmark of GNNs, frozen pretrained models (ChemBERTa, MolCLR) and fingerprints on six TDC ADMET endpoints; no family…
The Quagmire of Immunogenicity-Guided Antibody Design: Labels Limit Progress
Re-examines immunogenicity labels (ADA%) for antibody design: extends a 217-entry dataset to 231 therapeutics, finds a format lookup beats sequence predictors,…
Ark-ADMET: endpoint-specific selection of molecular representations and hyperparameters for gradient-boosted ADMET models under the Therapeutics Data Commons benchmark protocol
Endpoint-specific selection of fingerprints, descriptors, GIN/MiniMol embeddings and CatBoost hyperparameters matches or beats leaderboard scores on 7 of 21…
Single-cell perturbation prediction near experimental reproducibility with GENESIS
GENESIS, an energy-score-trained generative model, predicts single-cell response distributions from experimental covariates and approaches replicate…
Where single-cell foundation models help: a benchmark of annotation, perturbation and regulatory inference against simple baselines
Benchmarks five single-cell foundation models against simple baselines for annotation, in silico perturbation and GRN inference; FMs help mainly for…
Topic history
| Date | Rank |
|---|---|
| #3 | |
| #29 | |
| #20 | |
| #5 | |
| #21 | |
| #16 | |
| #14 | |
| #19 | |
| #13 |