#3 · ML / AI methods
Benchmark baselines and transfer conditions reshape reported gains in biological ML
The common thread is evaluation design for machine learning models in biology: whether apparent gains survive appropriate baselines, held-out contexts, and retrieval-oriented metrics. A baseline-aware framework for sequence models on noncoding RNA assays finds that model increments are sensitive to the chosen baseline. In synthetic-lethality prediction, degree-matched AUROC and screen retrieval dissociate, with near-chance models still giving 2.62-fold enrichment. A single-cell foundation model study reports that reliability transfer differs from prediction transfer, using SHAP-based reliability-attribution fingerprints. Transfer is examined in a frozen perturbation model on the 2025 Virtual Cell Challenge benchmark and in transcriptomic source selection, where similarity ranking adds information beyond tissue matching. Temporally withheld validation appears in Tclin-likeness modeling. COPD comorbidity biomarker and HNSCC pathway dependency studies apply learned models to public data with less emphasis on evaluation design. What is new is the explicit separation of discrimination, retrieval, and reliability from headline prediction accuracy.
Published papers
Temporally Validated Models for Target "Tclin-likeness".
Knowledge-graph ML models estimate Tclin-likeness, ranking non-Tclin human proteins by resemblance to approved drug targets; temporally withheld validation and…
Preprints
Centered cross-context perturbation transfer on the public 2025 Virtual Cell Challenge benchmark
Frozen cross-context perturbation transfer model centers an external KOLF2.1J source on H1 training targets and uses isotonic regression; scores on the 2025…
Prediction transfer does not imply reliability transfer in single-cell foundation models
Shows reliability transfer in single-cell foundation models differs from prediction transfer; introduces SHAP-based reliability-attribution fingerprints across…
Evaluating the added value of sequence models in noncoding RNA assays: a baseline-aware framework
Baseline-aware evaluation framework for sequence models on ncRNA assays (lncRNA CRISPRi, Cas13, miRNA repression); finds model increments sensitive to baseline…
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.
Network-guided identification of candidate biomarkers for COPD-associated comorbidities using deep learning and real-world transcriptomic data
CNN functional-distance model on a protein network nominates COPD-comorbidity biomarker candidates, tested on public transcriptomic data.
Predicting Pathway Dependency in Head and Neck Squamous Cell Carcinoma from Hallmark Transcriptomic Features
Transcriptome-based Elastic Net framework inferring Hallmark pathway dependency in HNSCC from GDSC cell lines, validated on TCGA-HNSC and GSE65858 with HPV and…
Transcriptomic similarity guides source selection beyond tissue lineage
Tests whether baseline transcriptomic similarity ranks source contexts for predicting held-out drug responses; ranking adds signal beyond tissue matching…
Topic history
| Date | Rank |
|---|---|
| #3 | |
| #3 | |
| #3 | |
| #5 |