#17 · ML / AI methods

Auditing shortcuts and interpretability in biological machine learning models

The dominant thread is scrutiny of what machine learning models actually learn from biological data. The PPI audit identifies shortcut biases (topology, self-interaction, taxonomy, function) in HIPPIE, IntAct, STRING and PDB datasets and offers an ILP-based approach to mitigation. Mechanistic interpretability of ESM-2 finds reduced attention on disordered regions, while embeddings and logits still recover radius of gyration and dynamic contact information. SeqMaestro pairs no-code interpretable models with testable biological hypotheses, and an oral-gut microbiome study uses SHAP site attribution to show complementary signals beyond accuracy. Other items apply structured or learned methods to specific data: CryoConvNeXt recovers low-abundance cryo-EM species missed by cryoSPARC, and a glycan motif framework raises true positives from 300 to 442 across 45 glycomics datasets. The interactome reconstruction item infers negative pairs from experimental search space. The shared emphasis is on auditing datasets and representations, not only aggregate model scores.