Background

Computational binding-prediction models for TCR–pMHC, antibody–antigen, and drug–target interactions report strong in-distribution accuracy but fail on out-of-distribution targets. Recent benchmarks show every leading TCR–pMHC predictor collapsing to within 0.66 ± 0.04 AUROC on unseen peptides; analogous collapses are reported in antibody–antigen and drug–target benchmarks.

The three problems have identical mathematical structure: a receptor space, a ligand space, sparse experimental positives, constructed negatives, and a shortcut induced by how the negatives were drawn. Recent work confirms none of these crosses domains, and no shared metric exists for comparing negative-source robustness across them.

The mechanism is shared across the three domains: training negatives are constructed by sampling strategies. Negative sampling strategies like peptide shuffling for TCR–pMHC, random mismatching for antibody–antigen, property-matched decoys for drug–target where each of which induces a domain-specific shortcut (V-gene composition, surface properties, molecular weight) that the model learns instead of binding rules.

Research Goals

Scholars will develop a unified framework that handles both hard-negative generation and cross-domain shortcut-resistance evaluation. Concretely, scholars will investigate per-domain generators, a shared benchmark that runs predictor AUROC across five negative-sampling strategies on the same positives, with coefficient of variation (CoV) as the headline robustness metric.

Skills Needed

Familiarity with any machine learning and natural language processing.

Skills Gained

Model training, hypothesis generation, statistical testing