Chemistry Mass Spectrometry Molecular Docking
Strong benchmark performance does not imply generalization. These papers test whether computational chemistry methods work outside the conditions they were evaluated on.
Spectral Transportability in Mass Spectrometry
Standard MS/MS benchmark embeddings are uninformative for cross-instrument transportability prediction
Zenodo 10.5281/zenodo.21249049
Intensity-weighted m/z bin fingerprints — the standard embedding for MS/MS benchmarking — saturate the Grassmannian: all 45 instrument pairs land within an 11.8% band near the theoretical maximum distance. An apparent correlation with degradation is entirely driven by physically incompatible EI-B pairs; after exclusion, no metric predicts degradation (rho = -0.211, 95% CI [-0.523, +0.145]). A preregistered embedding comparison across four representations shows the collapse is representation-specific: Word2Vec avoids saturation (54-57% of max) but still fails all three transportability criteria. Geometric discriminability on the Grassmannian does not predict cross-instrument performance degradation.
Zero-Parameter Baselines for Molecular Docking
Generation vs. Retrieval in Molecular Docking: Zero-Parameter Baselines Match Learned Generators Across 67 Targets
Zenodo 10.5281/zenodo.21281732
On TDC’s DRD3 docking target, a zero-parameter fingerprint retrieval method places first at 100 oracle calls (−8.00 kcal/mol vs SMILES-LSTM −7.59), beating every learned generator on the leaderboard. A budget-crossover atlas across 58 DOCKSTRING and 9 TDC targets (10 seeds, bootstrap 95% CIs) maps where this advantage ends: iterative case-based refinement wins 79% of targets at budget 100 (CI: 68–90%), while generation surpasses the retrieval floor only beyond a few hundred calls. A second finding: uniform random sampling from the library nearly matches targeted retrieval (AUC −11.66 vs −11.85), indicating that library composition dominates performance at low budgets. An oracle-ranker gap decomposition identifies per target whether the bottleneck is retrieval scope or budget exhaustion. All code and parameters pre-registered under commit SHA.