Interpretable molecular descriptors match high-dimensional embeddings on small datasets
Interpretable-by-Design Descriptor Portfolios Match a 2048-Dimensional Foundation Embedding on Low-Data Molecular Assays
Machine Learning
Summary
Predicting how molecules behave with limited data is tricky, and how you represent molecules can greatly affect accuracy. The authors explore whether a set of simple, easy-to-understand molecular features can perform just as well as a complex 2048-number model called CheMeleon. They found that their simpler, auditable features almost match CheMeleon's accuracy on nine different biological tests. While some differences remain for individual tests, overall, these transparent features work competitively, making models easier to understand and trust.
What this means in practice
- •For drug discovery chemists: Use audited, interpretable molecular features to predict drug activity when experimental data is limited.
- •For chemical safety analysts: Implement transparent descriptor sets to assess toxicity and metabolism endpoints with small datasets.
Authors
Yiqi Yao, Miquel Duran-Frigola
Abstract
In low-data structure-activity prediction, the choice of molecular representation can matter more than the choice of predictor, and tabular foundation models sharpen that effect. We ask whether a portfolio of compact, semantically named descriptor blocks can reach the accuracy of a 2048-dimensional CheMeleon embedding while staying auditable at the feature level, meaning that every input dimension carries a model name and a recorded training provenance. Starting from a fixed 11-dimensional physicochemical base, we greedily concatenate provenance-screened blocks using the labelled context alone. Across nine ADME/Tox assays and 50 evaluation cells, scored on common-coverage subsets restricted to the molecules that every representation covers, the portfolio reaches a mean test AUC of 0.762, against 0.764 for CheMeleon and 0.756 for Mordred. The pooled gap to CheMeleon is +0.003 AUC (task-bootstrap 95% CI [-0.020, +0.030]), which satisfies our predeclared pooled parity gate but not the per-assay gate. At 25 context labels the headline rule again satisfies the pooled gate; at 10 labels it does not. We also report four predeclared candidate-selection rules that we falsified. Post-freeze checks over ten seeds and three previously unseen assays support pooled competitiveness for compact, auditable representations; a same-width random-bundle control does not establish that greedy membership itself adds accuracy. Assay-level differences remain unresolved.