Establish the source of assay-level performance differences

Characterize why the interpretable descriptor portfolio matches the CheMeleon embedding on some molecular assays but substantially trails it on others, particularly Ames mutagenicity and hERG blockade.

Background

The portfolio achieves pooled parity with the 2048-dimensional CheMeleon embedding but fails the predeclared per-assay parity gate. Performance varies materially across endpoints: CheMeleon leads on Ames mutagenicity and hERG blockade, whereas the portfolio leads on intestinal absorption and P-gp inhibition.

The paper reports that these assay-level differences remain unresolved. Identifying the causes could guide the development of additional independent hub models or assay-specific descriptor portfolios.

References

Assay-level differences remain unresolved.

— Interpretable-by-Design Descriptor Portfolios Match a 2048-Dimensional Foundation Embedding on Low-Data Molecular Assays  (2609.30789 - Yao et al., 25 Sep 2026) in Abstract and Section 3.1, “Parity at 50 to 100 labels”