Determine whether greedy portfolio membership itself improves accuracy

Determine whether the specific descriptor blocks selected by the context-only greedy membership rule improve predictive accuracy beyond matched random compact descriptor bundles of the same width.

Background

The study uses a greedy search to add compact descriptor blocks to a fixed 11-dimensional RDKit physicochemical base. A same-width random-bundle control produced nearly identical pooled performance, and the confidence interval for the difference did not distinguish the greedy portfolio from random bundles.

Consequently, the observed parity with CheMeleon may arise primarily from the anchored compact construction rather than from the identity of the blocks selected by the greedy search. Whether greedy membership contributes an actual accuracy benefit remains unresolved.

References

A same-width random-bundle control does not establish that greedy membership itself adds accuracy.

— 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 Appendix, Section “Replication across seeds and on unseen assays,” subsection “Random same-width bundles”