Determine whether the lower greedy stopping threshold is universally optimal

Determine whether reducing the greedy portfolio stopping threshold from 0.005 to 0.001 consistently improves predictive performance across low-data molecular assays and context sizes.

Background

An exploratory stopping threshold of 0.001 improved the mean performance at 25 context labels on the development assays, but the gain was unstable at 10 labels and did not replicate on three unseen assays. The authors therefore do not establish the lower threshold as generally preferable.

A systematic determination of the optimal stopping threshold would require evaluating threshold choices across broader assay families, sample sizes, and independent datasets without relying on post hoc selection.

References

The per-assay gate still fails: the conservative threshold can exclude useful small additions, but this does not establish $\delta=0.001$ as universally optimal.

— 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 Section 3.2, “The few-label frontier”