Effective learning in the low-data hard-negative regime

Develop ligand-based virtual screening methods that achieve reliable enrichment above random ranking when trained with limited labeled actives from focused TopU hard-negative candidate distributions, particularly for targets with fewer than 50 labeled TopU actives.

Background

TopU-LBVS-low evaluates learning when training, validation, and test compounds are drawn from the same hard-negative distribution. The evaluation is divided into Tier 1 targets with fewer than 50 TopU actives and Tier 2 targets with at least 50 actives.

The reported baselines perform close to or below random enrichment, especially on Tier 1 tasks. Thus, the paper identifies low-data learning under hard-negative screening as an unresolved methodological challenge rather than a problem solved by the evaluated fingerprint, graph-neural-network, or pretrained molecular-model baselines.