Answers are grounded in this paper and cite their sources.
Synthetic accessibility filtering before cross-validation changes both the compounds available for learning and those used to evaluate ligand-based virtual screening models. We assessed SAscore filtering in PyRMD using SA-Filter Impact Benchmarking (SA-FIB), with 95 unfiltered target data sets and 438 filtered conditions across five thresholds. Historical residual-distance percentile cutoffs were transferred, while absolute epsilon values were estimated within each training subset. Mean paired F1 differences were −0.1078, −0.0361, −0.0147, −0.0074, and −0.0011 at SA ≤ 2.5, 3.0, 3.5, 4.0, and 5.0, respectively, relative to the corresponding eligible-cohort baselines. At SA ≤ 5.0, the F1 difference was not significant after Bonferroni correction, while ROC-AUC and PRC-AUC increased by 0.005085 and 0.003060. Combining duplicated processed pools into 92 units retained significance for ROC-AUC but not PRC-AUC. Reoptimizing percentile cutoffs within inner folds for the 92 eligible SA ≤ 3.5 data sets changed mean outer-fold F1 by +0.000294, with 30 improvements, 22 ties, and 40 decreases. Responses varied by target and metric. Historical parameter selection, reused outer folds, and differences between baseline and filtered evaluation populations limit interpretation to sensitivity within this PyRMD protocol. These findings support target-specific validation of synthetic accessibility filters.
No takes yet. Share an insight, caveat, or question.
Ferreira et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: