From ranking to decision: TASC-VS for calibration-aware top-k virtual screening

Virtual screening is most useful when it improves top-k prioritization under limited experimental budgets. We evaluated the full TASC-VS package on a curated 15-target subset of the LIT-PCBA virtual-screening benchmark. Methodologically, TASC-VS combines test-time adaptation (TTA), calibration, and uncertainty-aware ranking to support fixed-budget decisions. Coverage analysis showed that the subset is not full LIT-PCBA: it contains 23,250 evaluated molecule-target records and 181 positives, with positives in seven targets. Empirically, TASC-VS full showed modest, directionally consistent benefits over the evaluated comparators. On the seven positive targets, normalized enrichment factor at 1% (NEF1) and Boltzmann-enhanced discrimination of receiver operating characteristic (BEDROC) were 0.052 and 0.062 for TASC-VS full, compared with 0.0057 and 0.0279 for RTMScore-on-Vina-poses. Across all 15 targets, precision and false-positive count among the top 20 candidates (P20 and FP20) were 0.0433 and 19.13 for TASC-VS full, compared with 0.0067 and 19.87 for RTMScore-on-Vina-poses. Calibration improved reliability metrics, uncertainty penalization produced a small top-20 benefit, and TTA changed reliability slightly without driving the main ranking gain. Paired P20 analysis versus RTMScore-on-Vina-poses gave a mean benefit of 0.0367 (95% bootstrap confidence interval (CI), 0.0067-0.0733; raw Wilcoxon p = 0.026), but multiplicity-corrected values were not significant (Holm p = 1.0; Benjamini-Hochberg (BH) q = 0.322). An external Directory of Useful Decoys, Enhanced (DUD-E) stress test favored Vina over TASC-VS full. Thus, TASC-VS full provides calibration-aware top-k decision support on the curated LIT-PCBA subset; broader deployment requires separate validation.