Answers are grounded in this paper and cite their sources.
Accurate prediction of protein–ligand binding affinities is fundamental to structure-based drug design, yet existing benchmarks are often limited by outdated structural coverage, uncertain local structural fidelity, and physicochemical or protocol-dependent bias. Here, we present HiFiPL-Bench, a fidelity-first and bias-aware benchmark curated from RCSB PDB and BindingDB through hierarchical filtering, manual electron-density validation, affinity verification, and property-matched decoy generation. Across nine representative scoring functions, HiFiPL-Bench revealed reduced performance in scoring, ranking, and most docking evaluations relative to CASF-2016, whereas a controlled screening-like task using computationally generated property-matched decoys produced model-specific gains and losses that altered relative rankings rather than uniformly lowering performance. Model-specific overlap and deoverlap analyses further showed that documented development-data exposure differed substantially among methods and that deoverlapped sensitivity analyses affected models and evaluation tasks nonuniformly. Together, these findings indicate that performance estimates remain sensitive to starting-structure curation, task-specific preparation, data set composition, decoy design, benchmark protocol, and recoverable development-data exposure. HiFiPL-Bench therefore provides a reusable diagnostic framework for evaluating scoring-function performance in a structurally curated and bias-aware benchmark setting.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: