Abstract
<jats:p>Computational drug-discovery pipelines routinely rank candidate targets and compounds by predicted selectivity — the tendency to act on an intended protein while sparing a related offtarget. Cheap, scalable proxies for selectivity are attractive: sequence divergence between orthologs, pocket-restricted sequence identity, and comparative (“counter-”) docking are all in common use. We tested four such methods against benchmarks where the correct answer is independently known: a panel of arthropod acaricide targets with documented pollinator toxicity, and a set of dihydrofolate reductase (DHFR) ortholog pairs distinguished by a selective antifolate (trimethoprim, ≥103-fold bacterial-selective [6]) versus a cross-reactive one (methotrexate [6]). All four methods failed. Whole-protein and pocket-restricted sequence identity ranked knowntoxic targets as more divergent than known-safe ones; homology-transferred binding-site identity showed no resolution between a selective and a non-selective drug on the same protein pair; and counter-docking, given crystallographically defined binding boxes, produced score differences of ≤0.57 kcal/mol where ≥4 kcal/mol was expected — an order of magnitude below the dockingscore noise floor. We trace the failures to a common mechanism: selectivity is typically encoded by a small number of binding-site residues, and every cheap method dilutes that signal by averaging or scoring over the whole site or protein. We further present a worked case from a tick acaricide-discovery pipeline in which a crystallization-additive contaminant and an uncalibrated ortholog-docking comparison together produced a confident but spurious selectivity claim. We provide the DHFR benchmark as a reusable, drop-in test for any selectivity-prediction method, and argue that in-silico selectivity should be treated as unvalidated until it passes it. Where selectivity by target absence is available — an off-target lacking the protein entirely — it is the one mechanism robust to this failure mode.</jats:p>