Agreement and ranking stability across molecular docking workflows: a controlled reproducibility case study in hepatocellular carcinoma

Docking-based drug-repositioning studies often prioritise compounds using one workflow, yet cross-workflow agreement and ranking stability are seldom quantified. We treated docking output as a method-comparison problem and asked whether a published hepatocellular carcinoma ranking was reproducible under an alternative one. Using the same receptors, ligands and grid definitions, four repurposing candidates and the co-crystallised reference ligands were docked with AutoDock Vina and compared with published AutoDock 4 scores. Redocking reproduced the PI3Kα crystallographic pose (root mean square deviation 0.23 Å) and was borderline for FGFR4 (2.04 Å), where the deviation was confined to one peripheral ring. Across ten ligand-target pairs, scores were strongly associated (Pearson r = 0.79), but correlation is not agreement: Lin's concordance correlation coefficient was 0.78, with Bland-Altman limits of discordance of -2.05 to +2.66 kcal mol^-1, and rank concordance was modest (Kendall τ-b = 0.40). Atorvastatin shifted from rank 1 under AutoDock 4 to rank 4 under Vina at both targets, and leave-one-out analysis restored perfect concordance only when it was omitted. To isolate the scoring function, the AutoDock 4 force field was then applied to the identical files, grids, seeds and sampling depth; this matched protocol recovered the published ranking exactly (τ-b = 1.00), attributing the rank change to scoring rather than to preparation or search. Rescoring fixed poses with the opposing function moved atorvastatin two positions, and ten replicates at exhaustiveness 32 showed the Vina arm converged (standard deviation ≤ 0.04 kcal mol^-1) while the AutoDock 4 arm varied by up to 2.16 kcal mol^-1 for this compound alone. Agreement, rank stability and replicate reporting convey information that a single ranked table does not. The findings concern one published prioritisation and do not estimate how often such prioritisations fail across chemical space.