Tomdok: a fully reproducible multi-engine consensus virtual screening pipeline with data-driven safety annotation, validated by blind recovery of clinically validated inhibitors

Background. Virtual screening campaigns are frequently irreproducible: fragmented toolchains,undocumented parameters and absent safety annotation impede reuse and independentverification. Methods. We present Tomdok, an open-source, one-command pipeline (stages 00-09) thatintegrates target preparation (PDBFixer, pH 7.4), ChEMBL library construction with tunableclinical-phase depth and parent-based deduplication, data-driven hERG liability annotation,multi-GPU GNINA primary screening, three-engine consensus refinement (GNINA, AutoDockVina, LeDock) using a scale-invariant count-based consensus metric, redocking validation, rulebasedand data-driven ADMET profiling, CLEAN/FLAGGED safety-aware prioritisation,automated publication figures (PyMOL/PLIP), and generation of a journal-ready supplementarybundle. Every stage supports resume; every run is mirrored to a rotated log; outputs aredeterministic-by-seed. Results. As two case studies we screened 3,326 approved and Phase-III compounds against (i)the WRN helicase (PDB 7GQU), a synthetic-lethal target in microsatellite-unstable cancers, and(ii) mutant IDH1 (PDB 6VEI), a validated target in IDH-mutant glioma and AML. The WRNscreen yielded 60 consensus hits (43 CLEAN / 17 FLAGGED; self-docking best-pose RMSD1.46 A). The IDH1 screen yielded 31 consensus hits (20 CLEAN / 11 FLAGGED; best-poseRMSD 0.45 A) and, as its headline result, blindly recovered two clinically validated mutant-IDH inhibitors: vorasidenib (the co-crystallised ligand; triple-engine consensus) and enasidenib(approved mutant-IDH2 inhibitor; top affinity hit). A third, unproductive target (MAT2A, PDB7KCC) returned a single weak hit, illustrating target suitability as a first-class consideration.Each campaign completed in ~5 h on two consumer GPUs. Conclusions. Tomdok provides end-to-end reproducibility and safety-aware prioritisation forgenome-scale repurposing, and its blind positive-control recovery on IDH1 demonstrates thatconsensus docking can recognise clinically validated chemistry without prior knowledge.