Decision stability of a default docking pipeline: a seed, box-placement, and protomer audit on LIT-PCBA MTORC1

Agentic and automated drug-discovery workflows increasingly call molecular docking as a black-box subroutine and act on its ranked output, typically from a single run under software defaults. We ask a narrow, operational question: if the same default pipeline is re-run with a different random seed, a slightly displaced search box, or a different input protomer, which decisions-the top-k compounds a downstream agent would advanceactually change? Using AutoDock Vina 1.2.7 on the LIT-PCBA MTORC1 target (PDB 4jsx) with a fixed, score-blind panel of 60 ligands (20 actives, 40 inactives), we predeclared the decision rule and the seed and box-placement perturbation axes against a baseline, added a protomer arm whose selection criterion was amended during execution (disclosed with dates), and compared top-12 decision sets as symmetric differences (exits and entries), Jaccard indices, and binary decision changes, with pose quality checked by PoseBusters against the exact receptor used for docking. A positive control (redocking the co-crystallized ligand, symmetry-corrected RMSD 0.647 Å) validated the setup. Changing only the random seed replaced 3 of 12 top-12 members (6/60 binary decisions changed); a 2 Å box shift along x replaced 2, along y replaced 4; protomer changes at pH 7.0 replaced none. Eight ligands accounted for all flips, and the single PoseBustersinvalid pose was also a recurrent flipper. A threeseed majority-vote consensus cost 3× compute, dropped the invalid recurrent flipper, and still disagreed with every individual run; we report it as a priced mitigation with measured residual disagreement, not as certified stabilization. All results concern decision repeatability only: nothing here establishes that any decision set is correct, and a consensus can stabilize a wrong answer. All code, inputs, checkpoints, and negative results are provided in an accompanying archive prepared for the author; no public release has been made.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-02
DOI
https://doi.org/10.26434/chemrxiv.15011514/v1
Primary Topic
Computational Drug Discovery Methods
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

Decision stability of a default docking pipeline: a seed, box-placement, and protomer audit on LIT-PCBA MTORC1

Badri N. Dubey, Shereen Anand, Swananda Marathe, Anmol Kumar
2 links
ChemRxiv
Computational Drug Discovery Methods
preprint

Decision stability of a default docking pipeline: a seed, box-placement, and protomer audit on LIT-PCBA MTORC1

Badri N. Dubey, Shereen Anand, Swananda Marathe, Anmol Kumar
preprint en

Abstract

Agentic and automated drug-discovery workflows increasingly call molecular docking as a black-box subroutine and act on its ranked output, typically from a single run under software defaults. We ask a narrow, operational question: if the same default pipeline is re-run with a different random seed, a slightly displaced search box, or a different input protomer, which decisions-the top-k compounds a downstream agent would advanceactually change? Using AutoDock Vina 1.2.7 on the LIT-PCBA MTORC1 target (PDB 4jsx) with a fixed, score-blind panel of 60 ligands (20 actives, 40 inactives), we predeclared the decision rule and the seed and box-placement perturbation axes against a baseline, added a protomer arm whose selection criterion was amended during execution (disclosed with dates), and compared top-12 decision sets as symmetric differences (exits and entries), Jaccard indices, and binary decision changes, with pose quality checked by PoseBusters against the exact receptor used for docking. A positive control (redocking the co-crystallized ligand, symmetry-corrected RMSD 0.647 Å) validated the setup. Changing only the random seed replaced 3 of 12 top-12 members (6/60 binary decisions changed); a 2 Å box shift along x replaced 2, along y replaced 4; protomer changes at pH 7.0 replaced none. Eight ligands accounted for all flips, and the single PoseBustersinvalid pose was also a recurrent flipper. A threeseed majority-vote consensus cost 3× compute, dropped the invalid recurrent flipper, and still disagreed with every individual run; we report it as a priced mitigation with measured residual disagreement, not as certified stabilization. All results concern decision repeatability only: nothing here establishes that any decision set is correct, and a consensus can stabilize a wrong answer. All code, inputs, checkpoints, and negative results are provided in an accompanying archive prepared for the author; no public release has been made.

ChemRxiv
Indian Institute of Technology Dharwad (IN)
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Decision stability of a default docking pipeline: a seed, box-placement, and protomer audit on LIT-PCBA MTORC1 — Badri N. Dubey, Shereen Anand, et al. · ChemRxiv (2026) | TGRS Research Map | TGRS