CTMP: bounded no-reference protein–ligand pose repair with failure-aware virtual-screening audit
Year: 2026
Type: article
Abstract: Protein–ligand poses selected by docking or generative pipelines can retain useful global localization yet remain physically invalid, limiting their reliability for structural inspection and analysis. We introduce CTMP, a bounded no-reference post-selection repair method that uses only the supplied pose, molecular graph, and receptor coordinates. CTMP combines whole-ligand translation and rotation with acyclic-torsion adjustment under explicit movement caps; no ligand atom may move more than 2 Å from its supplied position. The transform preserves the molecular graph and returns a geometric certificate and identity audit. On 614 completed PoseX crystallographic self-docking cases, CTMP increased strict physical-core validity from 30.1% to 72.8% and joint RMSD-below-2-Å plus physical-core success from 13.0% to 31.1%. The number of poses below 2 Å remained 209/614, the median paired RMSD change was 0.000 Å, and overall residue-level contact F1 changed by $$-0.0011$$ (95% CI $$-0.0041$$ to 0.0019). In an independent 183-case frozen Vina-selection audit, physical-core success increased from 88.0% to 98.9% and joint success from 79.2% to 88.5%, whereas RMSD-below-2-Å success changed from 166/183 to 164/183. Across 15 LIT-PCBA targets and 2,464 requested molecules, strict physical-core validity increased from 38.6% to 91.6%, and guarded repair reached 93.4%; target-level ranking changes remained heterogeneous. CTMP therefore improves the structural usability of already selected poses without establishing improved rescoring accuracy, enrichment, activity discrimination, or prospective hit rates. CTMP separates post-selection coordinate repair from pose search and score optimization. Its contribution is a reference-free, graph-preserving, and auditable transform with explicit geometric bounds, evaluated through physical validity, pose localization, residue-level contact recovery, and frozen-selection analyses. The results show that substantial physical-validity gains can arise from repairing poses that are already localized, without candidate reranking or a general improvement in ranking metrics. This distinction establishes structural usability as an evaluation dimension separate from predictive performance and provides a portable repair-and-audit framework for docking, diffusion, and hybrid pose-generation workflows.
Source: Journal of Cheminformatics
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $2,390
OpenAlex ID: W7229526281
