Schrödinger Bridge Onsager Flow for Drug-Likeness Distribution Learning on Real, Structure-Derived DrugBank Data
Abstract Learning to reweight chemical libraries toward drug-like regions is a fundamental task in virtual screening. While entropic optimal transport (OT) offers a principled framework for this distribution matching, standard solvers like Sinkhorn-Knopp often struggle to capture sharp target marginals without extensive hyperparameter tuning. To address this limitation, we propose the Schrödinger Bridge Onsager Flow (SBOF), a kernel-based dual-ascent algorithm that circumvents the n × n coupling matrix by updating a single scalar potential per molecule. Evaluated on the DrugBank 5.1.10 release enriched with authentic RDKit-computed descriptors (n = 5, 357), SBOF converges to a total variation (TV) distance of 0.0141 (Pearson r = 0.9996). Crucially, a comprehensive sensitivity analysis reveals that SBOF retains a 91.9% improvement in TV distance even against an optimally tuned Sinkhorn baseline, demonstrating that its advantage is structural. To address real-world deployment where unseen molecules must be scored, we further propose SBOFNet, a 1.6M-parameter neural amortization that achieves a TV of 0.0380 in a single forward pass. Finally, we rigorously interrogate our data pipeline, confirming via a coverage-bias analysis that our 38.1% SMILES-matched subset is statistically representative of the broader DrugBank database.
Authors
- Najiba Tagougui (ORCID: https://orcid.org/0000-0002-8686-4568)
- Abdallah Ben Abdallah
Institutions
- University of Sfax (TN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-13
- DOI
- https://doi.org/10.21203/rs.3.rs-11448030/v1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint