FlashBind: Towards Accurate and Efficient Structure-based Virtual Screening

TMLR Paper9441 Authors

Abstract: Accurate prediction of protein-ligand interactions is central to computational drug discovery. Recent foundation models such as Boltz-2 have achieved remarkable accuracy in binding affinity prediction, yet their prohibitive computational cost remains a major barrier to large-scale virtual screening. Here we introduce FlashBind, a lightweight structure-based model that achieves a 50× speedup over Boltz-2 at inference time by replacing expensive structure prediction with a fast docking model and substituting costly PairFormer modules with a streamlined EGNN architecture. FlashBind matches Boltz-2 on standard virtual screening benchmarks and demonstrates superior generalization to enzyme-substrate specificity prediction. To evaluate real-world applicability, we apply FlashBind to target-based antibiotic screening against the essential bacterial proteins in E. coli and show that FlashBind substantially outperforms Boltz-2 and other virtual screening baselines. Notably, several top-ranked candidates exhibit potent inhibition of DnaG and effective bacterial growth inhibition against E. coli in wet-lab validation. Together, these results demonstrate that FlashBind bridges the gap between accuracy and efficiency, enabling ultra-fast, high-fidelity screening of massive chemical libraries for drug discovery.

Submission Type: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=2LlhJg8rl2

Changes Since Last Submission: Reformatted the manuscript to fully conform to the official TMLR template, ensuring the text block, margins, font sizes, and page size all match the required specifications. Also made minor presentational edits to keep the layout clean and well-organized. The scientific content is unchanged from the previous submission.

Assigned Action Editor: ~Lijun_Wu1

Submission Number: 9441