FlashAffinity: Bridging the Accuracy-Speed Gap in Protein-Ligand Binding Affinity Prediction

Songlin Jiang, Yifan Chen, Ze Cao, Wengong Jin

Abstract:

Accurate prediction of protein-ligand binding affinity is central to computational drug discovery. Recent foundation models, such as Boltz-2, have achieved remarkable accuracy, but their high computational cost poses a major barrier to large-scale virtual screening. We address this challenge by introducing a lightweight structure-based virtual screening model, \textbf{FlashAffinity}, that achieves similar performance as Boltz-2 in affinity prediction and binder classification tasks, while achieving a 50x speedup at inference time.
FlashAffinity replaces the expensive protein structure prediction models with a simple protein-ligand docking model and the PairFormer-based affinity scoring module with a cheap EGNN architecture.
In summary, this work bridges the gap between accuracy and efficiency, enabling ultra-fast virtual screening of massive chemical libraries.