On improving experimental binding affinity predictions with synthetic data

Abstract The success of deep learning binding affinity prediction models depends critically on expanding experimental data with reliable synthetic data. We extend the Structurally Augmented IC50 Repository (SAIR) with $\approx$80K absolute free energy perturbation (AFEP) calculations and present two distinct data splits, SAIR-FEP and SAIR-OOD (out-of-distribution), to simulate realistic drug discovery scenarios. We compare sequence-based proteochemometric (PCM) models and state-of-the-art, structure-based deep learning models and demonstrate that PCM models can be enhanced by physics-based descriptors. While structure-based deep learning methods capture finer geometric detail, their performance is highly sensitive to the input structure. By filtering for high-confidence, co-folded complexes, we show that the performance improves predictably, whereas training on all complexes blindly limits performance gains. Finally, using the SAIR-OOD split, we demonstrate that simultaneous training on synthetic and experimental data improves performance on publicly available, experimental benchmarks. For the OpenBind benchmark, our models yield the lowest root mean squared errors and a 1.85x improvement of Spearman ranking correlation when compared to the original AEVPLIG model training on experimental data alone. These results provide a clear strategy for using synthetic data to advance experimental binding affinity predictions.