Protein–protein and protein–peptide interactions are fundamental to biological processes, making the accurate prediction of their binding affinity crucial for drug design and mutational analysis. Here, we develop HyBind-NN, a multimodal graph neural network that integrates protein language models (PLMs) with 3D structural and dynamic datasets to predict protein–protein and protein–peptide affinity. First, we demonstrate that combining ESM-2 sequence embeddings with precise 3D Voronoi spatial geometry enables accurate affinity predictions across diverse structural datasets. Next, we show that the inherent limitations of static rigid-body structures can be mitigated through a multi-task learning framework. By utilizing residue-level root mean square fluctuations (RMSF) derived from molecular dynamics (MD) as an auxiliary training target, the model implicitly learns to capture the conformational entropy of flexible peptides without requiring computationally expensive MD simulations during inference. In our benchmarking study, we observe that this multimodal architecture outperforms both purely sequence-based and strictly structural state-of-the-art algorithms, achieving a mean absolute error of 0.89 for pKD (1.12 kcal/mol for ∆G) on the independent benchmark. Finally, we confirmed through ablation analysis that while the PLM provides the dominant predictive signal, geometric representations and dynamic regularization are crucial for resolving subtle conformational rearrangements. This study highlights the synergistic potential of combining PLMs with physics-aware architectures and demonstrates their application towards the robust prediction of intermolecular binding affinity.
Bogdanova et al. (Tue,) studied this question.