PPBind-3D/1D: protein–protein binding affinity prediction for structural and sequence data
Antibody Therapeutics · 2026-10-13
DOI: 10.1093/abt/tbag053 ↗Abstract
Abstract Background One key step of protein drug development is the screening of protein–protein binding (PPB) affinity. Various deep learning methods have been proposed to predict PPB affinity, but they are often challenged by data leakage, data scarcity, data heterogeneity, and structure-sequence gap. Methods To address these challenges in PPB affinity prediction, we propose a comprehensive framework by developing prediction models, PPBind-3D and PPBind-1D, to predict PPB affinity. PPBind-3D leverages structural information near the PPB interface to make its predictions. Additionally, PPBind-1D was developed using sequence data to address the lack of structural data in practical applications. Results PPBind-3D, equipped with monotonic neural network-constrained multi-task learning and Self-and-Mutual Attention, demonstrated state-of-the-art performance in PPB prediction across diverse datasets under both strict and random data partition settings. PPBind-1D, trained by aligning to PPBind-3D, overcame structural data scarcity to deliver robust sequence-based predictions while maintaining high accuracy, enabling practical deployment where structural data is unavailable. To facilitate real-world application, PPBind-3D was integrated into an online tool, which is available at https://rddc.tsinghua-gd.org/tool/dg-prediction. Finally, we demonstrated three application cases of our artificial intelligence models in the virtual screening of protein drugs, further confirming the framework’s utility and scalability. Conclusions Comparative and ablation studies demonstrate that PPBind-3D surpasses conventional benchmark approaches. Utilizing solely sequence information, PPBind-1D closely approximates the performance of PPBind-3D, thereby providing a viable alternative when structural data are inaccessible. Validation across three virtual screening case studies confirms the efficacy of both models, illustrating that our models can significantly facilitate high-throughput screening.
EUVIMED