CGDeepAff: Deep Learning-Based Approach for Protein-Ligand Binding Affinity Estimation Using CNN-GRU
Ekarsi Lodh, Shalini Majumder, Tapan Chowdhury · 2025
Protein-ligand binding affinity prediction is one of the most important phases in drug development since it directs the selection and optimization of therapeutic molecules. We introduce CGDeepAff, a unique deep-learning framework that combines CNNs and Gated Recurrent Units (GRUs) to estimate protein-ligand binding affinities accurately. This model incorporates various features to improve prediction accuracy, including numerical data, amino acid counts (aa counts), and protein sequences. On the CASF-2016 benchmark dataset, CGDeepAff surpasses previous models with Spearman's$\rho$of 0.861, Pearson's$R$of 0.855, and RMSE of 1.002. Finally, when compared to models like HAC-Net$(R=0.846, \rho=0.843$, RMSE = 1.205) and KDEEP$(R=0.82, \rho=0.82$, RMSE = 1.27), CGDeepAff outperforms them, especially in treating various protein-ligand complexes. These results indicate that CGDeepAff is a broadly generalizable method capable of being used in computational drug discovery. Thus, CGDeepAff constitutes a further step forward in the binding affinity prediction, possibly leading to a better tool for drug discovery.