RLBindDeep: A ResNet-LSTM based novel framework for protein-ligand binding affinity prediction

The prediction of the binding affinity of proteins and ligands in computational drug discovery with high accuracy is critical when evaluating the effectiveness of potential therapeutic compounds. This research work introduces RLBindDeep, a novel deep learning architecture based on the amalgamation of the ResNet and LSTM architectures, for improved accuracy in predicting protein-ligand binding affinities. Most traditional methodologies utilizing conventional molecular docking techniques suffer from poor accuracy owing to semi-flexible modeling approaches and limited considerations of complex interactions. On the other hand, RLBindDeep, which is formulated as a pose-independent binding affinity regression model that directly predicts experimental protein-ligand binding affinities from fixed complex structures, without performing docking or rescoring multiple poses, has performed well in extracting important features of the protein-ligand interaction. Specifically, the extracted features encompass ligand physicochemical descriptors (e.g., molecular weight, LogP, TPSA), protein-level features such as amino acid composition, and detailed interaction features including van der Waals, electrostatic, and hydrogen-bond energies. The model has been tested rigorously over the CASF-2016 benchmark dataset and has returned Pearson's coefficient R=0.875, Spearman's coefficient ρ=0.864, and Root Mean Square Error RMSE=0.993. This significantly outperforms existing state-of-the-art models, such as HAC-Net and AutoDock Vina. Improved accuracy and robustness in RLBindDeep further highlight the possibility of deep learning to revolutionize computational drug discovery processes, making strategies for drug development more efficient and targeted.