Protein-Ligand Binding-Affinity Feature Extraction and Prediction Using Deep Learning

Abstract

Protein–Ligand Binding Affinity (PLBA) prediction plays a vital role in drug discovery by quantifying the interaction strength between drug candidates and their biological targets. However, existing methods often struggle with generalization across diverse protein–ligand pairs. To address this, this work proposes a novel deep learning framework, PLBA Feature Extraction (FE) DL (PLBA-FE-DL), which integrates biologically-informed and taskspecific feature extraction with designed neural architecture. The key innovation lies in the combination of Enhanced Molecular Similarity Protein–Ligand Aligner (EMSPLA) for capturing structural interaction fingerprints, and self-supervised transformer-based encoders (ProtBERT and ChemBERTa) for deep semantic representation of protein sequences and ligand SMILES, respectively. These enriched representations are processed through Branched Neural Network (BNN) approach that separately learns protein and ligand features before merging them for interaction modeling. This design improves both learning capacity and generalization. Experimental results on PDBbind 2016 and 2020 datasets demonstrate the effectiveness of PLBA-FE-DL approach, achieving Pearson correlation Coefficients (R) of 0.956 and 0.93, respectively, along with reduced Root Mean Squared Error (RMSE) and Mean Squared Error (MSE). These results highlight advantage of combining advanced sequence-based embeddings with structure-aware alignment to enhance PLBA prediction.