AI-Driven Prediction of Protein-Ligand Interactions for Accelerated Drug Discover

This study proposes a lightweight artificial intelligence-based prediction framework to address the limitations of structure-based methods for predicting protein-ligand interactions. Unlike conventional approaches, the model relies exclusively on sequence-based inputs. A total of 12,309 ligand molecules were retrieved from BindingDB and encoded as 2,048-dimensional Morgan fingerprints derived from SMILES notation. Protein sequences were represented as 20-dimensional amino acid composition (AAC) vectors. These embeddings were used as inputs to an XGBoost regression model to predict binding affinity (pKi), followed by a second-stage regression model that calibrated predictions based on residual errors. The final calibrated framework achieved strong predictive performance, with an RMSE of 0.90, R² of 0.74, a binary classification accuracy of 88%, and an AUC of 0.948. Furthermore, shapley additive explanations (SHAP) highlighted the contributions of specific amino acid features —such as tryptophan (AAC_W) and histidine (AAC_H)—as well as ligand substructures, to the predicted binding affinity. By enabling rapid and accurate predictions without requiring 3D protein structures, this framework provides a versatile complement to structure-based methods and offers practical utility for applications such as drug repurposing and virtual screening.