Efficient Prediction of Protein-Ligand Binding Using ESM-2 and Mol2vec with Random Forest Model

In this study, we present a predictive model for protein-ligand binding interactions, leveraging ESM-2 and Mol2vec embeddings combined with a Random Forest classifier. Our approach was evaluated across three testing scenarios—"Molecule unseen," "Protein unseen," and "None seen"—to assess its generalizability to novel compounds and targets. Compared to deep learning models, our method demonstrated competitive AUROC scores and consistently high specificity, indicating robust predictive performance and a conservative strategy that minimizes false positives. While sensitivity was lower, strategies to address data imbalance and enhance model sensitivity are discussed. Our findings highlight the potential of combining efficient embedding techniques with interpretable machine learning models for scalable and effective binding interaction prediction