Unsupervised Machine Learning Approach for Identifying Biomechanical Influences on Protein-Ligand Binding Affinity

Abstract

Drug discovery is incredibly time-consuming and expensive, averaging over 10 years and $985 million per drug. Calculating the binding affinity between a target protein and a ligand through Virtual Screening is critical for discovering viable drugs. Although supervised machine learning (ML) can predict binding affinity accurately, models experience severe overfitting due to an inability to identify informative properties of protein-ligand complexes. This study used unsupervised ML to reveal underlying protein-ligand characteristics that strongly influence binding affinity. Protein-ligand 3D models were collected from the PDBBind database and vectorized into 2422 features per complex. Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), K-Means Clustering, and heatmaps were used to identify groups of complexes and the features responsible for the separation. ML benchmarking was used to determine the features’ effect on ML performance. The PCA heatmap revealed groups of complexes with binding affinity of pKd<6 and pKd>8 and identified the number of CCCH and CCCCCH fragments in the ligand as the most responsible features. A high correlation of 0.8337, their ability to explain 18% of the binding affinity’s variance, and an error increase of 0.09 in Decision Trees when trained without the two features suggests that the fragments exist within a larger ligand substructure that significantly influences binding affinity. This discovery is a baseline for informative ligand representations to be generated so that ML models overfit less and can more reliably identify novel drug candidates. Future work will focus on validating the ligand substructure’s presence and discovering more informative intra-ligand relationships.