Self-Supervised Machine Learning Approach for Identifying Biochemical Influences on Protein-Ligand Binding Affinity

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 is critical for discovering viable drugs. Although supervised machine learning (ML) models can predict binding affinity accurately, they suffer from lack of interpretability and inaccurate feature selection caused by multicollinear data. This study used self-supervised 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. LASSO Regression and hierarchical clustering were utilized to minimize multicollinearity between features. Correlation analyses and Autoencoder-based latent space representations were generated to identify features significantly influencing binding affinity. A Generative Adversarial Network was used to simulate ligands with certain counts of a significant feature, and thereby determine the effect of a feature on improving binding affinity with a given target protein. It was found that the CC and CCCN fragment counts in the ligand notably influence binding affinity. Re-pairing proteins with simulated ligands that had higher CC and CCCN fragment counts could increase binding affinity by 34.99-37.62% and 36.83%-36.94%, respectively. This discovery contributes to a more accurate representation of ligand chemistry that can increase the accuracy, explainability, and generalizability of ML models so that they can more reliably identify novel drug candidates. Directions for future work include integrating knowledge on ligand fragments into supervised ML models, examining the effect of CC and CCCN fragments on fragment-based drug design, and employing computational techniques to elucidate the chemical activity of these fragments.