Prediction of protein-ligand binding affinities using 3D machine learning methods

点击翻译按钮获取中文摘要

3D machine learning methods offer the potential to bridge the gap between docking and FEP, providing high throughput predictions of protein-ligand binding affinities with higher accuracy than classical docking scoring functions. Atomic environment vectors (AEVs) have been shown to be a promising representation for both rapid calculation of quantum mechanical conformational energies of small molecules, and in predicting protein-ligand binding affinities. In the 2022 rotation project, we applied these methods to predict the energy of of non-covalent interactions, in particular hydrogen bonds. Future directions include sharing parameters between elements, including charge and hybridization state as features, and transfer learning from the non-covalent interaction task to the protein-ligand binding task. It is also of interest to explore how AEVs can be used in molecular dynamics simulations following upon recent results from Roitberg on simulating a protein-ligand complex using ANI. These methods could be useful for including the entropic contributions to binding affinities as well as the enthalpic contributions from non-covalent interactions. This project would aim to test models on binding affinity and test the extent to which these models robustly learn non-covalent interactions and have improved generalizability over the prevailing ligand-based models for predicting binding affinity. This project falls into the EPSRC Biological informatics research area due to the use of computational modeling of biological systems using chemical and biological data. GSK is a collaborating company.