Deep learning a transferable model for drug-receptor binding-energy

Abstract

Machine learning has enabled numerous advances in computational chemistry which have previously been out of reach due to the complex forces that govern atomic in teractions. In particular, the construction of machine learning-based force fields for describing atomic interactions has been an attractive area of research because it may allow to close gaps between the efficiency of molecular dynamics simulations and accuracy of theoretical models. In this thesis, we explore the topic of predicting ligand-protein interaction energy using a deep learning, energy-based model that is able to capture molecular interactions based on an input 3D molecular graph. Building upon graph neural networks, our model can predict interaction energies by obtaining the underlying force field that governs the interaction between a protein and a drug-like molecule. We used protein-ligand interaction data from curated databases to optimize an energy function for our model. We adopt the score match ing approach to construct two loss functions and compare their performance. We show that in combination with an energy model, the use of score matching allows us to avoid estimating a partition function in our models, which is important for our large systems and leads to a better learning objective. We discuss training strate gies, and provide two metrics for evaluating the performance of the model. While we are able to successfully model atomic interactions to some extent, we found a major challenge in handling the memory requirements for representing protein-ligand data that substantially limited our ability to explore the configuration and hyperparam eter space in our models. As such, we complement our protein-ligand studies by also training on a artificial but simplified dataset, which allowed us to explore the hyperparameter space more in depth. We conclude by discussing possible strategies for the future development of these models.