Machine Learning with Molecular Dynamics to improve rapid protein-ligand predictions.

Exciting progress has been in made in ensemble-based, thermodynamically rigorous approaches to calculate the free energy of binding of small molecules to proteins and indeed recent work by us and others has demonstrated that these methods are capable of obtaining accuracy comparable to experiment. However, these approaches require large amounts of computer time and whilst that may be acceptable in some scenarios it prohibits the use of these approaches in scenarios where real time data is necessary (such as structural refinement or virtual screening). Thus, it would be desirable to develop approaches that are rapid, yet can deliver at the required level of accuracy. Deep learning and related machine learning technologies show great promise in this area, particularly where large data sets are available. Molecular dynamic (MD) simulations can provide huge amount of relevant data about protein-ligand interactions, but thus far these two disciplines have not really been combined. Our overarching question is: "Can machine-learning be combined with MD to improve rapid protein-ligand predictions?"

One of the key advantages of machine learning methodologies, as well as their speed, is their capacity to explain non-linear relationships, which is especially useful in the context of interactions between a protein and a ligand. The work we are proposing here will use MD data within a machine-learning context (neural networks in the first instance, and then deep neural networks) to improve affinity and pose predictions of small molecule binding to proteins. This is an exciting opportunity to improve the prospects for rational drug design.