Geometric Deep Learning for Binding Affinity Prediction

Drug discovery is an incredibly expensive and time-consuming process.

The average cost of bringing a drug to market is $1.3 billion, which is doubling every nine years, and takes ten to fifteen years.

The development of machine and deep learning techniques have promised to improve the efficiency of the drug discovery process and arrest this decline in productivity.

Structure-based drug discovery, one method of discovery, uses computational methods and the 3D structure of a protein to identify novel drugs that bind to the target.

Accurate scoring functions that can predict the binding affinity between a protein and a small molecule drug have been developed using machine learning over the last decade. These machine learning (ML)-based scoring functions have improved accuracy over pre-existing methods.

However, they have been primarily trained and validated using solved crystal structures of bound protein-ligand complexes.

This does not accurately represent a real-world drug discovery scenario where crystal structures for the bound protein-ligand complexes are not available. Sometimes there may be no crystal structure of the protein at all, and modelled structures must be used.

In addition, ML-based scoring functions are typically trained and validated using data from a single source and often fail to generalise to novel data sets.

To discern between predictions from the scoring functions that result from a bias in the training data and those that do not, the model's reasoning must be fully understood.

Current scoring functions do not provide this information and so have a reduced trust in the field to be used in drug discovery. Approaches such as attribution have shown promise in elucidating why the scoring function has made certain predictions.

Attribution can be used to check whether predictions have been made based on the interactions between the protein and drug instead of just on the composition of the drug alone, a common pitfall for scoring functions.

This DPhil aims to develop a binding affinity model using deep learning that has been explicitly designed to address the flaws currently in existing models in the field.

Geometric deep learning architectures, such as equivariant graph neural networks, have been successfully utilised for this problem and combined with attribution to show that the scoring function does learn biomolecular interactions instead of bias to make predictions.

These architectures will be built upon to extend the accuracy for modelled structures, not just experimentally determined crystal structures, to allow increased reliability for realistic applications of scoring functions.

A scoring function that is accurate, reliable, and easily interpretable will be valuable in any drug discovery project and will be available for all as open-source software.

The research will be completed in collaboration with the medical research charity LifeArc and their employees Dr Andy Merritt and Dr Kristian Birchall, who have valuable experience in accelerating healthcare innovation to make breakthroughs for patients. The project falls within the EPSRC AI and Data Science for Engineering, Health, and Government (ASG) research area.