Pedro Ballester: Machine-learning scoring functions for structure-based drug discovery: A 15-year perspective

Machine-learning scoring functions for structure-based drug discovery: A 15-year perspective

Molecular docking predicts whether and how small molecules bind to a macromolecular target from its 3D atomic-resolution structure. Docking supports two central problems in structure-based drug discovery: virtual screening, which aims at discovering drug leads in compound libraries, and lead optimisation, which improves their properties (e.g. binding affinity or potency). The success of these approaches depends critically on the accuracy of the scoring function applied to docked molecules.

Over the past fifteen years, machine-learning approaches to scoring have produced large measurable improvements in both virtual screening and binding affinity prediction. In this talk, Dr. Pedro Ballester will review such machine-learning scoring functions and share key lessons learned about how to enhance their performance for each of these applications. In particular, he will discuss:

• The key role of training data augmentation.
• When and how performance can be improved by building target-specific scoring functions.
• The shortcomings of current benchmarks and how they are being mitigated.
• The impact of docking pose errors on binding affinity prediction.
• The magnitude of the improvement introduced by deep learning algorithms.
• The need for releasing the code of scoring functions to reproduce and further evaluate them.
• What is currently possible in prospective virtual screening and binding affinity prediction guided by these tools.

Our Speaker:

Pedro Ballester

Dr. Pedro Ballester is a Royal Society Wolfson Fellow and Associate Professor at Imperial College London, where he has led the Artificial Intelligence for Healthcare group since September 2022. His research harnesses machine learning allied with domain knowledge to analyse and predict how small molecules modulate protein and cellular function. These approaches are applied to translational challenges in drug discovery and broader healthcare problems. Before re-joining Imperial (where he carried out his PhD), he was an Assistant Professor at INSERM in France, following postdoctoral fellowships at the University of Oxford, the University of Cambridge, and the EMBL-European Bioinformatics Institute in the UK. More about his group and projects can be found at ballestergroup.github.io.