Physics-Informed CNN Models for Scoring Protein-Ligand Interactions

• Grant number: 23/13721-0
• Support Opportunities: Scholarships abroad - Research
• Start date: August 12, 2024
• End date: July 31, 2025
• Field of knowledge: Biological Sciences - Biophysics - Molecular Biophysics
• Principal Investigator: Alessandro Silva Nascimento
• Grantee: Alessandro Silva Nascimento
• Host Investigator: Alan Aspuru-Guzik
• Host Institution: Instituto de Física de São Carlos (IFSC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Institution abroad: University of Toronto (U of T), Canada

Abstract

The recent impact of machine learning methods structural biology became evident in recent years when AlphaFold2 and RosettaFold achieved experimental-like accuracy for the modeling of protein structures. In addition, the AlphaFold2 team not only released the program to the community but also made the predictions to almost the entire UNIPROT database of protein sequences. At this point, an interesting question one could ask is: how can this tremendous amount of data be used for further scientific and technological advances? Structural models have long been used for the discovery/modeling of compounds able to modulate the macromolecule function. In this BPE research project, we propose the application of a deep learning model that could learn from the physical properties computed from the 3D coordinates of an experimental macromolecular structure or from an accurate model. The proposed model receives structural features computed in the 3D space, such as the electrostatic potential and the Lennard-Jones potential, for example, computed for a protein structure in discrete 3D grids. The fundamental hypothesis is that accurate deep learning models may be able distinguish incorrect ligand poses in a discrete grid in the same way they can distinguish objects in photos, for example. For this purpose, 3D grids will be generated with the LiBELa program, developed in our group, to compute grids for protein-ligand complexes from the PDBBind dataset, as well as from the Koes' CrossDocked2020 dataset. As outputs for the model, we foresee the possibility of scoring and ranking protein-ligand complexes and, in the limit, using this score to actually dock ligands with higher precision and more efficient execution than the traditional ligand docking. The project will be developed in Aspuru-Guzik's group. His group is a leading group in the field of machine learning and quantum computing and one of the major goals of the project involves the training of this proposer on the techniques of machine learning in Aspuru-Guzik's group.