CAREER: Geometric Deep Learning to Facilitate Algorithmic and Scientific Advances in Therapeutics

Harvard University, Cambridge MA

Investigators

Abstract

Imagine a vast universe of molecules and proteins, each with its unique structure and function. There are so many of them (around a trillion trillion trillion), and they all interact in complicated ways. This project will create artificial intelligence methods to better understand and analyze complex networks of data, especially those from the world of drug discovery. This project will develop geometric deep learning methods, a type of artificial intelligence that is good at understanding data that forms networks, such as how molecules interact with each other. These methods will adapt their understanding based on the specific context of each molecule, making them versatile and powerful. By aggregating billions of molecular observations using these methods, the project will create molecular search engines capable of identifying useful molecules across various criteria. These engines will be able to quickly and efficiently find the best molecules for specific purposes by considering dozens of factors all at once and discovering new possibilities that were previously impossible to explore just through experiments in a lab. This could lead to new drugs being discovered more quickly and cheaply. An integral part of the project is the education plan, which includes developing new curricula at undergraduate and graduate levels for molecular machine learning and preparing students for artificial intelligence-driven scientific roles. The outreach component focuses on increasing undergraduate research involvement, particularly among female and minority students, and educating them on the responsible use of AI in science. This project develops fundamental geometric deep learning algorithms for analyzing large, graph-structured datasets in therapeutic science, focusing on aggregating extensive molecular and protein sequence data to create adaptable molecular search engines. It aims to explore the vast molecular space, estimated at 10^60 molecules, and the plethora of protein sequences to unlock therapeutically valuable molecular interactions. The project's core is the development of innovative geometric deep learning algorithms. These algorithms will be context-aware, capable of adjusting to the molecular contexts in which they operate, and versatile enough to generalize to new tasks with limited data. They will leverage multimodal information to produce adaptable graph representations for various tasks and domains. This project will pioneer foundation graph models for general graph representations, crucial in molecular machine learning, paving the way to exploring larger molecular spaces inaccessible to experimental screening, significantly reducing costs, and establishing the foundation for geometric deep learning in therapeutic science. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.