Foundational Models for Drug Discovery

A foundation model (FM) is any model that is trained at scale on a broad dataset and can be adapted (e.g., fine-tuned) to a wide range of downstream tasks; current examples include BERT, CLIP and GPT-3. In this project, we investigate the challenges of building foundational models for drug discovery: capturing multi-modal information, explainability, and rapid adaptation to new lab experiments. Addressing these challenges will allow us to build robust foundational models that we can apply to downstream drug discovery tasks. In particular, we seek to leverage these foundational models to learn biochemistry fundamentals from biochemical interaction data. Biochemistry is the central discipline in the discovery of new medicines. Hence, a biochemistry foundational model will have great significance in tasks such as function prediction in proteins, binding prediction in chemical-protein interactions and antibody discovery. We will evaluate our biochemistry foundational model on several biochemistry datasets in the open-source domain.

Faculty Supervisor:

Blake Richards

Student:

Partner:

Valence Discovery Inc

Discipline:

Computer science

Sector:

Pharmaceuticals; Technology; Health and Related Sciences & Technology

University:

McGill University

Program: