Globally and locally consistent molecular representations through Fibered AutoEncoders

This project aims to explore ideas from a machine learning technique called Fibered AutoEncoders (FAE) for molecular representation learning. This technique will allow us to perceive the molecular space through a topological lens using the concept of fiber bundles, and it will give us a disentangled multi-level encoding of molecules. Since the molecular space is so large, this characteristic is essential to learn a latent molecular space that is easier to navigate and optimize.
Our specific aim is to learn a molecular latent space which stratifies molecular compounds by their scaffold (structural family) or. The benefits of such structuring include better predictive models, constraint and global optimization, and characteristic hopping in the molecular space.

Faculty Supervisor:

Courtney Paquette

Student:

Partner:

Valence Discovery Inc

Discipline:

Computer science

Sector:

Pharmaceuticals; Technology; Health and Related Sciences & Technology

University:

McGill University

Program: