Nonlinear Dimension Reduction and Machine Learning Algorithms for Coarse-grained Simulations of Biomolecular Complexes

Stephen Pankavich

pankavic@mines.edu

Alex Pak

apak@mines.edu

Biomolecules (e.g. proteins, lipids, nucleic acids, and carbohydrates) assemble into higher-order structures, so-called biomolecular complexes, to perform a variety of biological functions. Scientists have been interested in redesigning biomolecular complexes for applications across biotechnology, including in sustainability and healthcare. To do so, a molecular understanding of emergent biomolecular assembly is needed, which is difficult to probe from experiments alone. These insights can be revealed through multiscale dynamical simulations that leverage coarse-graining, a process to derive lower-resolution models from high-resolution (e.g. atomic) models. The project focuses on implementing multiscale dynamical simulations of macromolecular complexes by developing a holistic modeling framework that uses (i) supervised, nonlinear dimensional reduction techniques to coarse-grain atomistic variables and (ii) machine learning algorithms to restore the fine, atomic details of coarsened structures. This will require the exploration of new computational feature projection algorithms (e.g., Principal Component Analysis) and probabilistic generative models (e.g., Autoencoders).

All team members will work together as part of a collaborative research group. IMURF participants will regularly meet with professors and PhD students in a communal setting. Everyone on the project will meet weekly or bi-weekly as appropriate to discuss progress and share new ideas.

Grand Challenge: Engineer better medicines.

Steve Pankavich, pankavic@mines.edu | Alex Pak, apak@mines.edu