University of Washington
Machine Learning Guided Structure Prediction and Design of Macrocycles
SESSION 13: NEW FRONTIERS IN COMPUTATIONAL PEPTIDE DESIGN PART 2
Thursday, June 29, 2023, at 11:10 am - 11:30 am
Machine learning, ML, methods have demonstrated tremendous success in the structure prediction and design of larger proteins.1, 2 However, these established methods do not translate well to small peptides with non-canonical amino acids and crosslinks. We have previously described Rosetta-based computational methods for accurately designing constrained peptides and macrocycles. 3, 4. We have recently extended these approaches to design peptide binders against therapeutic targets and macrocycles with enhanced membrane permeability and oral bioavailability. 5
Here, we present our recent breakthroughs in machine learning-guided methods for structure prediction and de novo design of macrocycles for diverse structures and functions. In our benchmarking studies, these methods accurately predict the structure of macrocycles available in the Protein Data Bank.
Next, we extended these methods for redesigning and hallucinating structurally and chemically diverse macrocycles. Experimentally determined structures of these hallucinated macrocycles match very closely with the design models. In parallel, we have also developed generative models for macrocycles, which are based on variational autoencoders and denoising diffusion probabilistic models and show substantial improvements in time and accuracy.
These generative macrocycle models enable "unconstrained hallucination" for sampling diversity and "constrained hallucination" to incorporate arbitrary chemical and structural requirements in the de novo-designed macrocycles. Overall, these new ML-guided methods offer opportunities for accurate and custom design of peptides for various functional properties.
1. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFoldv. Nature 596, 583–589, 2021.
2. Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876, 2021.
3. Bhardwaj, G. et al. Accurate de novo design of hyperstable constrained peptides. Nature 538, 329–335, 2016.
4. Hosseinzadeh, P. et al. Comprehensive computational design of ordered peptide macrocycles. Science 358, 1461–1466, 2017.
5. Bhardwaj, G. et al. Accurate de novo design of membrane-traversing macrocycles. Cell, 2022
Gaurav received his doctoral degree in Integrative Biosciences from the Pennsylvania State University. During graduate school, he developed computational methods for studying evolution of highly divergent protein families. He did his postdoc with Dr. Kit Lam at University of California, Davis and Dr. David Baker at University of Washington, Seattle. His postdoctoral work focused on computational design of hyperstable constrained peptides and macrocycles with atomic-level accuracy.
Research Interests in the Bhardwaj lab focuses on Computational peptide design for enhanced cell permeability, oral bioavailability, and blood-brain barrier traversal. The lab is also interested in high-throughput design of targeted peptide therapeutics