Predicting Biomolecular Dynamics

For life to exist as we know it, biomolecules must convert between distinct states with specific timing. These dancing molecules are not just a fun quirk of biology – fundamental principles dictate that biomolecules need kinetics for the exquisite specificity they achieve in the messy interior of a cell.

What could we unlock for fundamental molecular biology and its applications (drug discovery, enzyme design, and more) if we had predictive power for the motions of biomolecules? The ability to accurately predict protein dynamics – multiple conformations, their probabilities, and the kinetics of transitions between them – is the next grand challenge for structural biology. My research seeks a quantitative and predictive understanding of biomolecular dynamics, and a deeper understanding of how evolution shapes dynamics and function.

The Wayment-Steele lab will be at the interface of computational and experimental structural biology: we will formulate and train models with principled biochemical questions in mind, and test their predictions with our own experiments.

We will initially focus on integrating deep learning and NMR dynamics experiments (relaxation dispersion), which are rich in information and undervalued in protein deep learning, but currently limited in throughput. My postdoctoral research focused on how AlphaFold and protein language models contain information on dynamics, yet deep learning in molecular biology is evolving rapidly.

The fundamental principles my lab will focus on for any modeling approach are: how do we use sparse, yet multimodal data in training; how do we construct meaningful prospective experimental tests, and how do we interpret what models are learning?

Maximizing the Impact

How will we maximize the impact of our results? De novo enzyme design will make for one stringent test in our own biological systems. Being situated in the vibrant Biochemistry department at UW-Madison will also allow for numerous opportunities to apply our work to more biologically relevant questions.

We will work with broader communities both at UW-Madison and beyond to conduct internet-scale crowdsourced challenges both for deep learning and molecular design. This allows us to recruit diverse expertise and creativity to these challenges, and hopefully allows many more worldwide to discover the delights of data on dancing molecules.