Computational Structural Biology at Vanderbilt

One can think of computational structural biology as two main approaches: knowledge-based and physics-based. Knowledge-based approaches, such as Rosetta, which incorporate elements of physical and chemical realities have made big strides in modeling protein structures. Physics-based methods such as MD simulations excel in revealing the dynamic modes of a structure.  However, these approaches are extremely expensive and scale rather poorly to large biomolecular systems. Biased potential MD simulations meant to lessen the computational cost are liable to sampling error. 

Advances in AI offer opportunities to develop a new generation of molecular simulation platforms to address these limitations. AI methods such as AF2 extract evolutionary information from multiple sequence alignments (MSA) to constrain the space of potential interactions and motifs that a given sequence can adopt. Reducing the confidence of these constraints by under-sampling the MSA can lead to structural heterogeneity and less probable (i.e., intermediate) configurations in AF2 predictions. This constrained space represents a rich potential distribution from which thermodynamic ensembles can potentially be derived. These ensembles can be obtained using AI-assisted structure prediction combined with conventional importance sampling techniques, such as Metropolis-Hastings, or they can be integrated into MD simulations to provide bias in metadynamics simulations. Alternatively, structures derived from reduced-MSA AF2 predictions can be directly combined with AI-assisted biased MD simulations or reservoir replica exchange MD approaches. In some cases, such as modeling the effects of protein mutations on functional free energy landscapes, reduced MSA-AF2 structures can seed short MD simulations for Markov state models, allowing the effects of mutations to be predicted through simple point energy calculations and energetic reweighting, massively increasing the scale of free energy analysis of biomolecules.

The possibilities are numerous and extend beyond the development of new algorithms. What will be the software environment in which these algorithms are developed and deployed? How will these platforms make their way into industry? New software platforms will be required. The Center seeks to be at the forefront of this revolution.