VAE-based multiscale simulation
We have recently shown that Variational Autoencoders (VAEs) can be used to learn efficient, collective Monte Carlo (MC) moves [1], in contrast to manual development of sophisticated move sets. VAE-based MC moves also provide an opportunity for seamlessly switching between resolutions of multiscale models, using coarser models to accelerate sampling while exactly recovering the correct ensemble at finer resolution. Our group is applying these techniques to predicting protein conformational ensembles, self-assembly of peptides, and polymer-based flexible electronics. Of particular interest in all of these scenarios is the critical role of water in driving interactions and determining mesoscale structure. Significant computational resources are required, however, to accurately model water in these contexts. We are using VAE-based MC moves to switch between solvent models, allowing for enhanced sampling with cheaper solvent models yet recovery of the exact details and effects associated with more expensive and reliable water models.
Techniques that we develop in this research area will enable the accurate mapping of self-assembly behaviors across a wide range of state conditions, which is a key step in designing soft materials with tunable responses to their environment (e.g., changes in temperature or concentration). Critically, we will simultaneously capture mesoscale structural shifts and atomistic details. The interplay between these scales can play a key role in determining the materials structure, as well as properties. For example, the electronic and ionic mobility within conducting polymeric materials is determined by both the size and shape of mesoscale domains as well as the ordering and fluctuations of atoms within each larger region.
Advanced solvation free energy calculations
Our group is developing rapid yet rigorous methods for determining solvation free energies through the merging of novel statistical mechanical formalisms and machine learning techniques. In molecular simulations, the ability to perform solvation free energy calculations opens up routes to computing many thermodynamic properties, including adsorption free energies, partition coefficients, binding free energies, and interfacial free energies. Such calculations remain difficult, however, for large biomolecules and biomolecular assemblies. We are leveraging information theory to rigorously write solvation free energies in terms of water’s response to a solute or interface, then coupling this to machine learning methods that learn how the probabilities of water degrees of freedom change. Excitingly, this allows us to see how specific shifts in water behavior contribute directly to free energy changes. This opens up opportunities in many application areas, from better understanding the function of ice-binding proteins, to rapidly designing interfaces for use in membranes by tuning their nanoscale features.
Computational design of membrane and chromatographic interfaces
Methodological developments in the group fuel a diverse set of application areas, which in turn inspire further research into theoretical methods. By combining VAE-based MC moves, which can accelerate simulations while preserving solvent information, with advanced solvation free energy calculation techniques that efficiently make use of that information, we have unique opportunities to study binding of both organic and inorganic solutes in membranes and chromatographic media.
A current project focuses on the determination of salt aggregation and binding propensities at interfaces of water-purification membranes. This type of fundamental knowledge is critical to reduce scaling in separations that purify contaminated waste streams with high saline or mineral content. With the ability to make such predictions rapidly, we are also seeking to determine ways in which the nanoscale geometry of the membrane interface may be adjusted to control or prevent salt deposition.
Another key thrust in this area will involve the prediction of free energies of binding of large proteins. It is currently computationally expensive to calculate free energies of binding of large proteins at interfaces, especially with accurate representation of water, which can play a large role in determining protein-interface interactions. We envision that our technical developments will make these types of calculations routine, opening the door for the computational design of interfaces, such as those on chromatographic media, for highly specific protein affinity. This will prove an important tool as new classes of proteins are developed for use as therapeutics or in consumer products. Alternatively, surfaces could be designed to resist binding by specific proteins, as is desirable in coatings to prevent biofouling.