(6cn) Understanding Multimodal Interactions through Deep Neural Networks and Statistical Mechanics

I am a fourth year PhD candidate in the Chemical and Biological Engineering department at RPI with an interest in combining MD simulations with AI/machine learning techniques to facilitate the design of proteins and multimodal surfaces for a range of applications. I am fascinated with the fundamental problem of understanding how water mediates complex interactions between surfaces containing multiple modes of interaction (e.g. hydrophobic or charge interactions). Despite the fact that these “multimodal” interactions play a central role in many biological processes, our fundamental understanding of the principles governing these interactions remains limited. In recent years, the advent of GPU computing has resulted in dramatic improvements in the computational speed of MD simulations. Additionally, the field of AI has experienced a renaissance, with neural networks being successfully employed for a variety of problems including reading text, classifying images, and even recently folding proteins. Thus, a new field combining the wealth of data arising from rapid MD simulations and the AI tools to mine this data is emerging. As a junior faculty member, my aim will be to leverage these powerful tools to address the fundamental question of how the structure of multimodal surfaces governs their interactions.

Teaching Interests:

As a graduate student, I have enjoyed the process of mentoring students at different levels. I have mentored three high school students through the process of formulating a research question, performing molecular dynamics simulations, and ultimately presenting their data at either a poster session or in a group meeting. I have also created more complex projects for three talented undergraduate students. In this case, because the students were more advanced, I guided them to learn the basic academic skills required as well as to learn to become partially independent researchers. Watching these students learn to internalize their projects to the point where they are excited to work independently and propose new ideas has been extremely rewarding. As a senior graduate student, I have recently played the role of mentoring first and second year graduate students through this process as well, requiring me to teach the creative skill of being a truly independent researcher, while also requiring me to teach thermodynamics and statistical mechanics at a deeper level. As a professor, I look forward to mentoring students as well as teaching these related courses (thermodynamics, separations, or statistical mechanics) or any other core courses formally in a classroom setting. I feel that my teaching experience has been personally rewarding and has helped me strengthen my own knowledge in these fields.

Future Directions:

There are two main areas of work that I would like to expand upon from my PhD research:

1. A Neural Network Approach to Predicting Patch-Based Protein Surface Properties
   Understanding the physical properties associated with protein surface patches is important for a wide variety of applications including protein solubility (important for protein crystallization or formulation of biologic therapeutics) and protein-x interactions (important for drug discover, protein-protein interactions, and protein surface interactions). Researchers have used variational autoencoders to evaluate the similarity of small molecules in latent space in order to predict molecular properties. We have recently developed an approach for decomposing protein structures (from PDBs) into protein surface patches (represented using a voxel representation with channels being used for atomic occupancies) and using a variational autoencoder with 3D convolutional layers to encode these patches into a latent space representation. In this representation, we can observe relationships between protein surface patches that capture classical heuristics (i.e. hydrophobic or charge classifications) as well as identify unexpected relationships. In the future, I plan to build upon this work to identify minimal sets of protein patches from which important classes of proteins can be reconstructed. I will then perform molecular simulations to quantify important parameters such as solvation (free energy of cavity formation) and protein-ligand interactions for this minimal set and use this data to predict these important parameters for the relevant classes of proteins. An advantage of this approach is that this latent space representation can also suggest mutations to proteins that might significantly alter the behavior of a given patch (with the caveat that that mutation may change the protein structure in unexpected ways).
2. Designing Patterned Surfaces through Self-Assembly of Small Molecules
   In my PhD, I have studied designing complex multimodal surface (including charged, hydrophobic, and hydrogen bonding characteristics for the application of chromatographic materials for protein separation. In addition to this application, there are many other applications for which designing patterned surfaces is important (design of biomaterials, biosensors, etc.). We have recently identified (using molecular dynamics simulations) a class of chromatographic ligands which self-assemble into patterned surfaces containing concentrated patches of hydrophobicity and charge on the nanometer length scale. The principles underlying this phenomenon are closely related to the self-assembly observed in surface-grafted block copolymer systems but differ importantly in that the ligands are much shorter and contain small chemical moieties instead of “blocks” resulting in shallow pattern formation instead of the formation of different phases. In the future, I will build upon this work to develop a statistical mechanical model for predicting the surface structure that an individual ligand will self-assemble into. Through this, I will map out the design space of surfaces which can be synthesized in this manner. Additionally, I will collaborate with experimentalists to evaluate important surface structures experimentally. This work has important implications for the design of new chromatographic materials and for the design of functionalized surfaces in general.