We explore machine-learning methodologies for predicting the outcomes of MD simulations by preserving their accurate time labels. This idea will greatly reduce the computational expenses associated with performing MD, making it broadly accessible beyond the current user-base of scientific researchers to high schools and colleges, where the computational resources are sparse.
Overview of Project Deep-MD
A major thrust of Biophyiscs investigations is to determine the function of proteins in living cells. Addressing this seminal question, three-decades of physics- and data-based computations as well as experiments have been dedicated to resolve protein structures from their sequence, and the area continues to grow. However, dynamics of the proteins (and not just their stationary structures frozen in time) are key for biological functions. Computational modeling of any complex dynamic system essentially boils down to a multivariate time series forecasting task, and hence a time series trajectory data of an evolving biological system is essential to analyze and computationally learn the underlying molecular physics. The specialities that biomolecular dynamics presents in our submitted data sets is that, the underlying data-structure of molecular geometries is high-dimensional (can grow up to 100 million Cartesian dimensions), and the associated time series switches between fast and slowly evolving regimes. Learning and predicting from this high-dimensional rapidly changing data will open new paradigms in the area of molecular simulation of protein dynamics, where our software, NAMD serves a community of nearly 20,000 users. Beyond serving as case studies for the application of LSTM and linear regression methods, applications of resolving these sets will allow real-world applications in the interpretation of data from Atomic Force Microscopy experiments.
Impact of Project Deep-MD
Molecular dynamics or MD simulations have emerged to become the cornerstone of today’s computational biophysics, enabling the description of structure-function relationships at atomistic details. These simulations have brought forth milestone discoveries including resolving the mechanisms of drug-protein interactions, protein synthesis and membrane transport, molecular motors and biological energy transfer, and viral maturation, encompassing a number of our contributions. More recently, we have employed molecular modeling to predict mortality rates from SARS-Cov-2 showcasing its application in epidemiology.