U. Deva Priyakumar

Center for Computational Natural Sciences and Bioinformatics

International Institute of Information Technology

Gachibowli, Hyderabad, 500 032

E-mail: deva[at]iiit.ac.in

 

 

Brief Bio-sketch:

Deva Priyakumar finished his PhD from Pondicherry University/Indian Institute of Chemical Technology in the area of computational chemistry. After a postdoctoral fellowship in the University of Maryland Baltimore in the area of biomolecular simulations and computer aided drug design, he joined as a faculty member in IIIT Hyderabad in 2008. He is currently an Associate Professor in CCNSB, IIIT Hyderabad and his research interests are in the application of molecular dynamics simulations, ab initio calculations and machine learning methods to study chemical and biological processes. Some of the recognitions include the INSA young scientist medal, Innovative young biotechnologist award, JSPS visiting fellowship, and distinguished lectureship award of the Chemical Society of Japan.

MACHINE LEARNING FOR DFT QUALITY ENERGIES AND FOR ANALYSING BIOMOLECULAR TRAJECTORIES

Recent advances in deep learning methods seemed to have resulted in resurgence of their applications in natural sciences during the last few years. Fundamentally, these data driven methods can broadly be classified as supervised and un-supervised methods. In the first part of the presentation, we will discuss the use of artificial neural network for predicting energies of small molecules. The ANN model was obtained based on a novel molecule featurization inspired by additive force fields (BAND: bag of Bonds, Angles, Nonbonds and Dihedrals). We will show that this model is applicable not only to the class of molecules that were used for the training, but also to more complex molecules. While there is certainly room for improvement, the apparent potential energy function can also be used to perform geometry optimization. In the second part of the talk, we will present the use of unsupervised machine learning along with graph theory to extract folding pathways from replica exchange molecular trajectories. A suitable vector representation was chosen for each frame in the macromolecular trajectory and dimensionality reduction was performed using PCA. The trajectory was then clustered using a density-based clustering algorithm, where each cluster represents a meta-stable state on the energy surface of the biomolecule. A graph was created with these clusters as nodes. We hypothesize that the most probable path of (un)folding from a starting to an ending state is the widest path (path which has maximum minimum edge weight) along the graph. Our method makes the understanding of the mechanism of unfolding in RNA hairpin molecule more tractable. As this method doesn't rely on temporal data it can be used to analyse trajectories from Monte Carlo sampling techniques and replica exchange molecular dynamics (REMD).

References:

[1].A. Chattopadhyay, M. Zheng, M. P. Waller, U. D. Priyakumar, J. Chem. Theory Comput., 14 3365, (2018)