ReLMM : Reinforcement Learning Optimizes Feature Selection in Material Models
In this work we have developed a data driven model to identify a near minimal subset of descriptors that are most correlated with the target variable while maintaining target variable prediction accuracy. We have demonstrated the capabilities of the model on a synthetic multiscale dataset and on a hybrid metal halide perovskite dataset. We compared the results of our model to other feature selection algorithms like LASSO and XGBoost and have observed our model perfoms on par, if not better than these models in terms of finding the near minimal subset.
Authors: Nikhil Thota*, Maitreyee Sharma Priyadarshini*, Rigoberto Hernandez
* Equal ContributionNestedAE: interpretable nested autoencoders for multi-scale materials characterization
We developed a physics informed neural network architecture for multi-scale materials modelling called NestedAE. We compared the performance of this architecture versus other state of the art machine learning models using a synthetic and metal halide perovskite dataset with bandgap and device efficiency properties.
Authors : Nikhil Thota*, Maitreyee Sharma Priyadarshini*, Rigoberto Hernandez
* Equal ContributionCorrelation Between Chemical Denaturation and the Unfolding Energetics of Acanthamoeba Actophorin
Actophorin belongs to a family of actin filament severing proteins. Site specific mutations in the protein have been found to increase the thermal stability of the protein compared to the wildtype. We used a previously developed schema called Adaptive Steered Molecular Dynamics with Telescoping Box (ASMD-TB) to analyze the resulting changes in the intramolecular interactions by computing the Potential of Mean Force (PMF) of Unfolding. We found that the relative stability of the mutant compared to the wildtype stems from the difference in the number of hydrogen bonds and salt bridges.
Authors: Nikhil Thota, Stephen Quirk, Yi Zhuang, Erica Stover, Raquel L. Lieberman, Rigoberto Hernandez
Implementation of Telescoping Boxes in Adaptive Steered Molecular Dynamics
We developed a schema that combines the Adaptive Steered Molecular Dynamics (ASMD) schema with a solvent box that varies in size between the ASMD stages. This solvent box can vary to maintain a constant cross section or constant volume as shown in the figure to the right. The method was benchmarked on the unfolding of Ala30.
Authors: Yi Zhuang, Nikhil Thota, Stephen Quirk, Rigoberto Hernandez
A Stoichiometric and Pseudo Kinetic Model of Loop Mediated Isothermal Amplification
We developed a stoichiometric model to comprehend the LAMP reaction network by classifying its reaction products and quantifying their populations. A simplified diagram showing the reaction pathways is shown in the figure to the right. We found that it is the smaller length reaction products that contribute more towards the exponential amplification of the target sequence.
Authors: Navjot Kaur*, Nikhil Thota*, Bhushan Toley
* Equal Contribution