Devang Thakkar

I am a senior undergraduate at Indian Institute of Technology Bombay, where I'm currently working on developing a multi agent architecture and implementing reinforcement learning algorithms for logistics networks for my Master's Thesis under A. Subash Babu.

I'm also working with theoretical models of phage populations in bacterial systems under Supreet Saini. Recently, I spent a semester at CentraleSupelec (formerly Ecole Centrale Paris) where my research was concerned with sample-specific coexpression networks under Chloe-Agathe Azencott. I have also spent time at TransUnion interning as a data scientist, working primarily with the Hadoop ecosystem.

Email  /  CV  /  Github  /  LinkedIn  /  Twitter

Research

The overarching theme of the problems I've worked on is the application of quantitative and computational approaches to problems in biology - especially regulatory networks and evolutionary dynamics. Over the course of these projects, I've worked on machine learning, network inference, population dynamics, multi agent systems and modeling of biological systems. Listed below are synopses of few of the projects I've recently undertaken.

Selecting Features from Sample Specific Coexpression Networks using Random Forests
Guide: Chloe-Agathe Azencott

Sample Specific Coexpression networks may be evaluated from the aggregate network by estimating the effect of each sample on the network as demonstrated in [A1]. The proposition here is that there might be other dissmilarity measures to calculate the edge weight that are at least as expressive. I used random forests to predict the edge weights using measures including L1, L2, and Mahalanobis distances, and identified important features using permutation feature importance. (Code)

Understanding Pathway Perturbation
Self-Undertaken Project

Several techniques have been used to infer gene regulatory networks (GRNs) and important pathways - primarily mutual information [B1], [B2], regression [B3] and heat diffusion [B4]. I used ARACNe to create a GRN from a TCGA BRCA (breast invasive carcinoma) dataset and identify the significant pathways. I further used a dissimilarity approach [B5] to calculate a score (explaining the degree of perturbation) for the FOXM1 pathway for each of the samples. (Code)

Multi Agent Systems for Reverse Logistics Networks (Stage I Thesis)
Guide: A. Subash Babu

Not much attention has been paid to reverse logistics networks from the perspective of optimization of customer initiated returns managent [C1], [C2]. I propose a novel mother-daughter model to combat operational/fixed costs in existing methods, implemented using a multi agent architecteure. I am currently using a temporal difference learning method to help the agents learn the optimal positioning, route, and schedule.

Modeling the optimal propensity of lysogeny for coexisting populations
Guide: Supreet Saini

Temperate phages make a developmental decision between lysogeny and lysis in order to avoid the extinction of not only their own species but also of their bacterial hosts. [D1] We are trying to estimate the optimal lysogenic propensity as a function of the environmental stresses for individual species and the multiplicity of infection in order to maximize coexistence.

About me

I spend my free time reading fiction, walking, and cycling. I've grown a liking for graphic novels over the last couple of years, primarily after having read Alan Moore, Neil Gaiman, Alison Bechdel, and Joe Sacco (Goodreads). Over the last few years, I've also been trying to get better at painting and visual design, but there's still a long way to go (Behance). I've also spent a year as an editor for Insight, IIT Bombay's official student media body, leading two print editions and working on a variety of articles (planning to add links soon). I also like word games and casual software development, which led me to develop an online version of Snatch, a multi-player word game we play at IIT Bombay.


Giving credit where it's due: Jon Barron's webpage