I am a Computational and Systems Biology PhD student at MIT in the Reddien lab. I am currently investigating claims of agelessness in the animal kingdom. Demographic studies have failed to find evidence of aging in several organisms, including highly regenerative planarian flatworms and hydra. However, some theoretical arguments claim that aging should affect all multicellular life. Are the theoretical arguments wrong and some organisms really are ageless, or are the demographic studies too limited in either time or sample size to detect aging in the examined populations? My project seeks to evaluate the likelihood of the existence of agelessness through two major avenues. The first investigates the evolutionary optimality of agelessness through mathematical models. The second approach investigates mechanistic barriers to agelessness, including mutational decay that we expect to affect all asexual lineages.

I paired up with John Mapes, a PhD student from LA Tech, to try to increase the accuracy of Disulfide Bond prediction techniques that currently exist. In proteins, the amino acid cysteine forms a strong disulfide bond with other cysteines. If one can predict which cysteines will bond in a protein, predicting the 3D structure of that protein becomes much easier. We ran PSI-BLAST on the amino acids neighboring the cysteines to generate a Local Similarity Matrix (analogous to a PSSM) to use as features for our data mining models. In addition to the LSM, we also got features from the predicted secondary structure, euclidean distance off of a Modeller prediction, and Correlated Mutations. Once we gathered all our features, we ran the data through several data mining models such as Neural Nets, SVM, Random Forests, KNN, etc. Our methods achieved record accuracies. Now, we are working on exciting new methods to predict cysteine oxidation states. We are experimenting with a feature matrix called RAM (Residue Adjacency Matrix) which operates off of raw sequential distance of each amino acid to the target position. This feature source scales better than a PSSM and has a larger effective range.

I participated in MSRP BIO (MIT's Summer Research Program in Biology) during my junior summer in 2018. During this REU, I worked in Dr. Tomaso Poggio's Lab in the McGovern Institute for Brain Research at MIT's Building 46. My project focused on understanding how and why robustness evolves in neural networks during their training regiments. "Robustness" is a measure of how resistant a neural network is to having its neurons ablated (activations clamped to zero). Like natural brains, neural networks can suffer loss of a certain amount of neurons without a dramatic detriment to function. Studying how robust a neural network is to neuron ablation is a good tool to understand how important any individual neuron is to the function of the whole network or, alternatively, the redundancy of neurons in the network. Previous research has focused on the robustness of fully-trained neural networks. By comparing the outputs of the network before and after neuron ablations at every training step, we get an estimate of the network’s robustness over time. We did several experiments in which we adjusted the architecture of the neural network, the type of data that we trained on, and the number of training examples we made the network fit. In all of our experiments, the network starts relatively robust, and then robustness dramatically decreases in the first few training steps. The network then recovers to a certain robustness which depends on various facets of the training regime such as the number of neurons in the network and the redundancy of the input data. The results that we obtained lead us to believe that robustness is related to both the data complexity and the effective capacity of the network.

Assisted with the mapping of various strains of the HIV virus via TCL and Python scriptsAutomated the process of submitting and retrieving proteins from the ITASSER serverSimulated solvating and ionizing proteins for further analysisAutomated the running of NAMD scripts on the proteins to fold them correctlyCreated phylogenetic trees of the virus samples in order to understand how the virus evolves

I had the opportunity to participate in a paid data science internship for CenturyLink during the spring of my junior year in 2018. During my work there, I attempted to connect several databases that existed in the CenturyLink Data Lake. When records from different databases are connected with each other, then the data becomes exponentially more powerful. For instance, if we had a database that contained customer information and a database that contained billing information that were not connected, then we would know where each customer was located and how much each customer was paying, but we would not know how much revenue each geographic region was providing for the company until we connected the databases. I developed several tools which would automatically search the databases and try to find connections to other databases in the Lake. These connections were then scored based on their likely strength and a graph was created that contained databases, tables, columns, and the links which connected them together.

I attended the BRITE (Bioinformatics Research and Interdisciplinary Training Experience) REU at Boston University during my sophmore summer in 2017. As part of this REU, I developed Micro Droplet Rate/Region Ocular Processor for the Densmore Lab. This program takes videos of microfluidic droplet generators and automatically calculates the droplet generation rate and the droplet diameters. It operates by applying Canny's Edge detection algorithm to each frame and finding the average pixel value. The graph of these average pixel values is a sine wave. The program smoothes this resulting wave and finds the period of it in order to calculate the droplet generation rate. A rectangle is fit to the frames that correspond to the local maxes of the wave. The length of these rectangles is taken as the diameter of the droplet. For the second part of my REU at Boston University, I developed a program called DAFD (Design Automation based on Fluid Dynamics). The goal of DAFD was to take a user's desired droplet generation rate and droplet diameter and suggest chip parameters to build a droplet generator that would correctly produce their droplets. To achieve this, we made 2500 droplet generators with known chip parameters and used μ-DROP to analyze the droplet generation rate and the average droplet diameter. Then, we fit two interpolation curves (one for generation rate and one for droplet size) to the collected data set. We then used a minimization algorithm to minimize the error of our solution set to both the rate curve and the size curve.