I am an undergraduate student of statistics
at the University of Toronto. My academic interests include
molecular dynamics, statistical mechanics, and
computational biology.
This summer I am working in
the Computational Cancer Genomics Group
led by Dr. Sushant Kumar at the Princess Margaret Cancer Center, where I am investigating a possible link between low-complexity regions in proteins and cancer.
Throughout my first two years in university,
I was working in the computational biophysics
lab led by Dr. Sarah Rauscher, where I participated in two projects:
1. An X-Ray crystallography experiment that applied an
electric field to a protein crystal was conducted
by Hekstra et al. to provide a new method of studying
protein function. The lab recreated that experiment in
silico, using molecular dynamics - a computer simulation
method for analyzing the physical movements of atoms
and molecules, and provided extensive sampling. My goal was to understand the dynamics
of that protein crystal, for which I deployed Markov state models - a statistical
approach to clustering the protein's conformational ensemble into
distinct low-energy states and calculating the probability of transition between them.
2. Dimensionality reduction is a statistical method that attempts to find a more compact representation of data that preserves its geometry. One of the popular approaches for dimensionality reduction is autoencoders - feed-forward neural networks that become narrow in the middle and then expand again. By defining the loss function as the error of data reconstruction and training the network, we attempt to produce data of reduced dimension in the latent space - the bottleneck in the middle. Hernández et al. adapted this idea for the analysis of molecular dynamics simulations. We verified their architecture on several simple proteins and a PDZ domain. Additionally, we augmented the architecture of the Variational Dynamics Encoder with an Equivariant Graph Neural Network.
The following project was completed under the
supervision of Dr. Duncan Dauvergne in the Department of
Mathematics in the Summer of 2023.
Multi-Particle Diffusion Limited Aggregation (MDLA) is a
probabilistic model of infection spread - an "infected"
particle is placed at the origin of an integer-coordinate grid.
At each of the other coordinates, a "healthy" particle is placed
with probability μ, after which it starts performing a simple random walk.
When a healthy particle attempts to land on an infected particle,
it instead becomes infected and freezes in place, causing the growth of the infection cluster.
In one dimension, this model is very well-studied: Kesten and Sidoravicius
showed that the growth
of the infection cluster is almost surely sublinear for
any μ in (0, 1). However, there were (at least at the time of completing the project) not a lot of results
published regarding the model in two or more dimensions.
One of the possible approaches to studying the model in two dimensions
is to consider the process on slabs - a finite number of
integer number lines stacked on top of each other. Throughout the summer,
we proved that for MDLA on k slabs, the process will almost surely have sublinear growth for
any μ in (0, 1/k).
Published in Nature Communications, 2024
Recommended citation: Klyshko E, Kim JS-H, McGough L,
Valeeva V, Lee E, Ranganathan R, Rauscher S (2024) "Functional protein
dynamics in a crystal." Nature Communications 15(1):3244
I have worked as a teaching assistant for the following courses:
This summer I will be a teaching assistant for STA107H5 - An Introduction to Probability and Modelling.