Learning Biopolymer Networks from Small Data

Authors: Emily Hopper, Minghao Rostami

Field of Study: Mathematical Sciences: Data Science and Statistics

Program Affiliation: Binghamton University Projects for New Undergraduate Researchers (BUPNUR)

Faculty Mentors: Minghao Rostami

Easel: 13

Timeslot: Morning

Abstract: Actin is a biopolymer that forms a network to both structurally support cells and carry out fundamental cellular processes. These networks grow based on a variety of conditions such as the availability of the Arp2/3 protein complex for branching. Producing real images of actin networks has proved difficult in the past, so it is implausible to collect a large enough dataset to train an image classifier in identifying the growth conditions of a given network. This research explores the implementation of multiple types of noise-based perturbation, such as a Perlin noise algorithm, to augment a small set of experimentally produced actin network density profiles with the goal of reaching a believable training set that can be used to train a convolutional neural network (CNN)-based image classifier and convince it of the augmented data’s authenticity, ultimately aiding in the classification of actin networks based on growth conditions.