Senior Thesis

Abstract

Recent breakthroughs in single-cell RNA sequencing technologies have improved our understanding of the molecular basis for cellular heterogeneity and dynamics. A crucial step in single-cell data analysis is to identify which cells correspond to which cell types. Traditionally, cell classification has been performed manually by experts, which is expensive and slow. In recent years, deep learning models have been proposed to automatically classify individual cells. However, many of these models yield little to no performance gains compared to simpler and more interpretable models such as logistic regression and support vector machines. In this work, we present a model-agnostic component that uses contrastive learning to improve internal model representations in deep neural network architectures for single-cell classification. We demonstrate that our contrastive component improves classification accuracy and significantly improves AUPRC of logistic regression, especially for rare cell types.

Report

Feel free to check out the brief here or the longer report here.