Erica Brown
I’m a rising junior at Brown University, concentrating in Applied Mathematics-Biology. My research interests primarily surround the applications of machine learning to precision medicine
Predicting cell-type-specific responses to chemical perturbations assists targeted treatment development with minimal off-target effects. Experiments profile cells’ transcriptomic responses to various perturbations, however, scaling them to numerous cell types and perturbations is costly. Working at Ó³»´«Ã½ has introduced me to some of the most intelligent and inspiring people that I’ve ever met. This summer has strongly developed my passion for scientific research, and I know that I’ll carry what I’ve learned from my mentors here for the rest of my career.Computational methods can predict cellular responses in unmeasured contexts, but this requires considering cell-type-specific gene programs. Single-cell foundation models adapt large language models to transcriptomic data, aiming to learn these programs for use in low-data settings.
We evaluate the utility of gene embeddings for generalizing chemical perturbation effect predictions to unseen cell types. Using MIX-Seq data on cancer cell lines, we use gene embeddings in a metric learning framework to (1) predict transcriptomic responses to perturbations and (2) classify chemicals by their mechanisms of action. We compare their performance with empirical gene co-expression data and metric learning-derived gene relationships using cell-line-specific training data. Overall, while embeddings are more informative than co-expression data, they fall short of the performance achieved with cell-line-specific training data
Project: Evaluating Single-Cell Embeddings for Chemical Perturbation Response and Mechanism of Action Prediction
Mentor: Pinar Demetci, Schmidt Center
