ICML Workshop (2025)
Soorin Yim, Kyungwook Lee, Dongyun Kim, Sungjoon Park, Doyeong Hwang, Soonyoung Lee, Amy Dunn(Jackson Laboratory), Danniel Gatti(Jackson Laboratory), Elissa Chesler(Jackson Laboratory), Kristen O'Conell(Jackson Laboratory), Kiyoung Kim
Abstract
Single-cell RNA sequencing (scRNA-seq) enables high-resolution profiling of cellular heterogeneity, offering a promising foundation for predicting phenotypes such as disease status. We propose a pooling strategy that utilizes cell type annotations by first aggregating cell representations within each cell type, followed by integration of cell type representations into a sample-level representation. Evaluated across three scRNAseq datasets of varying sizes and biological contexts, our model consistently outperforms baseline models in phenotype classification. Our model is particularly effective in datasets with missing or sparsely represented cell types. These results underscore the importance of carefully incorporating cell type information for robust phenotype prediction from scRNA-seq data.