Single-cell RNA sequencing, named Nature's Method of the Year in 2013, now has over 1,000 computational tools built around it, more than 150 of which use deep learning. The data structure is simple: a matrix where rows are individual cells and columns are genes, with values representing RNA expression levels per cell. The human genome has roughly 20,000 to 25,000 protein-coding genes out of 3 billion base pairs, and measuring which genes are active in which cells, at single-cell resolution, is the entire game.
Deep learning matters here because the data is high-dimensional, sparse, and noisy by nature. Bulk sequencing averages signals across thousands of cells and obscures the differences that drive disease, development, and drug response. Single-cell methods fix that problem, and deep learning fixes the computational problems that follow: cell type annotation, batch correction, dimensionality reduction, and trajectory inference across millions of cells. The Human Cell Atlas Project, led by researchers including Aviv Regev of the Broad Institute and Genentech, is the infrastructure context for why this scales.
The original article is worth reading in full because it moves past the overview and into specific deep learning architectures applied to each analytical task, with enough technical detail to evaluate the tradeoffs. If you work in computational biology, genomics, or applied ML, the taxonomy of methods alone justifies the time.
[READ ORIGINAL →]