Deep learning for single-cell sequencing: a microscope to see the diversity of cells

Deep learning has become an increasingly relied-upon approach in single-cell sequencing analysis, according to a review from The Gradient. The article describes how the field evolved from the first single-cell RNA sequencing paper in 2009, which Nature later named Method of the Year in 2013, through multimodal single-cell data integration (Method of the Year 2019) and spatially resolved transcriptomics (Method of the Year 2020).

Single-cell sequencing data is represented as a matrix in which each row corresponds to a sequenced cell annotated with a unique barcode and each column corresponds to a gene, with numerical entries giving expression levels for scRNA-seq data. The review states that scRNA-tools, a database compiling software for single-cell RNA analysis since 2016, included over 1000 tools by 2021, many involving deep learning.

The article cites three reasons deep learning suits single-cell data: high dimensionality, since thousands of genes are measured per cell; non-linearity between gene expression and cell-to-cell heterogeneity; and heterogeneity across diverse cell populations. Unlike conventional machine learning, which requires feature engineering by domain experts, deep learning autonomously captures relevant characteristics and addresses noise and sparsity, the article says.

Among architectures, autoencoders are highlighted for dimensionality reduction while preserving heterogeneity. Variants include denoising autoencoders, which add noise to the initial network layer to reduce overfitting, and variational autoencoders, described as generative models that encode data into vectors of means and standard deviations. Named models include scGen, scVI and scANVI.

For imputation and denoising, the article distinguishes deep learning and non-deep learning methods. It says dropout events produce sparsity with many zero values that may not reflect true expression. Non-deep learning examples include KNN-based imputation by Wagner et al. and the SVAER algorithm by Huang et al. Deep learning examples include the deep count autoencoder (DCA) introduced by Eraslan et al. in 2018, scVI from Lopez et al., which uses a zero-inflated negative binomial distribution, and scScope, which adds a recurrent network layer.

The article also discusses the Human Cell Atlas Project, described as the genetic equivalent of the Human Genome Project. Aviv Regev of the Broad Institute of MIT and Harvard and Genentech Research is quoted comparing the atlas to Google Maps, while Satija of the New York Genome Center is quoted on RNA measurement.