Supervised Deep Learning with Gene Functional Annotation for Cell Classification
Zhexiao Lin, Yuanyuan Gao, Wei Sun
Abstract
Gene-by-gene differential expression analysis is a widely used supervised approach for interpreting single-cell RNA-sequencing (scRNA-seq) data. However, modern scRNA-seq datasets often contain large numbers of cells, leading to the identification of many differentially expressed genes with extremely small p-values but negligible effect sizes, thus making biological interpretation difficult.
Introduction
A key step in the analysis of single-cell RNA-seq (scRNA-seq) data is gene-by-gene differential expression (DE) analysis, often followed by identification of biological processes enriched among the DE genes. However, because modern scRNA-seq datasets can contain very large numbers of cells, DE analyses frequently produce extremely small p-values for many genes even when the corresponding effect sizes are negligible. As a result, researchers often need to apply additional ad hoc filtering to obtain interpretable results.
Materials and Methods:
Although the applications considered in this paper focus on binary phenotypes, the SDAN framework is not restricted to binary outcomes. SDAN is formulated as a supervised representation-learning framework in which phenotype information enters through the supervised loss term. For multi-class phenotypes, the binary cross-entropy loss can be replaced with the standard softmax cross-entropy loss, with the classifier outputting a probability vector over multiple classes. For continuous outcomes, the supervised objective can be replaced with a regression loss such as mean squared error.
Discussion:
Biological knowledge, such as gene–gene interactions, is often noisy and may vary across tissues and conditions. Consequently, when incorporating such knowledge into deep learning methods, it is important to retain sufficient flexibility to learn the subset of biological information that is relevant to the specific task. SDAN addresses this challenge by combining a supervised classification loss with an unsupervised graph loss, thereby favoring gene sets that both align with biological knowledge and discriminate among cells.
Citation: Lin Z, Gao Y, Sun W (2026) Supervised deep learning with gene functional annotation for cell classification. PLoS Comput Biol 22(6): e1014327. https://doi.org/10.1371/journal.pcbi.1014327
Editor: Peng Wei, The University of Texas MD Anderson Cancer Center, UNITED STATES OF AMERICA
Received: January 27, 2026; Accepted: May 12, 2026; Published: June 1, 2026.
Copyright: © 2026 Lin et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: “Su et al. COVID-19 dataset: Gene expression data were downloaded from ArrayExpress: https://www.ebi.ac.uk/biostudies/arrayexpress/studies/E-MTAB-9357 on 2/22/2022. Patient information was extracted from Table S1 of Su et al.: https://data.mendeley.com/datasets/tzydswhhb5/5 on 2/22/2022.” SEA-AD dataset: The snRNA-seq data were downloaded from cellxgene website https://cellxgene.cziscience.com/collections/1ca90a2d-2943-483d-b678-b809bf464c30 on 2021/11/21. Donor meta data and clinical data were downloaded from the SEA-AD website https://portal.brain-map.org/explore/seattle-alzheimers-disease/seattle-alzheimers-disease-brain-cell-atlas-download on 2022/1/7. Cancer immunotherapy dataset: The gene expression data of Sade-Feldman et al. 2018 were downloaded from https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE120575. on 2023/11/6. Cell and sample information was extracted from Supplementary Table 1 of Sade-Feldman et al. 2018. The gene expression and meta-data of Yost et al. 2019 were downloaded from https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE123813. Codes for data processing and running SDAN for all the datasets are available at https://github.com/Sun-lab/SDAN.
Funding: This work was supported by the National Institutes of Health (HG013177 to WS; GM105785 to WS and ZL). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
