scSemiGCN: boosting cell-type annotation from noise-resistant graph neural networks with extremely limited supervision

Bioinformatics. 2024 Feb 1;40(2):btae091. doi: 10.1093/bioinformatics/btae091.

Abstract

Motivation: Cell-type annotation is fundamental in revealing cell heterogeneity for single-cell data analysis. Although a host of works have been developed, the low signal-to-noise-ratio single-cell RNA-sequencing data that suffers from batch effects and dropout still poses obstacles in discovering grouped patterns for cell types by unsupervised learning and its alternative-semi-supervised learning that utilizes a few labeled cells as guidance for cell-type annotation.

Results: We propose a robust cell-type annotation method scSemiGCN based on graph convolutional networks. Built upon a denoised network structure that characterizes reliable cell-to-cell connections, scSemiGCN generates pseudo labels for unannotated cells. Then supervised contrastive learning follows to refine the noisy single-cell data. Finally, message passing with the refined features over the denoised network structure is conducted for semi-supervised cell-type annotation. Comparison over several datasets with six methods under extremely limited supervision validates the effectiveness and efficiency of scSemiGCN for cell-type annotation.

Availability and implementation: Implementation of scSemiGCN is available at https://github.com/Jane9898/scSemiGCN.

Publication types

  • Research Support, Non-U.S. Gov't

MeSH terms

  • Neural Networks, Computer*
  • Signal-To-Noise Ratio
  • Single-Cell Analysis*
  • Supervised Machine Learning