Building a learnable universal coordinate system for single-cell atlas with a joint-VAE model

• DOI: 10.1038/s42003-024-06564-0
• Publication Date: 2024-08-12T05:02:11Z
• Authors: Haoxiang Gao; Kui Hua; Xinze Wu; Lei Wei; Sijie Chen; Qijin Yin; Rui Jiang; Xuegong Zhang
• Keywords: 0303 health sciences; 03 medical and health sciences; QH301-705.5; Gene Expression Profiling; Humans; Computational Biology; Biology (General); Single-Cell Analysis; Transcriptome; Article

A universal coordinate system that can ensemble the huge number of cells and capture their heterogeneities is vital importance for constructing large-scale cell atlases as references molecular cellular studies. Studies have shown exhibit multifaceted in transcriptomic features at multiple resolutions. This nature complexity makes it hard to design a fixed through combination known features. It desirable build learnable model major serve controlled generative data augmentation. We developed UniCoord, specially-tuned joint-VAE represent single-cell lower-dimensional latent space with high interpretability. Each dimension either discrete or continuous feature, supervised by prior knowledge unsupervised. The dimensions be easily reconfigured generate pseudo profiles desired properties. UniCoord also used pre-trained analyze new unseen types thus feasible framework annotation comparison. provides prototype enable better analysis generation highly orchestrated functions heterogeneities. designed create singlecell data, capturing enhance