Probabilistic tensor decomposition extracts better latent embeddings from single-cell multiomic data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell sequencing generates multiomic data, often represented as higher-rank tensors.
• Existing analysis tools frequently reduce tensor data to matrices, losing feature correspondences.
• This limits the comprehensive analysis of complex single-cell multiomic datasets.

Purpose of the Study:

• To introduce SCOIT, a probabilistic tensor decomposition framework for single-cell multiomic data analysis.
• To extract meaningful cell, gene, and omic embeddings from complex datasets.
• To enhance downstream analyses like cell clustering, gene expression studies, and data imputation.

Main Methods:

• Developed a probabilistic tensor decomposition framework named SCOIT.
• Incorporated Gaussian, Poisson, and negative binomial distributions to handle data sparsity, noise, and heterogeneity.
• Applied the framework to eight diverse single-cell multiomic datasets.

Main Results:

• SCOIT achieved superior cell clustering performance compared to nine state-of-the-art tools.
• Gene embeddings enabled effective cross-omics gene expression analysis and regulatory network studies.
• Cross-omics imputation performance improved significantly (3.38-39.26% PCC increase), handling missing omic profiles.

Conclusions:

• SCOIT effectively dissects cellular heterogeneity using multiomic single-cell data.
• The framework provides powerful tools for integrative omics analysis and imputation.
• SCOIT demonstrates flexibility by accommodating datasets with partially available omic profiles.