SCOIT

SCOIT applies probabilistic tensor decomposition to single-cell multiomic data to capture higher-order correspondences and extract latent cell, gene, and omic embeddings for integrative analyses including clustering, cross-omics imputation, and gene regulatory network inference.

Key Features:

• Probabilistic tensor decomposition: Represents single-cell multiomic data as higher-order tensors to capture relationships among omic layers such as genomics, transcriptomics, and proteomics.
• Statistical models: Integrates Gaussian, Poisson, and negative binomial distributions to model sparsity, noise, and heterogeneity in single-cell data.
• Embedding decomposition: Decomposes multiomic tensors into cell, gene, and omic embedding matrices that provide latent representations for downstream analysis.
• Cellular heterogeneity analysis: Uses cell embeddings to improve cell clustering, outperforming nine state-of-the-art tools across multiple metrics.
• Cross-omics gene expression and regulatory network analysis: Produces gene embeddings that enable cross-omics analysis of gene expression patterns and construction of integrative gene regulatory networks.
• Cross-omics imputation: Facilitates simultaneous imputation across omic layers, improving Pearson correlation coefficients by 3.38% to 39.26% over existing methods.
• Handling incomplete omic profiles: Operates on datasets in which subsets of cells possess only a single omic profile.

Scientific Applications:

• Cellular heterogeneity analysis: Dissects cellular heterogeneity and identifies distinct cell populations via latent cell embeddings and clustering.
• Cross-omics imputation: Recovers missing data across omic modalities to improve cross-modal correlations and downstream analyses.
• Integrative gene regulatory network inference: Constructs integrative gene regulatory networks across omic layers using gene embeddings.
• Empirical validation on diverse datasets: Applied to eight diverse single-cell multiomic datasets derived from various sequencing protocols to evaluate performance.

Methodology:

Represents single-cell multiomic data as higher-rank tensors and applies probabilistic tensor decomposition using Gaussian, Poisson, and negative binomial models to factorize into cell, gene, and omic embedding matrices and enable simultaneous cross-omics imputation.