SCREAM: Single-cell Clustering using Representation Autoencoder of Multiomics.

Area of Science:

• Single-cell biology
• Computational biology
• Genomics

Background:

• Single-cell multiomics technologies provide deep insights into cellular heterogeneity.
• Integrating diverse omics data presents challenges like high dimensionality and noise.
• Developing robust methods for multimodal single-cell data integration is crucial.

Purpose of the Study:

• To introduce SCREAM, a novel deep learning framework for robust integration and clustering of multimodal single-cell data.
• To address challenges in single-cell data integration, including high dimensionality and modality-specific noise.
• To enable accurate cell type identification from complex single-cell multiomics datasets.

Main Methods:

• SCREAM utilizes Stacked Autoencoders (SAEs) to create latent representations for individual omics modalities and their fusion.
• The framework employs Deep Embedding Clustering (DEC) to refine the integrated latent space and cell cluster assignments.
• SCREAM is a deep learning approach designed for multimodal single-cell data analysis.

Main Results:

• SCREAM demonstrated superior performance compared to eleven state-of-the-art methods on SNARE-seq and CITE-seq datasets.
• The method achieved the highest or near-highest Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) scores.
• SCREAM provides biologically meaningful multiomics embeddings for downstream analyses.

Conclusions:

• SCREAM is a highly accurate and robust method for cell type identification using multiomics data.
• The framework effectively integrates and clusters multimodal single-cell data.
• SCREAM offers valuable latent representations for diverse biological investigations.