Efficient Generation of Paired Single-Cell Multiomics Profiles by Deep Learning

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell multiomics technologies like CITE-seq and SNARE-seq offer simultaneous measurements but face challenges in complexity, noise, and cost.
• Existing single-omics datasets are vast but underutilized, presenting an opportunity for advanced computational methods.
• The need for cost-effective and less complex methods for generating paired single-cell omics data is critical for broader research application.

Purpose of the Study:

• To develop a deep learning framework, single-cell multiomics generation (scMOG), for in silico generation of single-cell assay for transposase-accessible chromatin (ATAC) data from single-cell RNA sequencing (scRNA-seq) measurements and vice versa.
• To assess the accuracy and biological relevance of cross-omics data generated by scMOG.
• To evaluate the utility of scMOG-generated data in downstream analyses, including tumor sample identification.

Main Methods:

• Development of scMOG, a deep learning-based framework for cross-omics data generation between scRNA-seq and scATAC-seq.
• Validation of scMOG's ability to accurately generate paired multiomics data, even when one omics is missing or outside the training set.
• Application of scMOG to human lymphoma data for tumor sample identification and investigation of its performance in proteomics data generation.

Main Results:

• scMOG accurately performs cross-omics generation between RNA and ATAC data, producing biologically meaningful paired multiomics data.
• Generated ATAC data, alone or combined with measured RNA, showed equivalent or superior performance in downstream analyses compared to experimentally measured data.
• scMOG demonstrated superior effectiveness in identifying human lymphoma tumor samples compared to experimentally measured ATAC data.
• The framework showed robust performance in generating surface protein data from other omics measurements.

Conclusions:

• scMOG offers a powerful computational approach to overcome the limitations of experimental single-cell multiomics profiling.
• The framework effectively generates high-quality, biologically relevant cross-omics data, enhancing the utility of existing single-cell datasets.
• scMOG shows significant potential for improving diagnostic accuracy, as demonstrated in human lymphoma analysis, and has broad applicability across different omics types.