The performance of deep generative models for learning joint embeddings of single-cell multi-omics data

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell multi-omics studies integrate diverse data types to overcome sparsity and uncover complex biological patterns.
• Deep generative models (DGMs) offer a promising approach for joint embedding and analysis of multi-omics data.
• The impact of sample size on the performance of DGMs for multi-omics integration remains an open question.

Purpose of the Study:

• To empirically assess the quality of deep generative model-based integrations for single-cell multi-omics data across varying sample sizes.
• To evaluate the robustness of popular multi-omics integration tools using different cell numbers.
• To provide recommendations for experimental design in single-cell multi-omics studies.

Main Methods:

• Literature review of deep learning methods for multi-omics integration.
• Empirical evaluation of eight popular DGMs using CITE-seq (RNA and surface protein) and 10x Multiome (chromatin accessibility and RNA) datasets.
• Analysis of DGM performance based on biological and technical metrics across different cell numbers.

Main Results:

• The study systematically examines the performance of DGMs for multi-omics integration under varying cell numbers.
• Results highlight the impact of sample size on the ability of DGMs to learn effective joint embeddings.
• Comparative analysis of popular tools reveals differences in their robustness to cell numbers for specific multi-omics data types.

Conclusions:

• The findings offer insights into the sample size requirements for successful deep generative model-based multi-omics integration.
• Recommendations are provided to guide the experimental design of future single-cell multi-omics studies.
• The study identifies areas for future development in deep learning approaches for analyzing complex biological data.