Unifying multimodal single-cell data with a mixture-of-experts β -variational autoencoder framework

Area of Science:

• Computational Biology
• Single-cell Genomics
• Bioinformatics

Background:

• Multimodal single-cell assays provide complementary biological insights but face integration challenges due to modality mismatch and data sparsity.
• Existing methods struggle with integrating diverse datasets and cohort coverages.

Purpose of the Study:

• To develop a scalable and flexible framework, UniVI (Unified Variational Inference), for seamless integration of multimodal single-cell data.
• To enable robust analysis across various study designs, including paired, tri-modal, and mosaic data.

Main Methods:

• UniVI employs a scalable mixture-of-experts beta-variational autoencoder architecture.
• It learns a shared latent space while preserving modality-specific structures using coupled encoders/decoders and a cross-modal alignment objective.
• The framework supports optional supervised heads and accommodates different data types like RNA, protein, chromatin accessibility, and DNA methylation.

Main Results:

• UniVI successfully integrates paired (CITE-seq, Multiome, SHARE-seq) and tri-modal (TEA-seq) data, producing coherent embeddings and improving label transfer.
• It demonstrates robust performance even with cell-type imbalance and modality-exclusive populations.
• The framework enables cross-modal reconstruction, denoising, and analysis of mosaic designs, including mutation-aware fine-tuning.

Conclusions:

• UniVI offers a flexible and interpretable framework for multimodal single-cell data integration.
• It effectively handles diverse data types and complex study designs, facilitating deeper biological discovery.
• The method supports practical applications like reference-to-query projection in partially observed datasets.