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

UniVI, a mixture-of-experts β-variational autoencoder, achieves integrated, coherent embeddings across multimodal single-cell data, including up to three modalities, with improved label transfer and reconstruction.

• Why it matters: Integrating multimodal single-cell data is challenging due to modality mismatch, sparsity, and uneven coverage, limiting comprehensive understanding of cell states and functions.
• What they did: UniVI employs modality-specific encoders/decoders with a shared latent space and alignment objectives, handling paired, trimodal, and mosaic datasets across human and mouse tissues, with optional supervision.
• The result: The framework produces robust, interpretable embeddings that enable cross-modal reconstruction, denoising, and genotype-associated neighborhood discovery, even under severe imbalance or modality-exclusive populations.