Analyze / Multimodal

RNA + ATAC + methylation Product-of-Experts model

Upload aligned omics matrices with features as rows and samples as columns. The app intersects matching sample IDs, preprocesses each modality, and can train a real Product-of-Experts multimodal VAE with separate RNA, ATAC, and methylation encoders/decoders. It reports shared latent coordinates, modality reconstruction error, disagreement, and research anomaly scores.

Research only: Outputs are exploratory signals, not medical advice, diagnosis, or treatment.

Demo mode

Use built-in demo RNA + ATAC + methylation matrices

The demo includes a synthetic shared biological shift across expression, chromatin accessibility, and methylation.

RNA matrix

Genes × samples. Values can be counts, CPM, TPM, or normalized expression.

File is samples × genes Preprocessing Top variable genes

ATAC matrix

Peaks × samples. Peak IDs should look like chr1:10000-10500.

File is samples × peaks Preprocessing Top variable peaks

Methylation matrix

CpGs/regions × samples. Beta values, M-values, or region averages.

File is samples × regions Preprocessing Top variable regions

Shared latent settings

Model engine Latent dimensions

Epochs

Batch size

Learning rate

Loss weights

λ RNA

λ ATAC

λ METH

β KL

Use the VAE engine for the true multimodal model. If PyTorch is not installed on the server, use the SVD fallback or install torch first.