Multi-Omic Variational Autoencoder (MVAE)

The MVAE aggregates information from available modalities using a Product-of-Experts (PoE) factorization. Each modality-specific encoder produces a Gaussian expert; their precision-weighted product forms the joint posterior. This formulation naturally handles missing modalities—absent experts contribute a uniform distribution, allowing inference to proceed without imputation. The PoE yields a sharper, more confident joint posterior than mixture-based alternatives.

Each omics layer receives a dedicated decoder network that reconstructs its data type from the shared latent code. This design accommodates fundamentally different data distributions:

• RNA-seq: Negative binomial or zero-inflated likelihoods
• ATAC-seq: Bernoulli or Poisson likelihoods for binary peak calls
• DNA Methylation: Beta-binomial or continuous-valued outputs Decoders operate independently, allowing the architecture to scale to new modalities without retraining existing components.

The shared latent variable z is regularized toward a standard Gaussian prior via the Kullback-Leibler (KL) divergence term in the Evidence Lower Bound (ELBO). This imposes a smooth, continuous structure on the embedding space, enabling:

• Interpolation between biological states
• Synthetic data generation by sampling from the prior
• Cross-modal translation by encoding from one modality and decoding into another The KL weight can be annealed during training to balance reconstruction fidelity against latent organization.

A defining capability of MVAEs is generative imputation of entirely absent omics layers. Given a subset of modalities, the model encodes to the joint latent space and decodes the missing modality. This is clinically valuable when expensive assays like proteomics or whole-genome bisulfite sequencing are unavailable for all patients. The imputation quality depends on the correlation structure learned during training—strongly coupled modalities yield more accurate cross-modal predictions.

To ensure robustness to missing data at inference time, MVAEs employ modality dropout during training. Random subsets of modalities are masked for each sample, forcing the PoE inference network to produce coherent posteriors from any combination of inputs. This is distinct from standard dropout applied to neurons—entire data streams are zeroed out. The subsampling ratio is a critical hyperparameter controlling the trade-off between full-modality reconstruction quality and partial-modality generalization.

The standard ELBO is extended with modality-specific reconstruction weights to balance the contribution of each omics layer during training. Without weighting, high-dimensional modalities like transcriptomics (~20,000 genes) dominate the loss, drowning out sparser layers like proteomics (~100-500 proteins). Weighting schemes include:

• Inverse dimensionality: Weight ∝ 1/feature_count
• Learnable log variances: Treat reconstruction weight as a trainable uncertainty parameter
• Equal weighting: Normalize each modality loss to unit variance

The shared, lower-dimensional mathematical manifold where MVAE projects embeddings from genomics, transcriptomics, and proteomics. This space enforces that semantically similar biological states—such as a specific cancer subtype—occupy proximal coordinates regardless of which omics layer was input. The encoder networks learn to compress modality-specific noise while preserving cross-modal signal, enabling cross-modal retrieval and unified visualization of heterogeneous data.

A critical regularization strategy during MVAE training where entire data layers are randomly zeroed out per sample. By forcing the decoder to reconstruct all modalities from a subset of inputs, the model learns robust joint representations that do not collapse when clinical assays are missing. This directly enables the MVAE's hallmark capability: missing modality imputation at inference time without requiring complete patient profiles.

The generative task of computationally converting one omics layer into another using the MVAE's encoder-decoder architecture. For example, predicting proteomic abundance solely from transcriptomic input by passing RNA-seq data through the shared latent space and decoding via the protein-specific head. This capability is essential for inferring expensive or destructive assays from cheaper, routine measurements in clinical settings.

The direct application of MVAE's probabilistic framework to predict entirely absent data layers. Given partial input—say, DNA methylation and gene expression—the model samples from the learned joint posterior distribution to generate the most likely proteomic or metabolomic profile. This addresses the pervasive problem of incomplete patient records in real-world biobanks and enables retrospective multi-omic analysis on legacy cohorts.

MVAE functions as a deep generative model capable of sampling entirely new, statistically coherent multi-omic profiles from the learned latent distribution. These synthetic patients preserve complex cross-modal correlations—such as the relationship between a promoter mutation and downstream RNA expression—without exposing real individual-level data. Critical for privacy-preserving research and augmenting rare disease datasets where sample sizes are insufficient for supervised learning.

MVAE architectures can be extended with conditional inputs that explicitly model technical confounders such as sequencing platform, lab site, or sample preparation protocol. By learning a latent representation that is invariant to batch effects while preserving true biological variability, the model harmonizes multi-omic data collected across different institutions. This enables large-scale meta-analyses without the spurious signals that plague naive concatenation approaches.