Multi-Omics Autoencoder

Each omics layer (genomics, transcriptomics, proteomics) is processed by a dedicated encoder network before fusion. This allows the architecture to learn the unique statistical properties and noise distributions of each data type independently.

• Genomics: Often uses 1D convolutional layers to capture local sequence motifs or fully connected networks for variant data.
• Transcriptomics: Typically employs dense layers with dropout to handle high-dimensional, sparse gene expression matrices.
• Proteomics: May use graph neural networks if protein interaction data is available, or dense layers for abundance data.

This modular design prevents the dominant variance of one modality from drowning out the signal in others.

The central constriction layer forces the network to learn a compressed, non-linear representation of the integrated multi-omics data. This bottleneck is the core of the autoencoder's utility.

• Dimensionality Reduction: Compresses tens of thousands of molecular features into a low-dimensional vector (e.g., 32–256 dimensions).
• Denoising: By learning to reconstruct the original inputs from this compressed code, the model is forced to ignore non-biological technical noise.
• Biological Signal Extraction: The latent space organizes samples by true biological state (e.g., disease subtype) rather than by batch or platform effects.

The point at which modality-specific representations are combined is a critical architectural decision. Common strategies include:

• Early Fusion: Concatenating raw or lightly processed features before a joint encoder. Simple but can struggle with disparate dimensionalities.
• Intermediate Fusion: The most common approach. Modality-specific encoders produce latent vectors that are concatenated and fed into a shared bottleneck layer. This balances modality-specific learning with joint representation.
• Cross-Modal Attention: An advanced mechanism where features from one modality (e.g., gene expression) are used to compute attention weights over features from another (e.g., protein abundance), explicitly modeling inter-omics relationships before fusion.

A key practical challenge in multi-omics is that not all assays are performed on every patient sample. A robust autoencoder must handle this missingness gracefully.

• Zero-Masking: Missing input features are set to zero, and the model learns to ignore them. Simple but can introduce bias.
• Dropout-based Imputation: During training, entire modalities are randomly dropped out, forcing the decoder to reconstruct them from the remaining modalities. This trains the model to perform cross-modal imputation at inference time.
• Product-of-Experts: The joint latent distribution is modeled as a product of individual modality-specific Gaussian experts, naturally handling missing modalities by simply omitting the missing expert's contribution.

Extending the autoencoder to a Variational Autoencoder (VAE) transforms the deterministic bottleneck into a probabilistic latent space. Instead of a single point, the encoder outputs a mean and variance vector parameterizing a Gaussian distribution.

• Smooth Latent Space: The KL divergence loss regularizes the latent space to be continuous and smooth, enabling meaningful interpolation between patient samples.
• Generative Capability: New, realistic multi-omics profiles can be generated by sampling from the latent distribution and decoding. This is used for synthetic patient data generation.
• Uncertainty Quantification: The variance in the latent space provides a measure of the model's confidence in its representation of a given sample.

For biomarker discovery, the latent space can be explicitly shaped by a clinical outcome. This is achieved by adding a supervised loss term.

• Latent Space Classifier: A small neural network is attached to the bottleneck layer and trained to predict a phenotype (e.g., responder vs. non-responder).
• Conditional VAE: The decoder is conditioned on both the latent code and a known label, allowing for controlled generation of class-specific molecular profiles.
• Feature Attribution: After training, techniques like SHAP or integrated gradients can be applied to the encoder to identify which input features most strongly influence the clinically relevant latent dimensions, directly nominating biomarkers.