Multi-Omics Integration

The primary architectural distinction in multi-omics analysis defines the direction of data fusion.

• Horizontal Integration (Meta-Analysis): Combines the same omics layer (e.g., transcriptomics) across multiple independent cohorts or studies. The goal is to increase statistical power and identify robust, reproducible biomarkers by overcoming small sample sizes.
• Vertical Integration (Cross-Omics): Combines different omics layers (e.g., genomics + proteomics) measured on the same set of samples. The goal is mechanistic insight—understanding how a genetic variant cascades through transcription to alter protein abundance and ultimately cellular phenotype.
• Mosaic Integration: A hybrid approach that stitches together different omics layers measured on partially overlapping or entirely different sample sets, often using optimal transport or imputation to bridge the gaps.

A critical design choice in deep learning for multi-omics is the stage at which data modalities are combined.

• Early Fusion: Raw or pre-processed features from all omics layers are concatenated into a single input vector before feeding into a model. This allows the model to learn cross-modality interactions from the ground up but suffers from the 'curse of dimensionality'.
• Mixed (Intermediate) Fusion: Each modality is first processed by its own dedicated sub-network (e.g., a Graph Neural Network for a protein-protein interaction network, a CNN for histology images). The learned latent representations are then merged in a shared hidden layer.
• Late Fusion: Independent models are trained on each omics layer, and only their final predictions or decision scores are combined via averaging or a meta-learner. This is robust to missing modalities but cannot learn complex cross-modal feature interactions.

A foundational concept where disparate high-dimensional omics data are projected into a common, low-dimensional embedding space where coordinates have a unified biological meaning.

• Canonical Correlation Analysis (CCA) finds linear projections that maximize correlation between modalities. Deep variants like Deep CCA learn non-linear alignments.
• Variational Autoencoders (VAEs) force each modality into a shared latent bottleneck, regularized by a prior distribution, enabling generative tasks like multi-omics data imputation.
• Contrastive learning aligns modalities by pulling paired samples (same cell, different assay) together and pushing unpaired samples apart, without needing a reconstruction objective.
• A well-constructed latent space allows for cross-modal prediction—inferring protein expression from transcriptomics alone.

Moving beyond matrix factorization, graph-based methods explicitly model biological entities as nodes and their relationships as edges.

• Similarity Network Fusion (SNF) constructs a patient-similarity network for each omics type and iteratively fuses them into a single consensus network. This excels at patient subtyping and survival analysis by capturing both shared and complementary signals.
• Knowledge Graph Embeddings project structured biological knowledge (e.g., gene-disease associations, drug-target interactions from curated databases) into a continuous vector space. This allows geometric operations like link prediction to hypothesize novel relationships.
• Heterogeneous Graph Neural Networks operate directly on multi-relational graphs where nodes can be genes, proteins, drugs, or diseases, learning context-specific representations by message passing across diverse edge types.

A family of statistical methods that decompose a high-dimensional omics matrix into a product of lower-rank matrices, revealing the latent structure driving variation.

• Multi-Omics Factor Analysis (MOFA) is a seminal unsupervised framework. It infers a low-dimensional set of latent factors that capture the principal sources of variation across all data modalities from the same samples. Each factor explains a percentage of variance in each omics layer, directly pinpointing which molecular processes are coordinated.
• Joint Non-negative Matrix Factorization (Joint NMF) projects multiple data types onto a common basis of metagenes or metaproteins, enforcing non-negativity for interpretability.
• iCluster+ integrates diverse data types (continuous, binary, categorical) under a joint latent variable model with sparsity constraints, enabling simultaneous clustering and feature selection.

A distinct paradigm that integrates omics data not through statistical correlation, but by mapping it onto a mechanistic, genome-scale reconstruction of an organism's biochemistry.

• Genome-Scale Metabolic Models (GEMs) are mathematical representations of all known metabolic reactions in a cell, encoded with gene-protein-reaction (GPR) rules.
• Flux Balance Analysis (FBA) uses these GEMs to simulate steady-state metabolic flux under defined constraints. Transcriptomic or proteomic data are integrated to context-specificize the model by constraining reaction bounds (e.g., turning off reactions for unexpressed genes).
• This approach provides a mechanistic, rather than purely correlative, explanation for phenotypes like growth rate or drug resistance, directly linking genotype to metabolic phenotype through a biophysically constrained framework.