Graph Neural Network Fusion

A network data structure containing multiple node types (genes, proteins, metabolites) and multiple edge types (physical interaction, co-expression, pathway membership) that serves as the foundational input for GNN-based multi-omic fusion. Unlike homogeneous graphs, heterogeneous graphs preserve the semantic diversity of biological relationships, allowing separate learnable weight matrices for each node-edge type combination. Construction typically draws from curated databases such as STRING, Reactome, and Gene Ontology.

The fundamental computational unit that updates each node's embedding by aggregating feature information from its local neighborhood and applying a shared transformation. In the genomic context, a GCN layer propagates signals across protein-protein interaction networks so that a gene's learned representation reflects not only its own omic measurements but also the expression states of its interaction partners. The message-passing formula is: H^(l+1) = σ(D̂^(-1/2) Â D̂^(-1/2) H^(l) W^(l)).

An integration approach that constrains GNN architecture or training using prior biological databases to ensure mechanistic plausibility. Rather than learning all connections from data, the graph topology is explicitly defined by known interactions from resources like KEGG or BioGRID. This injects domain expertise into the model, reducing the search space, improving generalization on small datasets, and producing embeddings that align with established biological pathways rather than spurious correlations.

An extension of GCNs that employs self-attention mechanisms to learn differential importance weights for each neighboring node during message aggregation. In multi-omic fusion, a GAT can learn that a gene's protein interaction partner with highly correlated expression deserves more attention than a weakly expressed neighbor. This dynamic weighting is critical for identifying key regulatory hubs and produces more interpretable attention coefficients that can be visualized as edge importance scores.

A GNN variant specifically designed for multi-relational data where different edge types carry distinct semantics. R-GCNs assign separate transformation matrices to each relationship type (e.g., one matrix for 'phosphorylates' edges, another for 'co-expressed with' edges), then aggregate messages across all relation types. This architecture is particularly suited for heterogeneous biological graphs where physical binding, genetic interaction, and pathway co-membership represent fundamentally different biological signals.

The initial feature engineering step where multiple omic measurements for the same gene or protein are combined into a single input vector before graph propagation. For example, a gene node's initial features might concatenate:

• RNA-seq expression (continuous scalar)
• DNA methylation status (beta value)
• Chromatin accessibility (ATAC-seq peak signal)
• Copy number variation (discrete category) This fused vector becomes the node's h^(0) input to the first GNN layer.