Heterogeneous Biological Graph

The defining characteristic of a heterogeneous biological graph is its ability to represent diverse biological entities as distinct node types within a single unified structure.

• Gene nodes: Represent coding and non-coding genomic loci with attributes like chromosomal position and strand orientation
• Protein nodes: Encode amino acid sequences, domains, and post-translational modification sites
• Metabolite nodes: Capture small-molecule structures, concentrations, and pathway membership
• Disease nodes: Represent phenotypic states with ontological annotations from resources like MONDO or DOID
• Drug nodes: Encode compound structures, targets, and mechanism-of-action classifications

Each node type carries its own feature vector dimensionality, requiring type-specific encoders before graph convolution.

Edges in these graphs are typed and directed, each encoding a specific biological relationship with distinct semantic meaning. Unlike homogeneous graphs where all connections are equivalent, heterogeneous edges require relation-specific message passing.

• Physical interaction edges: Experimentally validated protein-protein binding from databases like STRING or BioGRID
• Co-expression edges: Correlation of transcript abundance across conditions, weighted by Pearson or Spearman coefficients
• Regulatory edges: Transcription factor to target gene relationships derived from ChIP-seq peaks or motif analysis
• Pathway membership edges: Links between proteins and the Reactome or KEGG pathways they participate in
• Disease association edges: Connections between genes and phenotypes curated from GWAS catalogs or ClinVar

Each edge type defines a unique metapath—a composite semantic relationship traversing multiple node and edge types.

Metapaths are predefined sequences of node and edge types that capture higher-order semantic relationships in heterogeneous graphs. They serve as structured context windows for graph neural network aggregation.

• A metapath like Gene–(encodes)→Protein–(interacts)→Protein–(participates_in)→Pathway captures functional module membership
• Another like Drug–(targets)→Protein–(associated_with)→Disease encodes therapeutic indication logic
• Graph attention mechanisms learn to weight different metapaths dynamically based on the downstream prediction task
• Metapath-based random walks generate training sequences for heterogeneous skip-gram embedding methods
• The selection of metapaths injects domain knowledge priors into otherwise purely data-driven graph learning

Standard graph convolutions assume all neighbors contribute equally. Relational Graph Convolutional Networks (R-GCNs) extend this by learning separate weight matrices for each edge type, enabling type-specific message transformation.

• For each relation r, a distinct transformation matrix W_r projects neighbor features before aggregation
• The update rule: h_i = σ( Σ_r Σ_j∈N_r(i) W_r · h_j / c_i,r ) where N_r(i) are neighbors connected via relation r
• Basis decomposition reduces parameter explosion by expressing each W_r as a linear combination of shared basis matrices
• Block diagonal decomposition further regularizes by enforcing sparsity patterns in the relation-specific weights
• This architecture directly enables the model to learn that protein-protein binding and co-expression convey fundamentally different biological signals

Heterogeneous biological graphs are often anchored to curated knowledge bases that provide the ground-truth edge structure, ensuring the graph reflects experimentally validated biology rather than purely inferred correlations.

• Reactome and KEGG provide pathway topology edges with directed, stepwise reaction semantics
• Gene Ontology supplies hierarchical term annotations that can be encoded as is_a and part_of edges
• STRING contributes confidence-scored protein-protein interaction edges with experimental and text-mining provenance
• DisGeNET and Open Targets link genes to diseases with literature-supported association scores
• DrugBank and ChEMBL provide drug-target binding edges with affinity measurements

This anchoring transforms the graph from a statistical construct into a mechanistic model of cellular biochemistry, enabling knowledge-guided fusion that respects known biology.

Not all interactions in a biological graph are equally relevant for every prediction task. Heterogeneous graph attention mechanisms learn to dynamically weight edges based on the query context.

• HAN (Heterogeneous Graph Attention Network) applies node-level attention within each metapath and semantic-level attention across metapaths
• HGT (Heterogeneous Graph Transformer) computes attention scores that are conditioned on both source and target node types, learning type-specific key, query, and value projections
• Edge weights can incorporate experimental confidence scores from databases like STRING as attention priors
• Conditional attention modulates edge importance based on a context vector—for example, a disease state embedding that up-weights known disease-associated interactions
• This dynamic reweighting allows the same graph structure to serve multiple downstream tasks (variant effect prediction, drug repurposing, patient stratification) without architectural changes

The application of graph convolutions to integrate multi-omic data structured as heterogeneous biological graphs. Nodes represent genes, proteins, or metabolites, while edges encode known interactions such as physical binding or co-expression. Message-passing layers propagate signals across node types, enabling the model to learn latent representations that capture both local neighborhood topology and global network structure for downstream tasks like phenotype prediction.

An integration approach that constrains multi-omic model architecture or training using prior biological databases. By incorporating curated knowledge from resources like Reactome, Gene Ontology, or STRING, the model's connectivity is restricted to mechanistically plausible interactions. This reduces overfitting to spurious correlations and ensures that learned representations align with established biochemical pathways, improving both interpretability and generalization to unseen biological contexts.

The computational inference of causal regulatory interactions between transcription factors and target genes by integrating multi-omic data. Heterogeneous graphs combine chromatin accessibility (ATAC-seq), transcription factor binding (ChIP-seq), and gene expression (RNA-seq) to distinguish direct regulation from indirect correlation. Graph-based models can orient edges directionally, reconstructing the hierarchical control logic that governs cellular differentiation and disease states.

A feature representation that explicitly encodes the activity levels of predefined biological signaling cascades by aggregating multi-omic signals at the pathway level rather than the individual gene level. In a heterogeneous graph, pathway nodes serve as hierarchical aggregators that pool information from member genes and proteins. This produces compact, biologically meaningful embeddings that are robust to measurement noise and directly interpretable in terms of known disease mechanisms.

The computational process of mapping feature vectors from different biological assays into a common coordinate system. In heterogeneous graphs, node-type-specific encoders transform each modality into a shared latent space where semantically similar biological states occupy proximal positions. This alignment enables operations like cross-modal retrieval—querying with a gene expression pattern to find structurally similar protein interaction subgraphs—and supports missing modality imputation through graph completion.

A dynamic, patient-specific computational model that integrates longitudinal multi-omic data into a heterogeneous graph representation. Nodes capture an individual's molecular state across time points, while edges encode both intra-modal dynamics and cross-modal regulatory relationships. This graph-based digital twin simulates molecular physiology to predict individual responses to therapeutic interventions, enabling virtual drug screening and personalized treatment optimization.