Link Prediction

Traditional methods rely on local proximity heuristics like Common Neighbors, Jaccard Coefficient, or Adamic-Adar to score potential links based on shared neighbors. Modern approaches use learned scoring functions—graph neural networks that encode both topological structure and node features into dense embeddings. A GNN-based decoder can capture non-linear, multi-hop dependencies that heuristic methods miss entirely, dramatically improving recall on sparse biomedical graphs where true associations are rare.

Biomedical knowledge graphs contain diverse node types (drugs, proteins, diseases, pathways) and edge types (binds, treats, inhibits, expresses). Heterogeneous graph neural networks like RGCN or HAN apply type-specific transformation matrices during message passing, preserving the distinct semantics of each relation. This prevents information conflation—a drug-disease 'treats' edge carries fundamentally different meaning than a drug-protein 'inhibits' edge, and the model must learn separate projection spaces for each.

Link prediction is framed as a binary classification problem requiring both positive and negative edges for training. Since biomedical graphs are sparse, negative samples—pairs of nodes with no known connection—must be generated artificially. The sampling strategy critically impacts model calibration. Uniform random sampling is fast but produces easy negatives; adversarial sampling or hard negative mining selects drug-disease pairs that are structurally similar to true positives, forcing the model to learn more discriminative boundaries.

In the transductive setting, all nodes appear during training and the model predicts missing edges among known entities—suitable for repurposing existing drugs to known diseases. The inductive setting requires the model to predict links for entirely new nodes (e.g., a novel compound) using only their initial features, such as molecular fingerprints or protein sequences. Inductive link prediction demands encoders that generalize to unseen graph structures, often leveraging GraphSAGE or attention-based aggregators that don't depend on fixed node IDs.

After encoding node embeddings, a decoder computes the likelihood score for a candidate link. Common choices include:

• DistMult: Simple bilinear product, fast but limited to symmetric relations
• ComplEx: Extends DistMult to complex space, capturing asymmetry
• RotatE: Models relations as rotations in complex space, excellent for hierarchical biomedical ontologies
• ConvE: Uses 2D convolutions over reshaped embeddings for richer feature interactions The decoder choice directly impacts the model's ability to capture the directionality of drug-disease 'treats' relationships.

Standard random edge splitting can produce over-optimistic results due to structural redundancy in biomedical graphs. If a drug's training and test edges share the same protein target neighborhood, the model memorizes local topology rather than learning generalizable patterns. Rigorous evaluation requires time-stratified splits (predicting future drug-disease associations based on historical data) or drug-wise/disease-wise splits where entire entities are held out, testing true generalization to unseen biological contexts.

A critical validation error that produces over-optimistic performance estimates and invalidates a repurposing model's real-world utility. In biomedical graphs, leakage occurs when edges from the test set share nodes or structural dependencies with the training set in ways that are not accounted for during splitting. For example, randomly splitting edges rather than using a time-aware split can allow the model to memorize a drug's entire interaction profile during training. Proper evaluation requires strict temporal or cross-validation strategies that simulate true prospective discovery.