Knowledge Graph Embedding

These models interpret relations as geometric translations in the embedding space. The core hypothesis is that for a valid triple (head, relation, tail), the embedding of the tail should be close to the embedding of the head plus the relation vector.

• TransE: The foundational model where h + r ≈ t. Highly efficient but struggles with complex 1-to-N relations.
• RotatE: Models relations as rotations in complex space, capturing symmetry and inversion patterns critical for drug-target interactions.
• PairRE: Encodes each relation as a pair of vectors, allowing independent modulation of head and tail entity dimensions for improved expressiveness.

These models capture latent semantics through tensor factorization, measuring the compatibility of entity-relation-entity triples via bilinear scoring functions rather than distance minimization.

• RESCAL: A bilinear model that captures pairwise interactions between all latent features via a full rank relation matrix. Computationally intensive but highly expressive.
• DistMult: A simplified bilinear model restricting relation matrices to diagonals, reducing parameters but unable to model asymmetric relations.
• ComplEx: Extends DistMult into complex-valued space, using Hermitian dot products to elegantly handle asymmetry and antisymmetry in biomedical relations.

Modern embedding systems use GNNs as encoders to generate entity representations that incorporate multi-hop neighborhood structure before applying a decoder for link prediction.

• R-GCN: A relational graph convolutional network that aggregates messages from neighboring nodes with relation-specific weight matrices, capturing the heterogeneous nature of biomedical knowledge graphs.
• CompGCN: A composition-based GCN that jointly embeds entities and relations, applying relation-specific transformations during message passing to reduce parameter explosion.
• NodePiece: Tokenizes the graph vocabulary into a fixed set of anchor nodes, enabling compositional entity representations that generalize to unseen drugs and diseases in zero-shot repurposing scenarios.

The choice of loss function and negative sampling strategy directly determines the quality of learned embeddings and their ability to rank true drug-disease associations above false ones.

• Margin-Based Ranking Loss: Enforces a minimum margin between the scores of positive triples and their corrupted negative counterparts, providing a hard constraint on embedding geometry.
• Binary Cross-Entropy: Treats link prediction as a classification problem, applying a sigmoid to the scoring function and optimizing against a 1/0 label for known associations.
• Adversarial Sampling: Dynamically generates harder negative examples during training by selecting corruptions that the current model scores highly, preventing the model from learning trivial distinctions.
• Self-Adversarial Sampling: Weights negative samples by their current model score, allowing the model to focus on more informative negatives without explicit candidate generation.

Biomedical knowledge graphs are inherently heterogeneous, containing multiple entity types (drug, disease, gene, pathway) and relation types (treats, binds, inhibits). Embedding methods must preserve this rich relational structure.

• Entity Type Awareness: Models incorporate entity type embeddings or use type-specific projection matrices to ensure drugs and diseases occupy distinct semantic subspaces.
• Relation Path Reasoning: Multi-hop paths like Drug → binds → Target → associated_with → Disease are encoded as compositional relation sequences, enabling indirect link inference.
• Literal-Enriched Embeddings: Augment structural graph triples with numerical attributes such as molecular weight or binding affinity values, grounding symbolic embeddings in quantitative biochemical reality.

Rigorous evaluation protocols prevent over-optimistic performance estimates that could mislead downstream drug repurposing efforts. Proper splitting strategies are essential.

• Time-Based Splitting: Splits the knowledge graph by publication date, training on historical associations and testing on newly discovered links to simulate real-world prospective prediction.
• Entity-Level Holdout: Removes entire drugs or diseases from the training set to test true inductive capability, preventing data leakage from overlapping molecular scaffolds.
• Mean Reciprocal Rank (MRR): The primary ranking metric that measures the average reciprocal rank of the correct tail entity across all candidate entities, penalizing low-ranked true associations.
• Hits@K: Reports the fraction of test triples where the correct entity appears in the top K ranked predictions, providing an intuitive measure of practical retrieval utility.

The primary machine learning task for knowledge graph embedding. It estimates the likelihood of a missing connection between two nodes in a heterogeneous biomedical network.

• Objective: Predict novel drug-disease edges that do not exist in the current graph.
• Scoring Function: A mathematical function f(h, r, t) that assigns a plausibility score to a triple (head entity, relation, tail entity).
• Example: Predicting that Metformin treats a specific rare cancer based on its proximity to related genes and pathways in the graph.

These models measure the latent similarity of entities and relations using tensor factorization, rather than geometric distance.

• RESCAL: A bilinear model that captures pairwise interactions via a full rank tensor, but is parameter-intensive.
• DistMult: A simplified bilinear model restricted to diagonal matrices, making it computationally light but unable to model anti-symmetric relations.
• ComplEx: Extends DistMult into complex-valued space, effectively handling asymmetry while maintaining linear space complexity.

Modern approaches that use message-passing neural networks to generate embeddings by aggregating information from a node's local neighborhood.

• R-GCN: A relational graph convolutional network that applies distinct weight matrices for different relation types during neighbor aggregation.
• CompGCN: A composition-based GCN that jointly embeds entities and relations, reducing parameter explosion in highly multi-relational biomedical graphs.
• Advantage: These models naturally capture the inductive bias of local graph structure, often outperforming shallow encoders on complex biological interactions.

A critical training technique that corrupts true triples to create negative examples, teaching the model to distinguish valid facts from false ones.

• Uniform Sampling: Randomly replaces the head or tail entity, but often generates trivially false negatives that don't help learning.
• Bernoulli Sampling: Replaces entities based on the relation's mapping property (1-to-N vs. N-to-1) to reduce false negatives.
• Adversarial Sampling: Dynamically selects harder negative samples that the current model is most likely to misclassify, improving embedding quality.

Rigorous validation is essential to avoid over-optimistic results that fail in translational medicine.

• Mean Reciprocal Rank (MRR): The average inverse rank of the first correct entity across all test triples.
• Hits@K: The fraction of test triples where the correct entity appears in the top-K ranked predictions.
• Data Leakage Risk: A critical error where inverse relations (e.g., treats and treated_by) cause test set information to leak into training. Always use a filtered setting that removes all known valid triples from the ranking before evaluation.