Standard knowledge-graph embedding models store one learned vector per entity in a lookup table, so an entity absent at training time has no representation and the graph cannot be extended without retraining. This concept covers the alternative: computing each node embedding from the node's own biological and textual attributes, so representations generalise to unseen entities. Learners work through the full pipeline on a biomedical knowledge graph — encoding amino-acid sequences, nucleotide sequences, SMILES strings and natural-language descriptions with frozen domain-specific language models; fusing the per-modality vectors into one node vector with attention-based and relation-guided fusion modules, including how to handle nodes that lack a modality; pretraining those vectors with graph contrastive learning inside each node type using augmentations such as feature masking and edge dropping; and finally scoring triples with a relational graph convolutional encoder and a tensor-factorisation decoder trained on sampled negative edges. The payoff to reason about is transfer: a representation built from attributes rather than identity can be evaluated on a different relation set and on nodes the model never saw, which is what makes tasks like drug-target interaction prediction and drug repurposing tractable on a graph that keeps growing.
This paper addresses two coupled problems in biomedical knowledge graphs: existing graphs such as PrimeKG describe entities only by their position in the graph structure, and existing knowledge-graph…