Authors
Clint Cuffy, Logan Mills, Bridget T McInnes
Published in
Journal of biomedical informatics. Pages 105123. Nov 12, 2026. Epub Nov 12, 2026.
Abstract
Literature-based discovery (LBD) infers implicit relationships that no single document states but that are latent across the literature. Computational LBD has been dominated by methods that embed a biomedical knowledge graph with graph-embedding algorithms. These methods draw on the graph alone, and a concept added to the graph has no representation until the embedding is retrained. We investigate whether contextual transformer encoders address these limitations for closed discovery on the Hallmarks of Cancer (HOC) benchmark.
We study two formulations of LBD on a shared encoder backbone: a discriminative model that treats discovery as knowledge-graph completion and classifies candidate edges, and a generative model that predicts the embedding of the implicit linking concept and ranks candidates by cosine similarity. We compare both against established baselines, a shallow KGC model (CD-2) and multilayer perceptron (MLP) and similarity-based models, using a ranking protocol and non-parametric significance testing.
Both formulations replicate multiple held-out discoveries and are statistically on par with prior work, while representing any concept that has a text descriptor without retraining, which the graph-embedding baselines cannot do. A general-domain encoder (BERT) is competitive with, and on one dataset exceeds, its domain-adapted counterpart (BioBERT), and input scaling helps the generative but harms the discriminative formulation.
Contextual encoders are viable for LBD, and how the task is posed shapes discovery-replication performance.
PMID:
43020402
Bibliographic data and abstract were imported from PubMed on 13 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 0
- Comments 0