Biomedical knowledge graph gains evidence and study conditions: SPIDER labels 79.0% of apparent conflicts as context differences
Synopsis
Biomedical knowledge graphs often collapse opposing relations between the same entity pair into one record, leaving readers unable to separate genuine scientific disagreement from differing study conditions. SPIDER uses a multi-agent workflow to ground graph relations in PubMed evidence and extract study conditions; starting from 48 integrated databases and ontologies and 1.95 million abstracts, it added 2.31 million literature-supported entity pairs. Across 89,937 evidence-supported comparisons of potentially incompatible relations, 79.0% were classified as context-dependent differences and 6.1% as genuine disagreement, with a macro-F1 of 80.3% against expert annotations.
Interpretation
Relations in a biomedical knowledge graph can now carry their supporting literature and study conditions individually, rather than only an aggregated relation label. Earlier graphs, even when linked to publications or curated records, did not necessarily associate each individual finding with its supporting evidence and specific study conditions, so opposing conclusions for the same entity pair across tissues, models or interventions were hard to distinguish. On a graph integrating 48 databases and ontologies, literature grounding raised the share of entities with at least one supporting publication from 14.6% to 65.0% and the share of entity pairs from 34.0% to 79.4%.
Most apparently contradictory relations between the same entity pair come from differing study conditions rather than truly opposing conclusions. Potentially incompatible relation types only flag candidates and do not establish scientific disagreement; judging that requires the supporting text and conditions. Among 89,937 comparison units with literature evidence for both relations, 79.0% were classified as context-dependent differences, 6.1% as genuine disagreement, 2.9% as relation-mapping errors and 12.1% as unresolved; against annotations by eight medical professionals on 618 sampled units, four-class accuracy was 83.3% and macro-F1 over the context-difference and genuine-disagreement categories was 80.29%.
The extracted structured study conditions alone improve a large language model's answers to conflict-oriented questions. With triples alone, the answering model showed no statistically significant gain on SPIDER-Bench. On SPIDER-Bench, triples plus structured conditions raised accuracy from 50.5% to 77.0% and macro-F1 from 50.3% to 77.0%, approaching the configuration that also supplied literature evidence.
The evidence-linked knowledge layer improved the same answering model on several biomedical reasoning tasks, with gains that varied by task. The comparison was the same model answering directly, or supported by five other biomedical knowledge graphs. BioMixQA accuracy rose from 87.46% to 95.50%, CIViC from 54.27% to 68.44%, Open Targets GWAS mean reciprocal rank from 0.8195 to 0.8630, and DICTrank macro-F1 from 0.4664 to 0.4942 in the three-class setting and from 0.3476 to 0.3718 in the four-class setting.
Perspective
The demonstrated scope starts from a research-oriented subgraph built from 48 integrated databases and ontologies, grounds relations using PubMed titles and abstracts, and performs condition-aware classification on 89,937 comparison units with literature evidence for both relations; SPIDER-Bench and the four downstream benchmarks were compared under the same answering model.
The classification proportions should be read alongside the expert-validation results and within the screened, evidence-supported comparison set; expert evaluation found a tendency to classify relation-mapping errors as context-dependent differences. The current framework mainly uses PubMed titles and abstracts and is limited by a predefined entity set, incompatibility patterns and condition templates, so details available only in full text or supplements may be missed, and incomplete conditions can leave comparability uncertain. The GLTP–ichthyosis case provides biological context only; it does not establish functional coupling between GLTP and ABCA12 or explain how GLTP variants cause disease.
