← All projects

SemNet Explorer – knowledge graph built on PubMed

Connecting biomedical knowledge graphs and language models for evidence-grounded mechanistic reports.

SemNet Explorer: An Evidence-Grounded Knowledge Graph–LLM Framework for Multi-Scale Mechanistic Reporting Across Biomedical Domains

Big Data and Cognitive Computing · 10(6), 171

Read the paper (opens in a new tab)

Study overview

  • Connected SemNet 2.0 knowledge-graph evidence to language-model reports using bounded two-hop paths from an anchor concept through a mediator to a disease.
  • Ranked anchors using path strength, HeteSim relatedness, and mediator count, then organized them into seven shared, pairwise, and disease-specific regions for global and anchor-level reporting.

Key findings

  • The shared three-disease core and AD–ALS region together contained more than 80% of anchors in each molecular, disease, and pharmacological layer. These are graph distributions, not clinical prevalence estimates.
  • In repeated LLM-judge comparisons, global grounding improved expression more consistently than judged content; local mediator grounding improved both measures more consistently, especially for molecular and drug concepts.
  • Comparisons with earlier expert-curated SemNet studies were qualitative and sometimes only partial. LLM-judge preferences do not independently validate the proposed biology.

Figures

SemNet Explorer workflow from PubMed-scale knowledge graph evidence to global and anchor-level LLM reports
Figure 1 — Overall workflow for evidence-grounded tri-disease comparative reporting with SemNet 2.0.Source: He et al. (2026), Figure 1 · CC BY 4.0
Log-scale bar charts of anchor counts across seven shared and disease-specific regions in molecular, disease, and pharmacological layers
Figure 4 — Anchor distributions across seven disease-overlap regions. Most anchors lie in the shared AD–ALS–FTD core or the AD–ALS pairwise region.Source: He et al. (2026), Figure 4 · CC BY 4.0
Heatmaps comparing LLM-judged content accuracy and expression quality for three anchor-level evidence-grounding modes
Figure 8 — Anchor-level grounding ablation. Heatmaps show judge win rates relative to the base prompt for content accuracy and expression quality, not independent biological validation.Source: He et al. (2026), Figure 8 · CC BY 4.0
← All projects