Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 55 of the 55 references it cites were found in a registry or an index; 8 of the 10 sources its claims cite were checked for support.

Biomedical knowledge graph-optimized prompt generation for large language models

Open access

Abstract

MOTIVATION: Large language models (LLMs) are being adopted at an unprecedented rate, yet still face challenges in knowledge-intensive domains such as biomedicine. Solutions such as pretraining and domain-specific fine-tuning add substantial computational overhead, requiring further domain-expertise. Here, we introduce a token-optimized and robust Knowledge Graph-based Retrieval Augmented Generation (KG-RAG) framework by leveraging a massive biomedical KG (SPOKE) with LLMs such as Llama-2-13b, GPT-3.5-Turbo, and GPT-4, to generate meaningful biomedical text rooted in established knowledge. RESULTS: Compared to the existing RAG technique for Knowledge Graphs, the proposed method utilizes minimal graph schema for context extraction and uses embedding methods for context pruning. This optimization in context extraction results in more than 50% reduction in token consumption without compromising the accuracy, making a cost-effective and robust RAG implementation on proprietary LLMs. KG-RAG consistently enhanced the performance of LLMs across diverse biomedical prompts by generating responses rooted in established knowledge, accompanied by accurate provenance and statistical evidence (if available) to substantiate the claims.

Overview

CoChat Summary

## TL;DR

This paper introduces KG-RAG, a token-optimized retrieval-augmented generation framework that uses the SPOKE biomedical knowledge graph to ground LLM outputs in established biomedical knowledge. Across benchmark tasks, it reduced token use by more than 50% versus prior KG-RAG approaches and improved Llama-2 performance on a challenging MCQ dataset by 71%.

## Key findings

  • On a 100-question retrieval benchmark, KG-RAG achieved 97% retrieval accuracy, compared with 75% for Cypher-RAG and 61% for full-text index retrieval.
  • Under prompt perturbation (lowercasing entity names), Cypher-RAG retrieval accuracy dropped to 0%, while KG-RAG maintained 97% and full-text index dropped slightly to 58% from 61%.
  • KG-RAG used an average of 3,693 tokens per response, versus 8,006 for Cypher-RAG and 10,590 for full-text index, corresponding to 53.9% and 65.1% token reductions, respectively.
  • On the MCQ dataset, KG-RAG increased Llama-2 accuracy from 0.31 ± 0.03 to 0.53 ± 0.03, described as approximately a 71% boost; on the same KG-RAG setup, GPT-3.5-Turbo outperformed GPT-4 on MCQ (0.79 ± 0.02 vs 0.74 ± 0.03, P < 0.0001).

## Methods

  • Framework design: disease entity extraction from user prompts, disease-node matching in SPOKE, neighborhood/context retrieval from the graph, context pruning with sentence-transformer embeddings and cosine-similarity thresholds, then prompt assembly and LLM generation.
  • Data/sample: SPOKE, a biomedical knowledge graph with 42 million nodes, 160 million edges, and 41 integrated biomedical databases; benchmark datasets included 311 True/False questions, 306 MCQs, and 100 retrieval-comparison questions, plus 165 validation questions for hyperparameter tuning.
  • Analysis: compared KG-RAG with prompt-only, Cypher-RAG, and full-text index approaches using accuracy, Jaccard similarity, robustness to prompt perturbation, and token usage; LLMs tested were Llama-2-13b, GPT-3.5-Turbo, and GPT-4 with temperature set to 0.

## Limitations

  • The paper states the approach is currently limited to biomedical questions centered on diseases, though it includes a proof of concept extending to more than 87,000 non-disease nodes; this is an explicit limitation.
  • The paper says it was developed on SPOKE only and future work will extend it to other biomedical KGs and domains; this is an explicit limitation.
  • The paper notes that context quality depends on the underlying graph and that SPOKE is “not entirely error free” or ready for clinical use; this is an explicit limitation.
  • Evident from the text: the evaluation focuses on curated benchmark questions and retrieval tasks rather than a formal clinical validation of the KG-RAG system in real-world biomedical use.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1093/bioinformatics/btae560

Cites

55 of 55 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1. Schema for the Knowledge Graph based Retrieval-Augmented Generation (KG-RAG) Framework. The direction of the arrows indicates the flow of the pipeline in this framework.
Figure 1. Schema for the Knowledge Graph based Retrieval-Augmented Generation (KG-RAG) Framework. The direction of the arrows indicates the flow of the pipeline in this framework.CC BY · PubMed Central
Figure 2. Detailed schema of KG-RAG. Dashed boxes show the details of “Disease entity recognition” and “Prompt-aware context extraction” from the Knowledge Graph.
Figure 2. Detailed schema of KG-RAG. Dashed boxes show the details of “Disease entity recognition” and “Prompt-aware context extraction” from the Knowledge Graph.CC BY · PubMed Central
Figure 3. Prompting examples. Example prompts (top boxes) and prompt-based (middle boxes)/KG-RAG-based (bottom boxes) responses using GPT-4. (A) A biomedical prompt to search the drugs used for weight management in patients with “Bardet-Biedl syndrome” and (B) Another prompt to compare the statistical associations of PNPLA3 and HLA-B genes with the disease liver benign neoplasm. Both (A) and (B) show the associated “prompt aware graph” context.
Figure 3. Prompting examples. Example prompts (top boxes) and prompt-based (middle boxes)/KG-RAG-based (bottom boxes) responses using GPT-4. (A) A biomedical prompt to search the drugs used for weight management in patients with “Bardet-Biedl syndrome” and (B) Another prompt to compare the statistical associations of PNPLA3 and HLA-B genes with the disease liver benign neoplasm. Both (A) and (B) show the associated “prompt aware graph” context.CC BY · PubMed Central
Figure 4. Hyperparameter analysis and RAG comparison. (A) Hyperparameter analysis performance curves using prompts with single (top) and two (below) disease entities mentioned in it. The x-axis denotes the “Context volume” (number of associations from KG) and the y-axis denotes the mean performance (Jaccard similarity) across the prompts. The red curve denotes “PubMedBert” and the blue curve denotes “MiniLM” transformer models. (B) The comparative analysis between KG-RAG, Cypher-RAG, and Full-Text Index using Apache Lucene in terms of retrieval accuracy (top) and token usage (bottom). Insight shows an example where Cypher-RAG fails to retrieve context from the KG when the input prompt is slightly perturbed, but KG-RAG and Full-Text Index remain robust in context retrieval. However, only KG-RAG could extract provenance information linked to the assertion from the context, whereas Full-Text Index could not. It is evident that KG-RAG has significantly lesser token usage when compared to Cypher-RAG and Full-Text Index method (bottom). Error bar in the token utilization bar plot (bottom) represents standard error of the mean (SEM).
Figure 4. Hyperparameter analysis and RAG comparison. (A) Hyperparameter analysis performance curves using prompts with single (top) and two (below) disease entities mentioned in it. The x-axis denotes the “Context volume” (number of associations from KG) and the y-axis denotes the mean performance (Jaccard similarity) across the prompts. The red curve denotes “PubMedBert” and the blue curve denotes “MiniLM” transformer models. (B) The comparative analysis between KG-RAG, Cypher-RAG, and Full-Text Index using Apache Lucene in terms of retrieval accuracy (top) and token usage (bottom). Insight shows an example where Cypher-RAG fails to retrieve context from the KG when the input prompt is slightly perturbed, but KG-RAG and Full-Text Index remain robust in context retrieval. However, only KG-RAG could extract provenance information linked to the assertion from the context, whereas Full-Text Index could not. It is evident that KG-RAG has significantly lesser token usage when compared to Cypher-RAG and Full-Text Index method (bottom). Error bar in the token utilization bar plot (bottom) represents standard error of the mean (SEM).CC BY · PubMed Central
Figure 5. LLM performance on True/False and MCQ datasets. Performance (Accuracy) distributions of LLMs on (A) True/False and (B) MCQ datasets. Figure panel shows the distributions corresponding to both prompt-based and KG-RAG-based approaches. Vertical-dashed line indicates the mean value of the distribution. The higher the value, the better the performance.
Figure 5. LLM performance on True/False and MCQ datasets. Performance (Accuracy) distributions of LLMs on (A) True/False and (B) MCQ datasets. Figure panel shows the distributions corresponding to both prompt-based and KG-RAG-based approaches. Vertical-dashed line indicates the mean value of the distribution. The higher the value, the better the performance.CC BY · PubMed Central

Verification

Claims the paper makes

4 Valid · 2 Not valid · 0 Not checked · 50 Own results and interpretations

25 more claims

Create a free account to see all 56 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 7
  • It may state two separate claims as one. 21

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 35 references

Cited by

See all 135

Metadata

Authors
Soman, K., Rose, P. W., Morris, J. H., Akbas, R. E., Smith, B., Peetoom, B., Villouta-Reyes, C., Cerono, G., Shi, Y., Rizk-Jackson, A., Israni, S., Nelson, C. A., Huang, S., & Baranzini, S. E.
Affiliations
  • University of California, San Francisco (US)
  • San Diego Supercomputer Center (US)
  • Institute for Systems Biology (US)
  • Bakar Computational Health Sciences Institute, University of California , San Francisco, CA 94158,
  • Mate Bioservices, Inc. Swallowtail Ct. , Brisbane, CA 94005,
Corresponding author
Baranzini, Sergio E
Venue
Bioinformatics
Published
2024-09
Volume
40
Issue
9
ISSN
  • 1367-4803 (ISSN-L)
  • 1367-4811
OpenAlex
W4402567773
Cited by
135
References
35
Citations by year
2026: 61 · 2025: 63 · 2024: 9
FWCI
10.72
Type
journal article
Language
en
Open access
Locations
  • Bioinformatics (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • eScholarship (California Digital Library) (submittedVersion, cc-by)
  • PubMed Central (submittedVersion, cc-by)
Funders
National Science Foundation (NSF_2033569)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Advanced Graph Neural Networks
Keywords
knowledge graph retrieval-augmented generation, prompt generation, large language models, biomedical knowledge graphs, context extraction, Llama 2, GPT-4, GPT-3.5-turbo
MeSH terms
Algorithms, Humans, Natural Language Processing, Computational Biology

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Biomedical knowledge graph-optimized prompt generation for large language models. https://app.cochat.ai/papers/10.1093/bioinformatics/btae560
The paper
Soman, K., Rose, P. W., Morris, J. H., Akbas, R. E., Smith, B., Peetoom, B., Villouta-Reyes, C., Cerono, G., Shi, Y., Rizk-Jackson, A., Israni, S., Nelson, C. A., Huang, S., & Baranzini, S. E. (2024). Biomedical knowledge graph-optimized prompt generation for large language models. Bioinformatics. https://doi.org/10.1093/bioinformatics/btae560

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.