Abstract
Large Language Models (LLMs) have demonstrated exceptional performance in various natural language processing tasks, utilizing their language generation capabilities and knowledge acquisition potential from unstructured text. However, when applied to the biomedical domain, LLMs encounter limitations, resulting in erroneous and inconsistent answers. Knowledge Graphs (KGs) have emerged as valuable resources for structured information representation and organization. Specifically, Biomedical Knowledge Graphs (BKGs) have attracted significant interest in managing large-scale and heterogeneous biomedical knowledge. This study evaluates the capabilities of ChatGPT and existing BKGs in question answering, knowledge discovery, and reasoning. Results indicate that while ChatGPT with GPT-4.0 surpasses both GPT-3.5 and BKGs in providing existing information, BKGs demonstrate superior information reliability. Additionally, ChatGPT exhibits limitations in performing novel discoveries and reasoning, particularly in establishing structured links between entities compared to BKGs. To overcome these limitations, future research should focus on integrating LLMs and BKGs to leverage their respective strengths. Such an integrated approach would optimize task performance and mitigate potential risks, thereby advancing knowledge in the biomedical field and contributing to overall well-being.
Introduction
The rapid advancement of Large Language Models (LLMs) like ChatGPT has sparked significant interest in their potential to transform biomedical research and clinical practice. However, the high-stakes nature of the medical field demands a level of accuracy and traceability that general-purpose LLMs often struggle to provide. This study evaluates the performance of ChatGPT (GPT-3.5 and GPT-4.0) against traditional Biomedical Knowledge Graphs (BKGs), specifically focusing on their ability to handle question answering, knowledge discovery, and structured reasoning.
Methodology
Researchers conducted a comparative analysis using three primary systems: GPT-3.5, GPT-4.0, and iDISK, a specialized knowledge graph for dietary supplements. The evaluation framework spanned three dimensions:
- Question Answering (Q&A): Assessing accuracy on 50 health-related questions from Yahoo! Answers.
- Knowledge Discovery: Testing the models' ability to suggest novel drug repurposing candidates for Alzheimer’s Disease (AD).
- Knowledge Reasoning: Evaluating the generation of structured "triplets" (Subject-Predicate-Object) and the validity of provided citations.
Verification was performed against "ground truth" databases including ClinicalTrials.gov, PubMed, and comprehensive BKGs like iBKH and ADInt.
Key Findings
The Retrieval vs. Discovery Paradox
While GPT-4.0 excelled in retrieving existing information—outperforming both iDISK and GPT-3.5 in Q&A accuracy—it demonstrated significant limitations in true "discovery." Analysis revealed that nearly all "novel" drug suggestions for Alzheimer's made by ChatGPT were already documented in existing knowledge graphs. This suggests that LLMs function primarily as sophisticated retrieval engines rather than engines for scientific deduction.
The Reliability Gap and the "Citation Trap"
A major concern identified was the prevalence of "hallucinated" citations. Unlike BKGs, which provide direct links to source data, ChatGPT often generates plausible but entirely fabricated academic references.
As illustrated in Figure 3, the model attempts to establish a structured link between Caryophyllus aromaticus and Alzheimer's Disease. However, the accompanying citations often lack validity, a phenomenon the authors term the "Citation Trap."
Challenges in Structured Reasoning
Knowledge Graphs are inherently designed to represent complex, multi-step relationships. ChatGPT, by contrast, struggled to maintain consistency when asked to provide structured links.
Figure 6 highlights this inconsistency. In one prompt, the model provides a detailed association for Loperamide with a specific citation, while in another, it claims no such link can be found due to its training data cutoff. This variability poses a significant risk for clinical applications requiring deterministic and traceable logic.
Conclusion and Future Directions
The study highlights that while LLMs offer unprecedented conversational access to information, they cannot yet replace the rigorous, structured truth provided by Knowledge Graphs. The authors propose a "neuro-symbolic" future: integrating LLMs as natural language interfaces for BKGs. This hybrid approach would leverage the linguistic strengths of LLMs while grounding their outputs in the verifiable facts of structured databases, ultimately advancing biomedical knowledge safely and effectively.
Summary updated