[Paper Review] From Large Language Models to Knowledge Graphs for Biomarker Discovery in Cancer

Md. Rezaul Karim, Lina Molinas Comet+4 more|arXiv (Cornell University)|Oct 12, 2023

Biomedical Text Mining and OntologiesBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper proposes a knowledge graph (KG) enhanced with a domain-specific ontology (OncoNet Ontology) and fine-tuned using large language models (LLMs) to improve biomarker discovery in cancer. By integrating BioBERT- and SciBERT-based information extractors with LLM-fine-tuned updates, the system enables interactive question answering and symbolic reasoning, significantly improving entity recognition and reducing concept drift in dynamic biomedical knowledge.

ABSTRACT

Domain experts often rely on most recent knowledge for apprehending and disseminating specific biological processes that help them design strategies for developing prevention and therapeutic decision-making in various disease scenarios. A challenging scenarios for artificial intelligence (AI) is using biomedical data (e.g., texts, imaging, omics, and clinical) to provide diagnosis and treatment recommendations for cancerous conditions.~Data and knowledge about biomedical entities like cancer, drugs, genes, proteins, and their mechanism is spread across structured (knowledge bases (KBs)) and unstructured (e.g., scientific articles) sources. A large-scale knowledge graph (KG) can be constructed by integrating and extracting facts about semantically interrelated entities and relations. Such a KG not only allows exploration and question answering (QA) but also enables domain experts to deduce new knowledge. However, exploring and querying large-scale KGs is tedious for non-domain users due to their lack of understanding of the data assets and semantic technologies. In this paper, we develop a domain KG to leverage cancer-specific biomarker discovery and interactive QA. For this, we constructed a domain ontology called OncoNet Ontology (ONO), which enables semantic reasoning for validating gene-disease (different types of cancer) relations. The KG is further enriched by harmonizing the ONO, metadata, controlled vocabularies, and biomedical concepts from scientific articles by employing BioBERT- and SciBERT-based information extractors. Further, since the biomedical domain is evolving, where new findings often replace old ones, without having access to up-to-date scientific findings, there is a high chance an AI system exhibits concept drift while providing diagnosis and treatment. Therefore, we fine-tune the KG using large language models (LLMs) based on more recent articles and KBs.

Motivation & Objective

• Address the challenge of concept drift in biomedical AI systems due to rapidly evolving cancer research findings.
• Integrate heterogeneous biomedical data—textual, omics, clinical—into a unified, semantically rich knowledge graph (KG) for cancer research.
• Enable domain experts to perform interactive question answering (QA) and validate biomarker-disease relationships using symbolic reasoning.
• Develop a domain-specific ontology (OncoNet Ontology) to support semantic validation of gene-disease relations in cancer.
• Mitigate outdated knowledge in KGs by continuously updating with facts extracted from recent scientific literature using large language models (LLMs).

Proposed method

• Constructed a domain-specific knowledge graph (KG) by integrating biomedical entities (genes, proteins, drugs, diseases) from scientific literature and structured knowledge bases.
• Designed the OncoNet Ontology (ONO) to model semantic relationships between cancer-related entities, enabling symbolic reasoning and validation of gene-disease associations.
• Employed BioBERT and SciBERT models for named entity recognition (NER), relation extraction (RE), and entity linking (EL) on unstructured biomedical texts.
• Fine-tuned the KG using LLMs (e.g., LLaMA-based models) to extract and incorporate new facts from recent scientific articles and knowledge bases, reducing concept drift.
• Enabled interactive question answering (QA) via natural language queries (NLQs) mapped to logical queries (DLQs), with reasoning over the KG using inference rules.
• Integrated neuro-symbolic AI principles by combining connectionist (LLM-based) and symbolic (ontology-based) AI to improve explainability and knowledge validation.

Experimental results

Research questions

• RQ1 How can a knowledge graph be constructed to support interactive biomarker discovery and question answering in cancer research?
• RQ2 To what extent can domain-specific language models (BioBERT, SciBERT) improve the accuracy of entity recognition and linking in biomedical texts compared to general-domain models?
• RQ3 Can large language models be effectively used to update and enrich a knowledge graph with up-to-date facts, thereby mitigating concept drift in dynamic biomedical domains?
• RQ4 How does the integration of symbolic reasoning via a domain ontology enhance the reliability and interpretability of biomarker-disease relationship predictions?
• RQ5 What is the impact of combining neuro-symbolic AI approaches on explainability and validation of AI-driven diagnostic and therapeutic recommendations in oncology?

Key findings

• BioBERT achieved the highest F1-score (91.43%) in disease normalization and 91.36% in gene/protein recognition, outperforming SciBERT by 2.08% and BiLSTM-CRF by 7.85%.
• The BioBERT model improved F1-score by 3.25% over BERT on average, demonstrating the advantage of domain-specific pretraining on biomedical texts.
• LLMs were successfully used to extract and integrate new facts from recent scientific literature, enabling dynamic updates to the knowledge graph and reducing concept drift.
• The KG enabled interactive question answering for complex queries such as 'Which POTSF biomarker is highly responsible for breast carcinoma and has PubMed evidence?'
• Symbolic reasoning over the KG allowed domain experts to validate AI predictions by tracing logical inference paths, enhancing trust and interpretability.
• The integration of LLMs with the KG significantly improved the system’s ability to keep pace with evolving biomedical knowledge, ensuring up-to-date biomarker discovery.