Using knowledge graphs in drug discovery (Part 1): how they link to large language models

Knowledge graphs, long established in data management and symbolic reasoning, are experiencing a renaissance in the era of generative AI, particularly in drug discovery. These structured data representations, which predate the current AI boom, have evolved from their origins as tools for data interchange and symbolic reasoning into sophisticated platforms for knowledge management and discovery.

What are knowledge graphs?

At their core, knowledge graphs serve as interconnected networks of information where relationships between entities are explicitly defined and meaningful. In drug discovery, these entities might include compounds, proteins, genes, diseases and clinical outcomes – all connected through verified scientific relationships. This structured approach to data organisation offers two key advantages:

• Systematic reasoning – enables understanding of complex relationships between data points
• Standardised framework – allows seamless data interchange across different systems and organisations.

The fundamental strength of knowledge graphs lies in their ability to represent relationships between data points in a way that mirrors natural language structure. As Andreas Kollegger explains, “You can read the graph… almost read a sentence along it.” This natural alignment between graph structures and language has become particularly valuable with the rise of large language models (LLMs), creating what Kollegger describes as “a very nice property that maps nicely to natural language.”

Moving beyond traditional NLP

The integration of knowledge graphs with large language models represents a significant evolution from their traditional use with natural language processing (NLP). In the earlier NLP era, knowledge graphs primarily served as structured repositories that NLP tools could query and update. The current integration with LLMs is fundamentally different and more powerful. While NLP tools previously handled the basic processing of text and relationships,  now LLMs serve as sophisticated processors with enhanced reasoning capabilities, able to:

• understand context
• identify connections between complex data points
• generate insights in ways that resemble human cognitive processes.

This evolution has enabled a more nuanced approach to handling unstructured data – from research papers and documents to images and other scientific content. When working with unstructured data, LLMs can now perform what Kollegger terms “distilling the essential facts.” For example, when analysing scientific literature, the system can identify key entities and their relationships, effectively creating a structured knowledge representation from unstructured text.

One of the key innovations in this space is the way LLMs interact with knowledge graphs. When presented with a question, an LLM can break down complex queries into component parts, identifying what information it needs to discover. The system then uses the knowledge graph to find relevant data points and, crucially, the relationships between them. This process creates what Kollegger describes as a path between points of interest, essentially building a specialised mini knowledge graph specific to the query at hand. This capability is particularly powerful because it allows the system to answer questions that might require information from multiple sources or complex chains of reasoning.

Why it matters for drug discovery

The combination of knowledge graphs and LLMs has proven particularly valuable in drug discovery, where relationships between compounds, proteins, pathways and clinical outcomes are complex and multifaceted. The knowledge graph provides a curated, validated foundation of scientific relationships, while the LLM offers the flexibility to explore and reason about these relationships in novel ways. This hybrid approach addresses one of the fundamental challenges in drug discovery: the need to both maintain scientific rigor and enable creative exploration of new possibilities.

Furthermore, this integration helps address a critical concern in scientific research – the verification and validation of results. While LLMs can generate insights rapidly, the knowledge graph serves as a source of validated, structured information that researchers can audit and verify. As Kollegger notes, this creates a form of “pre-computed” knowledge, where the hard work of validating scientific relationships has already been done through research and academic studies.