Can I not just use an LLM for that?

But hang on… surely, I can just use the power of an LLM to manage, retrieve and analyze my data I hear (some of) you cry? Well, no, not completely. And certainly not in scenarios where evidence-based decision-making is crucial, such as the Life Sciences and other domains where decisions need to be calculated in an explainable manner with provenance, and where wrong decisions can result in dire consequences.

Let’s take a closer look at some of the specific tasks we believe the creation of, and application of ontologies are still vital…

1. Ontology generation

LLMs can help uncover ‘knowledge’, yet ontologies are needed to capture that for future use. While various technologies can support the semi-automated curation of ontologies, human validation is crucial for candidate classes, terms, and relationships.

Transforming LLM output into an ontology through a lightweight model enables versatility and reuse in downstream applications. Even if AI could create a flawless ontology, it is crucial to recognize that the value of a standard lies in its consensus among humans. While AI may generate a well-structured ontology, it may well struggle with nuanced distinctions, and diverge from what others use. The human in the loop is vital.

2. Alignment to get standards

While adept at recognizing types of things, like genes, LLMs struggle with aligning varied, synonymous, representations of instances to specific identifiers (e.g., understanding that BRCA1 and FANCS are equivalents). Short of providing an ontology as part of a prompt (limitations include prompt token length), the ability to annotate textual data to type, instance and accompanying relationships, or hierarchies, is still something that requires ontologies. Ontologies ‘know’ things and are validated by humans.

3. Search, single DB

LLMs’ limitations in search tasks are well-documented: hallucinations, outdated information, security and privacy, a lack of provenance and auditability/reproducibility to name a few. The Retrieval Augmented Generation (RAG) architecture, the widely accepted grounding approach, is gaining prominence.

Successful RAG systems hinge on precise information retrieval (IR), often using embeddings to indicate relatedness, not the crucial “how” for explainable decisions. In this domain, the consensus is that lexical/ontological or hybrid approaches excel over purely vector-based methods.

4. Search, democratizing siloes

Data seldom resides in a single enterprise source. It exists in diverse sources, formats, and syntaxes. To democratize data, aligning disparate siloed data to common standards is crucial, enabled by ontologies ensuring source-level interoperability. Similarly, converting natural language queries to ontological entities is vital for querying diverse data effectively.

Ontologies play a pivotal role in such cases, aiding seamless retrieval from siloes that key against IDs captured in ontologies, such as knowledge graphs. Whilst LLMs have a role in aspects of such a solution, e.g., converting natural language queries to a relevant query syntax (i.e., SSQL for SciBite Search or Cypher for Neo4J), summarising results from an accurate IR set – ontologies are also paramount.

When we model data with ontologies, we can ask extremely precise questions, get definitive answers and are able to use reasoning to both deduce and explain answers to queries. Ontologies were born out of a need to unambiguously identify the kinds of entities in the world (and in our data) and the relationships that hold between them. LLMs, on the other hand, provide a statistical approach to identifying these that can be presented through natural language – enabling us to see what things are potentially related, but not how.

Large language models are not going anywhere and will benefit us all, their ability to support operational tasks is clear for all to see, however, in situations where evidence-based decision-making is paramount, particularly within R&D in the Life Sciences, ontologies still have a massive role to play.

Just like ontologies or any piece of software for that matter, large language models are just tools to help you do things. No one thing solves every problem on their own. We must all remember to start by understanding the problem, not shoehorning a solution into a place it does not fit.