Talk to Your Biomedical Data: Turning Fragmented Research into Traceable Answers

Biomedical researchers don’t lack data. What they lack is a fast, trustworthy way to ask this data questions. Genomics, proteomics, clinical trials, scientific literature, and patents accumulate across hundreds of public and proprietary sources that were never designed to work together. Researching a single therapeutic target can take hours of pulling facts from separate databases. 

At Graphwise, we’ve spent years solving exactly this problem — building connected biomedical knowledge graphs and the conversational tools that let researchers query them in plain language.

The cost of disconnected data

Most research data lives in silos. In roughly 90% of cases, data from instruments, platforms, research labs, and internal systems of record stays disconnected. Teams struggle to pull together domain-specific insights quickly. And every delay in research pushes back the moment a drug reaches the market and a new treatment reaches patients.

Knowledge also disappears when experts with 20 or more years of experience take what they know with them when they leave. Teams end up repeating experiments and organizations can invest hundreds of millions of dollars in the same area without getting the expected result. Even when the right data is there, it’s not always clear how it bears on a decision, because hypotheses, assumptions, and evidence aren’t properly linked. Choosing which five therapeutic targets to advance into preclinical work, after evaluating hundreds of candidates, requires connecting facts that typically live in separate systems.

Regulated work raises the stakes further. To respond to authorities, teams have to align enterprise-wide scientific standards, such as controlled vocabularies, ontologies, and connected metadata, and then reuse that data in the right context. When data lacks structure and domain context, conversational AI built on top of it can hallucinate. For companies that rely on R&D to bring drugs to market, unreliable answers slow innovation.

Our approach to Scientific Knowledge Management

Scientific Knowledge Management is about making all available research data, internal and external, work for the teams trying to innovate. We approach it by bringing proprietary and public sources together into cohesive biomedical knowledge graphs, then letting AI agents reason over them. Four capabilities do most of the work

The first is domain-specific reasoning. AI agents are only as useful as their grasp of biomedical context, and that context comes from well-modeled domain ontologies. We draw on our Life Sciences expertise and data inventory to help organizations build and extend those ontologies, which shortens the path between discovery and insight.

The second is a semantic layer that sits across siloed data and harmonizes it. The data becomes findable, accessible, interoperable, and reusable (FAIR) by design. In chemical and materials science, this keeps AI answers relevant, precise, and traceable, so formulation cycles move faster. Across R&D, it means fewer duplicated experiments and more efficient teams.

The third is governance and explainability. It comes from running agentic workflows on domain-aligned data: a single source of truth grounded in reference data models and ontologies. Every fact an agent produces can be traced back to its origin. That reduces hallucinations and speeds up responses to regulators.

The fourth capability is an agentic AI backbone. With unified data underneath, agents can reason autonomously over past experiments. In biotech and medical research, this includes extracting scientific claims automatically to support hypothesis generation and validation. Experts can then spend their time on research instead of normalizing and connecting data.

Three client engagements

At the Oxford Drug Discovery Institute, a team working on Alzheimer’s therapeutic target discovery needed to evaluate 54 pre-selected genes as potential therapeutic targets. The information they had to work with was sparse and scattered across multiple databases, much of it public. We brought their internal and public clinical and research data together in one system, with visual dashboards, navigation views, and analytical tools. Researching a single target used to take anywhere from several hours to one or two days, and that time dropped drastically. The team could now prioritize the most promising candidates much faster.

A top-five global pharma company needed a preclinical discovery knowledge graph for one disease of interest. We combined their internal data, including gene expression and experiment results, with public sources such as UniProt, Ensembl, and Reactome, along with other relevant terminologies and ontologies. All of it was aligned to the company’s proprietary domain model, which we also helped build. The resulting graph holds more than one billion interconnected facts, and over 30% of them were inferred from the ingested data. For researchers, that means faster access to relevant insights in their daily work.

The third engagement, top-15 global pharma company, focused on chemistry, manufacturing, and controls (CMC). Their workflows were built around documents. Relevant information was buried among millions of siloed data points, and turning it into reports for regulatory submissions was a struggle, especially as requirements kept changing. We built an enterprise-wide CMC data hub that pulls data and metadata from SharePoint, Veeva, TetraScience, and ELN and LIMS systems, harmonized and validated against the company’s own taxonomy and ontology. We then experimented extensively with RAG and GraphRAG approaches to build conversational interfaces that answer competency questions. Parts of the regulatory submission process are now semi-automated.

The Linked Life Data inventory

Much of this work runs on our Linked Life Data inventory — a data fabric for open data in Life Sciences and Healthcare. We also offer it to clients as an accelerator for building a knowledge graph from scratch, targeted at a specific problem. The inventory rests on a methodology developed over several years for converting datasets from their native formats into a semantic, ready-to-use form. A verification process then keeps all those datasets harmonized with each other.

The key step is semantic mapping: recognizing when different datasets describe the same entity. Once those links are in place, the datasets can be connected into a single, well-linked network. We usually begin with a harmonization model, which sits as a top layer above the sources and maps its key entities to their counterparts in each dataset. Much of this mapping can be automated. For widely used sources like UniProt, we can reuse structures the original model already defines. Wherever open data has gaps, proprietary data can fill them.

Each dataset also carries metadata describing its content, licensing, and ownership, based on two W3C standards, DCAT v3 and VoID. Teams can see licensing constraints upfront and plan their knowledge graph builds around them. Automated workflows download and process the data, generate the metadata, and package everything together with the RDF data, ready to load and merge into other systems. 

The inventory currently covers more than 250 public datasets spanning genomics, proteomics, metabolomics, molecular interactions, pharmacology, clinical and medical data, scientific publications, and patents. That range supports use cases across the whole pharma value chain, including early research and discovery, preclinical development, manufacturing, clinical trials, and marketing and distribution.

Building a target discovery knowledge graph

Target discovery shows how the inventory works in practice. We defined a model for the use case and populated it with datasets covering proteins, pathways, drugs, cells, compounds, genes and variants, and diseases. Since the semantic mappings were already established in the inventory, these datasets linked up with each other from the start.

We then added facts extracted by NLP pipelines from more than 80 million publications and patents. The structured part of the graph also serves as training data for AI models. Some predict gene-disease associations based on how similar the objects are. Others run similarity searches to find comparable genes and proteins across species.

From knowledge graph to conversational assistant

A biomedical knowledge graph like this solves a lot of data problems: siloed sources, slow information retrieval, and inefficient drug discovery. Yet, researchers still need an easy way to explore it and advanced search with stacks of filters rarely fits complex biomedical questions. Our Talk to Your Biomedical Data assistant  lets them work through a research problem one plain-language question at a time. Under the hood, the knowledge graph lives in GraphDB, and GraphDB’s Talk to Your Graph functionality connects it to a large language model. The LLM is given the graph’s semantic model, which lets it turn each question into a SPARQL query, retrieve the relevant data, and summarize what it finds.

The answer then passes through the Graphwise Extractor. It recognizes the concepts mentioned in the text and tags each one with its URI in the knowledge graph. In this way, the researcher can click through and explore further. This step handles annotation only and is not part of a multi-hop RAG architecture. Alongside the text, the interface draws a visual subgraph of the relevant slice of the knowledge graph. Ask which genes are involved in Pompe disease, for example, and you get the gene-disease pairs plus a graph of those connections. A follow-up question about which biological processes those genes are involved in groups them into clusters by shared processes and pathways.

The setup is flexible. The demonstrator uses OpenAI, but any LLM will work, whether it’s a public service or a model hosted on-premise in the client’s own infrastructure. Switching only requires configuring the GraphDB instance. The assistant also comes with two configurations for two kinds of users. A standard user may know little about the domain and asks general questions. An expert knows the entities and relationships in detail and can ask much more precise ones.

Building and maintaining the system

How long does it take to get something like this running? Less time than most teams expect. If the datasets you need are already in the inventory, the knowledge graph can be ready within a week. Configuring Talk to Your Graph in GraphDB takes a couple of days more, mostly spent tuning the prompts, and then the system starts returning meaningful answers. A custom interface is optional. Some use cases need one, and for others, what’s built into GraphDB is enough.

Then there’s the question of keeping it all current. Public sources change constantly. UniProt, for instance, updates every two weeks. The good news is that the semantic model rarely needs to change with them. It captures the main entities and the relationships between them, while most updates touch specific data properties your use case doesn’t depend on. The trick is to decide early which properties actually matter, then manage the mapping of incoming data to the model as its own ongoing process. And if the model itself does need to change, you update it in the system configuration. The LLM picks up the new version automatically.

What it delivers

Across all our biomedical knowledge graph projects, we follow a semantic backbone approach. It combines built-in domain models and ontologies, verified data unified across silos, and knowledge hubs that sit on top. The results are consistent enough to point to a few typical outcomes. Preparing regulatory compliance documentation often takes 30 to 50% less time. Duplicated experiments and duplicated work drop by 15 to 30%. For tasks like mapping constantly changing regulatory requirements to internal workflows and business processes, manual hours fall by more than 80%. Institutional knowledge also stays in the organization when people move on.

The effect on innovation speed is harder to put a number on, since it depends so much on the use case. Still, getting a successful drug to market sooner can be worth hundreds of millions of dollars in potential revenue.

Biomedical data will keep growing and keep changing. Whether it moves research forward depends on three things: it has to be connected, it has to carry its meaning with it, and researchers have to be able to question it directly. A knowledge graph built on harmonized data covers the first two. A conversational layer that shows where each answer comes from covers the third. Together, they let researchers spend less time assembling evidence and more time acting on it.

Do you want to dive deeper?