It has been an interesting observation to see the data challenges large pharmaceutical companies are facing this year as a result of expediting the deployment of agentic workflows and frameworks within early drug discovery R&D departments. I am commonly asked what best practices a pharma company should follow with regards to deploying an agentic framework that will be easy to scale, agile to new models, and avoid hallucinations to research questions. In almost every one of those conversations the difficulty comes back to problems with the data foundation the agents are being asked to reason over.

‍

The first is trust. I have found that scientists do not trust an agentic workflow if they are unable to validate the claims being made by drilling back to the underlying source. When an agent returns an answer to a complex scientific question based purely off its internal weights and there is no route back to the source evidence, the scientist has no way of separating a well-founded claim from a hallucination, and so the output never makes it into a decision. It is therefore critical to base AI models and agentic workflows in structured, normalised data that allows the powerful reasoning capabilities of the model to be grounded in traceable data.  

‍

The second element relates to the ease of data integration. Data sources that have not been modelled to the Findable, Accessible, Interoperable, and Reusable (FAIR) principles cannot be easily mapped to internal data so every new dataset a research team wants to bring in becomes a bespoke integration project rather than a configuration change. The way in which this data is exposed to the agent is also important; data can be directly merged with internal sources within a client’s architecture, or it can be merged at the application layer through an API or MCP server. Choosing the right approach is key to delivering the right outcomes and ensuring maintainability where the state of the art moves every few months.

‍

At Biorelate we have promoted the fact that our causal relationship data is an excellent grounding semantic layer for agentic frameworks that is easily integrated and improves the quality and trustworthiness of the output [link to the other articles on the content hub that show this]. Causal data is particularly important in drug discovery and it allows scientists to build mechanistic models that explain why a target plays a role in disease, for example.

‍

Biorelate has successfully deployed our data within pharmaceutical companies' agentic architecture and the results were astounding. Our partner was able to establish the architecture within 3 business weeks, standing up a production environment which merged the Biorelate data with their proprietary datasets, leveraging the open source ontologies to expedite accurate data mapping. They deployed the custom dataset within a graph database (Neo4j) and exposed this via MCP to an internal agentic framework to perform complex scientific queries.

‍

The challenges of deploying complex architectures within large IT departments are a thing of the past, and with the technology barrier to entry being lowered so significantly in the past two years, combined with agentic workflows that can easily read structured, normalised, and FAIR datasets, the drug discovery R&D departments within pharma companies have a unique opportunity to accelerate the way scientific research and early drug discovery can operate today and be future proofed for success during this AI revolution.

‍

If you would like to learn more about agentic architecture best practices within drug discovery R&D please contact Biorelate to speak with a solutions consultant.

‍