Is Life Science R&D Ready for Agentic AI?
Agentic AI has huge potential to augment how scientists work, but its deployment is an endeavor that demands expertise.
Complete the form below to unlock access to ALL audio articles.
Meet the latest artificial intelligence (AI) buzzword: agentic AI. There’s been plenty of discussion about this emerging technology across the tech industry, but how can agentic AI be of use to life sciences research and development (R&D)? And how should organizations approach it?
The field of AI has evolved rapidly in a very short time, particularly within deep learning, where GenAI chatbots powered by a Large Language Model (LLM) are now almost an entry-level use case. Agentic AI differs by employing agents that can autonomously complete tasks by “reasoning” through a problem and providing its own chain of logic to answer a query.
This is not hypothetical: OpenAI has already launched the Operator, an agent for completing web-based tasks, and Deep Research, an agentic capability that conducts multi-step research for complex tasks. Meanwhile, CrewAI is a recent example from a smaller startup firm of a framework that supports the development of agents.
When it comes to drug development, there are countless time-consuming tasks that are ripe for “agentification” – from target identification and preclinical toxicology to clinical trial, approval and post-marketing surveillance use-cases. But in a highly regulated and evidence-based industry like life sciences, any application of AI – agentic or otherwise – must be carefully considered and approached with caution.
How can an agent work in drug discovery?
Let’s consider target prioritization as an example. A target prioritization agent could take a therapeutic area as an input and provide a list of prioritized potential therapeutic targets by performing the same kind of multi-step analysis a researcher would perform. To do this, it will need to access at the absolute minimum:
- A database describing disease-gene relationships
- A database capturing the drugability of targets
- Literature sources for providing evidence or novel hypothesis
These three elements are the tools and functions. Broadly, tools are data sources, which can be external or internal, while functions specify how data is manipulated and shared back to the user.
There are two approaches to how tools and functions can be made available to agents. An organization’s approach will depend on how much freedom it wants to give an agent. On one end of the spectrum, it could take a fully autonomous approach, which would involve sending a mix of tools and functions to an LLM, and letting the LLM reason and decide which tools it will use in any order it wants.
On the opposite end of the spectrum, the organization could put guardrails in place – giving the agent a prompt and the defined linear flow it should follow to complete a task. A plan for our target prioritization agent might be:
- First, go to the database of disease-gene relationships and retrieve a set of genes
- Next, go to the drugability info and retrieve information on each gene from step 1
- Then, review the literature for:
- evidence supporting the most drugable targets and associated adverse events
- other targets associated with the disease that were not captured in step 1 and can therefore be assumed as novel
For even more complex tasks or queries, multiple agents may also be combined. For example, an agent that specializes in identifying areas of unmet medical need could be used in conjunction with the target prioritization agent to look for prioritized targets for areas of unmet need.
How to use an agent safely and effectively in life sciences
Agentic AI has huge potential to augment how scientists work and save time. But in practice, evidence-based decision making is central in life sciences. For safe and effective adoption, R&D organizations should follow established AI best practice.
Firstly, data is critical. Agents can only be as good as the data they have access to, so sources should be verified and trusted. Agents should also have access to high-level descriptions of the tools they have at their disposal (e.g., “Use XYZ service to review whether a protein target can be modified by small-molecule compounds”).
Data sources must follow FAIR (findable, accessible, interoperable, reusable) principles. Agents will often have access to multiple tools with data in different formats and following different standards. It is vital to be able to harmonize the output from one system with the output of another. Using ontologies to equivocate data in this way alleviates these issues by converting unstructured text to machine-readable identifiers.
Traceability, explainability and transparency are key. It is vital to understand why an agent made a particular decision – what tools were used, that all relevant data were retrieved and what was the “reasoning” that led to the output from the agent? In short, can we be sure that the entire relevant search space has been examined. For that reason, human researchers must always be kept in the loop when designing and evaluating agent inputs and outputs.
Agent intelligence meets human intelligence
The potential for agentic AI to transform life sciences R&D is incredibly exciting, but to make it a success organizations need to employ a mix of scientific and tech expertise. On the one hand, they need people who can understand how agents work and how to apply them; on the other hand, they need scientists who can parse research questions and understand the data sources and nuances of the domain. Deploying agentic AI is not just a technical challenge, it’s a scientific endeavor that demands scientific expertise for success.
By following the best practices outlined here, agentic AI can be embraced successfully and responsibly – and can bring huge improvements to R&D workflows.
