Agentic AI in drug discovery

Last week I took part in a panel on agentic AI in drug discovery at TandemAI's symposium in Cambridge, MA. The room was packed for talks covering autonomous workflows on structured experimental data, enterprise lab automation, agents in medicinal chemistry and drug discovery, and a venture perspective.

The panel closed the afternoon with four of us discussing what agents change in drug discovery and where they fall short. These are some of the key points I took away.

Agents remove friction

Drug discovery teams run complex workflows across multiple tools and environments. Agents let a scientist describe the workflow in conversation instead of working at the command line, so a medicinal chemist or biologist can run it without waiting for a computational chemist. One panelist described using AI to go from a LinkedIn post about a new tool to a report for the team on whether it was worth testing. Others pointed to querying data across disparate sources in natural language. The productivity gains here come from removing effort.

Data quality remains a challenge

Agents speed up the data workflow, but don’t fix data quality. The risk becomes agents making decisions based on incomplete or incorrect data. Examples here include stereochemistry that isn’t defined or is recorded inconsistently. Assay data from different sources is generated under different conditions, so it should be combined with caution. One panelist added that a SMILES string loses information about the molecule, and agents will need better descriptors before they can perform well.

Reproducibility and the experiment

The conversation turned to whether a swarm of agents could converge on the best hypothesis. An agent can present a confident hypothesis but offer a better one each time it is challenged. A panelist noted that repeating the same task can give different results. Taken together, these mean an agent can be wrong and sound just as certain, which is a problem for hypothesis generation until it is resolved.

The nature of experimentation will change as agents use more predictive models to prioritise virtual compounds at scale. Instead of running wet lab experiments only on compounds we expect to succeed, I anticipate a shift to testing compounds to improve the models or validate a hypothesis, including the negative experiment we expect to fail.

Humans own the decision

Asked when they would trust an agent with a consequential decision, such as ordering a large, expensive synthesis or planning and starting a GLP tox study, the panelists gave the same answer in different words. Humans need to review and decide wherever governance and risk are significant. The reasons were costs and ethics. Agents can quickly rack up computational and experimental costs on compounds that were never going to succeed. And animal studies carry ethical weight. As one panelist put it, if the decision fails, the person is held accountable, not the agent.

In my own experience with everyday research workflows, I have more confidence in the outcome when I help design the workflow or approve the plan an agent proposes before it runs. I sign off on experiments and conditions that are appropriate for the task. I trust it less when I ask a question and the agent goes off and does its own thing, spawning increasingly complex workflows and taking several hours to accomplish something I could have done more quickly and comprehensively.

The role of the computational chemist is evolving

As agents take over workflow execution, computational chemists will spend less time running services for medicinal chemistry and more time interpreting results and shaping hypotheses. Other discovery team members will use agentic workflows to ask questions previously addressed by computational chemists. And computational chemists will use agents to be more involved in design and library evaluation. Panelists differed on the trajectory. One expected the role to be redefined within five years. Another expected the contribution to change while the discipline stays.

What does this mean for adoption?

Two points from the panel bear directly on software adoption:

First, panelists agreed that not every scientist knows what the software their company licenses can do. An agent that knows each tool's capabilities could deliver the value those licences were meant to provide. That raises the bar for vendors. My view is that software an agent cannot access will not be used, so APIs and MCPs will matter more to adoption than they do today.

Second, the advice on agentic rollout was to pick specific use cases, build trust, and add layers gradually. Data security and whether agents should operate only within an organisation’s firewall or have internet access came up more than once during the afternoon. One concern raised was that an agent working for a service provider might carry one client's data into work for another, although self-hosted models and sandboxed environments could reduce these risks.

If agents become the way scientists reach their software, vendors should ask themselves two questions. Can an agent find and use our software without a scientist's help? And what have we built in, such as warnings when inputs fall outside what a method was designed for, so results can be trusted when no scientist is checking?