From chemist to AI agent builder: inside the rise of agentic AI in drug discovery

Srijit Seal’s trajectory through drug discovery reads like a compressed timeline of the field’s own transformation. Trained as a chemist, he completed a PhD in machine learning and toxicology under Prof Andreas Bender at the University of Cambridge, followed by a postdoc at the Broad Institute of MIT and Harvard with Dr Anne Carpenter – a period that coincided with the earliest stirrings of agentic AI in scientific workflows. After a stint at Merck, he joined Human Chemical in March 2026 as Principal Scientist AI/ML, where he now builds AI agents for chemical safety and toxicology. He is also a visiting scientist at Uppsala University in Sweden, where much of the collaborative work underpinning a major new paper took shape.

That paper – “AI agents in drug discovery: applications and case studies,” – is the anchor for our conversation. Co-authored by Seal, Dinh Long Huynh, Anne Carpenter, Andreas Bender, Ola Spjuth and the broader AIAgents4Science (AIA4S) Consortium, it comprises case studies from startups across the US, UK and Sweden to map out where agentic AI is delivering tangible results in drug discovery right now.

The capability step-change

When asked what had changed in the last six to twelve months, Seal’s answer was blunt: speed of build and breadth of acceptance. Ideas that would have taken months to prototype in early 2025 can now be assembled in days, thanks to improvements in tool calling, skills, connectors and the reasoning capabilities of frontier models. But Seal is careful to distinguish ease of construction from actual scientific capability. The underlying models still require non-expert users to understand their computational environment – package management, variable handling, environment configuration – even if the wrappers have become far more accessible.

“There are huge opportunities to empower lab scientists and toxicologists with tools that help them search, analyse data and generate hypotheses.” – Srijit Seal

The bigger shift, he argues, is cultural. Six months ago, a significant portion of the field was still dismissing agentic AI as hype. That resistance has weakened substantially. The release of increasingly capable models in early 2026, alongside the emergence of tools like Claude Code, have moved the conversation from ‘is this real?’ to ‘how do I use it?’ As Seal put it, telling someone you work on agentic AI in 2026 invites the response: who doesn’t?

Why generalist models fall short

A recurring theme in our discussion was the inadequacy of general-purpose large language models for specialised drug discovery tasks. Seal was direct: you cannot hand a compound or a target to a general model and expect it to construct the mechanistic understanding, access the relevant toxicology models, or find the papers with the correct effect sizes. Drug discovery demands domain-specific context – ie, which assays are regulatory-accepted, which in vitro data are replicable, which animal studies have adequate statistical power, etc. General models lack the training data equivalent of GitHub for chemistry; public databases like ChEMBL and PubChem exist, but in-house pharmaceutical data is fundamentally different in character and quality.

“You can’t just throw a compound or target into any large language model and expect it to understand the molecule, access relevant toxicity tools or even know the field. There’s a lot of mechanistic understanding that you need.” – Srijit Seal

This is precisely the gap the AIA4S Consortium paper addresses. Rather than theorising about what agents might do, the authors assembled real case studies from startups that are building and deploying specialised agentic systems. Coincidence Labs demonstrated multi-agent literature analysis for medicinal chemistry, compressing weeks of patent and synthetic aperture radar (SAR) data extraction into automated workflows.

Human Chemical – Seal’s current employer – showcased an agent for in silico toxicology that combines predictive graph neural network (GNN) models with literature synthesis and regulatory data retrieval, illustrated through a detailed endocrine disruption risk assessment of the synthetic odourant Cashmeran. Kiin Bio presented Virtual Scientists, integrating over 100 tools across biology, chemistry and clinical domains, executing an end-to-end idiopathic pulmonary fibrosis discovery workflow in under two hours versus the typical two to three weeks. Other case studies covered automated protocol generation (Potato’s Tater agent), drug repurposing for rare diseases (Augmented Nature), automated small-molecule synthesis (onepot.ai), focal graph-based target discovery (Plex Research) and discovery-to-deal asset evaluation (Convexia Bio).

The preclinical toxicology bottleneck

Seal reserved particular emphasis for what he sees as one of the most impactful near-term applications: investigative toxicology in preclinical development. Many compounds are killed before reaching the clinic because unexpected toxicity findings cannot be adequately explained. Scientists default to terminating programmes rather than pushing forward with an incomplete understanding. Much of the investigative work, Seal noted, still relies on manual literature searching – often through Google – which cannot scale to the hundreds of papers that may be relevant to a single toxicity signal.

Agentic systems change this equation. At Human Chemical, besides working on chemical risk assessment for a range of chemical industries, Seal is also building the case for agents in drug discovery to autonomously generate complete hypotheses explaining observed toxicities in specific species, integrating predictions, literature, regulatory assessments and metabolite analysis into coherent narratives. These are not summaries, but investigative workflows that would take a human toxicologist weeks to assemble.

“I wouldn’t have expected the agents to come up with an entire hypothesis of why we saw XYZ toxicity in ABC species and build out an understanding of that. The investigative work is really a bottleneck – a lot of compounds get killed even before they enter the clinic for toxicity understandings that you can’t make sense of.” – Srijit Seal

If these tools can reduce the number of compounds unnecessarily abandoned and decrease reliance on animal testing by enabling better interpretation of existing data, the impact on both attrition rates and getting better with human-relevant data could be substantial.

The workforce question

We touched briefly on the workforce implications. Seal’s view is that these tools empower scientists rather than replace them, creating opportunities through efficiency gains. I offered a counterpoint: every industrial revolution has produced job displacement and while new roles emerge, workforce rebalancing historically takes decades, not years. We agreed that resistance to adoption – as Seal illustrated with historical examples from experience with computers and English-language education in India – tends to be more costly than adaptation. But the transition is neither painless nor instantaneous.

The startup advantage

“Chemical space contains around 10^60 molecules. At the end, it’s an optimisation game, balancing drug target, chemistry, safety, efficacy and manufacturing. There are opportunities for AI to improve each of these areas.” – Srijit Seal

One of the paper’s most striking features is its startup-heavy author list. Seal explained this was deliberate. The consortium specifically sought out startups at the frontier of agentic AI deployment because large pharma, at least until recently, was not hiring in the agentic AI space and was not building these systems internally. Startups have the freedom and velocity to experiment with novel architectures that pharma’s risk-averse culture does not easily accommodate. The paper also included venture capital perspectives – Zetta Venture Partners contributed – to capture where investment capital sees the most promising applications.

What readers should take away

Seal’s closing message was an encouragement to try. Specialised agentic tools from various providers are increasingly available with free trials. Their architectures differ substantially from what the field has seen before and their capabilities – in literature synthesis, toxicity prediction, hypothesis generation and protocol design – are expanding rapidly. But they will only improve if practitioners engage with them, identify gap and feed back what they need.

“They can help you compress analysis from weeks to hours. They can help you develop assays, protocols and hypotheses. We’ll only be able to develop them if people come back to us and say, here’s what we want.” – Srijit Seal

The paper and the consortium behind it, represent a deliberate effort to move the conversation from theoretical potential to documented practice. For drug discovery scientists still on the fence, the evidence base is now substantial enough to warrant hands-on engagement.

Srijit Seal’s paper, “AI agents in drug discovery: applications and case studies,” is available as a pre-proof in Drug Discovery Today (DOI: 10.1016/j.drudis.2026.104650). The AIAgents4Science Consortium can be found at broad.io/aiagents4science.