From AI Advisors to Operational Lab Workers
AI agents in life sciences are software systems that combine large models with validated scientific tools so they can design, run and interpret autonomous lab experiments, moving from passively describing drug discovery workflows to actively executing drug discovery automation tasks across wet-lab and computational pipelines while staying inside real-world scientific protocols. The shift underway is not cosmetic. For two decades, digital tools mainly advised scientists: suggesting structures, ranking hits, predicting routes. Now agents are expected to behave like operational workers, decomposing scientific goals into steps, calling the right tools and triggering physical experiments. That expectation is reshaping how R&D groups buy infrastructure, structure data and even design their lab spaces. It also raises a hard question: once AI systems execute protocols end-to-end, how do we govern their decisions without slowing the very speed we want?
BioNeMo: Turning General Models into Working Scientists
The clearest signal that agents are meant to work, not comment, is NVIDIA’s BioNeMo Agent Toolkit. BioNeMo takes the sprawling landscape of biomolecular models—protein-structure prediction, molecular docking, generative chemistry, genomic analysis—and packages them as documented “skills” an agent can call on its own. Debuting with nearly 50 BioNeMo toolkit partners, including Eli Lilly, Thermo Fisher Scientific and Dassault Systèmes, it offers a common toolbox for AI agents life sciences teams can trust. Kimberly Powell described the pattern bluntly: general-purpose frontier models might be the brains, but they cannot, by themselves, break “go design me a binder” into the five to seven specialized tasks that matter. The harness—the agent runtime that remembers context and enforces rules—plus BioNeMo’s skills turn those brains into workers that can carry out real scientific work rather than merely narrate it.
Boltz and Conversational Drug Screens: Agents as Colleagues
If BioNeMo is the toolbox, Boltz Bio shows how scientists will use it day to day. Since AlphaFold 2 cracked protein-structure prediction in 2020 and 2021, a wave of biomolecular models has followed, including Boltz, Chai, OpenFold and Protenix. Boltz has turned its own models into a drug discovery automation API: BoltzProt-1 for protein design and BoltzMol-1 for small-molecule hit discovery. Crucially, the company built this interface “for agents as much as for people”, exposing SDKs and plugins so coding agents can drive the pipelines. Running the Boltz API through Claude Code and its connector, one team asked in plain English for a small hit-discovery screen against EGFR using commercially available compounds—and the agent handled job submission, polling, data download and ranking. The run was quoted at $0.20 (approx. RM0.94) for eight compounds but ultimately billed around $0.10 (approx. RM0.47) because the system only charges for scored molecules. That is what operational looks like: an AI colleague that both talks chemistry and writes the code to execute it.
MilliporeSigma, SYNTHIA and Siemens: Bridging Design and the Autonomous Lab
Where Boltz focuses on in silico hits, MilliporeSigma is wiring AI directly into benches, solvents and automation. AIDDISON, the drug discovery platform the company launched in December 2023, sits on top of a portfolio that now touches nearly every bench, after Merck KGaA added Sigma-Aldrich and its reagents catalog in 2015 for about $17 billion (approx. RM80.2 billion). SYNTHIA, built from the academic Chematica program, maps retrosynthetic routes against roughly 12 million commercially available building blocks tied to the Sigma-Aldrich web shop and ranks them by cost, steps or how green they are. An API connects AIDDISON’s designs to SYNTHIA’s manufacturability checks and shopping lists. In September, Merck KGaA signed a memorandum of understanding with Siemens to connect digital experiment design to physical lab work across discovery, development and manufacturing, the first collaboration to draw on its newly acquired Dotmatics technology. In its fullest form, this stack aims at an autonomous lab that designs, runs and interprets its own experiments with little human input—a “lab in a loop” where positive and negative data continually refine the next experiment.
What Autonomous Experiments Mean for Drug Teams
All of these moves sit on a clear trigger: the maturing of biomolecular AI since AlphaFold 2, and now AlphaFold 3, turned structural prediction and interactions into routine computation. Drug developers face a choice between running open models on their own GPU clusters and pipelines, or renting access through workflow providers such as Tamarind Bio, Rowan or BioNeMo. With agents on top, the choice is more strategic than technical: are we comfortable letting software submit assays, call synthesis plans and queue reactions in an automated lab? Recent demonstrations suggest the answer is becoming yes. OpenAI and Molecule.one recently reported a “near-autonomous AI chemist” that paired GPT-5.4 with the Maria agent and an automated lab to run more than 10,000 reactions and improve a stubborn sulfonamide coupling, gains human chemists then reproduced at the bench. The next phase will test governance: Powell argues agents need validated domain-specific tools researchers already use, described in enough detail that the agent knows each tool’s inputs, outputs and failure modes. Until we get that level of clarity, agents will remain impressive advisors. Once we do, they will be expected to clock in like any other lab worker.
