From smart predictions to AI agents that act
AI agents in drug discovery are software systems that not only predict molecules or analyze data but also plan, trigger and interpret physical experiments through connected tools, autonomous lab workflows and domain-specific chemistry and biology models. This shift marks a move from passive analytics to active, closed-loop decision-making in how new drugs are found and made. For years, “AI in pharma” meant scoring compounds or predicting protein structures. That era is ending. The center of gravity is moving from models in isolation to agents wired into robots, assay systems and biomanufacturing stacks. NVIDIA’s BioNeMo Agent Toolkit, Boltz’s agent-first API, and MilliporeSigma’s SYNTHIA ecosystem show the same pattern: giving AI agents the handles they need to touch the real lab, not just simulate it.
NVIDIA BioNeMo: Toolboxes, not talk, for AI agents
The clearest sign that AI agents drug discovery is leaving the whiteboard is NVIDIA’s BioNeMo Agent Toolkit. NVIDIA describes it as an open, harness-agnostic platform that gives AI agents the building blocks to specialize for science, bundling models for protein-structure prediction, molecular docking, generative chemistry and genomic analysis as documented skills an agent can call on its own. With adoption from nearly 50 partners, including Eli Lilly, Thermo Fisher Scientific and Dassault Systèmes, this is not a toy sandbox; it is how production teams expect agents to touch validated research tools. Jensen Huang’s line captures the intent: frontier models are the brains; BioNeMo is the scientific toolbox. Kimberly Powell is explicit that the point is not another general chatbot, but a way for any agent “operating system” to orchestrate specialized workflows with governance and memory across complex pipelines.
Boltz and drug screening automation through natural language
If BioNeMo is the toolbox, Boltz is turning one of the core tools of drug screening automation into something agents can drive end-to-end. Since AlphaFold 2 cracked protein-structure prediction, an ecosystem of biomolecular models has sprung up; Boltz is now wrapping its BoltzProt-1 protein design and BoltzMol-1 small-molecule hit-discovery pipelines behind a simple API. The company admits it built the system “for agents as much as for people,” with coding agents as the primary interface its own scientists use. In one public demo, a coding agent reached the Boltz API through Claude Code and its connector, using plain English to run a small hit-discovery screen against the EGFR kinase with commercially available compounds. The company estimated $0.025 per molecule, $0.20 for the eight-compound demo. That is the economic logic of agentic automation: make each experiment cheap enough that software can iterate aggressively.
MilliporeSigma’s SYNTHIA and the rise of the ‘lab in a loop’
Boltz and BioNeMo mostly live on the design and analysis side; MilliporeSigma is wiring AI retrosynthesis tools into the messy world of flasks and reactors. Its AIDDISON platform pairs generative AI with machine learning and computer-aided drug design, screening upward of 60 billion compounds for drug-like properties using more than two decades of pharmaceutical R&D data. SYNTHIA, descended from the academic Chematica project, maps retrosynthetic routes using expert-coded reaction rules and a catalog of roughly 12 million building blocks tied to its reagents shop, ranking routes by cost, steps or greenness. An API connects AIDDISON and SYNTHIA so a designed molecule arrives with a manufacturability check and shopping list. MilliporeSigma’s parent has also signed a memorandum of understanding with Siemens to connect digital experiment design to physical lab work across discovery, development and manufacturing, drawing on recently acquired Dotmatics technology to synchronize design, execution and data capture.
Autonomous lab workflows: why now, and what still blocks ‘lights-out’ science
The sudden acceleration toward autonomous lab workflows is not accidental. AlphaFold 2’s success in 2020–2021 was an unlocking event that proved AI could solve narrow, deep scientific problems and sparked a wave of biomolecular models. Karen Madden points to AI and generative AI, alongside CRISPR, as technologies that open up practical work that was previously out of reach. On the ground, MilliporeSigma shipped the Opentrons-powered AAW Automated Assay Workstation in 2025 to run routine assays with less manual handling, and this year introduced bio-based replacements for workhorse HPLC solvents, showing how reagents, robots and sustainability are being woven together. Madden argues that automation is valuable not just to spare sore hands from pipetting but because it creates more reproducible, higher-quality data—and to use AI, you need that data.
From there, the industry is edging toward what Madden calls a “lab in a loop”: AI and in silico modeling design the experiment, the physical lab generates positive and negative data, and the next experiment is better informed. On the frontier, one group paired GPT-5.4 with Molecule.one’s 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. Yet the vision of a lights-out lab—“you set it up, walk away, and come back in a week and everything’s done”—remains aspirational. The hard work now is integrating chemistry knowledge, lab hardware control and bioprocess workflows, stitching together digital design, wet-lab execution, data capture and scale-up so agentic systems can act reliably, not recklessly.
From predictive models to agentic automation: the real test
The move from predictive analytics to agentic automation is the most important test AI in life sciences has faced. Models that suggest structures or rank molecules are useful; agents that can autonomously plan and run experiments will change how organizations are structured and how risk is taken. NVIDIA is betting that giving agents validated tools researchers already use, with clear inputs, outputs and troubleshooting, is what will convince skeptics. MilliporeSigma is betting that owning the full stack—from filters and solvents to AI retrosynthesis and lab automation—lets it stitch together the lab of the future, step by incremental step. Boltz is betting that if a coding agent can screen targets through natural language at cents per molecule, many more ideas will be tested earlier in discovery. None of these bets is guaranteed, but together they signal a simple conclusion: the era of AI agents drug discovery being stuck in simulation is over. The next decade will be defined by how safely, and how fast, we let them run the lab.
