From predictive models to agents that do the work
AI agents in drug discovery are autonomous software systems that chain together scientific models, tools and lab automation so they can design, plan and execute parts of pharmaceutical R&D workflows with minimal human intervention, rather than only predicting properties in isolation or suggesting next steps for scientists. For years, AI in the lab mostly meant machine learning models that scored compounds or predicted protein structures; now, a distinct shift is under way toward agent-based systems that call specialized tools, talk to lab robots and close the loop from design to experiment. NVIDIA’s BioNeMo Agent Toolkit embodies this turn: instead of another frontier model, it offers a “scientific toolbox” for any agent harness to call skills like protein-structure prediction, molecular docking, generative chemistry and genomic analysis so that general-purpose AI agents can carry out real scientific work, not just describe it.
Why BioNeMo’s agent skills matter more than another model
The key development is not a shinier model but that nearly 50 partners, including Eli Lilly, Thermo Fisher Scientific and Dassault Systèmes, are already validating NVIDIA’s BioNeMo Agent Toolkit in real life-sciences workflows. By packaging life-sciences software and models as documented skills an AI agent can call on its own, BioNeMo turns abstract model performance into usable pharmaceutical AI automation. “Frontier models are the brains. BioNeMo is the scientific toolbox,” said Jensen Huang, and that distinction matters. A harness-agnostic design means any agent framework can plug in, while governance controls let enterprises decide what tasks agents are allowed to run. This is how AI agents drug discovery moves from demos to deployment: not through monolithic “AI chemists” that claim to do everything, but through modular skills that slot into existing pipelines and satisfy compliance officers as much as they excite data scientists.
Boltz and conversational autonomous drug screening
If BioNeMo is building the toolbox, Boltz is showing what happens when you hand that toolbox to coding agents. Boltz’s new API exposes BoltzProt-1 for protein design and BoltzMol-1 for small-molecule hit discovery as services that are explicitly “built for agents as much as for people”. Internally, most of Boltz’s scientists, chemists and protein engineers already reach its models through coding agents like Claude Code, Codex and Gemini, and those integrations are now public via SDKs and command-line tools. In a test run, a user asked Claude in plain English to carry out a small hit-discovery screen against the EGFR kinase using a handful of commercially available compounds, with the agent orchestrating BoltzMol-1 to perform the autonomous drug screening workflow end to end. It estimated $0.025 (approx. RM0.12) per molecule, $0.20 (approx. RM0.92) for the eight-compound demo. That kind of conversational interface, tied to low per-molecule pricing, is what democratizes computational drug discovery beyond big-pharma GPU clusters.
SYNTHIA, AI retrosynthesis chemistry and the rise of the autonomous lab
The real proof that AI agents drug discovery has left the lab notebook comes from chemistry and automation. MilliporeSigma’s SYNTHIA, which grew out of the Chematica program, performs AI retrosynthesis chemistry by mapping routes with expert-coded reaction rules against roughly 12 million commercially available building blocks tied to the Sigma-Aldrich catalog and ranking them by cost, step count or sustainability. On the design side, AIDDISON pairs generative AI, machine learning and computer-aided drug design to screen upward of 60 billion compounds for drug-like properties. An API connects the two, so a molecule proposed in AIDDISON flows directly into SYNTHIA for a manufacturability check and a shopping list. Put together with the Opentrons-powered AAW Automated Assay Workstation shipped in 2025, a benchtop robot that runs routine assays with less manual handling, this becomes a “lab in a loop”: AI designs experiments, automation executes them and the data feedback improves the next cycle.
From skepticism to scaled pharmaceutical AI automation
The industry is not gliding smoothly into this future; scientists’ relationship with AI remains complicated, and distrust in parts of the community is high. Yet the direction of travel is clear. Agent plugins have already become the main way Boltz’s own experts run its models, and leaders there see “little experts” getting embedded into electronic lab notebooks to take on routine tasks. In its fullest form, an autonomous lab designs, runs and interprets its own experiments with little human input. Automation is not a convenience add-on: “Automation isn’t just nice because your hand doesn’t get sore from pipetting,” said MilliporeSigma CTO Karen Madden; “It creates more reproducible, higher-quality data. And to use AI, you really need that data”. The next logical step is digital twins of bioreactors and processes so agents can model outcomes before any run reaches the bench. Whether tools like BioNeMo push the field into this second phase of pharmaceutical AI automation remains open, but pretending that AI agents are still toys is no longer a serious position.
