From Predictive Models to Executable Drug Discovery Workflows
AI agents in drug discovery are autonomous or semi-autonomous software systems that combine language models with domain tools so they can plan, call scientific services and update data structures to execute end‑to‑end discovery and development workflows rather than only analyze information. After years of focus on single predictive models, pharma is starting to wire those models into continuous workflows that cover target selection, design, screening and trial setup. This shift marks a move from machine learning as decision support to agentic AI pharma teams can assign concrete tasks, such as configuring assays or translating protocols. Instead of stopping at a ranked list of compounds or a PDF protocol, agents now consume structured inputs, call specialized models and lab systems, and return machine‑readable outputs that downstream software or robots can act on. The practical question is no longer if AI can model biology, but how reliably it can close the loop between design and execution.
Toolkits for Agentic AI in Pharma: NVIDIA BioNeMo and Boltz
A new layer of infrastructure is making agentic AI pharma workflows more practical. NVIDIA’s BioNeMo toolkit turns life‑science models into documented “skills” that AI agents can call to do protein-structure prediction, molecular docking, generative chemistry and genomic analysis. NVIDIA describes frontier models as the “brains” and BioNeMo as “the scientific toolbox,” separating the agent harness from the domain tools. The BioNeMo toolkit has launched with nearly 50 partners, including Eli Lilly, Thermo Fisher Scientific and Dassault Systèmes, signalling strong interest in standardized agent skills for real scientific work. In parallel, Boltz has released an API that exposes its BoltzProt-1 protein design and BoltzMol-1 hit-discovery pipelines. According to Boltz, “We built this for agents as much as for people,” with SDKs and integrations so coding agents can drive autonomous drug screening workflows over simple API calls rather than bespoke pipelines.
Machine-Readable Protocols and Agentic Clinical Development
Agentic AI is also reshaping the slow, document-heavy front end of clinical trials. Medable’s Digital Data Flow Agent focuses on translating static clinical trial protocols into structured, machine-readable formats. The agent reads a protocol and renders it in CDISC USDM 4.0, a standardized JSON structure that turns what was once unstructured text into data software can act on. Medable pitches this as infrastructure for what regulators call real-time or continuous clinical trials, where protocol changes should propagate instantly across eCOA apps, documents and operational systems. Earlier, Medable used AI to translate a study protocol into a configured eCOA mobile app, with questionnaires, workflows and translation into about 25 languages, cutting deployment timelines that had been 12 to 16 weeks. The new agent goes deeper, offloading protocol data entry so clinical experts can focus on study design and interpretation rather than re-keying amendments.
Connecting Retrosynthesis, Autonomous Labs and Agent Workflows
On the wet‑lab side, agentic AI is converging with lab automation and retrosynthesis tools to bridge computational design with physical chemistry. MilliporeSigma, the Life Science business of Merck KGaA, is wiring its broad portfolio of research chemicals, reagents, filtration, water systems, software, automation and bioprocessing into an automated lab stack. Its Opentrons-powered Automated Assay Workstation runs routine assays with less manual handling, and the company is investing in AI retrosynthesis and autonomous lab frameworks so digital designs can flow into execution robots, solvent systems and data capture. This connects with agentic workflows where an AI agent can propose a synthetic route, call retrosynthesis services and trigger automated assays. Siemens’ SYNTHIA and similar retrosynthesis engines fit naturally into this loop, allowing agents to plan practical routes for candidate molecules and coordinate with robotic platforms to synthesize and test them, tightening the feedback cycle between design and experimental validation.
Toward Autonomous Drug Discovery and Industry-Scale Agents
Taken together, these efforts signal a shift from isolated machine learning models to agentic AI pharma platforms that can autonomously plan and execute discovery tasks. BioNeMo’s documented skills give agents a consistent way to apply cutting-edge biomolecular models; Boltz’s API lets them drive autonomous drug screening without standing up GPU clusters; Medable’s Digital Data Flow Agent turns wordy protocols into data pipelines; and autonomous lab stacks connect AI-driven design with physical assays. Adoption by companies such as Eli Lilly, Thermo Fisher Scientific and Dassault Systèmes shows this is no longer a speculative trend. The emerging pattern is agents supervised by humans but entrusted with tactical work: translating protocols, orchestrating screening campaigns and coordinating lab automation. As toolkits and standards mature, the competitive question for drug developers will be how much of their discovery and development workflows they are willing to hand to software agents.
