The new bottleneck: Experimental validation

Biology is entering a new bottleneck. When AI models can now generate hypotheses, candidate molecules, and experimental plans faster than labs can test them, the limiting step is now experimental validation at scale.

Today, Medra is introducing the AI Experimentalist, the scientific reasoning layer of our Physical AI Scientist platform. The AI Experimentalist works with Medra’s Physical AI Lab to design, execute, interpret, and improve experiments in a continuous closed loop.

Launching the AI Experimentalist

The AI Experimentalist works in concert with the Physical AI Lab, Medra's wet-lab execution layer. Together they form the Physical AI Scientist platform: a closed-loop system that tightly couples scientific reasoning and experimentation.

Through a chat interface, scientists can define goals, set constraints, and generate experimental plans against a specified target or outcome. Proprietary models, open-weight models, and scientific agents, including NVIDIA Nemotron open source models and NVIDIA BioNeMo Agent Toolkit, are accessible through integrations within the platform.

The AI Experimentalist translates wet-lab observations into the next experimental workflow, enabling experiments to iterate continuously with minimal human handoff needed.

From goals to executable campaigns

Scientists can start with a high-level goal such as: “Build an EGFR blocking antibody assay cascade.”

The AI Experimentalist uses campaign context, internal knowledge, publications, available assays, antibody panels, and analytical approaches to generate an executable experimental plan.

Experimental designs can be guided by target, cost, speed, assay availability, and coverage of the search or design space.

‍
Iterating with the Physical AI Lab in the loop

The AI Experimentalist continuously learns from each experiment. Because it is directly connected to the Physical AI Lab, it can interpret results, identify unexpected outcomes, propose next steps, and deploy revised protocols with minimal downtime.

With each readout, the AI Experimentalist can adjust protocol parameters and continue executing, accelerating time to results.

Case Study: Driving Down Time Spent on Experimentation

The AI Experimentalist can optimize the workflows inside an experimental campaign for goals such as cycle time, throughput, resource use, and reproducibility, not just the campaign itself.

In our antibody screening workflow, generating binding data typically takes three days, making cycle time a major bottleneck to experimental throughput. The process requires plasmid cloning, cell-free expression, preparation of material for biolayer interferometry (BLI), and binding analysis.

The AI Experimentalist identified the cloning step as a potential source of delay and redesigned the assay around linear DNA expression templates for cell-free protein synthesis. While the change removed a two-day cloning step, expected protein production from linear templates is lower than with the plasmid-based workflow.

Instead of treating the lower yield as a stopping point, the platform leverages the Physical AI Lab in the loop and tests multiple linear-template designs and protocol conditions in parallel, comparing the resulting yields, and iterating toward a workflow that recovers comparable expression for binding characterization while preserving the time savings.

The result: a three-day workflow is reduced to approximately 14 hours. The gain did not come from a single shortcut, but from the compounding effect of removing cloning, testing alternatives in parallel, tightening handoffs, adapting protocols, and feeding experimental results directly into the next run.

The AI Experimentalist is powered by NVIDIA Nemotron and BioNeMo Agent ToolKit

The AI Experimentalist uses a multi-agent architecture with a model-agnostic agentic harness, allowing Medra to incorporate capabilities from different frontier AI models, scientific agents, and predictive systems.
‍
NVIDIA Nemotron provides open-weight models that Medra can fine-tune for platform-specific tasks such as protocol editing, protocol optimization, and experimental decision-making. Medra has deployed NVIDIA’s family of open-weight Nemotron models, including Nemotron Ultra, Super, and Nano, giving the platform flexibility to serve the right model for each task. The BioNeMo Agent Toolkit extends this architecture with scientific agents and tools for biological reasoning. For example, when designing EGFR blocking antibodies, the AI Experimentalist can call structure prediction models through BioNeMo to help score and rank candidate panels.

Because Medra’s harness is model-agnostic, the platform can also integrate predictive models and scientific agents developed in-house, by partners, or by customers.

Toward continuous, autonomous, closed-loop science

The AI Experimentalist enables a new era of closed-loop science, where scientists define the goal and the platform continuously designs, executes, measures, and improves the experiments needed to reach it.

In the EGFR nanobody campaign, the time savings came from many small compounding optimizations: removing cloning, testing linear DNA templates in parallel, optimizing expression conditions, and immediately feeding results back into the next run.

That compounding effect is only possible when scientific reasoning is directly connected to wet-lab execution.

With Medra’s Physical AI Scientist platform, powered by NVIDIA Nemotron and BioNeMo Agent Toolkit, experiments can move from static workflows to adaptive systems that learn with every result.

Curious how the AI Experimentalist could accelerate your discovery programs? Book an intro with our team to learn more.