Physical AI Meets Scientific AI: Medra's Continuous Science Platform Accelerates Discovery

Medra Launches a Continuous Science Platform to Accelerate Data Generation and Discovery

AI in science is stuck on a simple constraint: data is scarce. Protein-structure datasets used for training models have been collected over decades and still represent a fraction-about 0.3%-of what fuels today's largest multimodal systems. Medra is addressing that gap with a Continuous Science Platform that integrates robotics ("Physical AI") and reasoning models ("Scientific AI") into a closed loop for running and improving experiments.

The aim is straightforward: compress years of experimental iterations into months by creating, labeling, and learning from data as an ongoing process-not a sporadic event.

Why this matters for researchers

Model quality in biology and chemistry depends on high-quality, diverse, and well-annotated data. If you want stronger predictions, you need faster experimentation cycles, precise provenance, and clear feedback loops between planning and execution. That's the bottleneck Medra is targeting.

Continuous Science: Physical AI + Scientific AI

Medra's platform is a self-improving system with two parts working in lockstep. Physical AI executes experiments with general-purpose robots guided by agentic vision-language models. Scientific AI analyzes the resulting data, proposes next steps, and updates protocols-tightening the loop with each cycle.

Physical AI: execution with provenance

The system automates up to 70% of instruments commonly used in labs and captures images, logs every motion, and records actions with high granularity. This creates an additional metadata layer-Infra-data-that links intent, context, and execution details to outcomes. The result: traceable runs, cleaner datasets, and fewer blind spots during troubleshooting.

Scientific AI: reasoning that learns from Infra-data

Reasoning models ingest Infra-data alongside entries from electronic lab notebooks and scientific literature. They suggest new experimental actions, refine parameters, and converge on optimal protocols faster. Over time, the loop reduces variance, improves hit rates, and surfaces edge cases you'd otherwise miss.

Where it's being used

Medra is collaborating with leading biotech and pharma teams on antibody design, gene therapy workflows, and cell-based assays. The company has released case studies with Addition Therapeutics and Lila, showing how closed-loop experimentation can lift throughput and shorten cycle times.

What this can mean for your lab

• Higher experimental throughput without rebuilding your entire stack
• Full provenance via Infra-data for every step, from pipetting to imaging
• Protocol optimization that adapts as conditions shift
• Better reproducibility and faster failure analysis
• Compatibility with a broad set of existing instruments
• Cleaner data pipelines for downstream model training

Day-to-day loop

• Plan: Scientific AI proposes runs based on prior results and literature
• Execute: Physical AI carries out protocols and logs granular metadata
• Measure: Outcomes are automatically captured and linked to actions
• Learn: Models update parameters and recommend the next experiments

Practical considerations

• Not every task is automatable; expect staged adoption across instruments and assays
• Map SOPs to machine-executable steps and validate guardrails early
• Align data governance, biosafety, and IT before scaling

Learn more

Explore Medra's Continuous Science Platform at medra.ai. For background on the datasets that have shaped structural biology, see the RCSB Protein Data Bank.

If you're upskilling your team on data workflows that support closed-loop experimentation, this program is a useful starting point: AI Certification for Data Analysis.