Drug Discovery AI
M

Medra

Medra builds a Physical AI Scientist for preclinical drug research: robotic lab workcells and scientific AI that design, run, interpret and improve experiments in a closed loop. Scientists state a goal and constraints, by text or voice, in the AI Experimentalist, the reasoning layer launched in June 2026. It turns the request into a protocol, and the Physical AI Lab executes it on standard instruments, using a vision language action model and cameras to watch the bench, catch errors and resume runs. Every action, video and sample's metadata is logged. The protocol library covers work such as nucleic acid extraction, RNA transfection, NGS library prep, iPSC differentiation, CRISPR editing, tissue staining, cell culture and protein purification, and the company says its robotics can operate more than 75% of the instruments scientists already use.

Partners install workcells at their own sites or run projects remotely at Medra Lab 001, a 38,000 square foot San Francisco facility with hundreds of robots that opened in April 2026. Genentech connects the platform to its laboratory information system and machine learning infrastructure, Addition Therapeutics and Cultivarium use it in discovery work, and a DARPA funded project runs the AI Experimentalist inside Lab 001. Integrations include NVIDIA Nemotron models and the BioNeMo Agent Toolkit.

Michelle Lee, a Stanford AI Lab robotics PhD, founded Medra in 2022. It raised a $52 million Series A led by Human Capital in December 2025, with Catalio, Menlo Ventures, 776 and Fusion Fund. Engagements are scoped per lab, and data in customer installations stays under the customer's control.

AI Health Index verifiedNovember 19, 2026
Compare Medra with other vendors
Founded
2022
Headquarters
San Francisco, California
Website
www.medra.ai
Categories
drug-discovery
Indexed Products
Physical AI Scientist, Physical AI Lab, AI Experimentalist, Medra Lab 001
Buyer Segments
Biopharmaceutical discovery teams, Biotechnology companies, Academic research groups, Scientific AI developers
Assessment

Capability Axes

An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read

AI Capability
AA on AI CentralityThe artificial intelligence is the product. Remove the model and there is nothing left to sell.
Vendor Published

Models run every layer of the platform. The AI Experimentalist interprets a scientist's objective, turns text into a protocol and reasons across data, images and results to propose the next run, while the Physical AI Lab uses a vision language action model to read what is happening at the bench and drive the instruments, with computer vision catching errors mid run.

The workcells and instruments supply the capacity, but model driven planning, perception and control are what let one system run protocols as different as RNA transfection and protein purification without custom automation for each. Partners get the same platform on their own site or at Medra Lab 001.

CC on Autonomy and Oversight ModelAutonomy is claimed and oversight is asserted without a mechanism. Human in the loop appears as a phrase rather than a described control.
Vendor Published

Scientists set research goals and constraints through a chat interface and can edit protocols in natural language by text or voice. Within those limits the system interprets observations and proposes the next run, and the execution layer detects certain problems and resumes work on its own. Videos, metadata and action logs let scientists inspect everything afterward.

Which changes need a scientist's approval before they run, how a partner vetoes a scheduled action and when an unexpected outcome is escalated to an operator are not described. The closed loop is sold as running experiments without human intervention, so those checkpoints matter more than they would in conventional automation.

BB on Model and Technology TransparencyThe approach or the suppliers are named without the version and update discipline behind them. Naming a supplier is the entry to this band both here and on Model Supply Chain Disclosure, which ask different questions of the same disclosure: who receives the data, and what produces the output.
Vendor Published

The physical layer runs a vision language action model that interprets camera and sensor input and controls laboratory equipment, offered as Text to Protocol, Instrument Agent and Closed Loop Optimization capabilities. The planning layer is a framework for several kinds of models and scientific software. The June 2026 launch names NVIDIA Nemotron models for protocol editing and optimization and the NVIDIA BioNeMo Agent Toolkit, and partners can connect their own analytical environments.

The robotics model's training data, the default model behind each function and how a customer installation approves changes to the model mix are not stated.

BB on Model Supply Chain DisclosureSubstantial partial disclosure, or a chain that is structurally short: an in house build, a cleared model that cannot be quietly swapped, or a deployment where the transfer does not occur at all. Naming the model provider exits the band below into this one; the axis rises from here on the completeness of the party list and on the terms that govern data once it arrives.
Vendor Published

NVIDIA Nemotron is a named model component of the AI Experimentalist, used for protocol editing and optimization, alongside integrations with partner research systems. In a customer's own cloud or premises, the policy leaves production data under the customer's control, so the data path depends on which software and models that partner enables. Medra handles limited diagnostic and support material only by agreement.

Remote projects at Lab 001 are where the chain is least clear. The terms give no full list of production model processors there and no retention or training obligations for each component a remote project uses.

CC on Clinical and Operational EvidenceNamed customers, or vendor reported percentages with no method, denominator or reference standard. Scale of use is recorded here and is not treated as evidence of benefit.
Third Party Estimated

Genentech describes connecting the platform to its machine learning infrastructure and laboratory information system for discovery research. Cultivarium's chief executive, Henry Lee, contrasts Medra's flexible robotics with automation he calls inflexible and brittle, Addition Therapeutics deploys the workcells, and DARPA is funding a project that runs the AI Experimentalist inside Lab 001.

The company's own measures describe capacity: a 38,000 square foot lab with hundreds of robots and robotics able to operate more than 75% of the instruments scientists already use, with no instrument inventory or method behind the figure. Repeatability, error rates and research time saved have not been measured in a published study.

BB on AI Safety and PHI StewardshipCategorical commitments are published, such as no training on customer data, without the retention schedule behind them.
Vendor Published

In a customer installation, the customer controls production data and Medra says it does not store that data in its own systems. Limited support or diagnostic data reaches Medra only if the customer provides it and expressly agrees. The privacy policy describes encryption, access control and incident response for personal data the company handles, and says it does not intentionally collect sensitive personal data.

Default retention and training terms for data generated in a remote project at Lab 001 are not given. The platform targets preclinical research, so protected health information and identifiable sample data fall to whatever the individual project sets up.

Regulatory and Compliance
CC on HIPAA and BAA PostureA status is stated but not supported: compliance is claimed, or business associate status stated, without the underlying document, including where the only privacy notice published covers the website rather than the service that handles patients. The mirror case grades here as well: a substantive privacy document that reaches the service, with no statement of business associate status anywhere and no scope position taken. One of the two elements the band above asks for is present, which is a grade below it and not a grade at the floor.
Vendor Published

The data policy separates production material that stays in the customer's environment from limited diagnostics or support material shared with Medra, and the company says it does not intentionally collect sensitive personal data. The platform is built for research teams working with instruments and laboratory records, as Genentech's integration shows.

No business associate relationship is described, and nothing covers protected health information if a collaboration involves human samples linked to patients. That applies to both customer installations and remote work at Lab 001, where any such arrangement would have to be written into the project agreement.

DD on Security Certifications and Trust CenterControls are asserted with nothing independent behind them, or nothing is published. Read the note before concluding anything: this is the grade most often corrected on a second pass, because assurance material frequently sits on a parent domain or inside an old announcement rather than on the product pages.
Vendor Published

The data policy lists encryption, access control, vulnerability management, security monitoring and incident response for information Medra handles. Customers can keep software and production information in their own cloud or premises, support and diagnostic data follows a separate agreed path, and Lab 001 is reached through research partnerships.

Those controls are described in general terms. No SOC 2 report, ISO 27001 certificate or independent assessment of Medra's operations is named, and the scope of any external testing for hosted or remote work is not given, which matters for partners sending proprietary research programs to a shared lab.

CC on FDA and Regulatory StatusNo device claim is made and the product is scoped accordingly. Most administrative and operational products sit here and are not penalized for it, because this axis grades the appropriateness of the positioning rather than possession of a clearance. A stated limit on use, such as research use only or a clause that the product gives no medical advice, beside marketing that sells the product for clinical decisions also sits here, with both quoted, because the buyer has to ask which one governs.
Vendor Published

The Physical AI Scientist is a preclinical research platform. It coordinates work on laboratory instruments and produces data for scientists to interpret, and the commercial examples, from Genentech's discovery programs to the projects run at Medra's own lab and the DARPA collaboration, are all research rather than patient care.

No FDA authorization or regulatory position is stated for use in a clinical laboratory, and nothing qualifies the platform for diagnosis or treatment. Validation for regulated production settings, such as GMP manufacturing or clinical testing, would be a separate matter from the discovery deployments reported so far.

DD on AI Governance and Bias DisclosureNothing published on how model behavior is governed or tested. Multilingual operation with no subgroup performance sits here when the vendor markets recognition quality as a strength, because a caller the system failed to understand leaves no complaint and no record.
Vendor Published

Researchers set goals and constraints and can examine the record returned with each result, since the platform keeps actions, videos and metadata from the physical process for every sample. That shows how an output was produced. The reasoning layer is built as a framework that can combine third party and proprietary models, including NVIDIA Nemotron.

No review framework decides which components enter a deployment or how a change is evaluated across labs and equipment. Performance across instruments, data sources and research settings has not been reported, and no independent evaluation covers how the model interprets results or chooses the next experiment.

CC on AI Liability and RecourseMechanisms exist that let someone challenge an output, such as audit trails, source traceability or review before commit, with nothing standing behind the output and no route for the harmed party.
Vendor Published

A failed run can be traced. The platform records actions, images and metadata for each sample, its vision checks flag problems during a run, and the scientist interpreting results can see when execution diverged from the plan. Medra's policy allows limited support access when a partner agrees to it, which gives a route to investigate an unexpected result together.

The standard commercial material sets no obligation to correct an erroneous research finding and does not say whether a partner receives a replacement run or a financial remedy after a failed experiment. Those terms sit in the individual project agreement.

Integration and Deployment
CC on EHR and Interoperability DepthIntegration is claimed through standards or a middleware layer with no system named and nothing to verify.
Third Party Estimated

Genentech says the platform connects with its laboratory information management system and internal machine learning infrastructure, and Medra integrates with in house notebooks, models and workflows as well as the manual and automated instruments a lab already owns. Those interfaces carry results into laboratory data workflows.

No specific laboratory software product or interface standard is named in the Genentech example, and there is no connection to hospital or clinical record software, which fits a preclinical tool. What data moves and how much write access the platform has depend on each partner installation.

BB on Deployment Model and Data ResidencyOptions and residency are stated with isolation or the processing path left open.
Vendor Published

The software can be installed in a customer's own cloud or on its premises alongside a workcell, and in that setup the privacy policy keeps production data under the customer's control, with Medra's access limited to agreed support or troubleshooting. Partners working remotely send their projects to Medra Lab 001 in San Francisco, which fixes where the physical work happens.

For remote engagements, where model requests are processed and backups kept, and how one partner's data is separated from another's, depend on the negotiated setup and are not stated in the standard terms.

Commercial
CC on Commercial TransparencyNo price is published and the posture is discoverable: a buyer can establish how the product is sold and what drives the cost before contacting the vendor. Most of the index sits here.
Vendor Published

Two engagement models are offered: workcells installed at a partner's facility, or remote projects run at Medra Lab 001. Scope can include the robotics workcell, integration with instruments and laboratory software, and the scientific planning layer, and the newest lab access offer is presented as a beta for selected projects. Prospects book an introduction through the website.

No price is given for a workcell, software license, installation or individual project. Whether a relationship is billed on hardware capacity, usage or research milestones is not stated, and the deployment type and the partner's lab environment decide what is configured and charged.

CC on Setting and Specialty CoverageCoverage is claimed broadly without specifics, or stated clearly with nothing validating it yet.
Vendor Published

Preclinical life sciences research is the setting. Pharmaceutical and biotechnology companies deploy workcells at their own sites, and selected academic, biopharma and government partners work through Lab 001. The protocol library spans molecular biology, cell biology and protein work, from nucleic acid extraction and NGS library prep to iPSC differentiation, CRISPR editing and cell viability testing.

Named users so far are Genentech, Cultivarium, Addition Therapeutics and DARPA. Validated performance across research programs, equipment combinations and lab environments has not been shown, and chemistry, such as small molecule synthesis, is not part of the listed protocols.

Alternatives

Alternatives to Medra

Other products that do the same kind of work as Medra, and what each one does.

Lila Sciences makes very large claims about what its platform can do, and as of September 2026 it has begun publishing evidence for some of them.

Chemify makes the molecules it designs, in a laboratory it built itself. Rather than handing a design to somebody with a laboratory, it is trying to turn its own laboratory into a printer.

LabGenius designs antibody drugs for cancer using a machine learning engine wired into its own robotic lab, and develops them through a mix of partnerships and a wholly owned pipeline.

Protein design company in Emeryville, California, built on the ProGen family of protein language models, which grew out of research first showing that large language models can generate functional proteins across diverse families, published in Nature Biotechnology in 2023.

BigHat Biosciences, founded in San Mateo, California in 2019 by Mark DePristo and Peyton Greenside, designs and engineers antibodies with machine learning tied to its own automated laboratory, starting from an existing molecule, its own or a partner's, and improving it in rounds.

Protein engineering software company, based in Amsterdam and Zurich, that licenses an AI platform directly to R&D teams rather than working through discovery partnerships.

Commercial

Pricing

Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.

Entry Price Pricing Basis BAA Tier Implementation Source
Custom project quote
Medra scopes partner installations around the workcell, equipment and software environment, and selected partners can also run projects remotely at Medra Lab 001. No unit of charge, recurring fee or rate for remote projects is published. The standard policy keeps production data under the customer's control in its own environment. No business associate agreement or arrangement for protected health information is described. Customer installations can involve lab equipment and software integration. Fees for the workcell, installation, training and support are set in the project agreement. Vendor Published

Site deployments and remote lab access are separate engagement options. The company offers selected beta projects for its lab access. Pricing and the responsibilities of each partner are scoped to the project.