Cradle

The generative model is the product and the delivery mechanism is software rather than a service wrapper. The platform inverts conventional protein engineering by taking a desired property and producing candidate sequences, rather than iterating experimentally from a natural starting point, and the company's own framing of protein design as a digital service is accurate to what is sold.

The wet lab exists to expand the foundational models across additional modalities and properties, which is the correct direction of dependency for an A grade: the lab serves the model, not the reverse.

The platform proposes and the customer's scientists dispose, with the stated value being fewer experimental rounds rather than no experimental rounds, so wet lab confirmation remains the structural check. One disclosed data point cuts the other way and is worth knowing: in the Align to Innovate competition the company's enzyme models were generated in an automated fashion without expert intervention, which indicates the system can operate with minimal human steering when the problem is well posed. Held at B because no confidence thresholds, failure rates or documented guidance on when a design should not be trusted were located.

The approach is described concretely, covering generative design from specified properties, 3D structure prediction, thermostability and codon expression optimization, and models trained on each customer's own data rather than a single shared model. Held at B because no peer reviewed methods paper establishing core platform performance was located in this review and no model cards or published benchmarks were found.

The partial offset is that the company entered a third party competition where its output could be measured against others, which is a form of external exposure most peers avoid, though the result is company reported rather than independently published here.

One architectural choice provides separation by design rather than by policy, which is the substantive point here. Models are trained on each customer's own data rather than as a single shared model, so a partner's assay results shape a model fitted for them rather than improving a service sold to a competitor, and in this lane that is the commercially sensitive question rather than a privacy one: a protein engineering company's assay data reveals which targets and which properties it is pursuing.

Per customer fitting answers that structurally, provided the boundary holds, and the residual question is the one this index has now learned to ask of every local or per customer training claim. Customer data staying separate and model artefacts staying separate are different commitments, and nothing states whether anything learned in fitting one customer's model, including architectural choices, hyperparameters or general representations, informs the base from which the next customer's model starts.

The clinical form of this axis does not reach the workflow, which operates on protein sequences, structures and customer assay data with no patient records involved. On enumeration there is nothing: no hosting arrangement, no sub processor list, no retention position for customer assay data, and no statement of what happens to a customer specific model at the end of an engagement. Ask all four.

Operational evidence is strong and unusually well specified. As of the December 2025 update the platform served six of the top 25 global pharmaceutical companies across more than 50 R&D programmes, with customers named rather than anonymized, including Johnson & Johnson, AbbVie, Novo Nordisk, Novonesis, Grifols and Lundbeck, and the earlier Series B disclosure of 21 customers and 31 molecules in development gives a growth trajectory that can be checked against later statements.

More notable for this index, the company submitted to third party comparison: in the Align to Innovate competition its enzyme models outperformed or ranked in the top two among roughly 30 teams drawn from industry and academia. Entering an open benchmark where you can lose is rare in this category and the index treats it as a positive evidence signal.

Held at B rather than A because the competition result is company reported, no peer reviewed outcome study was located, the stated R&D speedups of 1.2 to 12 times are vendor generated, and no therapeutic designed on the platform with human data was identified.

Not applicable in the provider sense and rated accordingly rather than penalized. The platform operates on protein sequences, structures and customer assay data with no patient records in the workflow. The stewardship question that matters here is customer intellectual property rather than PHI, and it is addressed unusually directly on the commercial and governance axes.

Not applicable. Customers are pharmaceutical, biotechnology, agricultural, food and chemical R&D organizations licensing design software, not covered entities transferring protected health information.

A trust centre does exist at trust.cradle.bio, alongside a dedicated security page, and the underlying programme is described in more technical detail than almost anything else in this category. That corrects an earlier finding on this row.

What is substantive. The platform runs on a major cloud provider's enterprise infrastructure in European data centres, which is a residency statement rather than a vague assurance. The company describes a defence in depth approach spanning every architectural level from employee devices to the web interface and including the AI processing layer, and names specific modern controls: zero trust architecture, phishing resistant multi factor authentication and verified software stacks. Security is owned by a named executive whose background is engineering and applied machine learning research at a major technology company, which is a real accountability line. The cloud provider has published a case study describing the work, so part of this is corroborated by a third party rather than self asserted.

What holds it at B is a claim inconsistency across the company's own pages, and it is the exact construction this index tracks. The product pages state the company is fully compliant with a named attestation. The security page says its practices align with and exceed that standard alongside others. The case study says the company is adopting those standards for its platform. Those are three materially different claims about the same thing, and none of them states that a report has been issued, by whom, covering what scope, for what period. Adopting a standard, aligning with it and holding an attestation against it are separate states.

One distinction worth keeping from the earlier assessment: the company's commitment that customers retain full ownership of proteins engineered on the platform is a contractual term, not a security attestation, and the two answer different questions. Ask for the report, its scope and its period, and ask directly whether the audit has completed.

Nothing to assess rather than something assessed poorly. The platform is design software, is not a regulated medical device, and the company maintains no disclosed proprietary clinical pipeline. Regulatory standing for any protein designed on the platform sits with the customer that owns it, which is the direct consequence of the IP ownership term, and customer assets are not publicly attributed to the platform.

No formal governance framework, model card or bias disclosure was located, and a second search did not change that. What the company does answer well is a different question, and the distinction matters.

On tenant isolation it is unusually explicit. Where a shared model improving on every customer's data creates an obvious conflict, the company states that models are trained on each customer's own data and tailored to them, and its chief executive has put it plainly on the record: improvement from a customer's experimental cycles always happens in isolation for that customer, and the company never crosses the organisational boundary. Customers retain full ownership of proteins engineered on the platform. It also entered an open third party benchmark against roughly thirty teams from industry and academia, which is an evaluation it could have lost. All of that is credited and it is better than the lane norm.

None of it answers this axis. The domain relevant question is where the model degrades: a system trained largely on well characterised proteins will perform unevenly across sequence space, and no performance breakdown by protein family, fold class or data density was located. That question has a sharp commercial edge here specifically, because the platform is sold on customers improving it with their own experimental data. The practical thing a buyer needs to know is how it performs on their family before they have contributed anything, and that cold start figure is exactly what is unpublished.

Buyers should also confirm in contract whether any learning generalises across tenants and what happens to trained artefacts at termination, since the public statements describe intent rather than architecture.

One form of external exposure sits on this record and it is worth crediting because most peers avoid it. The company entered a third party competition where its output could be measured against others, which means it submitted to a test it did not design, on a task it did not choose, against competitors who would notice a weak result.

That is a weaker artefact than an independent head to head evaluation published by the organisers, because the result here is company reported rather than independently retrieved in this pass, and it is a meaningfully stronger signal than a self run benchmark.

The approach is also described concretely, covering generative design from specified properties, structure prediction, thermostability and codon expression optimisation, and models trained on each customer's own data rather than a single shared model, which tells a partner what stages exist and how separation is achieved. Held at C because nothing is measured or committed.

No peer reviewed methods paper establishing core platform performance was located, no model cards or published benchmarks were found, and no warranty, indemnity or remediation commitment attaches. Per customer training also raises the cold start question this index has recorded elsewhere: a model fitted on a customer's own data performs worst when that data is thinnest, which is at the start. Ask for the competition result and its independent source, and for performance during early engagement.

Not applicable in the provider sense, with no EHR touchpoint or clinical workflow surface. The research equivalent is the platform's own interface, which is designed for customer scientist teams to operate directly rather than requiring a computational biology intermediary, and that accessibility is the company's stated reason for calling itself enterprise software rather than a service.

Genuinely a deployed software product rather than a partnership, which is rare in this category: customer scientist teams use the platform directly across more than 50 concurrent R&D programmes, and the company has built out dedicated teams in the United States to support deployment alongside its European base.

Held at B because no on premise or private cloud option was located and no regional data residency terms were published, so proprietary sequence and assay data does move into the vendor environment. Compare Iktos, which offers on premise and virtual private cloud implementations and therefore scores higher on the same question.

No pricing is published, which is the ceiling on this grade, and unlike the partnership only platforms in this category a rate card here would be both possible and meaningful since the product is licensed software. What is published is better than most: a specific and checkable customer footprint (six of the top 25 pharmaceutical companies, more than 50 R&D programmes, with major customers named), a full funding history (24 million dollar Series A, 73 million dollar Series B led by IVP with Index Ventures and Kindred Capital, over 100 million dollars total), and one genuinely important commercial term stated openly rather than buried in negotiation, that customers retain full ownership of any proteins engineered on the platform. Publishing the IP position is the disclosure that most affects a buyer's decision here.

The widest applied coverage in this category after Schrodinger, and unlike most peers the breadth is across industries rather than only therapeutic areas. Disclosed use spans therapeutics, diagnostics, agriculture, food production and chemical manufacturing, with customers reflecting that range: Johnson & Johnson, AbbVie, Novo Nordisk and Lundbeck on the therapeutic side, Novonesis in industrial enzymes and Grifols in plasma derived medicines.

Modality is proteins and biologics rather than small molecules, which is the real boundary, and the company is expanding its wet lab specifically to extend the models into additional protein modalities.