Cradle turns experimental results into AI-guided protein engineering

Cradle provides AI software for engineering proteins against experimentally measured objectives. Scientists define a project, generate candidate sequences, test them in a laboratory and return the measurements to the platform. Cradle then adapts project-specific models for the next round. Its value proposition is a more productive experimental loop, with the customer retaining responsibility for the assays and the evidence that a candidate meets the intended requirements.

01 / ProductSoftware for an iterative protein-engineering program

The Cradle platform supports engineering enzymes, peptides and antibodies through a web interface and API. It describes multi-property optimization, configurable sequence constraints and data management. Rather than asking a scientist to choose one fixed prediction and stop, the workflow uses new experimental measurements to shape subsequent candidates.

Cradle’s overview places biopharma and industrial biotechnology in the same broad offer. The common requirement is a measurable protein property and a reliable route to testing variants. That could involve several competing objectives. A sequence that improves one measurement but undermines another may fail the project’s target profile, even if the single-property result looks impressive.

The company operates an Amsterdam wet lab to develop datasets and evaluate its modeling methods. That internal work is distinct from a promise to run every customer’s experiments. Cradle describes itself as a software company; customers evaluating the platform should plan their own experimental execution or an appropriate external laboratory relationship.

02 / AudienceUseful when a team can close the measurement loop

The strongest fit is a team that can repeatedly generate dependable measurements and act on them. An existing protein-engineering campaign provides a natural starting point because it already has objectives, assays and a baseline approach. A new campaign can also be considered, but the team must define what success means before interpreting a model’s suggestions.

The limiting resource may be assay capacity rather than candidate generation. If only a small number of variants can be measured, each experimental slot should contribute either a likely improvement or information that helps the next round. This is why a portfolio of candidates can be more useful than a list sorted by one predicted property. The proposed evaluation should make that trade-off explicit.

For broader AI discovery context, see the Insilico blueprint. For a company-led platform coupling computation with large experimental programs, compare the Recursion blueprint. Cradle’s offer is more directly a software layer for a team’s protein-engineering work. The comparison is about who operates the research loop and what the customer can license, rather than a claim about superior science.

03 / WorkflowA proposed campaign starts with a trustworthy assay

A proposed pilot might use an existing protein scaffold with two measured properties that the team wants to improve together. Define the acceptable ranges, assay protocols and reference controls. Preserve the relationship between each sequence, experimental batch and result. A model cannot reliably learn a sequence–property relationship if changes in the measurement process are mistaken for changes in the protein.

Import the approved historical results and inspect the platform’s data-quality feedback. Decide which results are comparable and which should remain separate. Cradle describes automated model adaptation and data-quality analysis, but the scientist should still resolve ambiguous units, missing measurements and failed assays. A failed expression result may contain useful information; it should not be silently discarded merely because it lacks a favorable numerical value.

Next, generate a candidate set under the project’s sequence constraints. Cradle describes controls such as blocked mutations and regional constraints, as well as multi-objective optimization. Review the proposed designs against the project’s scientific assumptions before ordering or constructing them. Retain an explicit reference set so that any apparent improvement can be interpreted against the same experimental conditions.

After laboratory testing, return the results and compare the observed profile with the prediction. Examine both successful candidates and informative failures. Define the next round using what was learned, and stop if assay instability makes the comparison unreliable. This is an editorial proposal for evaluating the workflow. Sequenced has not run the platform, generated proteins or verified the performance multipliers advertised on Cradle’s site.

04 / Commercial modelAn annual software license with project-based scope

The licensing page describes one annual software license without royalty or milestone fees. Standard licensing is based on the number of active projects, while Enterprise is intended for organizations deploying across multiple programs or divisions. The page does not publish a universal currency amount, so a reader needs a scoped commercial proposal rather than an assumed monthly seat price.

Cradle lists unlimited users, data uploads and storage, compute infrastructure, technical support and API access among included license features. It also states that customers own their sequence intellectual property and generated data and reports. These are useful commercial distinctions, but they do not mean the definition of an active project, service commitments or every special deployment requirement is irrelevant.

The Academic program is for nonprofit research with limited availability. The page also directs contract research organizations and service providers to discuss a separate partnership route when using Cradle as part of their service. A commercial provider should therefore not assume an ordinary research license automatically covers reselling platform-enabled work to its own customers.

05 / DistinctionsThe model adapts to the experiment rather than ending at generation

Cradle’s platform description emphasizes learning from successive experimental rounds and balancing multiple properties. The practical distinction is that measurements become inputs to the next design cycle. The team is not limited to repeatedly querying a general model with the same static assumptions, although the value of adaptation still depends on data quantity, consistency and relevance.

Its wet lab description explains that the company creates datasets and tests model iterations against earlier versions. This offers a concrete explanation for why a software business maintains experimental infrastructure. It should be read as Cradle’s account of its development process, not independent proof that a customer’s protein will improve at the same rate as a showcased campaign.

The platform also describes provenance tracking, project permissions and API access. Those features matter when computational specialists and laboratory scientists share responsibility for a program. A usable handoff preserves which sequence was proposed, which constraints produced it and which result came back. Otherwise, an increasingly sophisticated model can still be undermined by a simple sample-mapping error.

06 / LimitationsPrivacy and experimental validity need different evidence

The security documentation says customer data is used to train models specific to that customer’s protein-engineering projects, and that data and trained models are isolated between customers. This is a more useful description than saying the platform never trains on customer data: adapting to those measurements is part of the product. The important boundary is whose model learns and where it can be used.

That documentation also describes access controls and customer-data handling. A buyer should review the applicable agreement, export route and termination process for its own research records. Retaining ownership of an output does not by itself answer how the team will reproduce an earlier model run or continue a program after the commercial relationship ends.

On the scientific side, ask how the evaluation handles uncertainty outside the measured region of sequence space. Better predictions for familiar variants may not transfer to substantially different candidates. The public documentation link led to an authentication page during this review, so detailed customer-only operating limits were not independently inspected. Request the relevant model, data-format and project-limit documentation before making a deployment decision.

07 / DecisionEvaluate the complete round, not the generated sequence list

Cradle is most compelling to assess where a team has clear protein objectives and can run repeatable experiments. The pilot should cover data preparation, candidate generation, laboratory measurements and the next design decision. That provides evidence about whether the loop becomes more useful over time, while keeping the experimental budget and the interpretation of results under scientific control.