Generate:Biomedicines

The generative model is the origin of the molecules rather than a filter applied to them. Chroma samples novel protein structures and sequences directly and can be conditioned toward desired properties, which the company uses to design therapeutics that do not begin from a natural antibody scaffold, and the lead clinical asset GB-0895 is stated to have been developed using the platform's binding affinity and developability optimisation modules.

The extensive biohardware, covering DNA assembly, protein production, multiplexed assays and an in house cryogenic electron microscopy core, exists to measure candidates at scale and feed structural data back into the next cycle of model learning. The dependency runs the right way for this grade: the laboratory serves the model.

Oversight is empirical and industrial in scale. The platform is described as a continuous generate, build, measure and learn loop in which billions of molecules are measured through biohardware and structural data is generated to correct the next round of model hypotheses, so designs meet physical evidence continuously rather than accumulating unchecked, with the cryo-EM core producing over 500 high resolution maps in 2025 as one concrete measure of that verification capacity.

Held at B because no confidence thresholds, published failure rates or documented human approval gates between model output and programme commitment were located, so the check is architectural rather than specified.

The strongest transparency posture in this category and the direct counter example to the closed end of it. Chroma was published in Nature in November 2023 with a named author list, and the company released the code as open source software with model weights made freely accessible to academic researchers and non profit entities, stating the goal of accelerating progress in generative biology.

That means outside researchers can run the model, probe where it fails and contest the claims, which is the condition this index treats as the difference between a demonstrated capability and an asserted one. Platform architecture beyond Chroma is also described concretely, as three modular capabilities covering programmable binding, programmable function, and programmable composition and developability, with the specific modules used for the lead asset identified.

The model layer here is not merely named but partly obtainable, which is the strongest possible answer to this axis and it is worth stating why. Code is open source and weights are freely accessible to academic researchers and non profit entities, so for that audience there is no chain to enumerate at all: they hold the artefact. For a commercial partner the position is weaker, since the release is scoped and they are back to the publication plus whatever the platform provides.

The platform itself is described in modules rather than as one system, covering programmable binding, function, and composition and developability, so a partner knows what stages their material passes through. The clinical form of this axis does not reach the discovery workflow, which operates on protein structures, sequences and assay and structural data generated in house with no patient records involved, and that is a fact about the product rather than a claim about its controls.

Human subject protection in the ongoing trials is governed by trial protocol and site ethics oversight rather than by anything this company publishes, which is the correct arrangement and means a participant's protections come from the site. What is unstated is the partner side: no hosting arrangement, no sub processor list, and no position on whether partner supplied sequences or assay results improve models used for others. Ask that, since it is the commercially sensitive question in this lane.

The most advanced clinical position in this category alongside Insilico Medicine, and the evidence is checkable through securities filings rather than press releases alone. GB-0895, a long acting anti TSLP monoclonal antibody engineered for potential every six month dosing, entered global Phase 3 studies for severe asthma with the first Phase 3 patient dosed on 26 January 2026, and is separately in Phase 1b for COPD; GB-4362 received FDA Fast Track designation on 23 January 2026.

Partner validation comes from collaborations with Amgen and Novartis. The company also discloses operating specifics most peers do not, including 312 employees as of 31 December 2025 of whom 138 hold MDs or PhDs. Buyers should hold two caveats: no approval exists and Phase 3 entry is not efficacy, and the platform economics claims, reaching proof of concept in three to five years at roughly 25 to 60 million dollars against traditional benchmarks of six to eight years and roughly 380 million, are company generated projections rather than realized results.

Not applicable in the provider sense and rated accordingly rather than penalized. The platform operates on protein structures, sequences and assay and structural data generated in house, with no patient records in the discovery workflow. Human subject protection in the ongoing Phase 3 and Phase 1b studies is governed by trial protocol and site institutional review board oversight rather than by vendor PHI policy.

Not applicable. The company is a clinical stage therapeutics developer and research collaborator rather than a business associate of covered entities handling protected health information.

No attestation, trust centre or security page was located. The company listed publicly in February 2026, so the filing route applies, but only partly, and the reason is worth understanding because it recurs with any recent listing.

What exists is the registration statement, and its risk factors are substantive. The company enumerates the threat categories it considers, including ransomware, malware, denial of service and social engineering, notes that an increased amount of remote working raises its exposure and creates data accessibility concerns, and states plainly that it may not be able to anticipate every type of threat or implement measures effective against all of them. Most notably it discloses that it has experienced cyber attacks, security breaches and incidents and other system failures, while stating that to its knowledge none has caused a material interruption or incident. Disclosing that incidents have occurred is candour, most companies say nothing, and it is credited here.

What does not yet exist is the part that matters more for this axis. The annual cybersecurity disclosure requiring a registrant to describe its processes for assessing, identifying and managing cybersecurity risk, together with board oversight and management accountability, attaches to the annual report rather than to a registration statement. Having listed in February 2026, this company's first annual report is not due until early 2027, so no such description has been filed.

So the position is the warning half without the processes half. A risk factor is written to protect an issuer from shareholder claims; it is not an assurance to a customer and it evidences no controls. Graded C on that basis. Recheck when the first annual report is filed, at which point this becomes assessable on the same terms as the established listed companies in this category.

The strongest regulatory standing in this category alongside Insilico Medicine, and correctly structured, with the platform not presented as a regulated device and regulatory activity at the asset level. GB-0895 is in global Phase 3 for severe asthma with dosing underway and in Phase 1b for COPD, and GB-4362 received FDA Fast Track designation on 23 January 2026. Buyers should note the standard qualifications that Fast Track is a development and review process mechanism rather than an approval or a finding of efficacy, and that no product has been approved.

No governance framework, model card, intended use statement or release policy was located, and that absence now reads differently than it did, because the risk it leaves open has since been demonstrated rather than merely postulated.

Two domain relevant questions apply here. The first is coverage: models trained on the public structural archive inherit that archive's composition, so performance will be uneven across fold classes and protein families in proportion to how well each is represented, and no breakdown is published.

The second is dual use, and it is the serious one. This company released a general purpose de novo protein design model openly, with weights available, which is a genuine contribution to open science and is credited on the transparency axis. The model is named in the biosecurity literature as an exemplar of exactly the tool class that raises these concerns. In October 2025 a study published in Science reported a two year confidential project in which open source protein design software was used to generate variants of proteins of concern that evaded the screening systems nucleic acid synthesis providers rely on, with large numbers of synthetic variants going undetected, after which patches were developed and deployed across those providers globally.

That does not implicate this company's model specifically, and it should not be read as doing so. What it establishes is that the category risk is real, measurable and was serious enough to require coordinated repair of shared screening infrastructure. Against that backdrop, an open weight release with no accompanying access policy, no disclosed pre release hazard evaluation and no red teaming statement is a governance gap rather than a theoretical concern. Peers in this category do publish controls of this kind, including purpose disclosure requirements and licence terms restricting use, so it is achievable rather than unusual.

Peer review at a leading journal is paired here with a released artefact, and the combination is what separates a demonstrated capability from an asserted one. The generative model was published with a named author list, and the company released the code as open source with model weights made freely accessible to academic researchers and non profit entities, stating the goal of accelerating progress in the field.

Releasing weights means outside researchers can run the model, probe where it fails and publish disagreement, and a claim that survives that treatment has been tested in a way no vendor benchmark can match, because the vendor no longer controls what gets measured or reported.

The platform architecture beyond that model is also described concretely as three modular capabilities covering programmable binding, programmable function, and programmable composition and developability, with the specific modules used for the lead asset identified, so a reader can tell which component did what on a real programme.

Held below the top grade because nothing attaches commercially: no warranty, indemnity, service level or remediation commitment was located, and because the released weights are scoped to academic and non profit use, so a commercial partner evaluating the platform cannot obtain them and is back to relying on the publication. Ask what performance the commercial configuration achieves against the published one, and how the modules are validated together.

Not applicable. This is a discovery and clinical development organization with no provider workflow surface and no EHR touchpoint.

Not applicable in the software sense. The Generate Platform is operated internally and coupled to the company's own biohardware and cryo-EM infrastructure rather than deployed to customers, and the commercial relationship is collaboration rather than licensing, so tenancy and residency questions do not arise for a counterparty. The exception is Chroma itself, which is distributed publicly as open source with weights available to academic and non profit users, meaning that component runs wherever the user chooses to run it.

Public listing forces the disclosure and the company met it in detail. An S-1 was filed on 4 February 2026 ahead of a Nasdaq debut on 27 February 2026 under the ticker GENB, priced at 16 dollars per share for 25 million shares raising roughly 400 million dollars gross, with a 30 day overallotment option for up to 3,750,000 additional shares that would take gross proceeds to roughly 460 million.

Filings disclose programme by programme status, headcount and composition, platform throughput figures and stated use of proceeds covering the Phase 3 programme, other pipeline assets and platform investment. The unflattering half is also on the record: shares opened at 15 dollars, below the offering price. No rate card exists and none would apply, since the commercial surface is the pipeline plus collaborations with Amgen and Novartis.

Broad within one modality class and absent outside it. The platform addresses protein therapeutics across antibodies, bispecifics, cytokines and entirely new protein modalities that do not derive from natural scaffolds, with a disclosed pipeline spanning respiratory disease, oncology and immunology. Small molecules are out of scope entirely, which is the real boundary. Buyers should note the concentration risk that the company's clinical credibility currently rests heavily on a single asset, GB-0895, in severe asthma and COPD.