Nabla Bio

The model is the whole proposition and the company says so in its own headline, promising drug like antibodies straight from the computer. There is no screening library, no immunised animal and no natural repertoire to fall back on; JAM generates sequences that did not previously exist, and the wet lab downstream exists to test what the model produced rather than to produce candidates itself. Partners buy designs against targets they nominate. Remove the generative system and the laboratory has nothing to test.

The autonomy is high and the checkpoint on it is published, which is an unusual combination in this lane. Designs are generated computationally without experimental optimisation, so the model runs a long way unattended, but nothing counts until it survives the laboratory: designs are expressed and tested for binding to the native target, and the preprints report the measured pass rate at each iteration rather than describing the loop in the abstract.

The company states a design to experiment cycle of three to four weeks. That is a concrete account of where model output stops being a claim, and it is why this grade sits above the lane. What is still undocumented is the human layer inside it, specifically who selects which of hundreds of generated designs are expressed and on what basis, and whether any design has ever gone to a partner without laboratory confirmation.

The most current and most specific disclosure in this lane, sitting at the top of this grade rather than the bottom. The system is named and versioned, JAM and now JAM-2, with an openly published technical report for the successor. The architecture is described at module level, a Generator that completes partially specified protein complexes and a Filter that scores experimental viability, with output resampled back through the Generator across iterative rounds.

Test time scaling and in context reprompting are named as the specific techniques behind the reported improvement, which is a mechanism claim precise enough to argue with. Neighbours in this lane document the lineage of their methods and withhold the current system; this vendor documents the current system. Still withheld: training data composition, model scale, the weights themselves, and any external evaluation harness, and the technical report is self published rather than certified.

The system is presented as built in house, and the preprints position JAM against other named design systems rather than as a layer on top of one, so the implied supply chain is short and the claim is at least specific enough to be contested by people in the field. Beyond that nothing is disclosed. No compute provider, no training infrastructure, no open source component inventory, no third party model licence and no dataset provenance for what the Generator and Filter were trained on. The only infrastructure fact visible anywhere is incidental rather than disclosed: the technical report is served from Amazon storage in a United States region.

The best non certified evidence base in this lane, and the qualifier is doing real work. Two detailed preprints publish methods, affinities, developability metrics and, importantly, bind rates per design round rather than only the successes, which lets a reader see the failure rate as well as the result.

The January 2025 work reported the first fully computationally designed antibodies against multipass membrane proteins at double digit nanomolar affinity; the May 2025 work reported picomolar to low nanomolar binders against CXCR4 and CXCR7 with agonist function. Independent scientists have assessed the work in review articles and described it as significant progress. Three large pharmaceutical partners have paid, and one has renewed at greater scale.

What holds this below the top grade is that neither preprint has been certified by peer review, every contributor holds equity in the company, and no molecule has entered a human. Publication of the JAM results in a peer reviewed venue would move this grade, and a review article discussing the same preprint is commentary rather than an independent check.

No protected health information is in scope. Two questions do apply and neither is addressed on any published surface as of 29 August 2026. The first is partner separation: three competing pharmaceutical companies run programs here at once, the platform improves from experimental data generated in the company's own laboratory, and nothing states what a partner's target information is permitted to inform or how it is walled off. The second is more particular to this vendor than to the lane.

Nabla designs novel binders to human receptors and has reported designing agonists, molecules that switch a receptor on rather than off, which it describes as the first of their kind. De novo protein design is a field with an active biosecurity conversation about screening generated sequences, and no statement of screening practice, dual use policy or release control appears anywhere in the company's material. Recording that as a located absence rather than as an allegation: nothing suggests a problem, and nothing addresses the question either.

Graded neutrally because the obligation does not arise. The platform operates on protein structures, sequences and laboratory assay measurements. No protected health information enters it, there is no covered entity relationship, and no business associate agreement is required or published. The convention for discovery vendors applies here cleanly, without the strain it carries at neighbours in this lane that work from patient derived material.

Two dedicated passes on 29 August 2026 returned nothing. A targeted domain search surfaced only recruiting pages and business databases, and a direct review of all five navigation sections plus the footer found a single legal document, a privacy policy with no date that scopes itself in its first paragraph to website visitors. Its security section is the standard acknowledgement that no data security measure is guaranteed.

There is no trust centre, no certification, no attestation, no penetration testing statement and no incident disclosure policy. Three of the world's largest pharmaceutical companies entrust target information to this company and nothing published describes how it is held.

No product falls under a clearance pathway. The company designs preclinical biologics, which are neither medical devices nor authorised products at this stage, and any regulatory obligation would attach to the partner advancing a molecule rather than to Nabla. Two dedicated passes on 29 August 2026 located no warning letter, enforcement action or adverse standing. As elsewhere on this axis, the grade records an absent obligation rather than an absent disclosure.

No governance framework, model card, responsible use policy or evaluation charter appears on any of the five sections of the company website as of 29 August 2026. What partially stands in its place is method disclosure with the failures left in: the preprints report bind rates round by round, so the reader sees how often the system misses, and both carry an explicit declaration that all contributors are employees or former employees holding equity. Publishing your own failure rate and declaring your own conflict is not a governance framework, but it is the behaviour a governance framework is meant to produce, and most of this lane does neither.

No terms of service exist on any public surface as of 29 August 2026. The privacy policy is the only agreement offered and it excludes everything that matters, stating explicitly that it applies to the website alone. There is no warranty position, no limitation of liability, no indemnity, no intellectual property allocation for designed molecules and no service level.

That last omission is the notable one for this vendor: the ownership of a computationally generated sequence, and who carries the risk if a design infringes or fails, is the central legal question of de novo design, and it is settled entirely inside private collaboration agreements that nobody outside can see.

No electronic health record surface exists and none would be appropriate for a preclinical design company, so this is graded neutrally under the convention for this lane. Recorded in its place: there is no third party software integration either, no application programming interface, no connector and no partnership with a design suite vendor. The platform is not offered as software at all. Partners nominate a target and receive designed molecules, so there is nothing on offer for a buyer to integrate into an existing stack.

Nothing is published and, unusually, the axis may be close to genuinely inapplicable here. The company licenses no software and grants no platform access; a partner supplies a target and receives molecules, so there is no tenant, no instance and no customer environment to describe. The one legal document on the site states in its opening line that it applies only to the website, and the only named processor anywhere is Google Analytics.

No hosting, residency, segregation or retention statement exists for partner target information. Graded neutrally under the convention for this lane, which is now the fourth consecutive record where that convention rather than the evidence is setting this grade.

Partners are named rather than described as a top ten pharmaceutical company, the deals are dated, and the aggregate figure is stated: more than $550 million across the AstraZeneca, Bristol Myers Squibb and Takeda collaborations announced in May 2024, with the second Takeda agreement in October 2025 disclosed as double digit millions upfront against roughly $1 billion in potential value.

Funding is itemised with named investors across two rounds to a total near $37 million, and the company states plainly that it is revenue generating, which most companies at this stage avoid saying either way. Held below the top grade by three things: no individual deal value is separated from the aggregate, the totals are ceilings inclusive of milestones rather than money received, and the founding year is inconsistent between company material and business databases.

Deliberately narrow and pointed at the hard part of the problem. The work is biologics design at the discovery stage, in single domain and paired antibody formats, extended in the Takeda work to multispecifics, receptor decoys and other custom biologics.

The target class is the differentiator rather than the breadth: multipass membrane proteins including G protein coupled receptors, ion channels and transporters, which are validated as disease relevant but resist conventional discovery because they cannot easily be presented outside a membrane. Nothing here covers small molecules, target identification, translational work or clinical development, and the company makes no claim to.