Iambic Therapeutics

The models are the discovery engine, not an analysis aid alongside it. Two named systems do the work: NeuralPLexer predicts protein ligand complex structures and their conformational changes, and Enchant is a multimodal transformer that learns from small noisy data sets to predict biochemical and cellular activity. The lead clinical asset was identified using the platform, which is the only proof that matters on this axis in drug discovery.

The platform generates molecular designs which are then synthesized and tested, with the company describing a high throughput experimental platform converting designs into biological insights each week. That closed loop of model proposal and wet lab verification is the oversight mechanism, and it is a strong one, but no disclosure was located on how candidate selection decisions are made between model output and synthesis, so the human decision points are inferred rather than documented.

Rare among AI drug discovery firms in submitting its core method to peer review rather than describing it in marketing terms. NeuralPLexer was published as a cover article in Nature Machine Intelligence, with reported benchmark performance on PoseBusters for the third generation, and the architectural approach is stated specifically as physics informed machine learning integrating physics principles into the AI architecture to improve data efficiency. A reviewer can read the method and evaluate the claim independently.

The clinical form of this axis does not reach the platform and one part of the business does involve human subjects, so the record separates them. The platform operates on molecular and preclinical data rather than patient records, which bounds the chain honestly: there is no protected health information for it to carry, and that is a fact about what the product does rather than a claim about its controls.

The clinical trial is a different matter, and it is governed by the trial protocol and the site review board rather than by any instrument this company publishes. That is the ordinary and correct arrangement, and it means a participant's protections come from the site rather than from the vendor, so the vendor's own disclosure has nothing to say about them either way.

What is absent is enumeration on the platform side: no hosting arrangement, no sub processor list, no retention position, and no statement of whether partner submitted targets, structures or assay results contribute to improving general models. That last question is the one that matters commercially in this category, because a partner's undisclosed targets are the most sensitive material they hold. Ask who hosts the platform, for a sub processor list, and for an explicit statement that partner submitted material does not train shared models.

The strongest possible evidence class in this category is a platform derived molecule with human data, and that exists here. IAM1363 moved from program start to clinical trial initiation in a reported two years, first patient dosed March 2024, with early Phase 1/1b data presented at the 2025 ESMO Congress reporting tumor reductions in heavily pretreated patients who had exhausted standard of care. Trial NCT06253871 is registered and independently checkable. Buyers should read this as validation of the discovery platform's speed and molecule quality, not yet as clinical proof of the asset, which remains early phase.

Not applicable in the provider sense and rated accordingly rather than penalized. The platform operates on molecular and preclinical data rather than patient records; the clinical trial does involve human subjects, governed by trial protocol and site IRB rather than by vendor PHI policy. No AI specific safety framework was located.

No HIPAA or BAA posture published, and the buyer relationship does not typically require one. Partners are pharmaceutical R&D organizations contracting for discovery collaboration rather than health systems handing over patient data.

Converted from Not Rated after a second search that reached the company's own legal pages. No SOC 2, ISO 27001 or equivalent attestation was located and no trust centre was found. The company is private, so the annual report route that supplies a cybersecurity governance disclosure for the listed vendors in this category does not exist here.

Two instruments are published and neither answers what a pharmaceutical counterparty is asking. The terms page is website boilerplate governing content posted to the site and disclaiming warranties on the site itself. The privacy policy is a genuine drafted instrument rather than a template, and within its own scope it is reasonable: it names a rights request process with an identity verification step, commits to confirming receipt within a stated period, provides an authorised agent route, and gives both an email address and a toll free number. Its subject is personal information of individuals under California law. It does not reach the discovery platform, partner target selections or compound structures, which is where the sensitive material sits.

Compute is a further consideration a buyer should raise directly. The company has publicly described running its training workloads on third party accelerated computing cloud infrastructure. Certifications held by those providers cover the infrastructure the models run on. They say nothing about this company's own systems, its access controls, or the partner supplied structures and target selections it processes. Ask where partner material is held and retained, and who inside the company can reach it.

Retained from the prior assessment because it remains the practical point: partners here are handing over proprietary target and compound information rather than protected health information, so diligence on this axis runs through the collaboration agreement rather than a public trust page.

Search hygiene note. A generic query pairing this company's name with certification terms returned an unrelated software company's trust centre carrying multiple certifications, alongside general explainer material about the attestation itself. Neither shares a name with this vendor. Confirm the domain resolves to iambic.ai before crediting any certificate.

Regulatory engagement is real but at the asset level rather than the platform level, which is the correct structure in drug discovery: the platform itself is not a regulated device. An IND for IAM1363 was accepted by FDA, the Phase 1/1b trial is registered on ClinicalTrials.gov as NCT06253871, and the study has expanded from United States sites into the EU. A second program, a dual CDK2/4 inhibitor, is described at discovery or IND enabling stage.

Upgraded from Not Rated after a second search directed at publications and released code. The prior note was right that a peer reviewed benchmark disclosure partially addressed the domain question. Retrieval shows it goes considerably further than partially, and this is now among the better governance records in the category.

The core structure prediction method is published as a peer reviewed cover article in a Nature portfolio machine intelligence journal, co authored with academic and hardware industry collaborators. Successor models are documented in public technical reports, and the note should be plain that those are preprints rather than peer reviewed work.

The strongest element is the released evaluation harness. The company publishes and maintains an open source benchmarking package in its own repository which evaluates its structure prediction model and other related models on two datasets. Releasing the instrument by which your own model can be checked, on the same footing as the models you compare yourself against, is the same behaviour this index credits elsewhere in the category, and very little of the field does it.

It also evaluates against an independently developed benchmark rather than an internal one, and reports physical validity separately from geometric accuracy. That separation is the substantive disclosure. A predicted structure can sit within the required distance of the experimental answer while being chemically impossible, so a single accuracy figure conceals exactly the failure that matters for downstream design. Reporting the physically valid rate, and reporting correct ligand stereochemistry as a further separate figure, exposes failure modes a headline number hides.

One of the evaluation datasets is a recent structures holdout, which is a temporal split. A temporal split is the standard way to test whether a model performs beyond the region where training data already exists, and that is precisely the domain bias question this axis asks in this category: whether the platform is skewed toward well explored target classes. Performance is also reported on a named difficult class, G protein coupled receptors, which the company itself describes as challenging existing predictors, with a median score across a held out set.

Held at B rather than A on the same basis as the other B grades in this lane. There is no formal governance framework, no model card, and no statement of intended and unsuitable use. The disclosure is a byproduct of scientific and engineering practice, and it lives in papers, preprints and a code repository rather than anywhere a buyer would look.

Disclosure quality is also uneven across the product lines, which a single grade hides. The structure prediction work is documented to the standard described above. The molecular property prediction model is presented very differently, with claims to be the leading model in the field on performance benchmarks that are not named, and a claim to outperform in vitro experiment for a pharmacokinetic endpoint in unspecified cases. Check that the evidence on this axis comes from the business a buyer is actually purchasing.

The core method was submitted to peer review and published as a cover article in a named journal rather than described in marketing terms, and performance is reported on a named public benchmark. The second half matters as much as the first and this index should say why.

A result on a shared public benchmark is comparable: any other group can run the same test, and a reader can place this method against alternatives on identical footing rather than taking a vendor's chosen metric on its own terms. That is a stronger form of falsifiability than publishing a number, because it removes the vendor's control over what was measured.

The architectural approach is stated specifically as physics informed machine learning integrating physics principles into the model architecture to improve data efficiency, which is a claim a reviewer can evaluate rather than an atmospheric one. Held below the top grade because nothing attaches commercially.

No warranty, indemnity, service level or remediation commitment was located, and benchmark performance for a published method is not a committed error characteristic for the system a partner actually uses, which are different artefacts once a method becomes a product. Ask what the deployed system achieves against the published benchmark, how often that is re measured, and what happens when a predicted structure is wrong.

Not applicable. This is a preclinical discovery platform with no provider workflow surface and no EHR touchpoint.

The commercial model is collaboration rather than software deployment. Partners access specific technologies under agreement, as with the reported grant of NeuralPLexer access to a large pharmaceutical partner, but no hosting, tenancy, or residency terms are published.

No price list exists and none would be meaningful, since the commercial surface is partnership structure rather than licensing. What is disclosed is unusually specific for a private company: a multi year collaboration with a major pharmaceutical partner carrying upfront, research cost, and technology access payments with success based milestones stated as potentially exceeding 1.7 billion dollars plus royalties on net sales. Deal shape is public even though economics per program are not.

Deliberately concentrated rather than broad. The disclosed pipeline is primarily oncology, with programs in HER2 driven cancers and cell cycle targets, plus a neurological indication described at discovery stage. The platform is claimed to work across multiple target classes and mechanisms, and the small molecule focus is explicit, so buyers outside small molecule oncology should treat applicability as unproven rather than assumed.