Iambic Therapeutics

The models are the engine of the whole operation. Iambic's platform is built around two named model families: Enchant, a multimodal transformer that predicts preclinical and clinical endpoints for a candidate compound profile, and NeuralPLexer, a flow matching model for protein ligand complex structure prediction.

The company describes the platform as the interplay between Enchant guiding which compounds to make and which experiments to run, and automated chemistry and biology producing hundreds of new compounds a week whose results refine the models. The operating scale is stated: model fine tuning each week on hundreds of molecular properties and roughly 12 million inferences a month across its own and partnered programs. Remove the models and the automation has nothing directing it.

The oversight structure is described and the threshold is missing. Two mechanisms are published. Every Enchant prediction is said to carry uncertainty quantification, which the company uses to decide which laboratory experiments to run and how to rank compounds against a target profile, so a low confidence prediction is routed to an experiment rather than acted on. And the platform's cycle puts experiment after model as a matter of course: hundreds of compounds are made each week and tested across physicochemical, biological and metabolic assays, with the results refining the models. The company describes its approach as placing the AI in the hands of experienced drug hunters, who make the program decisions.

What is not published is where the line sits: no confidence value at which a prediction is not relied on, no hit rate for model proposals, and no statement of which decisions are taken on model output alone.

The models are identified, the training data characterized and the versions published with dates. Enchant is described generation by generation: v1 at 1 billion parameters in 2024, v2 at 7 billion in 2025, and v3 in September 2026 at 41 billion parameters, pretrained on 4.5 trillion tokens across 16 biomedical modalities covering more than 6,000 molecular properties, with a mixture of experts architecture and new capabilities in inverse design, omics and image based prediction and peptides. NeuralPLexer3, the structure prediction model, was published as a NeurIPS 2025 paper on flow model based biomolecular complex structure prediction, and a companion paper on scaling multimodal and multitask transformers for small molecule discovery is listed with it. Every prediction is described as carrying uncertainty quantification.

What a reader cannot check is the evidence for the accuracy claims: the scaling results the company reports across preclinical and clinical endpoints appear in figures on its report page rather than as figures in the text, and no benchmark table or held out test result is stated in words.

A structurally short chain: the models are the company's own and are named. Enchant and NeuralPLexer are built and run by Iambic, with their architectures, scale and training modalities described and the structure model published at a peer reviewed venue, and the laboratory data that refines them is generated in the company's own automated facilities. No outside model provider is named as sitting between a partner's targets and a prediction, and the platform is not delivered to partners, so the transfer runs inward rather than out.

The party list is not enumerated: no hosting or compute provider, no subprocessor list, and no published position on whether a partner's compounds and assay results train models used on other programs, which for a company running its own pipeline alongside partnered ones is the question a partner should settle in the contract.

A candidate designed on the platform is in a named clinical study with dated results, and the platform's methods are peer reviewed. IAM1363, a selective brain penetrant HER2 inhibitor discovered on the platform, is in a Phase 1/1b study in patients with advanced HER2 altered cancers, which the company reports has shown anti-tumor activity and tolerability, with the study presented and listed in its publications on 19 October 2025; two further programs, IAM217 against KIF18A and IAM-C1 against CDK2/4, are in IND enabling work. The methods behind them are published: NeuralPLexer3 at NeurIPS 2025 on flow model structure prediction, and a paper on scaling multimodal and multitask transformers for small molecule drug discovery in December 2025.

The company's claim that it reached a clinical candidate in roughly two years, against an industry standard it describes as years slower, is its own, and no baseline comparison is published. The company has published no efficacy readout with numbers.

Nothing published reaches the data the platform runs on. The only privacy document is a website privacy policy, which by its own terms applies to the company's online activities and to visitors to iambic.ai, covering technical information about a visitor's device, log files and form submissions, with website analytics data unlikely to be kept in identifiable form beyond a year. It says nothing about the material a partner places into the platform.

That material is the question here: Iambic runs collaborations with pharmaceutical partners, so partner targets, compound structures and assay results enter the models, and Enchant is described as continuously learning from new laboratory data. Nothing states whether partner data trains models used for other programs, what is retained, or what happens at the end of a collaboration; those terms sit in private agreements. De-identification does not arise for molecular data, and safety engineering is not addressed.

No statement of status and no privacy document that reaches the product. The website privacy policy applies to visitors to iambic.ai and to the company's online activities; it does not reach the discovery platform or the collaborations that are the business.

No business associate status is stated anywhere, and no scope position explains whether the company can receive protected health information, which is a live question for a company running its own clinical trials and training models on clinical trial data, a modality it names among Enchant's sixteen. Silence is not a stated position.

Nothing is published about security controls. The site carries no SOC 2 report, ISO 27001 certificate, penetration testing cadence, security page or trust center, and the website privacy policy describes cookie and log file practice rather than how the platform is protected.

That gap sits against a business whose partners place undisclosed targets and proprietary chemistry into the platform under multi year agreements.

No device claim is made and the platform is scoped accordingly. Enchant and NeuralPLexer are discovery technology used inside the company and with partners, not products offered as regulated software, and nothing presents them as devices. The regulatory activity belongs to the medicines: IAM1363 is in a Phase 1/1b study in advanced HER2 altered cancers, and IAM217 and IAM-C1 are in IND enabling work, positions Iambic holds as sponsor. No product is approved. The company's marketing language, molecular superintelligence, describes its platform ambition rather than a claim about a regulated capability.

Model behavior is described in general terms with no evaluation a reader can check. The company reports that prediction accuracy improves predictably with model scale across a diverse range of preclinical and clinical endpoints, and says the gain holds for endpoints where data is sparse, which is the regime where a model of this kind is most at risk of failing. That is the right question to raise, and the evidence for it is presented in figures on the report page rather than as results with groups and counts in the text.

No governance framework, model card, intended use statement or evaluation method is published, and no breakdown is given of where prediction is weakest by endpoint, modality or chemical series.

Mechanisms let a wrong prediction be caught, with nothing published standing behind one. Uncertainty quantification accompanies every prediction and drives which experiments run; predictions are tested in the company's own laboratories on a weekly cycle; and the models behind them are described in published papers a reader can examine. For the company's own programs the chain continues into IND enabling work and clinical trials under regulatory oversight.

No error rate is published with a method and denominator: the accuracy and scaling results the company reports sit in figures rather than in stated numbers, and no held out benchmark is given in words. No warranty, indemnity or remediation commitment is published for partners, whose agreements are private, and the company sells collaborations rather than software, so there is no customer facing service commitment to read.

No clinical record integration exists: Iambic discovers its own medicines and runs partnered programs on its own platform, and no clinical record system is involved. For a pharma partner the record systems that matter are its research systems: electronic lab notebooks, laboratory information systems, research data platforms and cloud data environments.

None is named and no integration is claimed; the data the models consume comes from Iambic's own automated laboratories and from public and partner datasets moved under a collaboration rather than through a connection between systems. For a discovery partnership that is an absence rather than a defect. How compound designs, assay results and predictions move between the two organizations is not described.

An internal platform, with location implied rather than committed. The models are not delivered to anyone: partners bring programs to Iambic, whose automated chemistry and biology laboratories and computation run in its own operation, based in San Diego.

Nothing states where the models are trained or served, whether outside cloud or compute providers are used for a system at the scale described, where partner material rests, or how one partner's compounds and results are separated from another's and from the company's own programs.

No price is published, and the commercial structure is discoverable. Iambic does not license the platform: it develops its own pipeline and enters multi year discovery collaborations, with agreements announced with AbbVie in September 2026 and Bayer in June 2026, and a partners page describing what a collaboration involves. The shape of the terms is public, combining upfront payments, research cost coverage, technology access fees, milestones and royalties, with one reported agreement carrying success based payments the company describes as potentially exceeding $1.7 billion plus royalties on net sales.

That ceiling is the full milestone stack across a multi program agreement rather than a contract value, and it is what one partner negotiated rather than a price a new partner can apply; no rate card, unit of charge or per program economics are published.

Coverage is named and validation sits behind part of it. The company states that its own and partnered pipelines cover oncology, neuroscience, and inflammation and immunology, and says it does not consider itself limited to those areas; Enchant v3 adds stated capabilities in peptides, omics based and image based prediction and inverse design, extending the modality range beyond small molecules.

Validation is concentrated in oncology, where the programs are: a HER2 inhibitor in a Phase 1/1b study, a KIF18A inhibitor and a CDK2/4 inhibitor in IND enabling work, each described with its target and indication. The neuroscience and immunology programs are named without published results, and the new modalities are capabilities rather than demonstrated coverage.