Immunai

The models are what is sold. Immunai does not develop drugs and does not supply software a sponsor installs; it takes clinical and preclinical samples, generates single cell multiomic data from them, and returns a recommendation about a target, a candidate, a patient population or a trial arm. Every stage between the sample and the recommendation is computational.

The published workflow makes the dependence explicit. Generated data is enriched against AMICA, an atlas of more than 300,000 patient samples across tens of millions of cells; the ImmunoDynamics Engine computes immune features linking mechanism to treatment response; hypotheses are validated in the laboratory and then distilled into a scorecard. Remove the models and what remains is a sequencing service producing data a sponsor would have to interpret itself, which is precisely the work the company exists to do.

The oversight structure is described with mechanisms rather than asserted. The workflow published on the platform page has five named stages and two of them are checks: a validation stage in which machine learning hypotheses are tested in the laboratory using functional genomics in disease relevant model systems before they become recommendations, and a recommendation stage delivered as a scorecard with the underlying data shown. The company states the position directly, that its engine is not a black box and that it provides the data behind its recommendations. Immunai advises a sponsor's decision rather than making it, and the sponsor remains the decision maker on trial design, patient stratification and candidate progression.

A second, unusual control appears in the privacy notice: automated decision making is disclosed as applying to personal data provided for research, and a route to opt out of it is named for individuals with that right.

What is missing is the threshold. Nothing published states the confidence at which a recommendation is withheld, what happens when laboratory validation contradicts a model hypothesis, which judgments require a person on Immunai's side, or what the company does when a recommendation proves wrong after a sponsor has acted on it.

The data underneath the approach is characterized in detail, and the approach itself has a name. AMICA, the Annotated Multiomic Immune Cell Atlas, is the data foundation, and its contents are published as counts rather than adjectives: 5,000 to 6,000 studies, more than 350 single cell studies, more than 800 cell types including 80 immune cell types, more than 500 diseases and more than 300,000 patient samples across tens of millions of cells, drawn from internal experiments, academic partnerships and public sources. The measurement types are named: whole transcriptome RNA, surface proteome, T cell and B cell receptor repertoire and spatial transcriptomics, from clinical blood and tumor or tissue samples and from in vitro and ex vivo work. The ImmunoDynamics Engine is named as the analytical framework that identifies and prioritizes immune activation features against a stated biological question, and twenty-six dated peer reviewed publications put the methods in front of reviewers.

What is not published is what produces the output at the model level. No architecture, model family, parameter count or benchmark is given for the ImmunoDynamics Engine, the foundation models referred to in the operating system layer are not identified, and no version identifier or update practice appears anywhere, so nothing tells a reader whether the engine running a 2026 engagement is the one behind a 2024 publication.

Immunai publishes a named subprocessor list, with each party's purpose and the country it operates in. Google Cloud Platform for data storage in the United States, Google Workspace in the United States and the European Union, Benchling as the lab and research data platform in the United States, Hetzner for storage in Germany, GENEWIZ/Azenta for genomics services in the United States, Spin.AI for backup, AWS for data transfer and storage, Cookiebot in Denmark, and two administrative tools. For a company whose inputs are human blood and tissue samples with clinical annotation, that list is the answer to who else touches the data.

The model layer is Immunai's own and is stated as such: the ImmunoDynamics Engine and the AMICA atlas are built in house, so no third party model provider sits between a patient sample and a recommendation, and none appears on the list. The terms governing data once it arrives are published alongside it: the company states it acts as a processor under its partners' instructions for partner data, names consent as the legal basis where it collects research data directly, states plainly that such data is used to train its AI platform, and documents the cross border path through the European adequacy decision for Israel, an intra-group transfer agreement and standard contractual clauses.

One element is referenced rather than published. Retention is governed by an internal policy the notice describes but does not reproduce, so a reader is told that data is not kept perpetually without being told for how long.

The evidence here is named deployments and published research, both dated, and there is a lot of both.

Twenty-six peer reviewed publications are listed with dates, and several analyze named clinical programs rather than laboratory models: long term remission after cilta-cel in multiple myeloma linked to T cell diversity and low myeloid suppression (December 2025), an anti-TIGIT antibody improving PD-L1 blockade through myeloid and regulatory T cells (February 2024), immunomodulatory differences between cisplatin and carboplatin based chemotherapy with atezolizumab in urothelial cancer (February 2024), first line volrustomig data in advanced clear cell renal cell carcinoma presented at a major oncology congress (October 2023), and a methods paper benchmarking bulk RNA sequencing deconvolution approaches (August 2023).

The commercial record is the second line and is unusually strong of its type. AstraZeneca began with clinical trial optimization in September 2024, agreed in October 2025 to advance an inflammatory bowel disease target Immunai had identified, and expanded the oncology collaboration again in May 2026. Teva signed in November 2024 for oncology and immunology clinical trials, and a vitiligo mechanism and biomarker program for an IL-15 targeted antibody is featured on the company's own site. Boehringer Ingelheim launched a T cell target discovery collaboration in June 2026. A partnership with the Parker Institute for Cancer Immunotherapy, announced in April 2025, assembled a real world checkpoint inhibitor dataset from 3,700 blood samples across 1,070 patients. A sophisticated buyer returning three times, and paying for a target the platform found, is harder to manufacture than a case study.

The weakness is that none of it measures the platform. No published evaluation gives the accuracy of target identification, patient stratification or response prediction with a method and a denominator, the publications report biology discovered with the platform rather than how often it is right, and no molecule of Immunai's own is in a clinical study.

Here the question actually bites, because the platform is built from human blood and tissue samples with clinical annotation rather than from molecules, and Immunai answers a good part of it.

The privacy notice reaches the research data rather than stopping at the website. It states the categories held: a pseudonymized patient identifier, demographic information such as gender, height and weight, and biological data including blood and tissue samples, disease and treatment information, and genetic data. It separates two roles: for partner and customer data Immunai acts as a processor under the controller's instructions, and for data it collects directly it acts as controller. Training use is stated outright rather than buried, with the purpose given as performing scientific medical research and training the company's AI platform, and the legal basis named as consent where the data is collected directly from individuals. Automated decision making is disclosed as applying to personal data provided for research, with an opt out route named. Anonymization and de-identification are described, with the caveat that anonymous information may then be used and disclosed without restriction. A data protection officer and a data security officer are named with contact details, and breach notification is committed to.

What is missing is the retention schedule. The notice says data is kept until no longer needed and points to an internal retention policy it does not publish, so the window cannot be checked. One question the platform's own design raises is answered only indirectly: the atlas is described as built from internal experiments, academic partnerships and public data, which does not name commercial partner engagements, and nothing states explicitly whether work done for one sponsor enriches the atlas that serves another. Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not published.

A substantive privacy document reaches the data, and no status is stated anywhere. The notice is detailed and current: it names the categories of research data held, including pseudonymized patient identifiers, demographics and biological data such as blood and tissue samples and genetic data; it distinguishes processing as a partner's processor from processing as controller; it names legal bases under the European regime, the Israeli privacy law and the Swiss regime; it names a data protection officer and a data security officer; and it lists subprocessors with locations.

What it never mentions is the United States health privacy rule. Business associate status is neither claimed nor disclaimed, protected health information is not referenced, and no statement says whether Immunai considers itself in scope. The company is headquartered in New York and handles clinical samples from United States sites, so it cannot sit outside the rule as a non United States company, and with no scope position taken, nothing else places it outside either. A reader gets a careful European privacy posture and no answer to the American question.

Controls are described and nothing independent stands behind them. The privacy notice states that Immunai has implemented technical, organizational and security measures including encryption, role based access and documented security policies and processes, names a data security officer with contact details, and commits to breach containment, assessment and notification to regulators and affected individuals.

What is absent is the outside check. No SOC 2 report, ISO/IEC 27001 certificate or HITRUST certification is published, there is no trust center or security page, and no independent penetration testing on a stated cadence is described. For a company holding human blood and tissue samples, genetic data and clinically annotated patient records from more than 300,000 samples on behalf of several large pharmaceutical sponsors, an attestation is the first assurance a partner's security review would ask for, and none is published.

The product makes no device claim and is scoped accordingly. Immunai sells analysis to drug developers: its outputs are target hypotheses, mechanism of action insight, biomarker candidates and patient stratification recommendations delivered to a sponsor, not a diagnostic result returned to a clinician about a named patient in care. The sponsor remains responsible for trial design and for any regulatory submission built on the work, and Immunai develops no therapeutic of its own.

The positioning is the part to watch. Immunai's own language describes actionable clinical decision support and the prioritization of trial arms and patient subgroups, which is close to the boundary where an analysis that selects patients for a treatment starts to attract regulatory attention. Nothing published states how Immunai scopes itself against that line, and no clearance, submission or pathway is named anywhere. Where the company's work has reached a regulator's field of view it has done so inside a sponsor's program, as with the first line volrustomig data in advanced renal cell carcinoma presented in October 2023, which is AstraZeneca's regulatory position rather than Immunai's.

A governance framework with named process sits behind the data the models learn from. Immunai describes the atlas as governed by harmonized metadata and ontologies, stringent quality control, cell type annotation, normalization and review by domain experts, and publishes what the resulting foundation contains: 5,000 to 6,000 studies, more than 350 single cell studies, more than 800 cell types including 80 immune cell types, more than 500 diseases and more than 300,000 patient samples across tens of millions of cells. The company also publishes its use of automated decision making on research data and names the route to opt out of it, which is a governance disclosure rather than a marketing one. Twenty-six peer reviewed publications, dated and listed, put the platform's outputs in front of reviewers, and one of them is a benchmarking paper on deconvolution methods rather than a biological result.

Evaluation is the gap, and for this platform it is pointed. A model built from 300,000 patient samples to stratify patients and choose trial arms has an obvious subgroup question, and no performance results are published by sex, ancestry, age, disease or treatment line. No model card, no accuracy figure for target identification or patient stratification, and no independent audit of model behavior appear anywhere. The scorecard the platform delivers to a sponsor shows the data behind each recommendation, which is traceability rather than evaluation.

Mechanisms exist that let a recommendation be challenged, and Immunai names them. Every recommendation is delivered as a scorecard with the underlying data shown, and the company states its engine is not a black box and that it provides the data behind its conclusions, which is source traceability a sponsor's scientists can interrogate. Machine learning hypotheses pass through laboratory validation using functional genomics before they become recommendations. Individuals whose research data is processed are told that automated decision making applies and are given a route to opt out where the law gives them one.

Nothing stands behind the output itself. No published accuracy figure exists for target identification, patient stratification or response prediction, so there is no error rate with a method and a denominator to hold the company to. No guarantee, indemnity or remediation obligation is published for a sponsor that designs a trial arm around a recommendation that proves wrong, and the opt out right runs from data protection law rather than from anything this company has undertaken. For a platform whose output shapes which patients receive which drug in a clinical trial, that is the disclosure a reader would most want and it is absent.

Clinical data does reach this platform: Immunai works from clinical blood and tumor or tissue samples with disease and treatment annotation, and the atlas holds clinically annotated cohorts. So the question is real rather than hypothetical, and the answer is that no record system is named on either side.

No electronic health record, clinical trial management system, electronic data capture platform or sponsor data environment is named as connected, and no interoperability standard is claimed. The delivery model is samples in and analysis out: specimens and accompanying clinical annotation arrive under partnership specific arrangements, which is an integration described by contract rather than by connector. One research system of record is named, Benchling, and it appears in Immunai's own subprocessor list as the lab and research data platform it uses internally rather than as a connection a sponsor could use.

Nothing published describes how clinical annotation moves from a sponsor's own systems into Immunai's, or how results come back.

Residency is stated outright rather than implied. The subprocessor table names where each party operates: data storage on Google Cloud Platform in the United States and on Hetzner in Germany, AWS for transfer and storage in the United States, Google Workspace across the United States and the European Union, Benchling in the United States for lab and research data, and genomics services in the United States. The cross border path is documented too: transfers from the European Union, the European Economic Area and the United Kingdom rest on the European Commission's adequacy decision for Israel, transfers within the Immunai group are governed by an intra-group data transfer agreement, and transfers to countries without adequacy rely on standard contractual clauses approved by the European Commission or the United Kingdom regulator.

What is left open is isolation, and it is the question a sponsor would ask first. There is no tenancy model to evaluate because the delivery model is samples in and analysis out rather than software deployed to a customer, and nothing published describes how one sponsor's data and results are separated from another's on shared infrastructure, or where a given engagement's samples and derived data physically rest among the named locations.

There is no price to publish: Immunai bills through bespoke multi year research partnerships scoped to a sponsor's programs rather than through a subscription, so there is no rate card and no unit of charge. What the announcements show before contact is the shape of the deal rather than its size.

The figures that are published are contract ceilings, not prices: up to $85 million for the October 2025 agreement with AstraZeneca covering an inflammatory bowel disease target Immunai identified, and a further expansion announced in May 2026. Both describe eligibility including milestones rather than sums received, and no upfront payment, milestone trigger or royalty term is disclosed, so realized value cannot be estimated from them.

Two cost structures are visible in the delivery model rather than in any published term. The engagement is samples in rather than software out, so specimen collection, viability handling, cryopreservation and shipping sit on the sponsor's side. And profiling cost scales with cohort size, because single cell RNA, surface protein and receptor repertoire sequencing are per sample costs, so first year economics turn on how many patients are profiled rather than on a license fee.

Coverage is named at two levels and part of it is validated. The stated scope is the immune system across drug development, with three solutions named: target discovery, preclinical evaluation and clinical trial optimization. The disease areas are named through the work rather than claimed in the abstract: oncology across several tumor types, inflammatory bowel disease, autoimmune disease, vitiligo, multiple myeloma, urothelial cancer, renal cell carcinoma and pleural mesothelioma all appear in named partnerships or dated publications.

Validation reaches a good part of that. Peer reviewed papers cover checkpoint inhibitor response, CAR T cell function, anti-TIGIT combination biology and chemotherapy immunomodulation, and the real world checkpoint inhibitor cohort assembled with the Parker Institute spans 1,070 patients and 3,700 samples. The atlas itself is stated to cover more than 500 diseases and 80 immune cell types.

Breadth outruns demonstration. More than 500 diseases in the atlas is data coverage rather than validated operating range, the three solutions are offered across the whole of drug development with published evidence concentrated in oncology, and nothing published establishes where the platform has been shown to work as against where it has been applied.