Aignostics

The models are the product. Aignostics sells its pathology foundation models for licensing (Atlas 2) and as open weights for academic use (RudolfV 2), and each application runs on them: Atlas H&E-TME profiles the tumor microenvironment with models built on Atlas, and PathoSearch retrieves reference cases by comparing Atlas 2 embeddings. Its services for target discovery, translational research and diagnostic development are analyses built with the same models. There is no scanner, image management system or laboratory service that would stand without them.

The artificial intelligence is the product.

Review by the user is described for the newest product, with no threshold at which a match is withheld. PathoSearch returns reference cases with a similarity score, and the company states that the pathologist opens the real case behind each match and makes the call; its terms bar any use for diagnosis or active patient care. Atlas H&E-TME outputs cell, tissue and spatial measurements for researchers, with quality control models flagging slide areas to exclude from analysis. Nothing states the similarity score below which a match should not be relied on, or how an error in a delivered analysis is corrected.

The oversight structure described, with the threshold missing.

The models are identified and their training data characterized in detail; the update practice is not published. Atlas 2, co-developed with Mayo Clinic, is a vision transformer of about 2 billion parameters trained on 5.5 million slides from 1.8 million cases, with more than 100 stain types, 150 tissue types and seven scanner types, described in an arXiv preprint (2601.05148) that compares it with 15 public models on 80 benchmarks. RudolfV 2, released as open weights in August 2026, is a ViT-g model of about 1.1 billion parameters trained on more than 300,000 slides from Charité and LMU Munich, with two distilled versions. The Atlas H&E-TME preprint names the version it evaluated (v1.2.0). What is missing is how a customer learns that the model under an application has changed: the TME page names Atlas while PathoSearch runs on Atlas 2, and no release notes, version history or change notice for the applications is published.

Models named and data characterized, without the update practice.

An in-house build, with a partial list of the services around it. Aignostics trains its own foundation models, so no outside model provider sits between a slide and a result. Its privacy policy names Sentry and PostHog as the monitoring and analytics services for PathoSearch, with their retention periods, and allows sign-in through Microsoft or Google accounts. The hosting provider and region for the Portal, Atlas H&E-TME and PathoSearch are not named, and no subprocessor list covers the image data itself. The training data come from named partners, Mayo Clinic, Charité and LMU Munich, but nothing published describes the terms under which those archives were shared.

A structurally short chain at the model layer, with the operating chain partly named.

Published research on the models and applications, led by the company and not yet independent. The Atlas H&E-TME preprint (arXiv 2606.12346, 2026), with Aignostics authors and coauthors at Charité and Mayo Clinic, reports macro F1 from 0.85 to 0.95 across cancer types on more than 1,500 cases and 200,000 annotations from over 25 laboratories and eight scanner models; on 30 slides from three cancer types it reports a macro F1 of 0.74 against an IHC-informed reference, compared with 0.71 for pathologists working from H&E. The Atlas 2 preprint reports results on 80 public benchmarks, which outside groups can rerun. A case study with Bayer describes a target discovery platform recovering a known lung cancer pathway. No study shows a drug development or diagnostic outcome improved in use, and none is peer reviewed or independent.

Published research short of independent validation.

Retention and deletion are stated for each product; training use of customer images is not stated for the tissue profiling product. PathoSearch's terms require users to de-identify images and bar protected health information, allow Aignostics to process an uploaded image to provide the service, and state that images and results are kept for up to 90 days and then deleted. Atlas H&E-TME customer data are described as deleted from the company's network after processing, and free trial data likewise; platform logs are kept 30 days in active storage and archived for at least 2 years. Nothing says whether customer slides analyzed by Atlas H&E-TME or in services work may train models, and nothing published describes how the hospital archives used for training, which the company describes as access to data on over 30 million patients, were de-identified or governed. No safety engineering is published.

Categorical commitments published, with training use and the main product's retention window unstated.

A plain scope position for the search product and a service level privacy policy, with no data processing terms published. PathoSearch's terms state that users must not upload protected health information or personal data, that images must be de-identified under applicable law, and how any personal data processed despite that is handled. The privacy policy names Aignostics GmbH as controller, gives a data protection officer, and covers the Portal, Atlas H&E-TME and PathoSearch as services, including the legal basis for processing image data under contract. No business associate position is stated and no data processing agreement for biopharma image data is published, although the company works with US partners and has a New York office.

A scope position stated with a service privacy document behind it, and the agreement not published.

A recognized certification with its scope named, and no artifact. The company announced ISO/IEC 27001 certification in August 2022, issued by TÜV SÜD Management Service and covering development of its diagnostic solutions, research services and operation of its Portal platform, and recertification to the 2022 version of the standard in August 2025. It also holds ISO 13485 for its quality system and names a chief information security officer. The certificate itself, its current scope and its expiry are not published, and there is no trust center, SOC 2 report or penetration testing statement.

A recognized certification named, with the artifact on request.

No device claim, and each product is labeled for research. Atlas H&E-TME and PathoSearch both carry research use only statements, and the PathoSearch terms add that it is not a diagnostic tool, has not been cleared or approved by any regulator and must not be used for diagnosis or patient care. The PathoSearch page and launch announcement nonetheless present it to pathologists for resolving hard cases and building a differential, which sits uneasily beside those terms. Atlas 2 is described as carrying regulatory documentation to support integration into other companies' medical devices, and the company holds ISO 13485 and offers services to develop companion diagnostics with partners. FDA's databases list no Aignostics device.

No device claim, with the products scoped accordingly.

Performance of the tissue profiling product is published by cancer type, with the weakest groups identifiable. The Atlas H&E-TME preprint (arXiv 2606.12346, submitted June 2026 and revised July 2026; led by Aignostics with coauthors at Charité and Mayo Clinic; the company links it as its whitepaper) evaluates version 1.2.0 on more than 1,500 cases from over 25 laboratories and eight scanner models, and reports macro F1 by primary site from 0.85 (prostate) to 0.94 (lung) and by metastatic site from 0.85 (bone) to 0.95 (lung).

The Atlas 2 preprint reports robustness to scanner, stain and center variation on public benchmarks. Results by individual laboratory or scanner are not broken out, and there is no result by patient sex, age or ancestry. No governance framework or monitoring after deployment is described, and PathoSearch has no published evaluation.

A published error rate for the tissue profiling product, and nothing standing behind the search tool. The Atlas H&E-TME preprint, which Aignostics leads, reports macro F1 for version 1.2.0 by cancer type across more than 1,500 cases and 200,000 annotations, with its method and reference standard described, so a buyer has an error characteristic to test against. PathoSearch is supplied as is during early access: its terms disclaim any warranty that results are accurate, cap Aignostics' liability at EUR 100, and bar any use in patient care. No warranty, indemnity or remediation commitment attaches to either product.

A published error rate with its method and denominator.

Outputs can be exported to image management systems, with no system named. Atlas H&E-TME delivers polygon outputs the company says can be integrated into a customer's preferred image management system, along with thousands of quantitative metrics per slide, and the privacy policy describes API access. PathoSearch takes a screenshot as input and needs no DICOM export. No image management system, laboratory system, standard or interface specification is named for any product; for biopharma research buyers the relevant links are to image archives and analysis pipelines rather than to clinical records.

Integration claimed with no system named.

Hosted applications with no location stated. Atlas H&E-TME runs on the company's platform, where data are transferred, analyzed and then deleted, and PathoSearch is a hosted web service. Atlas 2 is offered for licensing and RudolfV 2 as downloadable weights, which a customer could run in its own environment. No hosting provider, region or tenancy is stated for the hosted services, although the company operates in Germany and the United States and describes the TME product as GDPR compliant.

Hosted options with location implied rather than committed.

No price is published, and how each product is sold can be established. PathoSearch is free during early access, with general availability to be governed by a separate agreement. Atlas H&E-TME offers a free trial to life sciences companies and free access for academics through a research access program and the OpenTME dataset. Atlas 2 is licensed to biopharma through an inquiry form, RudolfV 2 is free for academic researchers, and target discovery, translational and diagnostic development work is sold as services. No rate, unit of charge or license term is published for paid use.

The sales posture is discoverable, with no price.

Coverage is named for each product, with validation behind the tissue profiling product. Atlas H&E-TME lists ten primary organs on its product page, and the preprint reports results by cancer type for version 1.2.0 across primary and metastatic sites from more than 25 laboratories, with the page adding resections, biopsies and fine needle aspirates. PathoSearch covers over 300 diagnostic entities and 27 organs from more than 35,000 reference cases, with no published evaluation. Buyers are biopharma research and development teams, academic researchers and, for PathoSearch, pathologists under research use terms; no product is for clinical diagnosis.

Coverage named with validation behind part of it.