Trust check

Apheris Review: Federated Data Sharing for Data Owners

Apheris review for data owners: how its federated AI networks keep data on your servers, who pays whom, what members co-own and what to check before signing.

Short answerChecked

Apheris is a Berlin company that runs federated AI networks for drug discovery, in which models train inside each member's own systems and only model weights are combined. It does not publish an offer to buy data: members contribute assay or structure data and, in Apheris's words, "in return, get models". No network page publishes a fee or price. Ask for the consortium agreement before you contribute anything.

Digital calipers on a quality inspection report on a factory workbench
Key points
  • Apheris describes itself as "a neutral operator of federated data networks for drug discovery AI". It trains shared models on members' data without moving that data.
  • A data custodian keeps the data behind its own firewall and decides, through asset policies, who may run which computation with which privacy controls.
  • Members are paid in models, not money: in the ADMET network, Apheris says "Members co-own the resulting models". No fee, price or payment to members is published.
  • The consortium agreements are not public. Exit terms, model rights after you leave and liability for custom models are the clauses to ask for.

What is Apheris?

Apheris says it is "a neutral operator of federated data networks for drug discovery AI" and that "Apheris operates the networks and runs no drug programs of its own" (Apheris). It was founded in 2019 by Robin Röhm and Michael Höh, is headquartered in Berlin, and lists its funding stage as Series A. Its imprint gives a Berlin commercial register number, "HRB 221842 B" (Apheris imprint).

TechCrunch reported in January 2025 that OTB Ventures had "co-led an $8.25 million Series A into Apheris" and that the round brought "its total funding to $20.8 million"; Röhm told the outlet at the time that "Its customers include Roche and several hospitals" (TechCrunch). We found no later funding round announced as of 20 November 2026.

One of its networks, the AI Structural Biology (AISB) Network, is what Apheris describes as "An industry-led collaboration of nine leading pharmaceutical companies, formed to advance AI for drug discovery." Apheris says "The first initiative is complete": AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson and Takeda "each fine-tuned OpenFold3 locally, inside their own environment" (Apheris). On 22 September 2026 the network "launched AISB Bind", in which AbbVie, AstraZeneca, Bristol Myers Squibb and Johnson & Johnson train a model to predict how small molecules bind their targets. Apheris says "Federated training begins in September 2026, with the final models expected in early 2027." (Apheris).

FactWhat we foundSourceStatus as of 20 November 2026
What it isFederated computing software and the operator of drug discovery data networks (AI structural biology, ADMET, antibody developability, virtual cell, real-world evidence).Apheris about page and network pagesOperating; networks listed on its site
What it asks data suppliers forNetwork members contribute proprietary experimental data that stays in their own environment, run a Compute Gateway there and set asset policies.Apheris documentation; ADMET Network pageNetworks open to new members
Published pay or priceNone published. Apheris says members provide access to data and "in return, get models". No membership fee or Foundry price is published.Business Wire release, 29 September 2026; Apheris pagesNo price found on 20 November 2026
Program statusAISB Network: Apheris says nine pharma companies; its first initiative, fine-tuning OpenFold3 with five members, is complete, and AISB Bind (AbbVie, AstraZeneca, Bristol Myers Squibb, Johnson & Johnson) launched 22 September 2026. Antibody Developability Consortium founding members announced 29 September 2026. Foundry software launched 29 October 2026.Apheris AISB page and announcements; Business Wire; Apheris Foundry announcementLive; AISB Bind training began September 2026, final models expected early 2027; antibody dataset targeted for early 2027
Funding and entity$8.25 million Series A, $20.8 million raised in total (January 2025 report); Berlin register number HRB 221842 B.TechCrunch; Apheris imprintFunding stage listed as Series A; no later round found

Does your data leave your servers in an Apheris network?

Apheris builds on what its documentation calls data residency: "Ensuring raw, sensitive data remains secure, this principle mandates that such data never needs to leave its origin." The software that touches the data is a Compute Gateway, and "Compute Gateways reside within the environment of a data provider/ data custodian organization and ensure only computations matching the data custodian's requirements can be executed on a given dataset". The coordinating service is not on your side: "The Compute Orchestrator is hosted on security-hardened infrastructure run by the Apheris team" (Apheris documentation).

What does leave is model output. On the ADMET network Apheris says training runs locally, "Only model weights leave; the data never does", and the combined weights are "tested against data-reconstruction and membership-inference attacks before release" (Apheris). Its documentation frames the privacy goal the same way: "Privacy as such becomes a property of a computational result rather than an underlying property of the dataset" (Apheris documentation).

For hospitals, the real-world evidence network says studies execute "on patient-level data inside the hospital that holds it. Only approved, aggregated outputs come back", and lists "Patient data under GDPR, HIPAA and site governance" (Apheris). Apheris cites those rules; whether a given study meets them is for the hospital and its counsel to decide.

What does a data custodian control in Apheris?

Control sits in asset policies. Apheris describes them as "a set of rules that define allowed operations by named users on specified datasets", covering who can compute on which dataset, which computations are allowed, parameter limits and optional privacy controls. Only users with the "Owner" role can create or change them, and a custodian can sign each policy with its own private key so that the gateway rejects a tampered version (Apheris documentation).

Before any job runs, the gateway checks that "the submitter is allowed to run computations on the data; the computation is permitted on the data; the computation parameters are within the allowed bounds" (Apheris documentation). One limit matters for contracts. Partners can bring their own models, and the documentation states: "As Apheris does not maintain custom models, Apheris is not responsible for any security or privacy leaks caused by using custom models." It recommends reviewing custom model code before allowing it.

  • Beneficiaries: named users, by full email address, who may run computations.
  • Permissions: the model and model version allowed, with its parameter bounds.
  • Privacy controls: noise, bounding or rounding where the chosen model supports them.
  • Signature: an optional digital signature, checked before each computation runs.

Can you sell data to Apheris?

We found no Apheris page offering to buy data. Its networks are consortia. In the antibody consortium, "Each founding member will contribute proprietary antibody sequences", Ginkgo Datapoints fills the rest from public sources to reach 10,000 antibodies, and members "apply resulting models and their own fine-tuned derivatives internally while retaining ownership of the proprietary sequences and assay data they contribute" (Business Wire). The founding members named are AbbVie, argenx, Lundbeck and Takeda, and the consortium "remains open to additional pharmaceutical and biotech companies" (Apheris).

On the ADMET network Apheris states that "Members co-own the resulting models and fine-tune them on their own chemistry", and reports "1M+ measurements of private experimental data spanning 710 assays" used without pooling (Apheris). That count is the company's own figure. Before you contribute, Apheris says "A federated-uplift assessment first estimates what a network would add for each program's endpoints, before you contribute any data" (Apheris).

So a biotech, CRO or lab with assay archives gets access to better models, not a check. If the goal is cash for records, read our guides on licensing R&D experiment records and biotech and CRO lab data, and our GMP batch records guide, which also names the ADMET network.

Who pays whom in an Apheris data network?

No Apheris network page publishes a membership fee, a price for Foundry or a payment to contributors. What the sources do show is the exchange. Apheris's own boilerplate says: "Members provide privacy-preserving access to their proprietary data and, in return, get models with higher performance and a broader applicability domain" (Business Wire). Ginkgo Datapoints generates the antibody data and trains the base model; Apheris delivers it into each member's environment.

Someone pays Apheris, since it sells software and runs the networks. In January 2025 Röhm told TechCrunch that a product launched in the last quarter of 2023 had found product-market fit and that the company had "multiplied its revenue by four since then"; that is the company's own claim (TechCrunch). In the AISB Network, Apheris says "Participation is per initiative" and "Members who take part receive the resulting model" (Apheris). Who pays how much, per member or per network, is not public. Foundry can also be used alone: "You do not need to join a federated network to use Foundry with public, third-party or your own models" (Apheris). Its price is not published either.

Is Apheris legit?

What we verified: a registered Berlin company with a published register number, named founders, investors named by TechCrunch, pharma members named in a wire release issued jointly with Ginkgo Bioworks' Datapoints unit, and public technical documentation that describes how custodian controls work.

What we could not verify: the consortium and membership agreements, any fee, Apheris's financial results, and the audit reports in its trust center, which refused our scripted fetch. Ask for those reports and the agreement under NDA before you contribute data.

What to check before you sign with Apheris

Apheris publishes how its software works but not its consortium or membership agreements. These checks pair what it has published with what you need in writing.

  • Rights granted and model ownership. Apheris says ADMET members "co-own the resulting models". Ask what co-ownership means in the agreement: who may commercialize the network model, and whether you keep using it if you leave.
  • Derived outputs. Ask who owns your fine-tuned derivative and its weights. The antibody consortium says members apply "their own fine-tuned derivatives internally". Check that the agreement says the same.
  • Exclusivity and term. Not published. Ask whether membership bars you from other consortia and how long contributed data stays in training.
  • Exit and termination. Not published. Ask whether weights trained on your data stay in the network model after you leave, and what is deleted.
  • Custom models and liability. The documentation disclaims responsibility for leaks from custom models. Decide who reviews third-party model code and where liability sits.
  • Who de-identifies. For patient data, Apheris cites GDPR, HIPAA and site governance but does not publish who de-identifies or approves outputs. Name the responsible party in the contract.
  • Fees, audit and payment. No fee is published and no payment to members appears. Ask for every fee in writing and for the trust center audit reports.

How we checked this

We read every source on 20 November 2026, starting with the primary sources: Apheris's own about page, imprint, network pages including the AISB Network, the AISB Bind announcement of 22 September 2026, the Foundry announcement of 29 October 2026 and its Gateway documentation (version 3.12 for concepts and privacy controls, the latest listed, and version 3.11 for asset policies); the joint Ginkgo Datapoints and Apheris release on Business Wire, saved because the site refuses scripted fetches; and TechCrunch's January 2025 report. We searched for later funding, partnerships and changes to the networks.

We could not read the trust center, which refused our fetch, or the full text of a paywalled Handelsblatt profile of November 2026, and we could not verify any fee, contract term or financial figure that Apheris has not published. Figures such as assay counts are the company's own.

How does an Apheris network differ from licensing data for money?

In a license, you deliver a copy of the data and are paid under a contract that limits its use. In an Apheris network, nothing is delivered: computation comes to you, and you are paid in model performance. Both need a written agreement that settles rights, outputs and exit. Our data licensing agreement guide lists the clauses, and our review of a data licensing offer shows how to test one.

Origin Data Partners is an independent data partner that helps business owners value and license their company records: the company pays Origin nothing, and a buyer may pay Origin when a deal completes. We help you license your data.

Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.

Common questions

What do Apheris reviews show?

Public reviews of Apheris as a business partner are scarce. The record is press coverage, such as TechCrunch's 2025 funding report, and named members in announcements, such as AbbVie, argenx, Lundbeck and Takeda in the antibody consortium. Customer terms and satisfaction are not public. Ask existing members about governance and exit before you join, and read the agreement itself.

What is Apheris Foundry?

Foundry is Apheris's software for running, benchmarking and fine-tuning drug discovery models, launched on 29 October 2026. Apheris says "Its modules and workflows run inside your own infrastructure and can be used directly or through your agents", and that a company can use it without joining a network. It is a product you buy, not a channel for selling data. No price is published.

Who are the founding members of the Apheris antibody consortium?

AbbVie, argenx, Lundbeck and Takeda, announced on 29 September 2026 with Ginkgo Datapoints. Each contributes proprietary antibody sequences toward a panel of 10,000 antibodies, with Ginkgo filling the remainder from public sources. The joint release says the initial dataset is targeted for delivery to members by early 2027. The consortium remains open to other pharma and biotech companies.

Do Apheris network members own the models?

In the ADMET network, Apheris says "Members co-own the resulting models" and fine-tune them on their own chemistry. For the AISB Network it says "The weights of the model an initiative produces belong to the participating members." In the antibody consortium, members use the resulting models internally while keeping ownership of the sequences they contribute. What co-ownership allows, and what survives your exit, is in agreements Apheris has not published.

Can a hospital join an Apheris network with patient data?

Apheris runs a real-world evidence network for multi-site studies in which, it says, analysis runs inside each hospital and only approved, aggregated outputs return. It cites GDPR, HIPAA and site governance. Whether a particular study is permitted depends on the hospital's consents, ethics approvals and contracts, which Apheris does not decide. Ask who approves outputs before a study starts.

Does Apheris see my raw data?

Its documentation says raw data stays behind the custodian's firewall and that a Compute Gateway inside your environment runs only computations your asset policies allow. Apheris hosts the orchestrator that coordinates jobs and combines model weights. Model weights do leave your environment, which is why Apheris tests combined weights against reconstruction attacks. Ask for the test results.

Sources and further reading

What is your company’s data worth?

AI companies pay $20K to $5M for company records. Six questions, no files.