Four Pharma Rivals Train Antibody AI Without Sharing Sequences

Ginkgo Datapoints and Berlin-based Apheris opened the Antibody Developability Consortium on September 29, recruiting AbbVie, argenx, Lundbeck and Takeda to train antibody AI on 10,000 sequences.

The four firms will feed private sequences into a shared wet-lab campaign run by Ginkgo. Apheris will train and deliver models without letting any member see another member’s raw sequences. A blinded contest in 2025 already showed why they will not try this on open data alone.

Last Year’s Contest Hit a Wall at 246 Antibodies

Developability is the set of biophysical traits that decide whether an antibody can be made, formulated and pushed toward the clinic. Binding can look fine in a discovery screen and still die later on aggregation, viscosity, weak expression or sticky off-target binding. Public models have had almost no standardised data to learn those failures from.

Ginkgo Datapoints ran the Antibody Developability Competition from September 8 to November 18, 2025. It asked 113 teams scored on 80 held-out antibodies after training on 246 public clinical sequences. Teams came from 25 countries, 38 companies and 39 universities. The paper, published in mAbs in 2026, is blunt: models overfit the public set and do not hold across assays.

CONTEST SCORES ON THE HELD-OUT SET

Cross-validation on the 246 training antibodies ran hotter than those held-out scores, which is the signature of a set that is too small and too narrow. Hydrophobicity was the only property that stayed useful. Titer, self-association and polyreactivity fell into the 0.31 to 0.39 band. Robin Röhm, Apheris chief executive and co-founder, put the public cupboard at fewer than 250 IgG sequences with paired developability measurements, and said that set was never designed for diversity.

The open Ginkgo releases around that work, labelled GDPa1 through GDPa4, now cover 246 IgGs, 18 VHH constructs, 80 further IgGs, and 160 bispecifics with 71 parental precursors. That is still a rounding error against a 10,000-antibody panel. The contest did not fail because the field lacks clever architectures. It failed because the public labels were never built to cover sequence space.

Ginkgo Already Built the Lab That Makes the Data

Ginkgo Datapoints is the data arm of Ginkgo Bioworks. It already sells antibody characterisation as a production line, not as a one-off academic screen. The company’s PROPHET-Ab high-throughput assay platform was built to grind out the large, tidy tables that machine-learning groups say they need and almost never receive.

The first public dump from that lab, described in a 2025 mAbs paper by Ammar Arsiwala and colleagues, reported 246 antibodies: 106 approved, 135 clinical-stage, and 5 preregistration or withdrawn. Antibodies were expressed in HEK293F and purified on Protein A, with a 20-IgG subset also made in ExpiCHO. Rich Cohen, a Ginkgo Datapoints senior director and a co-author on that paper, said the consortium is building the largest standardised antibody developability dataset the industry has seen, plus the models trained on it.

THE PUBLIC PROPHET-AB PANEL

• Titer: Expression measured with a Valita assay.
• Purity: Reduced capillary electrophoresis SDS.
• Aggregation: Size-exclusion chromatography.
• Thermostability: nanoDSF and DSF melting reads.
• Colloidal stability: Stand-up monolayer adsorption chromatography.
• Hydrophobicity: Hydrophobic interaction chromatography.
• Heparin binding: Heparin affinity chromatography.
• Self-association: AC-SINS, plus DLS-kD on a 10-antibody subset.
• Polyreactivity: Bead-based readout against CHO soluble membrane protein and ovalbumin.

Those ten developability assays released as GDPa1 are the public proof of the lab, not the confidential consortium menu. Apheris says the consortium panel of core biophysical properties is shared with prospective members in a private overview. Immunogenicity and multi-specific formats are out of scope for the first model. Ginkgo says the production line can screen 2,400-plus antibody designs in parallel.

Jason Kelly, Ginkgo’s co-founder and chief executive, framed the commercial split in plain terms: Apheris handles federated learning so firms can co-fund a data asset without sharing proprietary data, and Ginkgo Datapoints runs the assays, topping up unused slots from public sequences until the panel hits 10,000. He also asked other companies to get in touch if they want to join.

Today, along with @Apheris_AI, we're introducing the Antibody Developability Consortium, with @AbbVie, @argenxglobal, @LundbeckUS, and @TakedaPharma as founding members.

The Consortium helps pharmaceutical and biotech companies predict manufacturability and developability risks… pic.twitter.com/2d3Dz8jWKK

— Ginkgo Bioworks (@Ginkgo) September 29, 2026

Apheris Sells the Right Not to Share

Apheris, founded in Berlin in 2019 by Röhm and chief technology officer Michael Höh, runs federated data networks and has no drug programs of its own. Training happens inside each member’s environment. Only model weights are aggregated. The company calls that stance “neutral by design,” and pharma has used it for structure prediction, binding and ADMET. The antibody consortium is the large-molecule version of the same offer.

For AI to impact developability decisions in a drug program, it has to perform on a pharma’s own molecules. The Antibody Developability Consortium delivers the largest standardised antibody dataset and the foundation model trained on it. Apheris’ federated infrastructure brings that model to each member and lets them fine-tune it on their proprietary molecules inside their own environment.

Robin Röhm, CEO and co-founder, Apheris

Members receive a foundation model (weights plus reproducible code), federated access to train or benchmark their own architectures on the full set, and the raw assay data for every sequence they contribute plus the public fill-in sequences. They keep ownership of what they put in. Models they train or fine-tune are theirs to use internally. They do not have to run the model on Apheris’s Hub; packaged weights can live on their own stack.

The privacy layer is the product people are actually buying. Pharma already holds more developability data than the public 246, but those internal tables cluster around each firm’s own programs and were measured with assays that were never meant to be pooled. A shared model trained on one lab’s protocol is the workaround. The four founding members still compete on drugs. They are cooperating on the measurements that tell them which candidates will gum up a tank or a syringe.

How the 10,000 Sequences Get Picked

Headcount is the press-release number. Coverage is the design constraint. Apheris says a diversity algorithm, written with the pharma members and academic advisors, runs locally inside each member’s environment and flags which of that firm’s sequences would add the most to a joint training set. No member sees another member’s candidates. Ginkgo then produces and characterises the chosen IgGs in one lab, to one set of protocols.

Members contribute proprietary sequences. Ginkgo fills leftover slots from public sources so the panel reaches 10,000. Contribution volumes sit in the consortium agreement, which has not been published. Sequences are released only where they are physically required to make the proteins. Apheris says it walks those handling rules through each member’s IP team early.

Work is staged. Phase 1 takes 1 to 2 months to select the 10,000 diverse IgG sequences. Phase 2 takes 4 to 5 months for the first 5,000 IgGs produced and characterized, with a foundation model trained and benchmarked on that batch. The partners plan to deliver the initial dataset to members by early 2027, which lines up with that first 5,000, not with a finished 10,000-antibody catalogue on day one. Later phases may add viscosity and other properties if the Phase 2 model earns it.

Independent scientific oversight sits with Charlotte Deane, professor of structural bioinformatics at the University of Oxford, and Peter Tessier, professor of pharmaceutical sciences and chemical engineering at the University of Michigan. Both also listed themselves as scientific consultants for Ginkgo Bioworks on the 2025 contest paper. Deane’s group built the Therapeutic Antibody Profiler; Tessier’s lab has spent years on high-throughput developability screens. They are the right names. They are not, on the paper trail, complete outsiders.

MELLODDY Already Tried This on Small Molecules

Cross-company federated learning in drug discovery is not a new dare. The MELLODDY project, an Innovative Medicines Initiative effort that wrapped its main runs in the early 2020s, put 10 pharmaceutical companies on one audited platform: Amgen, Astellas, AstraZeneca, Bayer, Boehringer Ingelheim, GSK, Janssen, Merck KGaA, Novartis and Servier. Partners kept their warehouses private and still trained together.

The published scale was 2.6 billion confidential activity data points on 21-plus million small molecules and 40-plus thousand assays. Collaborative models were typically 4% better at labelling molecules as active or inactive, showed a 10% increase in applicability domain, and were typically 2% better at estimating activity values. Gains were broader on pharmacokinetics and safety-panel tasks. Returns from extra data volume saturated.

TWO FEDERATED BETS, SIDE BY SIDE

Those lifts were real and modest. They are the honest baseline for what pooling private labels can do when no one will ship the raw rows. Darren Green, former head of cheminformatics at GSK and a MELLODDY participant, now sits on Apheris’s scientific advisory board. The antibody project copies the privacy idea and changes the data strategy: instead of federating decades of mismatched in-house assays, it pays one lab to make a clean panel from scratch, then federates the training.

Binding Models and Toxicity Models Came First

Apheris has been turning that pattern into a product line through 2025 and 2026. The antibody consortium is the latest network, not the first. On September 22, seven days before this launch, the AI Structural Biology Network opened the AISB Bind binding-prediction initiative with AbbVie, AstraZeneca, Bristol Myers Squibb and Johnson & Johnson, with finished models expected in early 2027. That network had already fine-tuned OpenFold3 on more than 20,000 proprietary protein-ligand structures from five partners.

On February 25, 2026, Apheris launched an ADMET network with Lundbeck, Orion Pharma, Recursion and Servier among the founding members, aimed at absorption, distribution, metabolism, excretion and toxicity models. Apheris has said ADMET issues account for an estimated 40 to 45% of clinical attrition in small-molecule programs. Lundbeck now sits in both the ADMET network and the antibody consortium. AbbVie and Takeda already had Apheris relationships through structural-biology work.

THE BUILD TO THE ANTIBODY PANEL

1. September 8, 2025: Ginkgo Datapoints opens the public antibody developability contest on 246 sequences.
2. November 18, 2025: The contest closes; 113 teams are scored on 80 held-out antibodies.
3. February 25, 2026: Apheris launches the ADMET Network with Lundbeck among the founders.
4. September 22, 2026: AISB Bind starts federated training on small-molecule binding.
5. September 29, 2026: The Antibody Developability Consortium opens with AbbVie, argenx, Lundbeck and Takeda.
6. Early 2027: Members are due to receive the initial dataset and the first foundation model.

The company’s public dashboard already lists the antibody set as “in build with Ginkgo,” next to more than 25,000 proprietary co-folding structures, more than 100 million binding datapoints, and more than 900,000 ADMET measurements across 700 tasks. Each network is a paid reason for a drugmaker to install Apheris’s gateway. Ginkgo, for its part, turns the same bioreactors into training rows it can sell again.

What the First Antibody Model Leaves Out

The first model is an IgG biophysical screen, not a full developability oracle. Apheris’s own FAQ is explicit: immunogenicity is out, multi-specific formats are out, and viscosity is a maybe for a later phase if Phase 2 holds up. Complex antibody formats are flagged as a future exploration, not a 2027 deliverable. Anyone buying this as a shortcut to bispecifics or subcutaneous high-concentration products will be waiting past the first drop.

WHAT WE KNOW

• Founders: AbbVie, argenx, Lundbeck and Takeda are in; the consortium remains open to more drugmakers.
• Split of labour: Ginkgo selects, makes and assays the IgGs and trains the foundation model; Apheris delivers it and keeps sequences local.
• First drop: Phase 2 covers 5,000 IgGs in 4 to 5 months, with the initial dataset promised by early 2027.
• Public ceiling: The 2025 contest’s best held-out Spearman ρ values were 0.708 on hydrophobicity and 0.310 to 0.392 on the other four properties.

WHAT IS UNCONFIRMED

• How many sequences each founder must give: Volumes are in the private agreement.
• The exact assay list for the 10,000: The core panel is shown to prospective members, not posted.
• Whether viscosity ships next: It depends on how the Phase 2 model performs.
• Whether a 10,000-row set beats MELLODDY-scale lifts: No prospective result exists yet.

The louder constraint in this field is still the data, not the next architecture. One hundred and thirteen teams could not push titer or self-association past a weak rank correlation when the training set was 246 clinical antibodies. Four companies that compete in the clinic are now paying to have 10,000 IgGs made in one lab rather than mail each other a FASTA file. That is a business model. It is also an admission that even AbbVie-scale in-house tables do not cover enough sequence space on their own.

The first 5,000 antibodies still have to be expressed, purified and run through the core panel before anyone can score the foundation model. Until that drop lands in early 2027, the 10,000-antibody claim is a production schedule, and the four logos on the launch note are the customers of that schedule.