Apheris Announced the Results of their Federated Learning Co-Folding Initiative—Now Four Big Pharma Are Betting Bigger

By Allison Proffitt

September 22, 2026 | Four big pharma—AbbVie, AstraZeneca, Bristol Myers Squibb, and Johnson & Johnson—are joining their data to train an AI model that predicts how molecules bind to their targets, a step the companies hope will speed the drug discovery process.

The initiative, called AISB Bind, was announced today by the AI Structural Biology (AISB) Network, an industry consortium powered by Berlin-based Apheris and run in collaboration with the AlQuraishi Lab at Columbia University. AISB Bind is the next iteration in the AISB Network’s push to demonstrate the power of federated learning. Model training begins this month, with finished models expected in early 2027.

AISB Bind aims to make machine learning models really useful for pharma companies’ drug programs, Apheris co-founder and CEO Robin Roehm told Bio-IT World. “It looks at what other data modalities are in the private pharma repositories that can help us build superior models that have higher performance.”

Co-Folding Federated Groundwork

This is not the first challenge the AISB Network set itself. Apheris offers products for federated networks for drug discovery, enabling their customers to join or build their own federated learning networks while protecting sensitive data. The company undergirds the AISB (AI Structural Biology) Network and supports the ADMET Network with Recursion focused on small-molecule property prediction and the antibody developability network with Ginkgo Dataworks.  

Earlier this month, the company published a paper on their own site outlining the results of the Federated OpenFold3 project. While the paper is not peer reviewed, an Apheris representative pointed out that it is co-authored by the AlQuraishi Lab at Columbia University and named team members from each pharma partner.

The five pharma companies—AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson, and Takeda—spent close to a year fine-tuning OpenFold3, a publicly-available “co-folding” model that predicts the 3D shape a protein and a small-molecule drug candidate form when they bind together, on more than 20,000 proprietary structures pulled from their active drug programs. Members of the group discussed the work last May at the Bio-IT World Conference & Expo.

In the federated project, none of those data ever left the companies’ own environments; instead, each partner trained the model locally and periodically sent back only the resulting model weights to a central aggregator run by Apheris.

The payoff was substantial. Tested against 1,056 known structures the partners had held back specifically for evaluation, the resulting model—AISB-1-Fed—produced a high-quality interface prediction 52.1% of the time, up from 35.6% for the public OpenFold3 model it started from and correctly placed the drug molecule itself 46.8% of the time, up from 28.9%. It also beat every public alternative the team tested it against, including Boltz-2 and two versions of ByteDance's ProtenixV1, by roughly 11 percentage points on both measures.

To rule out the possibility that the gains simply came from newer public training data rather than the proprietary structures, the team built a control model trained on OpenFold3 plus all available public data, but none of the five companies' private structures. That control fell well short of AISB-1-Fed, confirming that the private, drug-relevant structures, not newer public data, were driving the jump in accuracy.

“What we showed is that we can deliver a step-change in performance,” Roehm said. “What we didn’t show is that the model now solves all the drug [design] problems. We made it significantly better—much better than what we could achieve from model architecture innovation—but there are still many programs in the pharma data repositories that even the federated model doesn’t get right.”

Roehm said the finding underscores the value of diverse data. “This is not model architecture innovation,” he emphasized. “The model architecture is the same model architecture that is out there, whether that’s Boltz, OpenFold ... all of those essentially are the similar model architectures.” What’s scarce, he said, is the kind of dense, drug-relevant structural data that only exists inside individual pharma companies’ own archives, that they have been hesitant to expose to competitors.

While public data is valuable, the paper points out that since the last major public dataset cutoff, the Protein Data Bank has added roughly 70,000 new structures, but fewer than 3,000 of those involve an actual approved or investigational drug. The rest are largely cofactors, metabolites, and other molecules of little use for training a model to prioritize real drug candidates.

Training models on internal data alone also is not an equally robust solution, Roehm argued. Each company holds a naturally skewed slice of the problem, shaped by whatever targets and chemistry it happened to have worked on. Roehm said pooling data from multiple companies forces a model to generalize beyond any single group’s history. “By pooling data from multiple pharma companies, what you force the model to learn is kind of this diversity that you cannot get if you only look into the historic data of the respective single pharma companies’ repositories,” he said.

Getting companies comfortable sharing that data at all required convincing them it was safe. Apheris ran its own internal attacks against the federated model—including “membership inference” attacks designed to determine whether a given structure had been part of the training data—and reported that even under deliberately favorable conditions for the attacker, the technique failed on the majority of attempts and could reveal, at most, whether a molecule the attacker already possessed had been used in training, nothing about molecules it didn't already know.

BMS with a Participant’s View

Bristol Myers Squibb was part of the co-folding project and will be joining again for AISB Bind. Payal Sheth, SVP Therapeutic Discovery Sciences at Bristol Myers Squibb, discussed the pharma’s motivations and goals with Bio-IT World.

BMS wasn’t blindly sold on federated learning—or Apheris—but knew the challenges of structure prediction outpaced the depth of even their data pool. “No single company's data covers enough chemical and biological space [for structure prediction],” Sheth told Bio-IT World via email. “Federation let us test whether combining data across companies helps, without any data leaving our environment, and Apheris's governance and security held up to our review. The co-folding initiative was a contained experiment we could evaluate on our own held-out data, and the results made joining Bind a logical next step in that learning process.”

Sheth testified to the challenge of mapping decades of legacy data to a shared framework. “It was a real investment. Mapping decades of curated structural data to a shared schema takes scientific judgment, not just reformatting,” she said. But the investment in the co-folding project will return dividends for the Bind project. “We saw it as building a durable capability, and much of it carries over: standards, infrastructure, and governance are now established. Affinity and activity data bring their own challenges, but we start from a stronger position and with lessons learned.”

While the project has been fruitful so far—“Results on our own structures are encouraging relative to the baseline,” Sheth said—BMS is looking to understand where models carry real weight.

Success looks like, “Measurable improvement in decisions inside real programs: better prioritization before experiments, more reliable ranking within series,” Sheth explained. “We would frame it as continued learning rather than a finish line. Each cycle, internal and within the consortium, teaches us where models help and where they break. If teams reach for these models routinely on hard problems because the evidence has earned that trust, that's success.”

New Model, Learned Lessons

AISB Bind is designed to apply that same federated approach to binding affinity. Where the first initiative focused narrowly on predicting a molecule's 3D structure, Bind adds binding affinity and high-throughput screening data, including a large body of confirmed non-binders, a kind of negative signal that's rarely available at scale, with the goal of ranking which molecules within a chemical series are worth pursuing during lead optimization and separating true binders from noise during virtual screening.

Roehm described the technical approach as a departure from how the industry typically builds these models, which is to predict a structure once and then bolt on a separate model for binding affinity. Apheris's approach instead uses binding data to generate synthetic structures, filtering for high-quality predictions to expand the pool of structural data the model can learn from. “We use binding data and we generate synthetic structures from this binding data,” he said, calling it “the most innovative approach” of the new project.

Notably, the roster has shifted: Astex and Takeda, both contributors to the first project, are not part of AISB Bind, while AstraZeneca is joining for the first time. Roehm said turnover like that is typical of the network, reflecting individual companies’ own AI capability growth rather than any breakdown in trust. Companies that sit out an initiative keep whatever model they helped build, he said, but stop receiving updates as the collaboration moves forward, an incentive structure designed to reward continued participation without penalizing anyone for leaving.

Sheth said BMS continues to apply the same criteria when evaluating each new federated project. “Our data and IP must stay protected, the governance has to work for all members, and federation has to keep demonstrably improving performance over what we could achieve alone and relative to the best available alternatives,” she explained. “Part of what we are testing is where the data thresholds sit, meaning how much additional data, and of what kind, produces a step change versus incremental gains. If returns flattened to the point where the marginal data no longer moved the models meaningfully, that would prompt a rethink of where we invest. We continue to treat each cycle as an experiment with well-defined success metrics.”

Where the first project took nearly a year to produce a model, Roehm says the company now plans to release an updated Bind model every four to six weeks, benchmarking and validating each one before moving to the next. He added, “We have many more shots on goal, ultimately, to deliver a superior model over the course of the program.”