The AI Structural Biology (AISB) Network, powered by Apheris, today launched AISB Bind, a new federated initiative through which AbbVie, AstraZeneca, Bristol Myers Squibb, and Johnson & Johnson, will collaboratively train an AI model to predict how small molecules bind to their protein targets, accurately enough to guide two high-leverage steps in early drug discovery: virtual screening and lead optimization.

How tightly a molecule binds to its target is one of the central determinants of whether it becomes a drug, and predicting it reliably is among the hardest problems in early discovery. Teams today depend on slow, costly rounds of synthesis and assay to find out. A model that can anticipate binding computationally would let them prioritize the right molecules sooner, test fewer dead ends, and shorten the path from idea to validated candidate.

AISB Bind builds on the network’s first initiative, which focused on protein-ligand structure prediction. There, a co-folding model fine-tuned collaboratively across more than 20,000 proprietary structures from active drug-discovery programs significantly raised the share of high-quality predictions from 36% to 52%, outperforming both public models and any model fine-tuned on a single partner’s data alone. This is the first clear demonstration that federated training delivers a step-change in co-folding accuracy. AISB Bind now extends the approach from structure to binding prediction. The aim is a model that holds up in industrial conditions, making reliable, calibrated predictions on new targets, new chemistry, and new programs.

In AISB Bind, the four founding partners collectively contribute three complementary types of data: experimentally determined structures, quantitative affinity and potency measurements, and high-throughput screening activity data. The screening data plays a distinctive role here. It contributes a very large body of experimentally confirmed non-binders, complementing the compounds with measured affinity and adding a negative signal that is rarely available for training at industrial scale. Training across both strengthens the model in two ways. First, it provides a more robust ranking of how tightly related compounds bind, which medicinal chemists rely on to prioritize within a chemical series during lead optimization. Second, it leads to a cleaner separation of target-specific binders from inactive molecules, which virtual screening across large libraries depends on. Combined with large public datasets and additional training data generated through distillation, this brings together an unusually large and diverse body of industrial molecular-interaction data, much of it never before available for model training. The network combines these heterogeneous measurements without manually harmonizing them across partners, using machine-learning techniques that preserve each assay’s context.

Because training runs through federated learning, each partner’s proprietary data stays within its own secure environment and is never exposed to others. Apheris hosts the network and runs the federated training, and each partner receives the resulting model to run and further tailor within its own environment. The model supports work across the design-make-test-analyze cycle, from virtual screening to lead optimization.

Training across the industry at this scale is compute-intensive, and it is powered by the reliable and scalable infrastructure of Apheris’s technology partner, Amazon Web Services (AWS). Federated training begins in September 2026, with the final models expected in early 2027.

“The measured affinities and structures needed to train a binding model are scarce, and most of what exists sits inside individual pharma companies, too sensitive to share. AISB Bind lets pharma partners train one foundation model across that data while it stays protected in each company’s own environment. Every partner receives the resulting federated model, trained on far more data than any one of them holds, and can fine-tune it and embed it into their own workflows and drug programs.”
— Robin Röhm, co-founder and CEO of Apheris

“The data most useful for binding prediction sits inside pharma and rarely reaches academic models, which is exactly why they falter outside familiar chemical space. AISB Bind lets us train and evaluate methods across that industrial data without it leaving any partner, and the collaboration between the labs and the companies pushes the methodology forward faster than either side could alone.”
— Mohammed AlQuraishi, Columbia University

“This consortium shows how collaborative innovation can advance AI-driven drug discovery to help make better medicines for patients. In lead optimization, researchers need models that can accurately prioritize compounds and identify the most promising paths forward for medicinal chemistry. The federated approach enables participating companies to refine these models using their own proprietary data while benefiting from broader training across diverse datasets. These capabilities can support faster, more informed decision-making and open new opportunities for AI-enabled approaches such as virtual ligand screening.”
— Frank Oellien, Lead, Computational Drug Discovery, Emerging Modalities, AbbVie

“The real value of this initiative is in the quality of the data pool it brings together. Our aim is to improve the prediction of molecular interactions with enough precision to matter in a discovery program. Being able to bring confirmed, high-quality affinity data together across companies, without any of it leaving our own environment, is a realistic and exciting way to get there.”
— Payal Sheth, Senior Vice President, Therapeutic Discovery Sciences, Bristol Myers Squibb

About the AI Structural Biology (AISB) Network

The AISB Network advances AI for drug discovery within a secure federated computing paradigm. Its goal is to give members drug-discovery models that perform in real industrial conditions, trained across partners’ proprietary data without that data leaving any partner’s environment.

About Apheris

Apheris builds the industry’s largest federated data networks and customizable AI drug-discovery models across co-folding, binding affinity, ADMET, in vivo PK, and antibody developability. Across these networks, leading pharma companies train models on proprietary, industry-wide data that never leaves their environment, keeping data and IP fully protected. The result is superior models, strongest exactly where public models fall short, delivered as modular, agent-callable building blocks through Foundry, an AI-first platform embedded directly in pharma R&D workflows. For more information, visit www.apheris.com.

Media gallery

About The Author
Business Wire