Pharmaceutical companies are fundamentally changing how they develop drugs by sharing their most valuable AI models with competitors and smaller biotech firms, rather than keeping them locked away. This shift toward federated learning, a technique that trains artificial intelligence models on proprietary data without ever exposing that data outside company walls, is reshaping drug discovery across the industry.

What Is Federated Learning and Why Does It Matter for Drug Discovery?

Federated learning works by sending the AI model to the data, rather than bringing sensitive data to the model. The model trains on each institution's proprietary information, and only the model's learned weights, or mathematical adjustments, move between organizations. This approach allows companies to improve shared AI tools while keeping their most closely guarded datasets completely private.

"The model goes to the data. So the model is then moved into that environment, trains on that data, and changes to the model weights or gradients are then integrated into the local model. And you can repeat that," explained Jonathan B. Gilbert, Senior Director of Ecosystem Growth and Contributor Partnerships at Eli Lilly and Company.

Jonathan B. Gilbert, Senior Director, Ecosystem Growth and Contributor Partnerships, Eli Lilly and Company

The advantage is clear: companies get access to AI models trained on billions of dollars worth of collective data, while their own trade secrets never leave the building. Eli Lilly has invested an estimated over a billion dollars in internal datasets that now power its shared models.

How Are Companies Actually Implementing Federated Learning?

Eli Lilly launched TuneLab in September 2025, a platform that gives biotechs free access to the same AI models the pharmaceutical giant uses internally, provided they contribute their own datasets to improve them. The response has been remarkable. As of early 2026, Lilly reports more than 75 partners across three continents and dozens of countries using TuneLab.

The platform focuses on roughly 40 models centered on small molecule drug properties and antibody development, areas where AI can predict whether a potential drug will work safely in the human body. Lilly prioritized making TuneLab easy to use, allowing partners to log into a website and make predictions without needing deep technical expertise.

Other organizations are building similar networks. Apheris, a company specializing in federated learning infrastructure, supports multiple drug discovery networks including the AI Structural Biology Network with nine pharmaceutical companies focused on protein interactions, and the ADMET Network with Recursion and other firms studying small molecule properties.

Steps to Maximize Federated Learning for Drug Development

• Fine-tune models with proprietary data: While federated models train on collective industry data, companies should layer their own program-specific information on top. This combination outperforms using either approach alone, converting general model performance into actual drug program impact.
• Establish trust and governance frameworks: Federated networks require clear agreements about data privacy, security protocols, and how models will be used. Each company has different firewall rules, compute windows, and approval processes that must be negotiated upfront.
• Treat federation as an ongoing operation, not a one-time project: Successful federated learning requires continuous deployment, monitoring, and refinement rather than a single implementation event.

Why Is Open-Source AI Becoming Essential in Drug Discovery?

The pharmaceutical industry is increasingly recognizing that no single company can build the best AI models alone. Woody Sherman, Founder and Chief Innovation Officer at PsiThera and chair of the OpenFold executive committee, noted that as biological data volumes explode, the industry needs shared platforms rather than dozens of competing proprietary systems.

OpenFold, a consortium of more than 40 biotech, pharma, and technology companies, develops open-source AI tools for structural biology and drug design. The group created OpenFold3, OpenStability, and RF4, tools that extend capabilities beyond what existing models can do.

"We're going to need these open platforms that we can all build on. We can't all build our own foundation models from scratch; it just doesn't make sense," stated Woody Sherman, Founder and Chief Innovation Officer at PsiThera.

Woody Sherman, Founder and Chief Innovation Officer, PsiThera

DeepMind's AlphaFold revolutionized protein folding prediction, but even AlphaFold2 had limitations. It struggled with protein complexes, ligands, and ions. AlphaFold3 improved on these weaknesses, though accuracy for nucleic acids remained poor and the model had limited ability to predict how proteins change shape under different conditions.

Mohammad AlQuraishi, Assistant Professor of Systems Biology at Columbia University and cofounder of OpenFold, emphasized that AlphaFold2's breakthrough was not just providing answers, but providing confidence in those answers through calibrated predictions.

What Challenges Remain in Federated Drug Discovery Networks?

Federated learning is not a plug-and-play solution. José-Tomás Prieto, Director of AI Programs at Apheris, stressed that successful networks require significant engineering rigor. Each participating company has unique network constraints, firewall configurations, compute availability windows, and internal approval processes that must be carefully coordinated.

The business model for federated learning has historically been weak, but Prieto compared it to a seatbelt: while companies might prefer to work independently, the regulatory and competitive pressure makes federated learning increasingly unavoidable. Proprietary data never leaves the companies that own it, and models are delivered to individual nodes for internal use.

"If there is something to learn about the AI world today, it is that you cannot necessarily model your way out of a data problem," noted José-Tomás Prieto, Director of AI Programs at Apheris.

José-Tomás Prieto, Director of AI Programs, Apheris

One significant advantage of federated approaches is that models are trained on messy, diverse real-world data rather than biased toward well-characterized public datasets. This produces more robust predictions for actual drug development scenarios.

The next five years will likely see federated learning become standard practice in drug discovery, with improved generalization, speed, and scale driven by open foundation models. The combination of shared AI platforms, federated training on proprietary data, and fine-tuning for specific programs represents the future of how pharmaceutical innovation will operate.