From structure to selectivity: Federated AI for collaborative binding affinity prediction

Recent advances in protein structure prediction have transformed early drug discovery, but they have also revealed a new bottleneck: understanding what a molecule looks like is no longer the primary challenge—understanding whether it will bind with sufficient strength and selectivity to matter biologically is.

Binding affinity prediction depends on diverse experimental evidence. Critical output from screening assays, biophysical measurements, simulations, and negative results are distributed across companies and collaborators and cannot realistically be centralized due to intellectual property, governance, and regulatory constraints. Consequently, progress is increasingly limited, not by model design alone, but by how models can safely learn from fragmented data.

In this session, we will discuss how federated AI approaches enable organizations to contribute to, improve and evaluate affinity models without sharing proprietary datasets. By allowing models to train and run where data already resides, institutions can contribute signal while maintaining control over sensitive information.

We will explore:

• Why the field is shifting from structure prediction toward potency and selectivity prediction
• Why affinity modeling requires learning across heterogeneous datasets
• The practical barriers to collaborative AI in drug discovery
• How federated computing enables multi-organization learning without data pooling

The broader question we will examine is whether the next phase of AI-enabled discovery will depend less on isolated model breakthroughs and more on new mechanisms for secure, collective learning across the scientific ecosystem.