Federated learning tests its limits in drug discovery

Federated learning is gaining traction in drug discovery as new model architectures enable learning across disparate datasets without requiring full harmonization. The next frontier is newer and more complex modalities, where data are scarcer and more fragmented. Yet it cannot compensate when too few observations exist across the industry, particularly for in vivo pharmacokinetics and clinical efficacy.

Federated learning trains a shared model across data held by multiple participants without pooling the underlying records, giving biopharmas a way to learn from proprietary datasets distributed across the industry without requiring companies to expose or commingle them. It can therefore address one kind of data scarcity — useful data fragmented across companies — but not another: too few relevant experiments having been performed across the industry...