Five Drugmakers, Including Recursion & Servier, Team Up On Privacy-Preserving ADMET Network
Five pharmaceutical companies, including Lundbeck, Orion Pharma, Recursion, Servier, and one undisclosed organization, have committed around 80% of their internal ADMET datasets to a privacy-preserving federated learning network coordinated by Berlin-based Apheris, according to reporting by GEN Edge. The initiative aims to improve model performance in predicting absorption, distribution, metabolism, excretion, and toxicity, while allowing participants to retain control over proprietary data.
The ADMET Network is a governed federated environment in which partners train a shared foundation model across distributed nodes while keeping raw datasets within their own secure infrastructure. The resulting base model can be fine-tuned and deployed locally, allowing integration into existing discovery workflows without exposing proprietary compounds or program-level data. The model is not anchored to a single company’s pipeline, reducing dependency risks among participants.
The network is structured as an open, continuous learning collaboration. Additional pharmaceutical and biotech companies are reportedly in advanced discussions to join. While initially focused on small molecules, Apheris anticipates potential expansion to other modalities such as PROTACs, peptides, and macrocycles.
ADMET liabilities are estimated to account for roughly 40–45% of clinical attrition, within an overall drug development failure rate that approaches 90%. Although machine learning models have shown promise in predicting ADMET properties earlier in the discovery process, training data remain fragmented and concentrated within large pharmaceutical companies. Competitive constraints have historically limited cross-company pooling of such datasets.
The effort follows Apheris’s earlier work with the AI Structural Biology Consortium, which applies federated learning to fine-tune OpenFold3 for protein-ligand interaction modeling using proprietary datasets from multiple pharmaceutical partners.
Recursion, one of the contributors to the ADMET Network, recently reported preliminary Phase Ib and II data for REC-4881, a small molecule developed using AI via its Recursion OS platform, which integrates multimodal phenomic datasets, where cell-phenotype shifts can flag toxicity and off-target liabilities linked to ADMET.
If the network achieves broader chemical space coverage and measurable improvements in out-of-distribution generalization, it may offer a solution to one of drug discovery’s structural constraints: access to sufficiently diverse, high-quality ADMET data without relinquishing intellectual property.
The ADMET Network’s launch coincides with the launch of BEACON, a consortium launched a few days ago and led by Conscience, that aims to standardize how AI models are evaluated across biology and drug discovery.
BEACON is developing shared benchmarking frameworks, running community challenges, and building an open platform to assess foundation models and AI agents in disease research and small-molecule discovery—addressing reproducibility and methodological clarity, and complementing federated efforts that focus on secure cross-company model training.
Topic: AI in Bio
