Apheris, a Berlin-based company focused on enabling governed, private, and secure access to data for machine learning, has announced the launch of the ADMET Network, a federated initiative that allows pharmaceutical companies to collaboratively train models for absorption, distribution, metabolism, excretion, and toxicity (ADMET) predictions without sharing proprietary data.
The team exclusively told GEN Edge that five founding members, including Lundbeck, Orion Pharma, Recursion, Servier, and one additional entity not disclosed at this time, have each committed a sizable 80% of their relevant data to train a global ADMET foundation model accessible to all partners.
Drug discovery touts a staggering 90% failure rate with an estimated 40-45% of clinical liabilities attributed to poor ADMET. While AI promises to accelerate timelines and reduce costs by achieving accurate ADMET predictions earlier in the development pipeline, the data required to power these models remains sparse, fragmented, and concentrated in limited regions of chemical space.
Many big pharma companies hold the largest relevant datasets, yet proprietary and competitive considerations constrain data sharing and broader collaboration.
Robin Roehm, co-founder and CEO of Apheris, explains that ADMET Network members can fine-tune the base model directly inside their own secure environments. Inference runs locally, models integrate into existing workflows, and teams retain full control over how predictions are applied in live drug discovery programs.
“That combination, broad industrial training plus local specialization, is what allows models to remain both general and program-relevant,” said Roehm.
He also highlights that the structure of the network is not anchored to any single pharma pipeline. Participation strengthens the shared model without creating strategic dependency on another company.
Najat Khan, PhD, president and CEO of Recursion, emphasizes that clinical trials don’t just fail because of the molecule. Downstream challenges, such as administering the correct dose, analyzing pharmacokinetics (PK) during off-target effects, and picking the right patient population, all impact the success of a drug.
“Two out of three of those things are connected to ADMET,” Khan said in an interview with GEN Edge. “ADMET is what changes molecules in chemistry to potential medicines for patients.”
Julian Schönauer, PhD, director, strategy and operations at Apheris, reiterates that the value of the ADMET Network is to improve model generalizability to new chemical space. Benchmarking success must move beyond statistical significance and provide a tangible impact on drug discovery.
“We work closely with both computational stakeholders and drug programs to ensure that we’re not only seeing improvements in the metrics, but also informing practical decision-making,” Schönauer told GEN Edge.
Compounded impact
The ADMET Network is designed as a continuous learning collaboration that is open to new members. Each addition increases the effective chemical coverage for all participants. “That cumulative effect is central to the network’s long-term value,” says Roehm.
The team is already in advanced conversations with several pharmaceutical and biotech companies who are looking to join the network in the upcoming months. While the network is initially focused on small molecules, Apheris anticipates expanding to additional drug modalities, including PROTACs, peptides, and macrocycles.
Apheris is not alone in applying federated learning to AI-driven therapeutic development. In September, Eli Lilly announced the launch of TuneLab, a collaborative platform that provides participating companies with access to Lilly’s proprietary AI drug discovery models in exchange for data contributions aimed at improving model performance.
Unlike the Apheris ADMET Network, which is structured to manage the IP and governance complexities associated with large pharmaceutical companies contributing substantial proprietary datasets, TuneLab’s models are built primarily on Lilly’s internal data, with many of its collaborators being earlier-stage organizations.
The ADMET Network follows Apheris’s flagship collaboration with the AI Structural Biology (AISB) Consortium, which launched last March. Using proprietary data from AbbVie and Johnson & Johnson in a confidentiality-preserving environment, the initiative fine-tunes OpenFold3, a co-folding model for protein-ligand interactions developed by the lab of Mohammed AlQuraishi, PhD, assistant professor of systems biology at Columbia University. The initiative was expanded to Astex Pharmaceuticals, Bristol Myers Squibb, and Takeda in October.
As the field eagerly awaits proof points in the clinic, Khan reiterates that implementing AI models to address bottlenecks throughout the value chain, from binding affinity, ADMET, to the clinic, will provide compounded scale advantage. At the same time, achieving clinical impact is a long-term journey.
“In any industry, you want to look at the green shoots. Some people call it leading indicators,” Khan told me last December. “Discovery timelines are getting shorter. The number of compounds being synthesized is one tenth of traditional approaches because we’re doing more screening in silico.”
She highlights AI’s potential to uncover novel targets and treatment options for patients with unmet need. Notably, Recursion wrapped up 2025 with favorable preliminary Phase Ib and II results for REC-4881, a small molecule that reduced polyp burden for familial adenomatous polyposis (FAP). This rare genetic disorder is marked by the development of hundreds to thousands of precancerous colorectal polyps and currently has no approved therapies.
REC-4881 is the first candidate developed using the company’s full AI stack, Recursion OS, which leverages large-scale phenomic datasets, including transcriptomics, proteomics, metabolomics, and other multimodal data, to understand complex biology.
One point is clear: Tackling the data gap is essential to unlocking the full power of AI. As the ADMET Network expands, so does the collective insight into human biology needed to accelerated drug discovery.
