Skip to content

Details

Agenda
9:00 am -- 9:05 am PT -- door open
9:05 am -- 9:15 am PT -- Talk 1: NVIDIA FLARE lighting talk
9:15 am -- 10:00 am PT -- Talk 2: Apheris AISB Talk + QA
10:00 am -- 10:45 am PT -- Talk 3: Lilly TuneLab Talk + QA
10:50 am -- Closing

Meeting Link:
Microsoft Teams meeting
Join: https://teams.microsoft.com/meet/2770939233648874?p=eZnvXZb31xc861VGzU
Meeting ID: 277 093 923 364 887 4
Passcode: Ls3mH2h3

Talk 1: Federated Confidential Computing with NVIDIA FLARE
Lightning Talk, introducing the on-prem Federated Confidential Computing
Speaker : Chester Chen, Senior Product Manager & Engineering Manager, NVIDIA

Talk 2: What Federated Training Delivered: Co-Folding and ADMET Results from Apheris Pharma Networks, Apheris

Apheris builds superior drug discovery models through federated data, powered by NVIDIA FLARE. In this talk, we go through our co-folding and ADMET results. For co-folding, we show how we designed the benchmarks for the federated model and how OpenFold3 was trained on five pharma companies’ proprietary structural data, inside their security boundaries, in under ten weeks. For ADMET, we show how the network is set up and how it learns across pharma partners’ private measurements. We also cover how the resulting models compare with public baselines, how purpose-built data generation extends the networks, and how Apheris Foundry lets teams run, customize, and integrate the models in their own environment.

Speakers:
Robin Röhm is the CEO and Co-Founder of Apheris, where he focuses on building enterprise AI applications and federated data networks for drug discovery. With a background in Medicine, Philosophy, and Mathematics, he brings a multidisciplinary perspective to the scientific and computational challenges of working with distributed, IP-sensitive data. His work centers on enabling pharma companies to run, fine-tune, and benchmark drug discovery models securely within their own environments and to collaborate through governed data networks without sharing raw data. Before founding Apheris, Robin built life sciences data start-ups working with sensitive genomic and patient data.

Ian Hales is Senior Director of AI Engineering at Apheris, where he leads the engineering and research teams driving the success of its federated drug discovery models and Foundry applications. With a PhD in Computer Vision and over a decade of industry experience in machine learning, he combines hands-on expertise in model optimisation and software engineering with a track record of building and scaling ML teams. His work focuses on the technical delivery of federated co-folding and ADMET models, from research and benchmarking to deployment within pharma environments. Before joining Apheris, Ian held engineering leadership roles at Graphcore, developing and optimising AI models for specialised hardware.

Talk 3: Federated Learning Across Chemical Space: Lessons from Lilly TuneLab, Eli Lilly & Company

Lilly TuneLab is an initiative designed to enable cutting-edge biotech partners to collaboratively improve predictive models in drug discovery without revealing the proprietary data that makes those models valuable. By allowing partners to contribute to shared model improvement while keeping their compounds and assay data private, TuneLab aims to unlock the collective signal embedded across the broader biotech ecosystem.
Achieving this requires federated learning because sharing raw chemical structures or biological assay results can expose commercially sensitive IP. NVIDIA FLARE provides the framework needed to coordinate training across institutions without centralizing data.
However, federated learning in drug discovery presents unique challenges beyond those in other domains. Partners operate across vastly different regions of chemical space, meaning local data distributions will be highly non-IID. Compounding this, assay variability — differences in protocols, readout formats, and noise characteristics across labs — introduces additional heterogeneity that standard federated aggregation strategies struggle to handle.

In this talk, we will describe how TuneLab addresses these challenges through deliberate choices: from underlying model architecture to federated training and evaluation strategy. We will share insights from deploying these approaches across a growing network of partners and discuss what we have learned about making federated learning practical at scale in real-world drug discovery settings.
Speaker
Sam Holo, PhD – Sam is a Federated Learning Data Scientist on the TuneLab technical team at Eli Lilly & Company. Prior to Lilly, he received a PhD from the University of Chicago and worked as a forward-deployed engineer at McKinsey & Company and Rhino Federated Computing.

You may also like