iiinova — where your digital self keeps evolving

Julius Upmeier zu Belzen

@juliusuzb.bsky.social

PhD Student working on neural networks for the human genome.

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

@fabiantheis.bsky.social , Thore Bürgel, @steinfeldtjakob.bsky.social , Benjamin Wild, Roland Eils, and everyone else at @bihatcharite.bsky.social who supported this work.

030

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

f you’d like to discuss the work, we’d love to hear from you! Huge thanks to all my co-authors: @larnoldt.bsky.social , Noah Hollmann, Luis Hermann, Khue M. Nguyen, Leonard Eckhoff, @lkohleick.bsky.social ,Sedra Abou Ghaloun, Hannah Schmidt, @stefanhgm.bsky.social , ...

130

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

This work was inspired by the omnigenic model proposed by Boyle, Li & @jkpritch.bsky.social (2017), as well as Ma et al.’s DCell model from Trey Idekers Lab.

130

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

We trained the model using data from the @ukbiobank.ac.uk and evaluated it externally in the All of Us Research Program. We’re extremely grateful to both programs and, especially, to their participants for making this research possible.

130

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

• Multi-task learning, improving prediction across phenotypes • Deep multimodal integration (demonstrated with NMR metabolomics) • Transfer learning for smaller disease-specific cohorts

120

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

The OGM enables: • Direct biological interpretability • Deep integration of covariates • Nonlinear modeling of variant–variant and variant–covariate interactions

110

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

The OGM achieves competitive predictive performance while learning biologically meaningful intermediate representations that can be directly interpreted. Those representations also make several things possible that are difficult with standard polygenic scoring approaches.

120

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

We have seen how powerful domain-specific architectures can be, from CNNs in computer vision to attention in language models. We saw an opportunity for a biologically-informed architecture for human genetics.

Neural network architecture schema: Variant and covariate inputs are mapped to gene nodes. These are mapped go GO-nodes, which are organised in a hierarchy that leads to a phenotype model.

150

Julius Upmeier zu Belzen @juliusuzb.bsky.social · 05/08/2026

🧬 New preprint! An interpretable omnigenic neural network architecture for the human genome We introduce the Omnigenic Model (OGM), a neural network architecture for human genetics. It uses biological system structure to aggregate genetic variation for disease risk prediction.

doi.org

1155

Reposted by Julius Upmeier zu Belzen

Lucas Arnoldt @larnoldt.bsky.social · 16/06/2025

NetworkVI is a group effort by @juliusuzb.bsky.social, Luis Herrmann, Khue Nguyen, supervised by @fabiantheis.bsky.social & Benjamin Wild & Roland Eils at @bihatcharite.bsky.social and @www.helmholtz-munich.de‬. Code: github.com/LArnoldt/networkVI 🧠 Open-source & ready to use!

github.com

GitHub - LArnoldt/networkvi

Contribute to LArnoldt/networkvi development by creating an account on GitHub.

032