Multi-scale modeling of human tissues from spatial transcriptomics with TERRA
bioRxiv2026Preprint
A graph-transformer foundation model that reads tissue structure from spatial transcriptomics across scales, from single cells to whole tissues.
🧬 AI × Bio | Machine Learning Scientist & Engineer
Sebastian Birk
Principal Research Scientist · Wellcome Sanger Institute (from October 2026)
I build machine learning systems for spatial biology — the models, and the infrastructure that makes them usable.
Two of those systems are mine end to end: NicheCompass, published in Nature Genetics, and TERRA. Both are first-authored, open source and in other people's hands.
My background is half research, half engineering: a doctorate at Helmholtz Munich and the Technical University of Munich, and seven years in data science consulting — the last five building production ML platforms on Azure and Databricks for pharma, latterly alongside the PhD. I still consult. I care about work that survives contact with real data and real users.
Always happy to hear from people working at the intersection of AI and biology.
- 🧬 Generative & foundation models for biology
- 🔬 Spatial transcriptomics & single-cell genomics
- 🧠 Representation learning & self-supervised learning
- 🕸️ Graph machine learning
- ⚙️ Scalable ML systems & scientific software
Currently
Agentic systems for spatial genomics — bringing histology, spatial and single-cell omics, clinical data and knowledge graphs into a single multimodal analysis loop.
Software All software →
- NicheCompass★ 124
End-to-end analysis of spatial multi-omics data. A graph deep-learning method that models cellular communication to learn cell-niche representations which stay interpretable, because each latent dimension corresponds to a spatial gene programme representing a known biological process — cell–cell communication, cellular metabolism or transcriptional regulation.
First author and lead developer
- TERRA★ 62
A self-supervised foundation model for spatial transcriptomics, built on a graph transformer with a Joint-Embedding Predictive Architecture. Pretrained on 112 million cells across 20 human tissues, it learns reusable representations at gene, cell and neighbourhood scale.
First author and lead developer
- MintFlow★ 27
Separates microenvironment-induced from cell-intrinsic gene expression, so that tissue microenvironments can be both mapped and reprogrammed.
Co-first author
Selected publications All publications →
Mapping and reprogramming human tissue microenvironments with MintFlow
bioRxiv2025Preprint
Disentangles the gene expression a cell owes to its microenvironment from what is intrinsic to it, making tissue niches something you can steer rather than only observe.
Quantitative characterization of cell niches in spatially resolved omics data
Nature Genetics 57(4), 897–9092025Journal
NicheCompass — a graph deep-learning method that learns interpretable cell-niche representations from spatial omics, grounded in known spatial gene programmes for cell–cell communication, metabolism and transcriptional regulation.
News
- G2T: reconstructing tissue from gene expression with embedding-distance flow matching.
- Learning discrete cell and niche codes with dual residual vector quantization.
- TERRA is out — a multi-scale foundation model for human tissues from spatial transcriptomics.
- Graph Tokenization Meets JEPA accepted at the ICML 2026 Workshop on Graph Foundation Models.
- SQUINT accepted at the ICLR 2026 Workshop on Machine Learning for Genomics Explorations.