Moritz Gerstung

Integrated in vivo combinatorial functional genomics and spatial transcriptomics of tumours to decode genotype-to-phenotype relationships - Nature Biomedical Engineering A method integrates perturbation mapping with 10X Visium spatial transcriptomics to map tumour genetic complexity and heterogeneity.

This is a great set of methods for studying combinatorial effects of cancer mutations on spatial phenotypes.

Clever experimental design by @breinigmarco.bsky.social hijacking Visium and elegant analysis by @lomakinai.bsky.social and @elihei.bsky.social

www.nature.com/articles/s41...

How does tumour heterogeneity arise? How can we predict cancer cell plasticity? In 2 new studies, we trace #glioblastoma heterogeneity to a spatial cancer cell trajectory w. multimodal cell atlassing bit.ly/4mkrWgs & predict plasticity w. snRNA/ATAC+deep learning bit.ly/3FbI6Ic 🧵

Spatial biology of cancer evolution - Nature Reviews Genetics In this Review, the authors discuss our latest understanding of the spatial aspects of cancer evolution, including the roles of cancer subclonal structure, tissue architecture, and interactions betwee...

Spatial biology of cancer evolution go.nature.com/44hbib6 #Review by Zaira Seferbekova, @lomakinai.bsky.social, Lucy R. Yates & @moritzgerstung.bsky.social
Free to read here: rdcu.be/c1jwe

Our experience with the multimodal seg is similar as yours- works great for many cell types or tissues, but not all.

Plus there can be contaminating transcripts on top.

Thank you! Yes, we believe segger has an edge over membrane-based segmentation because it also recognises the co-occurrence of transcripts. You can also use 10x multi-modal instead of nuclear segmentation as seggers input.

Classification and Personalized Prognosis in Myeloproliferative Neoplasms | NEJM Myeloproliferative neoplasms, such as polycythemia vera, essential thrombocythemia, and myelofibrosis, are chronic hematologic cancers with varied progression rates. The genomic characterization of...

and also myeloproliferative neoplasms.

Back then, the implementation was very clunky and could only be done by R experts.

ebmstate now makes the inference much easier with only a few lines of code.

www.nejm.org/doi/full/10....

Precision oncology for acute myeloid leukemia using a knowledge bank approach - Nature Genetics Peter Campbell, Hartmut Döhner and colleagues present an analysis of genetic mutations and clinical information from 1,540 patients with acute myeloid leukemia, demonstrating the utility of clinical k...

This type of model predicts a patient’s journey across several mid- and endpoints and relates the progression to hundreds of variables.

It was used to learn detailed prognostic models for acute myeloid leukaemia ..

www.nature.com/articles/ng....

Some classical statistics today.

ebmstate, an R package for multistate models with empirical Bayes covariate effect estimation.

Developed by Rui Costa during his postdoc in my group.

doi.org/10.32614/RJ-...

Segger: Fast and accurate cell segmentation of imaging-based spatial transcriptomics data The accurate assignment of transcripts to their cells of origin remains the Achilles heel of imaging-based spatial transcriptomics, despite being critical for nearly all downstream analyses. Current c...

From great collaborations come great things. Excited to share Segger, the solution to segmentation of spatial transcriptomics (ST) data, with the @steglelab.bsky.social and @moritzgerstung.bsky.social labs, spearheaded by the great Andrew Moorman and Elyas Heidari www.biorxiv.org/content/10.1...

Segger logo

Message passing intuition behind the segger’s link-prediction model and the network architecture

1/ New preprint! 🍳

@elihei.bsky.social and our team at @embl.org , @dkfz.bsky.social, and @mskcancercenter.bsky.social built #segger - a fast, accurate cell segmentation tool for spatial transcriptomics that assigns transcripts to their cell origins!

doi.org/10.1101/2025...

Also tagging Elyas Heidari @elihei.bsky.social here who led this fantastic work.

A great thanks goes to all the other authors and contributors from the Gerstung, Stegle and Peer labs who made this work possible.

GitHub - EliHei2/segger_dev: a cutting-edge cell segmentation model specifically designed for single-molecule resolved spatial omics datasets. It addresses the challenge of accurately segmenting indiv... a cutting-edge cell segmentation model specifically designed for single-molecule resolved spatial omics datasets. It addresses the challenge of accurately segmenting individual cells in complex ima...

Segger comes as robust open-source software and has already been tested by many in the community. github.com/EliHei2/segg...

Andrew Moorman and other members of @danapeer.bsky.social's lab helped carry out a rigorous assessment based on various 10x Xenium data sets with bespoke segmentation stainings, providing the ground truth to demonstrate segger's superior performance.

Segger is a super fast graph neural network algorithm, which makes cell segmentation much more reliable and faster.

Spatial transcriptomics holds great promise to understand biological tissue function, but the assignment of transcripts to cells has been a substantial bottleneck.

For this reason, Elyas Heidari, a student in my lab and in @steglelab.bsky.social built segger.

www.biorxiv.org/content/10.1...

Evidence Aggregator: AI reasoning applied to rare disease diagnostics Retrieving, reviewing, and synthesizing technical information can be time-consuming and challenging, particularly when requiring specialized expertise, as is the case of variant assessment for rare di...

New preprint! We worked with @msftresearch.bsky.social and @broadinstitute.org to see whether large language models (LLMs) can be useful to variant scientists in deciding whether genetic variants seen in a patient are responsible for their disease. tl;dr yes they can: www.biorxiv.org/content/10.1...

Looking forward to @moritzgerstung.bsky.social seminar @uob-ieu.bsky.social on 30th Jan 2025 at 1pm, ,“ Using AI to Predict Disease Risks” in person and on line bristol-ac-uk.zoom.us/j/94273829130

Amazing work. We‘re all walking laboratories for somatic evolution.

Somatic mutation and selection at epidemiological scale As we age, many tissues become colonised by microscopic clones carrying somatic driver mutations ([1][1]–[10][2]. Some of these clones represent a first step towards cancer whereas others may contribu...

Resharing here a recent X post. In this preprint, we introduce an improved version of NanoSeq, a duplex sequencing protocol with <5 errors per billion bp in single DNA molecules, and use it to study the somatic mutation landscape of oral epithelium in >1000 people. 1/ www.medrxiv.org/content/10.1...

Safe travels! We’re still waiting for the snow you got yesterday to arrive but chances look slim. Probably better travel wise.

Thanks. Was on a bit of a social media hiatus, but this place looks like a new hope.

Congratulations! Looking forward to seeing the outputs of your scientific ingenuity becoming multiplied.

Action Intent Links | Bluesky Authors, websites, and apps can use action intent links to implement "Share on Bluesky" buttons, or similar in-app actions. Logged-in users will be directed to the corresponding action view in the Blu...

We strongly suggest that academic publishers and other platforms that host research rapidly implement a Share to Bluesky button for their articles. Here's how:

docs.bsky.app/docs/advance...

#AcademicSky #HigherEd #Altmetrics

Example Paion report for a confirmed myxopapillary ependymoma diagnosis. Shown are H&E, prediction heatmaps and automatically selected representative image tiles.

3/3 Paion is a new algorithm to diagnose brain cancers from H&E images.

* distinguishes 102 types of brain cancer
* high confidence predictions in 50-70% of cases across 7 cohorts
* top1 acc > 85% in high conf subset
* 2 day turnaround v 3 weeks for molecular testing

Preprint coming soon.

2/3 Delphi-2M generative AI model for multi-disease risks

* a single GPT-derived model learns risks of 1257 diseases
* samples future life courses
* offers insights into multi-morbidity dynamics

www.medrxiv.org/content/10.1...

Multi-cancer risk stratification based on national health data: a retrospective modelling and validation study Data available in national electronic health databases can be used to approximate cancer risk factors and enable risk predictions in most cancer types. Model predictions generalise between the Danish ...

1/3 Cancer risk models based on national health registries

* trained 6.7M individuals from DK
* validate in UK
* disease history most important predictor for age >65
* family history predicts early onset cancers

www.thelancet.com/journals/lan...

5:20 start this morning for a virtual talk at the WEHI in Melbourne, covering:

* Cancer risk models
* Delphi multi-disease genAI
* Paion, a new brain tumor digital pathology algorithm

Summary below:

New at Science
Evo, a large language of life (LLLM) genomic foundation model, predicting & generating tasks from molecular to genomic scale
science.org/doi/10.1126/...

Delphi is cool: nanoGPT, adapted to generate (clinical ontology term, waiting time) tuples i.e. a continuous time Markov chain on clinical ontology terms, trained on patient histories. Predicts conditions on par with clinical indicators in current practice.