Lightnews — Scholar-powered news

Multimodal AI agents for capturing and sharing laboratory practice

1 2 6

Patricia Skowronek @patiskowronek.bsky.social · 1d

In labs, hands-on expertise is often lost because it's not written down. We leverage multimodal AI agents to capture & share expertise by analyzing video and speech to generate protocols, detect errors, and guide researchers. #AI #TeamMassSpec
📄 Preprint: doi.org/10.1101/2025... 1/🧵

We present a multimodal AI laboratory agent that captures and shares tacit experimental practice by linking written instructions with hands-on laboratory work through the analysis of video, speech, an...

1 1 6

Mann Lab @mannlab.bsky.social · 15d

16/
If this interests you:
🔁 Retweet the first post:
bsky.app/profile/mann...
⭐️ Give our github repo a star github.com/MannLabs/scP...
❓ tell us what you are going to do with #scPortrait

Mann Lab @mannlab.bsky.social · 15d

15/
This work was a fantastic collaboration:
@mannlab.bsky.social
@fabiantheis.bsky.social
@v-hornung.bsky.social
A big shoutout to all of our co-authors: Alessandro Palma,
Altana Namsaraeva, Ali Oğuz Can, Varvara Varlamova, Mahima Arunkumar, @lukasheumos.bsky.social, @georgwa.bsky.social

scPortrait integrates single-cell images into multimodal modeling

Mann Lab @mannlab.bsky.social · 15d

14/
Bottom line:
scPortrait turns microscopy images into a first-class modality for machine learning and multimodal foundation models, alongside RNA and proteomics 🔬💻
🔗 Preprint: doi.org/10.1101/2025...
🔗 GitHub: github.com/MannLabs/scP...

Machine learning increasingly uncovers rules of biology directly from data, enabled by large, standardized datasets. Microscopy images provide rich information on cellular architecture and are accessi...

Mann Lab @mannlab.bsky.social · 15d

13/
Our extensive documentation and tutorials make scPortrait easy to use and accessible.
And as part of @scverse.bsky.social it’s compatible with existing tools like scanpy, squidpy, bento-tools or Moscot 🚀🐍
mannlabs.github.io/scPortrait/i... #OpenSourceTools #Tutorial #CodeDocumentation

SPARCS, a platform for genome-scale CRISPR screening for spatial cellular phenotypes

Mann Lab @mannlab.bsky.social · 15d

12/
scPortrait already scales:
1️⃣ 120M+ single-cell images from image-based genetic screens
2️⃣ applied on patient-derived datasets to perform AI-driven morphology analysis
Refs:
1️⃣ BioRxiv2023 ➡️ www.biorxiv.org/content/10.1...
2️⃣ Nature 2025➡️ www.nature.com/articles/s41...

Forward genetic screening associates phenotypes with genotypes by randomly inducing mutations and then identifying those that result in phenotypic changes of interest. Here we present spatially resolv...

www.biorxiv.org

Mann Lab @mannlab.bsky.social · 15d

11/
We also ship a benchmark dataset of Golgi morphologies and use it to compare image featurization tools: #ConvNeXt, #SubCell, #CellProfiler

Mann Lab @mannlab.bsky.social · 15d

10/
✨ Embedding images into transcriptome atlases ✨
We use scPortrait to embed single-cell images from a @10xgenomics.bsky.social Xenium ovarian cancer dataset into the #SCimilarity transcriptome atlas (R2 = 0.65), recovering meaningful cell types

Mann Lab @mannlab.bsky.social · 15d

9/
✨ Morphology defined cell states ✨
Image embeddings generated with scPortrait resolve intra- vs extratumoral macrophages with distinct morphologies, linked to anti-inflammatory vs fibroblast-like programs

Mann Lab @mannlab.bsky.social · 15d

8/
✨ Transcriptomes from images ✨
Using optimal transport + flow matching, scPortrait generates gene expression directly from CODEX images, capturing canonical marker expression like TCL1A in germinal centers in the tonsil
#CODEX #flowmatching #OT

Mann Lab @mannlab.bsky.social · 15d

7/
With standardized single-cell image datasets in place, the key question is: what new biology can we unlock?
We highlight three use-cases for scPortrait

Mann Lab @mannlab.bsky.social · 15d

6/
The new .h5sc format provides fast random access to single-cell images for ML training.
It follows #FAIR data principles (findable, accessible, interoperable, reusable) and integrates with @scverse.bsky.social tools via AnnData.

Mann Lab @mannlab.bsky.social · 15d

5/
The scPortrait pipeline transforms raw input images step by step:
• stitch FOVs
• segment & extract cells
• output standardized .h5sc single-cell image datasets

From messy pixels → inputs ready for training 🖥️

Mann Lab @mannlab.bsky.social · 15d

4/
Problem: microscopy images are messy, fragmented, and hard to use for ML
Solution: scPortrait standardizes them into a new .h5sc single-cell image format turning 🔬microscopy images into a reusable resource for integrative cell modeling

Mann Lab @mannlab.bsky.social · 15d

3/
Microscopy images are:
📈 easy to acquire across scales (organism → subcellular)
🖥️ information-rich (cellular architecture, tissue structure, perturbation responses)
= 🚀 ideal fuel for foundation models of cell behavior

scPortrait integrates single-cell images into multimodal modeling

Mann Lab @mannlab.bsky.social · 15d

2/
AI has had major breakthroughs (#alphafold #chatgpt) & computational models can now detect patterns in complex datasets without external guidance 🧠🖥️

🧬 biological datasets often contain entangled information making them complex to interpret →🧠🖥️ + 🧬 = unlock new biology

1 1 2

Mann Lab @mannlab.bsky.social · 15d

1/
Tomorrow’s large scale cross-modality models will further unlock biology, but they need standardized inputs. With @sophia-maedler.bsky.social & @nik-as.bsky.social, we built #scPortrait, @scverse.bsky.social package to turn microscopy images into single-cell image datasets for multimodal modeling

1 2

Mann Lab @mannlab.bsky.social · 15d

Our preprint on scPortrait is out! We built a framework + format to turn microscopy into standardized single-cell image datasets. #scPortrait scales >100M cells, integrates with @scverse.bsky.social, & enables cross-modality modeling from morphology to transcriptomics
doi.org/10.1101/2025...

Machine learning increasingly uncovers rules of biology directly from data, enabled by large, standardized datasets. Microscopy images provide rich information on cellular architecture and are accessi...

1 2 8

Mann Lab @mannlab.bsky.social · 25d

Congratulations on the inaugural of #AITHYRA, the new Biomedical AI institute in Vienna. Thank you for the invitation to speak at the symposium and the fascinating discussions on AI × proteomics. #AIforLifeScience

4

Mann Lab @mannlab.bsky.social · 29d

Thanks CPR, reNEW, @ucph.bsky.social
Full story:
www.sciencedirect.com/science/arti...

Mann Lab @mannlab.bsky.social · 29d

Withdrawal of the stem-cell niche WNR (Wnt-Noggin-R-spondin) cocktail in organoid cultures decreased proliferation and stemness, as well as improved differentiation, moving organoids toward more tissue-like states.

Mann Lab @mannlab.bsky.social · 29d

scDVP’s high-resolution enabled us to observe fine-grained protein abundance changes along the colon crypt axis. For instance, the abundance gradient of CA1 was only visible in organoid transplants and in vivo tissue.

Mann Lab @mannlab.bsky.social · 29d

We further employed scDVP to explore differences in individual intestinal epithelial cells. This dataset comprised about 2,700 proteins across cell types, and confirmed clearer stemness-to-differentiation programs after organoid xenotransplantation.