Incredible work by Idan Milo, Nofar Azulay, Yuval Bussi, Raz Ben-Uri, Tal Keidar Haran, Ofer Elhanani, Yotam Harnik, Oran Yakubovsky, Tomer Salame, Ido Nachmany, Martin Wartenberg, Philippe Bertheau, @dmichonneau.bsky.social and Gerard Socie
For many more details, including the fate of epithelial cells (with an interesting shift from paneth to endocrine cells associated with loss of stem cells), clinical correlates and more check out the paper! www.science.org/doi/10.1126/...
Finally, we identify inter-crypt heterogeneity in deterioration with spatially confined immune microenvironments. Deteriorated crypts may neighbor healthy, suggesting local processes driving this compartmentalization and underscoring the importance of spatially-resolved analyses.
To examine this we need to distinguish immune cells from the host and donor. How? Cool trick! We imaged the X and Y chromosomes in sex-mismatched samples. Turns out that host T and plasma cells dominate the gut in the first month following transplantation and persist for months!
But why do some patients have more CD8s and others more macs? Turns out that time after tranplantation is a major correlate for immune composition. Could it be that immune reconstitution from the hematopoietic stem cell transplant differs across immune cell types?
How does this change in GVHD? We know that donor T cells attack the recipient, but is this the whole story? No! CD8s were only enriched in some patients. We suggest a major role for plasma cells, macs and neutrophils, which correlate with worse clinical manifestation.
We used MIBI to profile 59 diagnostic biopsies from patients with GI GVHD and 18 healthy controls. Healthy gut was similar in its organization across individuals: CD4s in the base of the crypts, plasma in the bottom of the villi and macs at the tip, with some differences between colon and duodenum.
Paper alert! A spatial atlas of human gastrointestinal acute GVHD reveals epithelial and immune dynamics underlying disease pathophysiology. We examined how immune cells are organized in the gut and what happens to them in GVHD. TL;DR Timing is key! www.science.org/doi/10.1126/...
We’re adding new features to CellTune frequently – follow @celltune.bsky.social for updates.
Exceptional software engineers – DM to join us!
Congratulations to Yuval Bussi, Dana Shainshein, Eli Ovits, Robert Schiemann, and all the lab members who contributed data, labels and ideas for CellTune.
Analyzing spatial proteomics? Give CellTune a try: celltune.org
Developing methods? Benchmark with our cell label database: celltune.org/CellTuneDepot
Questions? Feedback? Interested in hosting a CellTune workshop for your lab? Reach out: celltune.org/contact
To make it accessible, we built CellTune into an intuitive software, adding advanced capabilities for visualization, gating, annotation and more. CellTune was designed specifically for spatial proteomics and has been a game-changer for our lab – we’ve now used it in 6 projects. 5/7
To benchmark accuracy, we generated gold standard consensus labels from 3 manual annotators. CellTune outperformed every existing method in both precision and recall! It also identified more granular cell types! 4/7
CellTune’s core innovation is an optimized human-in-the-loop workflow. It trains a model and iteratively refines it by prioritizing information-rich cells for human curation. This approach improves classification accuracy and uncovers rare or novel cell types. 3/7
New tools for cell classification in spatial proteomics emerge frequently—but without reliable labels, we’re stuck in a loop comparing inaccurate new predictions to inaccurate old ones.
CellTune delivers accurate results by learning from you! 2/7
🚨Preprint Alert!🚨
Cell classification is one of the most difficult tasks in analyzing spatial data. We present CellTune - a powerful toolkit for accelerating biological discovery in spatial proteomics.
📄 www.biorxiv.org/content/10.1...
👇🧵1/7
If I were a student - I would definitely join 🤩 Best PI and super cool science.
This project was a constant uphill struggle. We had so many ups and downs along the way and so many times when we thought that we’ve reached a dead end. It was only possible due to my unbelievably talented and strong-willed team. So proud of you!
We did many cool experiments to show that the model generalizes across tissues, understand what CombPlex learns and what are the advantages and limitations of the approach. You can read all of this in the paper: rdcu.be/eeZH2
We then tested it experimentally. We devised protocols to perform combinatorial staining of proteins, both for fluorescence microscopy and for mass-based imaging. It worked! CombPlex accurately reconstructed the single-protein images of 22 proteins from 5 channels.
We first tested this idea in simulations in silico. We used publicly-available CODEX data to simulate an experiment in which 22 proteins are measured in 5 imaging channels. It worked! CombPlex accurately reconstructed the single-protein images.
Although it is mathematically infeasible to differentiate protein barcodes, images contain information to constrain the solution space. E.g. protein signals are sparse, continuous and have specific coexpression patterns. Maybe neural networks can learn these directly from images?
Why wasn’t combinatorial staining applied to proteins?
To differentiate targets using combinatorial staining, they need to be spatially-resolved to generate a unique barcode. Proteins are ~10000X more abundant than mRNAs, so this assumption breaks for standard imaging conditions.
Protein-imaging approaches use a separate measurement channel for each protein. In contrast, mRNA-imaging approaches use combinatorial staining, whereby each target is encoded by several channels (e.g., a sequence of colors). This increases the number of targets exponentially.
In the last decade, imaging of proteins in tissues has increased to dozens, using approaches such as cyclic fluorescence. Meanwhile, imaging of mRNAs in tissues has increased to thousands using combinatorial staining (e.g., MERFISH or seqFISH). Why this discrepancy?
Coupling CombPlex with instruments for high-dimensional imaging could pave the way to image hundreds of proteins at single-cell resolution in intact tissue sections.
This is amazing work by Raz Benuri, Lior Ben Shabat, Omer Bartal, Shai Bagon @ Ofer Elhanani. So proud of you!
See 🧵 for details.
In CombPlex, every protein is imaged in several channels, and every channel contains agglomerated images of several proteins. These combinatorically-compressed images are then decompressed to individual protein-images using deep learning.
TL;DR Multiplexed imaging methods use a separate channel for each protein, inherently limiting their scalability. We devised CombPlex (COMBinatorial multiPLEXing) to exponentially increase the number of proteins that can be measured using any imaging modality from C up to ~2^C.
Super excited to finally see this published! Our new method to use combinatorial staining and deep learning to push the multiplexing boundaries of spatial proteomics!
@yardenasamuels.bsky.social opening #MICC2025 in Athens - the birthplace of Democracy. Looking forward to an amazing conference!
Cool!!!!!
