Wish I were in Baltimore for this! Looks super exciting and forward thinking.

Last talk of this session! #NeurIPS2025

Next up! #NeurIPS2025

First up! Nine minutes!!

Please join us for these incredible papers and discussions with authors on important issues for the NeurIPS community 🤩

GitHub - ragnhildlaursen/NNMF: Neighborhood Non-Negative Matrix Factorization Neighborhood Non-Negative Matrix Factorization. Contribute to ragnhildlaursen/NNMF development by creating an account on GitHub.

We are really proud of this work. Please try out NNMF on all of your gene count spatial transcriptomics data, whether you need hard clusters or scalable, interpretable, spatially aware dimension reduction! Feedback welcome!! github.com/ragnhildlaur...

genes characterizing the 30 NNMF factors in the CRC data.

On these CRC data, we studied the factors based on their top ten genes. We found immune-dominated factors & factors capturing intra- and peri-tumoral stroma, among others. Importantly, some factors were shared x patients and some were patient specific, characterizing tumor-specific immune responses.

Top row: cell types in two patient samples; Bottom row: NNMF signatures in the same two samples, showing substantial complexity.

Then, we applied NNMF to MERFISH data publicly released by Vizgen (vizgen.com/data-release...) that includes 500 genes in ∼1.9 million cells from two human colon cancer samples. NNMF showed enormous complexity, where each factor included many cell types and identified detailed biological structure.

NNMF factor 1 and factor 7 across the 3D aligned brain slices.

On the same MERFISH mouse brain data, we aligned the eight parallel slices and ran NNMF on the 3D aligned data. NNMF easily labeled the important regions in 3D, and smoothed the factors across all three dimensions.

Manually annotated brain sample.

hard clustering for NNMF + Kmeans, BASS, and MENDER on two parallel samples.

All of the ten factors from NNMF in a single sample, colored by weights.

Next, we ran NNMF on MERFISH single mouse hypothalamus data with eight parallel slices on each individual slice (2D). NNMF + K-means produces hard clusters that match the manual clustering well. But the real story is how much detail and biological complexity soft clusterings add. Vasculature!

run time comparison for human brain and mouse merfish across BASS, MENDER, and NNMF.

Hard clusters from a manual annotation, NNMF's top signature, NNMF+K-means, BASS, and MENDER on the human brain data.

On human brain 10X Visium data and mouse brain MERFISH data, we compared MENDER and BASS to NNMF in terms of run time, and found that MENDER is fastest and NNMF is a close second. However, MENDER uses cell type labels for the hard clustering, not gene counts, and produces poor clusterings.

Benchmark comparison across five datasets and three metrics, for 15 different methods.

We use the very cool hard clustering benchmark system pubmed.ncbi.nlm.nih.gov/38491270/ and compared NNMF to fourteen state-of-the-art spatially-aware hard clustering methods, showing good performance of NNMF even in the hard clustering scenario.

Graphical description of Neighborhood NMF, including its input and how we determined the number of factors.

NNMF works by using standard NMF updates, but using Gaussian smoothing of the factor weights on each spot at each iteration that encourages similar weights for spots nearby in space. No matrix inversion needed!

example of factor weights on mouse brain sample, and the gene programs that define each factor. Then we use K-means to built a hard clustering.

NNMF is available in R, performs nonnegative matrix factorization on the gene counts that yields soft clusterings of every spot in spatial transcriptomics, and scales to many samples, arbitrary dimensions, & millions of spots. We run K-means on the soft cluster weights to get a NNSF hard clustering.

Neighborhood nonnegative matrix factorization identifies patterns and spatially-variable genes in large-scale spatial transcriptomics data Tissues consist of multi-cellular neighborhoods in which different cell types express correlated gene programs due to shared signaling environments. Methods for identifying these spatial neighborhoods...

Super excited to tell you about our preprint on Neighborhood Nonnegative Matrix Factorization (NNMF) for spatially-aware dimension reduction in spatial transcriptomics, led by Ragnhild Laursen in collaboration with Karin Pelka @pelkalab.bsky.social and her lab!

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

Exciting update!! @bioimagearchive.bsky.social is now hosting the first publicly available Incucyte data! If you have live-cell imaging data, please consider uploading to this amazing repository!! Thanks to Julia Carnevale and Alex Marson for experimental data —

www.ebi.ac.uk/biostudies/b...

Me three!

Feedback welcome! And please play with these data! There is a lot more signal there.

Thank you to @bioimagearchive.bsky.social for hosting these Incucyte image data -- this is a new thing for them, and they have been so kind in working through the details of submission (link coming soon!)! 🎉

GitHub - vanvalenlab/Caliban-2024_Schwartz_et_al Contribute to vanvalenlab/Caliban-2024_Schwartz_et_al development by creating an account on GitHub.

Try out Caliban and Occident on your own Incucyte data! More phenotypes and analyses added regularly.

github.com/vanvalenlab/...
github.com/bee-hive/occ...

GitHub - bee-hive/occident: Github repo for Occident Website hosting Live Cell Image Data Github repo for Occident Website hosting Live Cell Image Data - bee-hive/occident

With five new collaborations in the works, and a paper characterizing the differences using explainable AI already accepted as an oral presentation at #PSB2025 (lead by high school senior Marcus Blennemann), look for future work in this space!
www.biorxiv.org/content/10.1...

In summary, we found that, compared to the SH KO control condition, TCR T cells with the RASA2 KO have a longer dwell time and cripple cancer cells more effectively this way, whereas TCR T cells with the CUL5 KO proliferated more frequently upon activation, adding more T cells to the fight.

With a Markov model, we deconvolved when, in frame t-1, there is one cancer cell and one T cell in a window, and in frame t there is one cancer cell and two T cells. We were able to quantify how often this doubling of T cells attacking a cancer cell was due to proliferation or due to recruitment.

Most thrilling is that we can identify active T cells based on relative cell size and morphology, and watch T cells activate (differentially based on condition) after interacting with cancer cells.

Even more exciting, the speed of cancer cells decreased after interactions with T cells, as did their overall size (indicating stress).

While the # of T cell--cancer cell interactions increased similarly, these interactions & their effects were modulated by the CRISPR KOs. E.g., the time a T cell remained attached to a cancer cell (as estimated by a negative binomial and Markov model separately) was highest in RASA2 KO T cells.

Cancer cell and T cell morphology changes dramatically depending on state. These changes are visible in the brightfield imaging – active interacting T cells are larger and change to less circular shapes. Cancer cell begin to aggregate together when interacting with T cells.

We found that the number of T cells attached to cancer cells reduces the likelihood that the cancer cell will proliferate, with the beneficial KO T cells having greater effects on proliferation reduction.

We can study differences in cancer cell division events (lower in beneficial KO T cells) and average T cell speed (faster in beneficial KO T cells).

We found that T cell proliferation increased in the two beneficial KO T cells, in the CUL5 KO T cells in particular.

With the masked, tracked cells, we went to work to develop Occident. We were curious how well the RFP markers captured cancer cell number; we found that RFP lags as a proxy for cancer cell numbers.