Please check our SuppNote for more methodological details and our unsuccessful attempts with parametric models (and prove my PI wrong who claimed no one would read them!) https://t.co/isv554JpJI

[5/n] For differential usage analysis, we implement a conditional HSIC test that controls for spatial confounding by removing location effects from both isoform usage and potential regulators

[4/n] We validated SPLISOSM's three spatial variablity tests in simulation, confirming that they capture variability in the correct layer while producing calibrated, permutation-free p-values that are robust to data sparsity

[3/n] The statistical power of HSIC-based tests depends heavily on kernel choice. We introduce: (1) a full-rank graph-based spatial kernel with provably higher power and principled approximation if needed and (2) a new compositional kernel that handles sparse data with NA ratios

[2/n] SPLISOSM reformulates spatial variability (and differential usage) detection as *multivariate* kernel (*conditional*) independence tests. We employ the Hilbert-Schmidt Independence Criterion (HSIC) with theory-driven kernels for isoform compositions and spatial coordinates.

🧵How does SPLISOSM work? [1/n] Analyzing isoform patterns in ST data is challenging. Isoform usage is multivariate and compositional (sum-to-one), data is extremely sparse (<1 UMI per gene&spot), and for differential analysis, spatial autocorrelation generates false associations

[6/n] SPLISOSM works across ST platforms with isoform information (SiT, Visium fresh frozen, Slide-seqV2 etc). We envision SPLISOSM to uncover additional *transcript*-omic insights from existing data in health and disease😉. Code available at

[5/n] In glioma, microenvironment reshapes transcript diversity during tissue remodeling and immune infiltration. We found changes in antigen presentation, signal transduction, adhesion, cytoskeleton, and ribosome pathways (the spliceosome machinery even modifies itself!)

[4/n] Using *conditional* differential usage test, we linked isoform patterns to RNA binding proteins, revealing interesting regulatory mechanisms such as Arpp21 auto-regulation via miR-128-2, conserved SEPTIN8 selection in mouse and human brains, and even an "paradox" (Fig 5)

[3/n] We found 1000+ genes with spatial isoform variation in brain enriched in synaptic signaling pathways and neuropsychiatric disorders (functionally distinct from variably expressed genes). 25% of 3'end diversity affects protein coding. Many patterns conserved across species.

[2/n] Most genes produce multiple RNA isoforms, yet their spatial distribution is overlooked despite being captured by major sequencing-based ST platforms. SPLISOSM (SPatiaL ISOform Statistical Modeling) now completes the missing piece🧩 Methodology to come in a subthread⬇️

[1/n]📢How is your spatial analysis *transcript*-omic if transcripts are only aggregated to gene counts? Introducing SPLISOSM, our new computational method for learning the isoform landscape and its regulation from spatial transcriptomics data https://t.co/eBwnwA8Twx

[2/2] There are many cool applications in the paper. Please stop by at the poster and chat if you’re interested☺️ East Exhibit Hall A-C #1106 Wed 11 Dec 11 a.m. PST — 2 p.m. PST Paper: https://t.co/DWCwo4FTJR Code:

[1/2]📣I will be at #NeurIPS2024 this week to present our work on a multi-subspace extension of PCA for representation learning. We unify supervision and disentanglement through HSIC, and showcase how to learn separate and interpretable subspaces of interest for single-cell data

1/8 🧬 Excited to share our latest work on the diverse role of #macrophages in #atherosclerosis! We've developed a pipeline to quantify the disease relevancy across macrophage subpopulations using single-cell multiome data. Check out the paper at https://t.co/ahMCY2wQha

More details and examples in the paper https://t.co/38tE19vS8f. Don't forget to check out the package https://t.co/P6wJD1d6Wn and start your favorite *spatially informed* analysis right away!

[8/n] Finally a shoutout to the amazing team especially JB, @alexanderfuxi and my supervisors @david_a_knowles @RabadanLab! Excited to have the first major piece of my PhD out and it’s been an incredible journey so far. Hopefully more to come in 2024!🤞

[7/n] Most importantly, what new biology can we learn?👨‍🔬In colon cancer, we discover distinct spatial localization patterns of plasma cell subtypes (IgG+ in lesion and IgA+ in mucosa), which is reported of clinical interests, and link them to fibroblast-related matrix remodeling.

[6/n] Example 2⃣: Slide-seqV2 data is often too sparse to blend into single-cell atlases. Extending SCVI, we provide the first *spatially aware* joint embeddings of spatial and single-cell human prostate data, removing batch effects and reducing ambiguity in label transfer.

[5/n] When is spatial modeling useful? Always, but even more so with noisy data. Example 1⃣: When using deconvolution to map transcriptomic cell types to epigenomic CUT&Tag data, biologically coherent embryonic compartmentalization is revealed only under spatial regularization.