(7/N) What’s more, HiPoNet’s learned feature weights are interpretable. For example, immune-related markers like CD11b, CD118, and FOXP3 consistently emerge as important across views, aligning with known tumor-immune biology.

(6/N) We find that HiPONet consistently outperforms other methods on these datasets.

(5/N) We then evaluate HiPOnet on patient level classification tasks using single cell and spatial proteomics data:

(4/N) We first check that HiPoNet preserves topological features of datasets by predicting persistence from the datasets:

(3/N) HiPoNet models each cloud as a simplicial complex, and uses simplicial wavelet transforms to capture point-cloud level embeddings. Not only that---it uses multiview learning, i.e. different projections of the features to capture different underlying factors in the data.

(1/N) Thrilled to share that our paper HiPoNet (High dimensional Point cloud Network) to be presented at NeurIPS 2025! HiPoNet treats an entire high-dimensional point cloud as a datapoint! It captures multi-scale geometry and topology of the cloud perform classification and regression tasks.

Join us for an exciting session in @jointmath.bsky.social tomorrow! Here is our revised schedule!

Merry Xmas!

(9/n) GSPA-Pt can be used to map patient sample manifolds. We mapped 48 melanoma patient scRNA-seq samples and classified response from the patient embedding using logistic regression. The GSPA-Pt gene embeddings achieved the highest classification performance.

(8/n) GSPA-LR concatenates ligand (L) and receptor (R) for a pair representation. LR pair modules captures a diverse range of LR profiles, finer the cluster-level analysis. For example, in skin cells, Module 5 includes Ccl5–Ccr5 link present in AG and AG CPI, epithelial, myeloid and T cells.

(7/n) With the Nik Joshi Lab at Yale we presented a new dataset of 39K CD8+ cells from LCMV infections. Interestingly only GSPA-based gene localization finds genes associated with type 1 interferon signaling. DE requires clustering, but DE genes in clusters do not reveal this signature!

(6/n) GSPA facilitates a measure called gene localization, via comparison to a uniform distribution. For PBMCs, cell embeddings using all genes versus only top localized genes showed similar structure, suggesting that localized genes capture cell-type variation and geometry.

(5/n) We show that for (PBMCs), GSPA with compressed wavelets grouped cell-type specific genes from PanglaoDB. In embryoid-body lineages early trajectory genes linked to embryonic stem cells (NANOG, POU5F1) and late trajectory genes linked to hemangioblast-specific signatures (CD34, PECAM1, TAL1).

(4/n) Rather than transposing the cellxgene matrix for gene embeddings, we frame genes as signals on a cell–cell graph, decomposed using graph-based diffusion wavelets. Diffusion wavelets use differences of powers of P. Then we then reduce dimensionality with an autoencoder.

Happy thanksgiving!!

Up Mont Royal in the rain! For those of you who don’t know I’m on a one-semester sabbatical to Montreal!

Pictures from @logconference.bsky.social meetup Montreal!

Siddharth from our lab explaining protscape and scattering to CT Governor Ned Lamont and others at the AI summit hosted by @boehringer.bsky.social and BioCT arxiv.org/abs/2410.20317!!

Learning on graphs meetups this year!