Jiayu Su

Our expanded preprint, SONIC, is out! Spatial omics maps thousands of molecular patterns per sample. SONIC detects patterns amid noise and compares them across tissues, even when samples cannot be aligned.

1/ Human eyes are great at spotting spatial patterns under a microscope. 🔬 So why do computational methods struggle so much? Our new work reveals the mathematical answer + provides a scalable solution 📄

Physical frequencies provide a shared language for spatial patterns, while scalable compute enables large gene-centric atlases. We embedded >529,000 expression patterns from 58 mouse and human Stereo-seq embryo sections to trace gene dynamics across species, all on a MacBook Air!

Big thanks to Jun Hou, Haoyu, Dian, Xiao, David and Raul! This is a truly fun project in a fast-changing world. Gemini 3 Pro proofread v1, and GPT-6 Astra helped fixed more bugs. Now just waiting for the sub-n log n FFT to scale SONIC further 😉 Code:

7/ Scalability: 1 million spots in under 1 second Modern spatial data is massive. Leveraging the shared quadratic form, we optimized pattern detection (FFT for grids, implicit solvers for irregular shapes) to be orders of magnitude more efficient.

8/ Finally shoutout to my co-authors, especially Jun Hou Fung for extending the work to cell-type-specific SVG tools as well. And to Gemini 3 Pro: Your PI might skip the math in the Supplementary Notes, but G3P surely read every equation. 🤖📝

RT @NatureBiotech: Mapping isoforms and regulatory mechanisms from spatial transcriptomics data with SPLISOSM

[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

[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

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!)