📣Excited to share my last postdoc paper with
@soumya-boston.bsky.social on eQTL mechanisms depending on where the RNA is in the cell! @broadinstitute.org @harvardmed.bsky.social
TL;DR:Early RNA eQTL variants in the nucleus and late RNA eQTL variants in the cytosol have distinct molecular mechanism🧵

Thanks to our wonderful co-authors Lakshay Sood, Michelle Curtis, Laurie Rumker, Daniel Stein, Mukta Palshikar, Saba Nayar, Andrew Filer, Helena Jonsson, and Ilya Korsunsky, and to @soumya-boston.bsky.social for his mentorship!

11/11 Try it!
VIMA is fast (<2hrs on single GPU for 75-sample ST dataset), needs minimal parameter tuning (we trained the autoencoder with identical hyperparameters on all our datasets), and is pip-installable. Check out our repo (including demo) at github.com/yakirr/vima

10/11 Why is this exciting?
- VIMA finds new biology beyond traditional spatial case-control approaches
- VIMA works across diverse spatial modalities (MERFISH, CODEX, IHC)
- VIMA avoids cell segmentation & manual annotation

9/11 And also!
We used VIMA to link TNF inhibition strongly to loss of lymphoid aggregates in ulcerative colitis (CODEX), to find new patterns of fibroblast organization in rheumatoid arthritis (IHC), and more.

8/11 For example
In an Alzheimer’s spatial transcriptomics dataset, VIMA separated dementia cases from controls with high accuracy -- and the spatial structures that it found included both known signals and a novel oligodendrocyte-rich cortical layer 6 niche enriched in dementia.

7/11 VIMA works in many spatial modalities
We applied VIMA to three datasets spanning three really different spatial modalities. In each case we found signals that were either impossible or very difficult to find with the standard paradigm.

6/11 Simulations
We showed in large-scale simulations that VIMA (blue) can powerfully and accurately identify many different kinds of spatial signals, and that it does much better than simpler alternatives that either don’t use a VAE or don’t use microniches.

5/11 How does VIMA work?
- It learns fingerprints via a ResNet18-style conditional VAE that removes sample and batch effects, then builds a nearest-neighbor graph to define microniches (A-C).
- It rigorously tests for case-control associations at global & microniche levels (D-E).

4/11 How is VIMA different?
It uses a conditional VAE to extract “fingerprints” capturing core tissue biology, defines microniches (small, overlapping groups of patches with similar fingerprints), and uses high-dimensional statistics to identify disease-associated microniches.

3/11 Fancier approaches
Excellent new methods are coming out for annotating spatial neighborhoods with more sophistication, but they still bin spatial neighborhoods into mutually exclusive niches, and they’re also susceptible to batch effects that limit case-control comparisons.

2/11 The current paradigm
Traditional case-control approaches annotate spatial neighborhoods either with average cell-type abundances or average transcriptional profiles. This can overlook key signals because it ignores local spatial relationships within each neighborhood.

1/11 The problem
How can we detect disease-associated tissue features in spatial data without forcing cells into predefined types or spatial neighborhoods into discrete niches?

Excited to tell you about VIMA (Variational Inference-based Microniche Analysis), our deep learning + stats hybrid for case-control analysis of multi-sample spatial molecular datasets. 🧵👇 www.biorxiv.org/content/10.1... @soumya-boston.bsky.social @broadinstitute.org