The abstract is read at the publisher; the summary is JClub's.
SAFFRON shows that, across spatial transcriptomics cohorts, only two spatial foundation models capture global spatial gradients more accurately than naive baselines, while none capture local plaque microenvironment patterns.
- Why it matters: Spatial gene expression reflects both local microenvironment effects and global spatial gradients, but foundation-model embeddings are hard to interpret and it was unclear whether they truly learn multi-scale variation.
- What they did: The authors built SAFFRON, a sparse autoencoder framework using a Matryoshka SAE to decompose SFM embeddings into sparse, interpretable features and then test feature correlations with known spatial-variation sources.
- The result: Benchmarking several SFMs, they found global gradients are learned better by only two models, with gradient signals concentrated in a small subset of sparse SAE features; local A{beta} plaque microenvironment patterns are not learned more accurately than baselines.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.