Overcoming segmentation bottlenecks in brain tissue for spatial proteomic inference at population scale

In the quest to accurately infer spatial proteomics at a population scale, overcoming segmentation bottlenecks is crucial. This is particularly challenging in human brain tissue, where cell bodies exhibit complex morphologies. To tackle this issue, we present SPInfer, a workflow created using manually annotated cell bodies from 5,832 images of the human prefrontal cortex.

By benchmarking eight segmentation methods, including classical, convolutional, and foundation models, we found that adapting Cellpose-SAM to this task significantly improved cell segmentation accuracy from 0.37 to 0.84. This enhancement allows for the recovery of cell compositions that align well with matched single-nucleus RNA-seq and spatial transcriptomics data.

Across 14,765 images from 69 postmortem donors, our approach revealed that Alzheimer's disease pathology models, when using zero-shot masks, did not find significant associations. In contrast, fine-tuned masks identified four associations, three of which linked higher OLIG2-positive fractions to increased neurofibrillary tangle and plaque burden.

Furthermore, employing fine-tuned masks, we developed a graph attention network that could classify four cell types without the need for marker information. This network also performed differential expression analysis, revealing a correlation between cell morphology and gene expression. By fine-tuning foundation models to segment complex tissues, SPInfer offers a pipeline that paves the way for biological discovery in spatial proteomics at a population scale.