Summary
A bioRxiv preprint describes ProxiNet, a framework that learns cellular proximity from spatial transcriptomics and applies those relationships to dissociated single-cell RNA sequencing. The authors report spatial validation and use the approach to identify inferred neighborhoods in an Alzheimer dataset.
A bioRxiv preprint introduces ProxiNet, a computational framework designed to infer which cells are likely to be near one another in dissociated single-cell RNA sequencing data. The method learns patterns of pairwise cellular proximity from spatial transcriptomics datasets, then transfers those patterns to data from which the cells’ original positions have been lost.
The authors report that ProxiNet predicted proximity across brain regions and spatial technologies, including what they describe as zero-shot transfer between technologies. In spatial datasets with known coordinates, the inferred cellular neighborhoods recovered reproducible tissue organization and anatomical structure. The work is a preprint, and its abstract does not provide sample sizes or detailed performance measurements.
How ProxiNet adds spatial context
Spatial transcriptomics measures gene activity while retaining information about where cells are in tissue. By contrast, dissociated single-cell RNA sequencing profiles individual cells after tissue has been separated, providing transcriptomic data without preserving their original arrangement.
ProxiNet uses the spatial data as a reference. It learns transcriptomic signatures associated with pairs of nearby cells, then applies those learned relationships to dissociated data to identify inferred cellular neighborhoods. The result is a prediction of proximity based on gene-expression patterns—not a direct measurement of each cell’s position in the dissociated sample.
The authors also used gradient-based attribution to identify genes and broader transcriptional programs associated with the model’s proximity predictions. This offers a way to examine which expression patterns contribute to inferred relationships, rather than treating the neighborhood predictions as unexplained outputs.
Applications reported in the preprint
When applied to dissociated single-cell data, ProxiNet identified neighborhoods with differing cell-type compositions and candidate intercellular communication programs. In an Alzheimer dataset, the authors report age-associated remodeling of inferred neighborhoods. One Alzheimer-associated neighborhood at 8 months had distinct astrocyte and neuronal transcriptional states, alongside candidate communication programs.
These findings show how spatially learned relationships might add a neighborhood-level view to datasets that lack preserved tissue coordinates. The Alzheimer analysis is an application of the computational framework: the neighborhoods are inferred from transcriptomic signatures, while the preprint reports validation of the approach’s spatial organization using separate datasets with known coordinates.