Lung niche architecture

Model inputs: cell-type counts by depth bin, neighbouring cell-type pairs, cell count and spatial span. Mixture components summarize these features. Geometry is a covariate, not a learned anatomical outline. The neighbour term uses a composite likelihood, not an independent-edge joint probability model.

Source contribution: GSE250346 measurements, cell annotations and published cNiche/tNiche labels originate from Vannan et al., Nature Genetics (2025). That study already investigated spatial niches and airspace remodeling. These results are not new data generated by this website or an analysis endorsed by the source authors.

Added here: region grouping with the mixture model, derived comparisons and an interactive explorer. Cartoons show selected source regions; cell positions are measured, tissue shading is schematic. They are not model-generated anatomical reconstructions.

Not established: independent-cohort replication, superior performance over the published analysis, or validated disease-severity prediction. Published-label agreement on the same tissue is descriptive, not independent validation. Core-level association tests shown here do not account for repeated cores from the same patient.

Reproducibility: model implementation nichearchetype/spatial_em.py; current fit records results/resplit_20260912/model_airsac/fit_airsac.json and results/resplit_20260912/model_non_alveolar/fit_non_alveolar.json. Both use pair_weight=1.0 and count_and_span. Cell count enters both the size and geometry terms; that duplication needs an ablation, not a novelty claim. Source identities and label-file hashes are provided in the result panels.