Methods: We integrated bulk RNA-seq with clinical follow-up, scRNA-seq, and spatial transcriptomics (Xenium, Visium) across the benign-to-anaplastic thyroid spectrum. Xenium neighbourhoods were clustered on cell-type composition and thyrocyte differentiation state into nine niches, transferred to Visium and external multi-histology data by an Elastic Net model using each spot with its neighbourhood context, and linked to survival.

Results: Nine niches resolved into four tumour-core niches ordered along the thyrocyte differentiation axis (N^normal-like, N^diff-high, N^diff-mid, N^diff-low) and five TME niches (perivascular, lymphoid aggregate, fibrotic, myeloid, neutrophil). Niche composition shifted monotonically with pathology: benign and differentiated tumours were built from N^normal-like and the perivascular niche, whereas anaplastic tumours were dominated by N^diff-low with the myeloid, neutrophil and fibrotic niches. Adjacency was structured, the neutrophil niche preferentially enveloping the least differentiated core, and the transferred model reproduced this architecture in external data. Beyond compositional expansion, niches were qualitatively reprogrammed: within a given niche, classifiers separating anaplastic from non-anaplastic tissue selected ECM-remodelling and inflammatory programmes. Myeloid markers recovered this way concentrated at the tumour-TME interface, tracking with ECM-remodelling signalling and reduced local T cells. Aggressive niche scores were associated with shorter overall survival.

Conclusion: ATC progression reflects both quantitative expansion of the myeloid, neutrophil and fibrotic niches and qualitative reprogramming of the same niches, with myeloid activity at the ATC front linking ECM remodelling to T-cell exclusion. Niche scores are prognostic, and targeting them may improve TKI and checkpoint inhibitor efficac