Thymus-derived myeloid programs track microglial tolerance states across human cohorts

Microglial functional positioning varies across individuals and disease contexts, yet conserved biological axes supporting reproducible cross-cohort inference remain poorly defined. Intrathymic T cell development offers a paradigm in which chemokines, extracellular matrix (ECM)/adhesion factors, and axon guidance cues jointly govern immune cell migration and selection. We asked whether thymus-derived transcriptional programs provide a transferable basis for quantifying microglial tolerance-like positioning across independent human cohorts.

We applied consensus non-negative matrix factorization to Tabula Sapiens thymus single-cell transcriptomes to derive eight Myeloid Education Signatures (MES); under formal enrichment testing, two were significantly enriched for ECM/adhesion biology and the rest were treated as data-driven programs without a dominant axis. MES scores were projected into four independent microglia cohorts (SEA-AD, Olah, Tuddenham, multiple sclerosis; >100 donors) and related to a donor-level tolerance score (homeostatic minus activation gene-set activity). Glucocorticoid receptor (GR) activity, calibrated from an external perturbation dataset, was evaluated as a moderator. Cross-cohort synthesis used random-effects meta-analysis with sensitivity analyses and multiple testing control.

Seven of eight MES modules were inversely associated with tolerance positioning; the eighth (MES08, defined by CCL2/CXCL2 chemokine genes) showed the most heterogeneous coupling across cohorts. The association strengthened monotonically with microglial purity in three of four cohorts, was recapitulated in Visium spatial transcriptomics, and was robust to covariate adjustment, housekeeping-gene removal, and leave-one-dataset-out sensitivity. This coupling was preferential to thymus-derived modules: modules built identically from six other tissues (blood, spleen, liver, lung, bone marrow, lymph node) coupled 1.2 to 2.9 fold less strongly, across module-number and feature-selection choices. Under a formal per-module interaction model, GR activity did not moderate the association, a well-powered null reproduced with a canonical GR panel and an external cohort. Irisin pathway (five-gene FNDC5/PPARGC1A module) associations were modest and largely non-significant after correction.

Thymus-derived myeloid programs constitute a transferable, tissue-preferential axis inversely coupled to microglial tolerance-like positioning, without detectable glucocorticoid moderation.