Multimodal profiling reveals tissue-directed signatures of human immune cells altered with age

Steven B Wells; Daniel B Rainbow; Michal Mark; Peter A Szabo; Can Ergen; Daniel P Caron; Ana Raquel Maceiras; Elior Rahmani; Eli Benuck; Valeh Valiollah Pour Amiri; David Chen; Allon Wagner; Sarah K Howlett; Lorna B Jarvis; Karen L Ellis; Masaru Kubota; Rei Matsumoto; Krishnaa Mahbubani; Kouresh Saeb-Parsy; Cecilia Dominguez Conde; Laura Richardson; Chuan Xu; Shuang Li; Lira Mamanova; Liam Bolt; Alicja Wilk; Sarah A Teichmann; Donna L Farber; Peter A Sims; Joanne L Jones; Nir Yosef

doi:10.1038/s41590-025-02241-4

. 2025 Aug 13;26(9):1612–1625. doi: 10.1038/s41590-025-02241-4

Multimodal profiling reveals tissue-directed signatures of human immune cells altered with age

Steven B Wells ^1,^2,^#, Daniel B Rainbow ^3,^#, Michal Mark ^4,^#, Peter A Szabo ^2,^#, Can Ergen ^5,^#, Daniel P Caron ^2,^#, Ana Raquel Maceiras ^6,^7,^#, Elior Rahmani ⁵, Eli Benuck ⁴, Valeh Valiollah Pour Amiri ⁵, David Chen ¹, Allon Wagner ⁵, Sarah K Howlett ³, Lorna B Jarvis ³, Karen L Ellis ³, Masaru Kubota ⁸, Rei Matsumoto ⁸, Krishnaa Mahbubani ⁹, Kouresh Saeb-Parsy ⁹, Cecilia Dominguez Conde ⁶, Laura Richardson ⁶, Chuan Xu ^6,⁷, Shuang Li ^6,⁷, Lira Mamanova ⁶, Liam Bolt ⁶, Alicja Wilk ^6,⁷, Sarah A Teichmann ^7,^10,^11,^✉, Donna L Farber ^2,^8,^✉, Peter A Sims ^1,^12,^✉, Joanne L Jones ^3,^✉, Nir Yosef ^4,^5,^✉

PMCID: PMC12396968 PMID: 40804529

Abstract

The immune system comprises multiple cell lineages and subsets maintained in tissues throughout the lifespan, with unknown effects of tissue and age on immune cell function. Here we comprehensively profiled RNA and surface protein expression of over 1.25 million immune cells from blood and lymphoid and mucosal tissues from 24 organ donors aged 20–75 years. We annotated major lineages (T cells, B cells, innate lymphoid cells and myeloid cells) and corresponding subsets using a multimodal classifier and probabilistic modeling for comparison across tissue sites and age. We identified dominant site-specific effects on immune cell composition and function across lineages; age-associated effects were manifested by site and lineage for macrophages in mucosal sites, B cells in lymphoid organs, and circulating T cells and natural killer cells across blood and tissues. Our results reveal tissue-specific signatures of immune homeostasis throughout the body, from which to define immune pathologies across the human lifespan.

Subject terms: Mucosal immunology, Lymphoid tissues

Wells et al. profile RNA and surface protein expression to describe dominant tissue-specific effects on immune cell composition and function across lineages in the human tissues across age.

Main

The immune system leverages a dynamic network of specialized cells spread across the body to defend against infections and cancer, regulate inflammation and repair tissue damage. Myeloid cells—macrophages, monocytes and dendritic cells (DCs)—initiate innate immunity at mucosal and barrier sites, while adaptive immunity is mediated by antigen-specific T and B lymphocytes in lymphoid organs. Immune memory is established following antigen-driven activation and differentiation of T cells and B cells, resulting in heterogeneous subsets of circulating and tissue resident memory T cells (T_RM) and B cells that persist in diverse tissues¹. With age, immune memory accumulates, although responses can become dysregulated, increasing susceptibility to infections, cancer and autoimmunity. As human immune cells and the effect of age are mostly studied in blood², we lack a comprehensive understanding of the effect of age on the majority of innate and adaptive immune cells that are maintained in tissues.

Investigating human tissue immunity across diverse ages has been difficult to achieve. Obtaining tissues from organ donors enables the acquisition of blood and multiple tissues from individual donors and the isolation of viable immune cells across lineages for phenotypic, functional and multimodal single-cell profiling. Previous studies of lymphocytes from organ donors showed that T lymphocyte, natural killer (NK) cell and innate lymphoid cell (ILC) subset composition, tissue residence and certain functional attributes are specific to the tissue^3–9, indicating that localization has a dominant role in determining the maintenance and functional responses of lymphocytes. Whether these tissue-specific effects on human T cells are exhibited by other immune cell lineages, such as B cells and myeloid cells, and whether aging exerts general or tissue-specific effects have not been established.

Here, we present a comprehensive analysis of human immune cells using cellular indexing of transcriptomes and epitopes (CITE-seq) to simultaneously profile transcriptomes and >125 surface proteins of myeloid and lymphoid-lineage cells in 14 tissue sites of 24 organ donors aged 20–75 years. We identified a crucial role of tissue on immune cell composition, function, homing and differentiation across myeloid and lymphocyte lineages, including signatures specific for the gut, lungs and lymphoid organs. Across age, site-specific immune cell composition was largely maintained, although age-associated changes in function, signaling and metabolism were identified in certain subsets and sites, including macrophages in the lung, B cells in lymphoid organs and CD8⁺ T cells across sites. Together, our findings reveal the complex interplay between tissue, lineage, subset and age in immune homeostasis that is important for defining immune dysfunctions in disease.

Results

Multimodal profiling identifies immune cells across tissues

We isolated mononuclear cells (MNCs) from blood, multiple lymphoid organs, lungs, airways, intestines and other sites using established protocols^7,9 from 24 donors (10 females and 14 males, aged 20–75 years) (Fig. 1a). Organ donors originated from New York City (USA) and Cambridge (UK) and were free of chronic infection, cancer and overt disease (Supplementary Table 1). We performed single-cell RNA sequencing (scRNA-seq) on tissues from all donors, including CITE-seq with 127 proteins from 22 donors (Supplementary Table 2). We obtained 1.28 million immune cell events from 10 sites with >75,000 cells per site, including blood, bone marrow (BM), spleen, different lymph nodes (LNs) including lung-associated LN (LLN), mesenteric LN (MLN) and inguinal LN (ILN), lungs, comprising bronchoalveolar lavage (BAL) and lung parenchyma, and jejunum (JEJ), divided into the intraepithelial layer (JEL) and lamina propria (JLP) (Fig. 1a). We also purified low numbers of immune cells from liver, skin, colon epithelium and colon lamina propria from 9 donors (4 females, aged 25–75 years; 5 males, aged 20–55 years). These sites were therefore not included in the annotated dataset below and are provided as a separate reference (Extended Data Fig. 1a).

Fig. 1 — a, Plot of cell numbers across 24 donors and 10 tissue sites (blood (BLO); BM; spleen (SPL); LNs, including ILN, LLN and MLN; lung (LNG), comprising BAL and parenchyma (PAR); and JEJ (divided into JEL and JLP)) and donor metadata including 10× Genomics sequencing chemistry (3′, n = 2; 5′, n = 22), sex (male, n = 14; female, n = 10), CMV serostatus (positive, n = 16; negative, n = 8), age (range, 20–75 years) and location of tissue acquisition (US, n = 12; UK, n = 12). Bottom bars depict the number of cells profiled from each donor, and right bars depict the number of cells in each tissue across all donors. b–d, UMAP embeddings colored by donor (b), tissue site (c) or immune lineage classified by MMoCHi (d) in the BLO, BM, SPL, ILN, LLN, MLN, BAL, PAR, JEL and JLP. Lymph., lymphocytes.

Extended Data Fig. 1 — a) Plot of cell numbers and donor metadata from tissue sites collected but excluded from analysis. **b-h**) UMAP embeddings of the dataset colored by (b) site of collection, (c) 10x Genomics sequencing chemistry, (d) sex, (e) age, (f) CMV serostatus, (g) EBV serostatus, (h) MMoCHi immune subset classification.

For data integration, we leveraged multi-resolution variational inference (MrVI), which is designed for cohort studies. MrVI harmonizes variation between cell states (for unified annotation of cell states across samples) and accounts for differences between samples¹⁰. Visualization with uniform manifold approximation and projection (UMAP)¹¹ showed similar results across US and UK donors, sequencing technologies and other donor covariates such as sex, cytomegalovirus (CMV) and Epstein–Barr virus serostatus (Fig. 1b and Extended Data Fig. 1b–g). Although cells from blood and lymphoid organs (BM, spleen, LNs) clustered similarly, cells from mucosal sites (lung, JEJ) clustered distinctly (Fig. 1c). For annotation, we used MultiModal Classifier Hierarchy (MMoCHi)¹², which leverages both surface protein and gene expression to hierarchically classify cells into predefined categories (Supplementary Fig. 1 and Supplementary Table 3). MMoCHi defined six major immune lineages found across all tissues (Fig. 1d) comprising 13 T cell, 5 NK/ILC, 6 B cell and 7 myeloid subsets (Extended Data Fig. 1h). These results showed broad and consistent representation of major immune lineages in our dataset.

Immune cell subset composition is specific to the tissue

We next analyzed the subset composition and heterogeneity for each immune cell lineage across tissues based on the MMoCHi annotations above. T lymphocytes (610,429 cells) comprised low-frequency γδ T cells, which develop early in ontogeny, and predominant αβ T cells (Fig. 2a,b and Supplementary Fig. 2). CD4⁺ and CD8⁺ T cells (αβ TCR⁺) were subdivided into naive (T_N), terminal effector (T_EMRA) and memory subsets, including effector-memory (T_EM), central memory (T_CM) and T_RM, along with CD4⁺ regulatory T cells (T_reg)^6,13,14 (Fig. 2a,b). Surface proteins were essential for identifying T cell subsets that were not fully resolved by scRNA-seq (Fig. 2b and Supplementary Fig. 3), as shown before^9,12,15. For example, surface CD45RA expression was required to distinguish CD45RA⁺ T_N cells from CD45RA⁻ T_CM cells and CD45RA⁻ T_EM cells from CD45RA⁺ T_EMRA cells, and surface γδ or αβ T cell receptor (TCR) expression to accurately identify γδ T cells from CD8⁺ T cells, which can express TRDC¹² (Fig. 2b). In addition, T_RM cells were distinguished from T_EM cells based on surface expression of CD69, CD103 and/or CD49a⁶ (Fig. 2b and Supplementary Fig. 3). T cell subsets were differentially distributed across sites; CD4⁺ T_N, CD8⁺ T_N and CD4⁺ T_CM cells were enriched in blood and multiple LN, CD4⁺ T_reg cells were enriched in LN, while CD4⁺ and CD8⁺ T_RM cells predominated in JEJ and were present at lower frequencies in lungs, spleen and LN (Fig. 2a,b). T_EMRA cells were mostly CD8⁺ and enriched in the BM and spleen and, to a lesser extent, in lungs, while T_EM cells were distributed across most sites (Fig. 2a,b). Mucosal-associated invariant T (MAIT) cells, distinguished by TRAV1.2 expression, CD161 and other markers¹⁶, were predominantly found in the spleen, BM and lungs (Fig. 2a,b). TCR clonal analysis provided additional correlative evidence for subset delineation and tissue distribution (for example, with the highest clonality observed in the T_EMRA subset, as previously described^7,17) (Supplementary Fig. 2).

Innate lymphocytes (130,414 cells) were predominantly mature CD56^dimCD16⁺ NK cells expressing cytolytic markers (KLRF1, GZMB) and enriched in blood, BM and lungs (Fig. 2c,d). Immature CD56^brightCD16⁻ NK cells were present at lower frequencies across most tissues (Fig. 2c,d). We detected low frequencies of ILCs consisting largely of CD16⁻NCR2⁺IL7R⁻ ILC1s with high expression of tissue residency markers (CD69, CD49a, CD103) enriched in JEJ, and IL7R⁺KIT⁺RORC⁺ ILC3 (ref. ⁸) found in LN, spleen and JEJ (Fig. 2c,d). Putative CD127⁺CD62L⁺TCF7⁺ NK/ILC precursors, resembling CD56^brightCD16⁻ NK and ILC3s¹⁸, were enriched in blood but present in all tissues (Fig. 2c,d).

B cells (272,162 cells) were classified into 6 subsets largely confined to lymphoid organs (Fig. 2e,f), including IgD⁺ naive B cells, CD27⁺ memory B cells, germinal center B cells expressing AICDA (encoding activation-induced adenosine deaminase, which mediates somatic mutation and class switch recombination¹⁹), plasma cells expressing immunoglobulin genes and SDC1 (CD138) and plasmablasts expressing proliferation markers (MKI67, TOP2A) (Fig. 2f). CD11c⁺ memory B cells expressing TBX21 (encoding the transcription factor T-BET) resembled ‘atypical B cells’²⁰ and were found at low frequencies in spleen and BM (Fig. 2e,f). Variable frequencies of memory B cells in LN and spleen expressed CD69 (Fig. 2e,f), denoting tissue residency²¹. Plasma cells expressing IgA were enriched in the JLP, while IgG⁺ plasmablasts were enriched in lymphoid organs (Fig. 2e,f). B cell receptor (BCR) analysis indicated that plasmablasts exhibited the highest clonal expansion across lymphoid sites, while memory B cells and plasma (but not naive) cells expressed mutated BCR (Supplementary Fig. 4).

Myeloid lineage cells (225,268 cells) comprised C1QA⁺MS4A7⁺ macrophages, FCN1⁺CD14⁺ classical and FCN1⁺FCGR3A⁺ non-classical monocytes and DC subsets, including CLEC9A⁺ DC1, CLEC10A⁺ DC2, CCR7⁺ migratory DCs and CD123⁺LILRA4⁺ plasmacytoid DCs (pDCs) (Fig. 2g,h). Classical monocytes were most abundant in blood, while non-classical monocytes were found mainly in BM and lung (Fig. 2g,h). DCs were found at low frequency across all tissues; pDCs were enriched in the BM, migratory DCs in the LNs, while DC2s were found mainly in the LNs and JEJ (Fig. 2g,h). Macrophages were found predominantly in the lung and at low frequencies in the BM, spleen, LNs and JEJ (Fig. 2h).

Immune cell subset composition for the lineages above in blood, lymphoid organs and mucosal sites was specific to the site and conserved across donors (Fig. 2i and Extended Data Fig. 2). We defined additional subset heterogeneity based on RNA expression, identifying proliferating cells across lineages and functional subsets for CD8⁺ T_EM cells and CD8⁺ T_EMRA cells, also with tissue-specific distribution (Supplementary Figs. 5 and 6). This comprehensive, annotated map of immune cells across tissues can serve as a reference for future analysis, and we provided pre-trained models for label transfer of our cell-type annotation in the popV framework²² (Supplementary Fig. 7). This subset analysis showed consistent, tissue-specific composition between sites.

Extended Data Fig. 2 — **a-d**) Box plots showing the percentage of each immune cell subset within a tissue (colored box) compared to the average proportion across all other tissues (gray box, “Rest”) for (a) T cells, (b) NK/ILCs, (c) B cells, and (d) myeloid cells. Box plots show the median (center), interquartile range (IQR; box), and whiskers extending to 1.5×IQR. Dots represent individual donors. Statistical significance determined by Bayesian modeling using scCODA. * denotes adj. p-val (FDR) < 0.05. Abbreviations: BLO (blood), BM (bone marrow), SPL (spleen), LN (lymph node), LNG (lung), JEJ (jejunum).

Immune cell phenotype is strongly tissue-associated

To understand the influence of tissue localization on gene expression, we performed a two-step differential expression (DE) analysis. First, we compared the major immune lineages (CD8⁺ T cells, CD4⁺ T cells, γδ/ΜΑΙΤ cells, myeloid cells, NK/ILC and B cells) within each site (for example, blood, BM, spleen, LN, lung, JEJ) versus the other sites using pseudobulked linear mixed modeling (LMM), controlling for covariates (for example, sex, CMV serostatus)²³ (Fig. 3a and Supplementary Tables 4 and 5). We identified 13 clusters from significantly differentially expressed genes (DEGs) (C1–13) grouped by lineage and/or tissue (Fig. 3b and Supplementary Table 6). We then conducted a similar across-tissue DE analysis for each subset and evaluated whether these signatures were expressed by individual subsets within a lineage using gene set enrichment analysis (GSEA) (Fig. 3a). To visualize effect sizes, determine the contribution of compositional differences and identify tissue-specific signatures across lineages, we integrated this analysis with subset frequencies within a tissue, fold changes (FC) compared to the other sites and the average expression of gene clusters by subset (Fig. 3c–h, Extended Data Fig. 3, and Supplementary Tables 7 and 8).

Fig. 3 — a, Schematic showing cell lineage-level and cell subset-level DE analysis using linear mixed models (evaluated by dreamlet) in which cell lineage-level analysis compares a lineage in one tissue (for example, B cells from JEJ) against the same lineage in all other tissues, and cell subset-level analysis compares an immune cell subset in one tissue (for example, plasma cells in JEJ) against the same immune cell subset in all other tissues, followed by pre-ranked GSEA to assess gene cluster enrichment in subsets across tissues. b, Heatmap of lineage-level DEGs (adjusted (adj.) P value (false discovery rate, FDR) < 0.05, log₂(FC) > 1, average mean expression >2) showing z-score average gene expression across tissue–cell lineage combinations, labeled by gene cluster with selected genes. c–h, Evaluation of gene clusters in cell subset-level DE analysis for myeloid cells (c and d), T cells and NK/ILCs (e and f), and B cells (g and h), shown as bar plots (c, e and g) with normalized enrichment scores (NES) of indicated gene clusters in cell subset-level DE by pre-ranked GSEA and heatmaps (d, f and h) showing cell subset frequencies as a percentage of their lineage within each tissue, FC in subset frequency in indicated tissue versus other tissues and split violin plots showing average gene cluster expression in the indicated tissue versus other tissues. Plotted bars denote adj. P value (FDR) < 0.05; n.s., not significant. AMP, adenosine monophosphate; O-glyco., O-glycosylation; prot., protein; circ., circulating; unconv., unconventional.

Extended Data Fig. 3 — **a-h**) Evaluation of additional gene clusters in subset-level DE analysis. Bar plots (**a,c,e,g**) show normalized enrichment scores (NES) of indicated gene clusters in subset-level DE by pre-ranked GSEA. Plotted bars denote adj. p-val (FDR) < 0.05 and n.s. denotes non-significance. Heatmaps (**b,d,f,h**) show subset frequencies (Freq. (%)) as a percentage of their lineage within each tissue, fold changes (FC) in subset frequency in indicated tissue versus other tissues, and split violin plots show average gene cluster expression in indicated tissue versus other tissues.

Myeloid cells from all tissues exhibited transcriptional profiles not expressed by other immune lineages (clusters C1–C4), which also varied by tissue (Fig. 3b). C1 (chemokine, complement, lipid transport) and C2 (PPAR signaling associated with alveolar macrophages²⁴) were enriched in the JEJ and lung, respectively, particularly in macrophages (Fig. 3c,d). By contrast, C4 (anti-microbial peptide production and cell signaling) enrichment was explained by increased frequencies of BM and spleen monocytes relative to macrophages rather than transcriptional changes within and across subsets (Extended Data Fig. 3a,b).

Tissue-associated genes within clusters C5–C9 were expressed primarily by T cells and innate lymphocytes (Fig. 3b). C5 genes encoding stem-like transcription factors and markers (TCF7, LEF1, ITGA6)^25,26 and LN homing receptors (CCR7, SELL) were enriched in CD4⁺ and CD8⁺ T_RM cells and T_EM cells in LN (Fig. 3e,f). Conversely, C6 genes encoding molecules for mucosal residency (ITGA1, CXCR6, ITGAE)⁶ and gut homing (CCR9) were enriched in CD4⁺ and CD8⁺ T_RM cells, γδ T cells, MAIT cells and NK cells in the JEJ (Fig. 3e,f). CD8⁺ T_EMRA cells, γδ cells, MAIT cells and NK/ILCs in the JEJ showed reduced expression of C9 genes associated with cytolytic effector function (GZMB, PRF1, IFNG, NKG7) compared to other sites, while C9 was enriched in the lung (Fig. 3b and Extended Data Fig. 3c,d). C7 genes associated with protein metabolism were enriched in NK/ILCs and innate T cells in the JEJ, while C8 genes encoding chemokines (CCL4, XCL1) and innate lymphocyte functions (KLRC1, NCR1) were expressed by T cell and NK/ILC subsets across all sites (Fig. 3b and Extended Data Fig. 3e,f). Therefore, this analysis identified shared gene expression profiles across T and innate lymphocyte subsets that varied by site.

Tissue-enriched signatures for B cells (C10–C13) were mostly associated with specific subsets (Fig. 3b,g,h and Extended Data Fig. 3g,h). C10 included cell cycle genes and was enriched in B cells and plasma cells in the lungs (Fig. 3g,h), possibly because of higher plasmablast abundance. JEJ B cells showed increased expression of C12 (B cell differentiation and protein transport) and C13 (B cell activation) genes (Fig. 3b), which were derived from plasma cells comprising the majority of the B cell lineage in intestinal sites (Fig. 3g,h).

We also applied consensus single-cell hierarchical Poisson factorization (scHPF)^15,27 to identify gene co-expression patterns common to different immune lineages or sites (considering the JEJ, lung and LN) (Extended Data Fig. 4a and Supplementary Table 9). We found a proliferation module enriched in the lungs, a lymphoid-specific module in the LNs, an intestine-specific residency module and modules associated with effector and cytolytic functions prominent in the lungs but not JEJ or LNs (Extended Data Fig. 4a–d). The lymphoid module included genes associated with stemness (KLF7, LEF1, SOX4)^28,29 and lymphoid homing (CCR7, SELL, ITGA6) markers expressed mostly in T cells, some NK cells and ILCs, and not in B cells or myeloid cells (Extended Data Fig. 4b–e and Supplementary Table 10). The JEJ residency module included intestinal tissue residency genes (CCR9, ITGAE, CD101, CD160)^7,30 enriched in T cells, NK cells and ILCs (Extended Data Fig. 4f–h). These signatures were also reflected at the surface protein level (Extended Data Fig. 4e,i and Supplementary Table 11). Thus, tissue-specific gene expression modules spanned multiple cell types, suggesting shared tissue adaptations.

Extended Data Fig. 4 — a) Correlogram of hierarchically clustered Pearson correlations between factor gene scores generated by consensus-scHPF. Cell type models, tissue enrichment, and representative genes per module are indicated. **b-i**) Analysis of the consensus-scHPF Lymphoid module (**b-e**) and JEJ residency module (**f-i**). Heatmaps (**b,f**) show z-scored average expression for each module across immune cell subsets and tissues. Violin plots (**c,g**) show module gene expression across immune subsets and expanded tissue groupings within donors. Volcano plots (**d,h**) show DE genes (adj. p-val (FDR) < 0.05; shown in pink) from subset-level DE analysis by linear mixed model (dreamlet) in selected immune cell subsets across indicated tissues. Genes from the module are highlighted in blue. Box plots (**e,i**) show differential expression of surface proteins (as LFC) across tissues for selected immune cell subsets. Box plots depicted as median (center), interquartile range (IQR; box), and whiskers extending to 1.5×IQR. Dots represent individual donors with sufficient cell numbers for inclusion. Abbreviations: BLO (blood), BM (bone marrow), SPL (spleen), LN (lymph node), LNG (lung), JEJ (jejunum).

Source data

Resident immune cells express tissue-specific signatures

The tissue environment poses unique requirements for resident immune populations such as T_RM cells, plasma cells and macrophages for maintaining homeostasis^31–33. We integrated the DE analysis shown in Fig. 3 with surface protein expression to define site-specific signatures for resident immune cells (Fig. 4 and Supplementary Table 12). CD4⁺ and CD8⁺ T_RM cells expressed genes and/or surface proteins for tissue residency (CD103, CD101, CD49a), gut homing and localization (CCR9, CCR5 for CD4⁺ T_RM cells) and reduced PD-1 (for CD8⁺ T_RM cells) in intestines relative to lungs and lymphoid organs (Fig. 4a–f). Lung T_RM cells showed increased expression of effector or cytotoxicity (IFNG, GZMH, GZMA, PRF1) and regulatory (CTLA4) genes relative to the JEJ and lymphoid organs; T_RM cells in LN, spleen and BM had higher expression of stem-like markers (TCF7, KLF2) and certain integrins and costimulatory markers (ITGB2, CD27, CD28, ICOS) relative to the JEJ and lungs (Fig. 4a–f). These site-specific profiles for T_RM cells showed adaptations related to migration, localization and function.

Fig. 4 — a–l, Expression of tissue-enriched genes from cell subset-level DE analysis and corresponding surface proteins for CD4⁺ T_RM cells (a–c), CD8⁺ T_RM cells (d–f), plasma cells (g–i) and macrophages (j–l), shown as dot plots with the percentage of cells expressing selected genes enriched in one tissue versus the other sites (a, d, g and j) or surface protein expression (c, f, i and l) and heatmaps showing median log₂(FC) surface marker expression across tissues (b, e, h and k). Dot size represents the frequency of gene expression in group (a, d, g and j) or frequency (% positive) (c, f, i and l); color (red, blue) intensity indicates average gene expression.

Tissue plasma cells and macrophages also exhibited distinct tissue signatures. For plasma cells, we found several JEJ-enriched genes, including IGHA2 (consistent with predominant IgA⁺ plasma cells in the gut), the non-classical HLA molecule CD1d, plasma cell transcription factors (RUNX2, ID3)^34,35 and the tissue residency marker CD69 (Fig. 4g–i). Plasma cells in lungs, and to a lesser extent, LN and spleen, expressed higher levels of integrins (CD11c, CD18) (Fig. 4g–i). JEJ macrophages had increased expression of integrins (ICAM1, ITGA4), chemokines (CCL24, CCL3L1) and regulators of macrophage activation (for example, SLAMF7)³⁶, whereas lung macrophages had higher expression of CD11c (ITGAX), markers of efferocytosis (MRC1, MARCO) and lipid metabolism (PPARG, TREM2, FABP4) (Fig. 4j–l), consistent with alveolar macrophages³⁷. Lastly, macrophages in the spleen expressed markers of red pulp macrophages (SPIC, VCAM1)³⁸ (Fig. 4j). These observations revealed site-specific signatures for activation, migration, metabolism and cell–cell interactions involved in tissue residency.

Immune cell profiles change with age across tissues

We investigated age-associated effects across immune lineages and tissues. A global analysis of transcriptional variance within each major lineage revealed that tissue site explained the majority of variance, while age accounted for a much smaller fraction (Fig. 5a and Supplementary Table 13). Most of the top age-dependent genes in each lineage (Fig. 5b) also exhibited tissue-specific variation. Immune cell composition in the different sites was largely maintained across age, except for significantly decreased frequencies of CD8⁺ T_N cells in the blood and LNs, concomitant increases in T_EM cells in the blood and T_RM cells in the LNs and lower frequencies of classical monocytes in the BM (Fig. 5c and Supplementary Table 14). These results indicate that tissue-driven immune cell composition and profile are largely maintained with age.

Fig. 5 — a, Variance decomposition analysis (dreamlet), with box plots showing percentage of variance in gene expression explained by age (<40 or >40 years old), CMV serostatus (positive or negative), sex (male or female), processing site (US or UK), 10× Genomics chemistry (5′ or 3′) and tissue group (BLO, BM, SPL, LN, LNG or JEJ) for CD4⁺ T cells, CD8⁺ T cells, B cells, myeloid cells and NK/ILCs. Box plots show the median (center), interquartile range (IQR; box) and whiskers extending to 1.5× IQR. b, Variance decomposition analysis with stacked bar plots showing selected genes from the top (up with age) and bottom (down with age) 50 DEGs by cell lineage-level linear mixed model (dreamlet) across age for CD4⁺ T cells, CD8⁺ T cells, B cells, myeloid cells and NK/ILCs. c, Heatmaps of change in subset composition by age (<40 or >40 years old) across BLO, BM, SPL, LN, LNG and JEJ, shown as log₂(FC) calculated by generalized linear model (two-sided Wald test). Gray denotes insufficient cell numbers for comparison. d, t-SNE of immune cell subsets based on age-associated DEGs (log₂(FC) > 0.1, unadjusted (unadj.) P < 0.05) by subset-level linear mixed model (dreamlet). e–g, Bar-and-dot plots of top DEGs by age (adj. P value (FDR) < 0.05 for at least 1 tissue; among top 50) for classical monocytes in BLO, BM, LNG and SPL (e), non-classical monocytes in BM, LNG and SPL (f), and macrophages in JEJ, LN, LNG and SPL (g). Solid bars show median log₂(FC) across tissues; error bars, 95% CI. Statistically significant genes are indicated by circles outlined in black. h, Box plot showing the percentage of CD95⁺ macrophages in the lung in donors <40 or >40 years old. Statistical significance determined by generalized linear model (two-sided Wald test). i, Consensus scHPF macrophage *APOE*–*TREM2* signature with dot plots showing gene rank (left) and raw and covariate-adjusted (partial reg.) cell scores by linear mixed model in individuals <40 or >40 years old (right). j, Heatmap of enrichment of top 200 factor genes in cell subset-level DE by GSEA in cMono, ncMono and macrophages in BLO, BM, SPL, LN, LNG and JEJ. k, *APOE*–*TREM2* signature cell scores from a human lung dataset (two-sided Wilcoxon rank-sum test) (left) and enrichment of the factor in macrophage DE by GSEA across age (right). For all panels, q denotes adj. P value (FDR); *adj. P value (FDR) < 0.05.

To interrogate specific effects of aging on immune cells across lineages and sites, we conducted a separate DE analysis for each tissue group and immune subset with sufficient representation, comparing younger (<40 years) versus older (>40 years) individuals, while controlling for sex, CMV status and other covariates (Supplementary Tables 15–17). An embedding of similarities between age-related DEGs revealed that some immune cells exhibited changes specific to tissue site (for example, T cells in the LN, lymphocytes in the JEJ), while others showed subset-specific changes independent of site (for example, monocytes in blood, BM and lung, and memory B cells in LNs and lung) (Fig. 5d).

We identified genes regulated with age in each subset and tissue by integrating DE results with GSEA (Supplementary Fig. 8a). For myeloid lineage cells, mucosal sites had the most age-related DEGs, with similar trends in lymphoid organs (Fig. 5e–g). Age-related changes in classical monocytes included increased expression of genes associated with proliferation and inflammation (KRAS, CALM1)^39,40 and decreased expression of genes for macrophage differentiation (for example, RAB44)⁴¹ (Fig. 5e); non-classical monocytes showed age-associated increase in expression of genes for cell–cell interactions (LGALS1, ITGA9) and decreased expression of metabolism and mitochondrial regulation genes (LYRM7, SARS2) (Fig. 5f). Macrophages had decreased expression of genes associated with metabolism and mitochondrial fitness (MARC1, MTOR)⁴² and increased expression of genes associated with M2 macrophages (ID1, ADGRE1 NPR2)^43,44 and interferon signaling (MX2) (Fig. 5g) along with increased CD95 (Fas) at the protein level (Fig. 5h). Together, these age-related changes in monocytes and macrophages were subset-specific, enriched in mucosal sites and indicated decreased overall fitness.

We used scHPF to identify age-associated gene signatures for myeloid cells and GSEA to assess their expression in different subsets and sites (Supplementary Tables 18 and 19). We uncovered an APOE–TREM2 signature, including apolipoprotein genes (APOC1, APOC2 and APOE) and TREM2, a triggering receptor expressed on myeloid cells⁴⁵ that binds ApoE and facilitates macrophage functions, such as phagocytosis and chemotaxis, and induces metabolic changes⁴⁶ (Fig. 5i). This APOE–TREM2 signature was significantly downregulated with age in monocytes and macrophages in the lungs, lymphoid organs and blood (Fig. 5j), and in an independent published dataset from human lungs⁴⁷ (n = 29; see Methods) (Fig. 5k). Overall, this analysis showed subset and site-specific features of macrophage aging involving a major functional and metabolic pathway.

T and B lymphocytes express distinct aging signatures

We applied the above approaches to identify age-associated signatures in T cells and B cells. For CD4⁺ T cells, CD8⁺ T cells and B cells, there were relatively few genes (for example, those associated with oxidation and inflammation: SOD1, IL18BP, IL15) that changed over age in two or more subsets within each lineage (Extended Data Fig. 5a–d). Pathway analysis revealed increased inflammation, apoptosis and reduced TCR signaling across multiple T cell subsets with age (Supplementary Fig. 8b). Despite the paucity of age-associated gene expression changes across subsets, CD8⁺ T_EMRA cells exhibited multiple age-associated changes conserved across sites; these included increased expression of NK cell genes (NCAM1, KLRF1, GNLY) and the NK cell marker CD56, consistent with findings in blood⁴⁸, and reduced expression of genes associated with signaling (CD6, JAK3), proliferation and metabolism (TCF7, RPTOR) (Fig. 6a,b).

Extended Data Fig. 5 — **a-c**) Heatmaps of age-associated genes across multiple tissue-subset combinations among CD4⁺T cells (a), CD8⁺T cells (b), or B cells (c). Statistical significance determined by linear mixed model (dreamlet). * denotes adj. p-val (FDR) < 0.05. d) Bar-and-dot plots of DE genes by age in naïve B cells by subset-level linear mixed model (dreamlet). Bars show median LFC across tissues, error bars represent 95% CI, dot color indicates tissue with significant genes (adj. p-val (FDR) < 0.05) outlined in black. (**e-g**) Dot plots showing factor cell scores from the CD8⁺T cell Cytokine factor (e), CD8⁺T cell *GZMK*⁺ signature (f), and B cell *RAS*⁺ signature (g) on respective cell types in human blood samples from the Human Immune Health Atlas. Each dot represents a donor with statistical significance determined by two-sided Wilcoxon rank sum test across age.

By scHPF analysis, we identified two prominent age-associated transcriptional signatures shared across multiple CD8⁺ T cell subsets. A cytokine signature, containing genes for effector cytokines and chemokines (for example, CCL3, CCL4, XCL1, IFNG, TNF), was increased with age in all CD8⁺ T cell subsets and γδ T cells across all sites examined (Fig. 6c,d). This aging signature was also detected in published datasets from human lungs⁴⁷ (Fig. 6e) and in peripheral blood mononuclear cells (PBMCs) from the Sound Life cohort (n = 96, age 26–65 years) within the Human Immune Health Atlas⁴⁹ (Extended Data Fig. 5e). A second signature contained GZMK encoding the cytolytic molecule granzyme K, the transcription factor EOMES⁵⁰ and activation or signaling markers (PDCD1, HLA-DR, FCRL3)⁵¹ (Fig. 6f). This signature aligned with a granzyme K-containing, age-associated signature identified in T cells in aged mice and human blood⁵² and in PBMCs from the Human Immune Health Atlas⁴⁹ (Extended Data Fig. 5f). This GZMK⁺ signature was increased with age in CD8⁺ T_EMRA cells, γδ T cells and CD8⁺ T_EM cells across tissues but not in CD8⁺ T_RM cells in lungs and JEJ (Fig. 6g). The frequency and clonal expansion of CD8⁺ T_EMRA cells enriched in the GZMK⁺ signature was higher in older than in younger donors (Fig. 6h,i). These results show distinct age-associated signatures in mucosal resident T cells compared to circulation.

For B cells, there were more age-related DEGs in memory compared to naive B cells in LNs (Fig. 6j and Extended Data Fig. 5d). Memory B cells from older donors had increased expression of genes associated with inflammatory cytokines (IL18, IL18BP), cell adhesion (CD58, LGALS1) and cell death or autophagy (FAS, ITM2A), along with reduced expression of proliferation (CDCA7, IL2RA), lipid metabolism (ACSM3, PNPLA7) and differentiation markers (Fig. 6j). Select IL-18 pathway-associated genes in B cells and CD4⁺ T cells in blood were validated in the Human Immune Health Atlas cohort⁴⁹ (Supplementary Fig. 9a,b). Transcript and/or surface expression of IgM (IGHM) and IGHD were reduced in older compared to younger B cells, and the frequency of IgM⁺ B cells decreased with age (Fig. 6k,l and Supplementary Table 20). We identified by scHPF that a gene signature related to RAS signaling (RASA4B, RASGRF1, GAB2) downstream of the BCR⁵³ was downregulated with age in naive and memory B cells across all sites (Fig. 6m,n). We validated this age-associated signature in BM-derived B cells from an independent dataset (n = 39, age 2–84 years)⁵⁴ and in PBMCs from the Human Immune Health Atlas⁴⁹ (Fig. 6o and Extended Data Fig. 5g). Pathway analysis further revealed increased inflammation and reduced BCR signaling in LN B cells with age (Supplementary Fig. 8c). These results showed that B cells exhibited diminished signaling and functional dysregulation across tissues over age.

Host factors and subset heterogeneity affect T cell aging

CMV infection drives immune cell alterations, including increased accumulation of T_EMRA cells with age in blood, spleen and lungs^4,55. We investigated the impact of CMV serostatus on cell composition and immune aging and found no significant associations with CD4⁺ T cell, CD8⁺ T cell or B cell frequencies (Supplementary Fig. 10a,b). Two CD8⁺ T cell signatures were associated with CMV serostatus after regression of other covariates: the GZMK⁺ signature and a GNLY⁺ signature (GNLY, FGFBP2 and CX3CR1) (Supplementary Fig. 10c–f). The GNLY⁺ signature was enriched across all CD8⁺ T cell subsets in CMV⁺ donors, while the GZMK⁺ signature was variably enriched in different sites and subsets of CMV⁺ donors by GSEA (Supplementary Fig. 10d,f and Supplementary Tables 21 and 22). Therefore, CMV infection drives T cell gene signature changes that overlap with, but are distinct from, age-related immune alterations.

CD4⁺ T cells are highly heterogeneous and exhibit functional and phenotypic continuums^56,57, suggesting that age effects could differentially manifest within or across subsets. We applied an annotation-independent analysis of aging in CD4⁺ T cells in the lung, JEJ and LN, leveraging a per-cell estimation of age effects using counterfactual analysis with MrVI, which separately considers each cell and controls for covariates¹⁰. This analysis identified groups of cells with similar predicted age-associated changes in gene expression (‘modules’), which we interrogated by DE analysis across age (Fig. 7, Extended Data Fig. 6, and Supplementary Tables 23 and 24). In the lung, a fraction of CD4⁺ T cells (~25%, comprising T_EM cells, T_RM cells and T_EMRA cells) exhibited decreased cytotoxicity (GZMH, GNLY, GZMA) and increased cytokine receptor (IL18R1, IFNGR1) genes with age (Fig. 7a–c and Extended Data Fig. 6a,b). Similar upregulation of cytokine responsiveness with age occurred in some CD4⁺ T cells in blood, BM, LN and spleen, while decreased cytotoxicity was unique to the lungs (Fig. 7c,d and Extended Data Fig. 6b). CD4⁺ T cells in the JEJ (mainly T_RM cells) exhibited an age-related decline in T_H17-associated genes (IL17A, IL17F, IL22, RORC, CCR6)⁵⁸ and increase in pro-inflammatory cytokines (IFNG, TNF) (Fig. 7e–g and Extended Data Fig. 6c,d); age-associated downregulation of IL-17-associated genes was also observed in CD4⁺ T cells in lung and blood (Fig. 7g–h and Extended Data Fig. 6d). CD4⁺ T cells in the LN (also in spleen and lungs) exhibited reduced expression of genes associated with regulation (IL10, TIGIT, CTLA4, CD27) and increased expression of inflammation and activation markers (IL18BP, TNFRSF4, TNF) with age (Fig. 7i–l and Extended Data Fig. 6e,f). These results revealed age-associated transcriptional changes in tissue CD4⁺ T cells associated with site-specific functions.

Fig. 7 — a–l, MrVI analysis shows modules of genes changing over age for CD4⁺ T cells in the lungs (a–d), gut (e–h) and LNs (i–l), depicted as UMAP embeddings (a, e and i), colored by subset (left) and tissue module score (right). Age-associated DEGs identified by linear mixed model (dreamlet) in donors <40 or >40 years old are shown in volcano plots for module-positive cells in each tissue (b, f and j) and box-and-dot plots (c, g and k) in BLO, BM, LN, LNG and SPL (c), BLO, LN, LNG and JEJ (g), and BM, LN, LNG and SPL (k), and heatmaps (d, h and l) showing enrichment of manually selected module genes across BLO BM, LN, LNG and SPL (d), BLO, LN, LNG and JEJ (h), and BM, LN, LNG and SPL (l) in tissue-specific DE of module-positive cells by age using GSEA. In a, e and i, module score is computed as the sum of inferred log₂(FC) for genes upregulated in donors >40 years old minus the sum of log₂(FC) for genes downregulated in individuals >40 year old. In b, f and j, selected genes with high effect size (by estimated log₂(FC), unadj. P < 0.05) are labeled. Bars show median log₂(FC) across tissues; dot color indicates tissue with trending DEGs (unadj. P < 0.05) outlined in black. In c, g and k, genes with known effector or activation functions are shown, selected from the module or trending DEG list. Tissues are included only if sufficient signature-positive cells were detected. Separate analyses (columns) were conducted for upregulated and downregulated genes. *Unadj. P < 0.1.

Extended Data Fig. 6 — **a-f**) Modules from MrVI analysis of CD4⁺ T cells in the LNG (**a,b**), JEJ (**c,d**) and LN (**e,f**). Bar plots (**a,c,e**) show average predicted LFC ( > 40 y.o. / < 40 y.o.) of top marker genes predicted to change similarly with age in module-positive cells for each tissue model. Stacked bar plots (**b,d,e**) show subset distribution (%) of module-positive cells, colored by subset.

Discussion

We present a comprehensive analysis of the human immune system across tissues and ages through multimodal profiling of blood and tissues from organ donors spanning six decades of adult life. We found that tissue localization was a dominant driver of the immune landscape, determining immune cell composition, cell states and functional capacity. With age, these tissue-specific properties were largely maintained, although certain subsets and sites showed altered function, migration and regulation. Our results reveal that the human immune system is highly specialized for diverse tissue environments to maintain homeostasis and mount effective immune responses.

We demonstrated that each tissue imposed site-specific immune cell compositions and adaptations that varied by lineage, and these tissue effects were conserved across donors. Although we realized and reinforced site-specific features for T_RM cells at barrier sites and lymphoid organs^7,59, whether these adaptations applied to other immune cells remained unknown. Here, we found that site-specific signatures for T cells in the gut (high tissue residency, low cytotoxicity), lungs (high effector function, increased regulation) and lymphoid organs (stem-like features) were not exclusive to the canonical resident populations, were shared across NK cell and ILC subsets and were absent from B cell and myeloid lineages. The enhanced expression of stem-like transcription factors TCF-1 and LEF-1 by LN memory T cells suggests that lymphoid organs may serve as reservoirs for long-lived memory cells, given these factors’ requirement for memory T cell generation^59,60. Macrophages and plasma cells also exhibited site-specific features in the gut, lungs and lymphoid organs through distinct subset-specific pathways, such as alveolar macrophages in the lungs and red pulp macrophages in the spleen. These lineage-dependent tissue adaptations probably reflect niche localization and interactions with distinct structural and immune cells within each tissue.

Age-associated gene signatures identified for macrophages, T cells and B cells were intrinsic to the subset and site. The APOE–TREM2 gene signature, essential for crucial macrophage functions⁴⁶, was reduced with age by lung macrophages. APOE–TREM2 expression in microglia is associated with neurodegeneration in Alzheimer’s disease and in other macrophage types with cardiovascular diseases^61,62. TREM2 can have different effects on macrophage functions; promoting anti-inflammatory ‘M2-like’ function in some contexts and phagocytosis and sustained inflammation in others^63,64. The age-associated loss of TREM2 in lung macrophages could thus account for compromised immunity to respiratory pathogens and increased lung cancer susceptibility observed in the aging population. TREM agonists that enhance phagocytic function are being tested in clinical trials in Alzheimer’s disease⁶⁵ and could be considered in the rejuvenation of aging macrophages in other sites.

Other age-associated features were specific to lymphocyte lineages. T cells in circulation expressed higher levels of genes associated with inflammation, cytotoxicity and NK-like markers with age, as previously reported^48,66. Circulating T_EMRA cells and T_EM cells upregulated GZMK and other markers, similar to senescent GzmK⁺CD8⁺ T cells found in mice and human blood⁵². T_RM cells in the lungs and intestines did not exhibit this age-associated gene signature, suggesting that the tissue environment may insulate them from signals that promote cellular aging or that cellular aging is tissue-specific. However, both circulating (T_EM, T_EMRA) and T_RM cells had increased expression of genes for pro-inflammatory cytokines and chemokines with age, consistent with inflammaging implicated in cardiovascular diseases and metabolic dysregulation^67,68. Our findings suggest that human T cells may be more prone to innate functions such as cytokine-driven activation (for example, via IL-18) with age. We also identified an age-associated increase in IL-18 expression and reduced BCR-mediated signaling within tissue B cells, which is a feature of NK-like B cell subsets identified in disease contexts^69,70. Thus, aging may reflect a broader age-related shift to innate-like functions in both T cells and B cells.

Our findings have important implications for immune monitoring, therapeutic modulation and clinical advancement. The compartmentalization of immune subsets across tissues emphasizes the importance of site-specific immune monitoring in disease states, as exemplified in severe COVID-19, in which immune dynamics in the respiratory tract rather than blood correlated with infection outcome^71,72. The distinctness of gut-specific subsets provides rationale for targeted intestinal interventions, as demonstrated by rotavirus vaccines⁷³. The identification of stem-like profiles (marked by TCF7 and LEF1 expression) in LN T cells and NK cells has direct relevance to adoptive CAR-T immunotherapies, in which stemness is associated with remission⁷⁴. LNs may thus represent an optimal source of NK and T cells for engineering adoptive cell therapies against cancer, infections and autoimmunity^75,76.

Our study has several limitations. The low frequency of certain immune subsets in tissues, including DCs, macrophages in lymphoid organs and hematopoietic progenitors, precluded aging analysis and will require sorting for future studies. Similarly, an in-depth analysis of TCR and BCR across sites and age would require isolating memory T cells and B cells from each site. Finally, although we identified age-associated changes in 24 donors, additional donors would increase power and probably reveal additional aging signatures. In conclusion, this dataset, along with the models and analyses presented, can serve as a valuable and actionable resource, informing targeted immune modulation by site and age in future treatments for infectious, neoplastic and inflammatory diseases.

Methods

Ethics statement

This study uses samples obtained from deceased organ donors and does not qualify as human subjects research in the USA, given that the donors are deceased and not living, as confirmed by the Columbia University Institutional Review Board. In the UK, samples were collected and analysed under an ethically approved research protocol (REC 15/EE/0152).

Statistics and reproducibility

The study analyzed immune cells from multiple tissue samples obtained from 24 organ donors. No statistical method was used to predetermine sample size. No data were excluded from the analysis. Investigators were not blinded to allocation during experiments and outcome assessment, as this is a profiling study.

Tissue acquisition from organ donors

Tissues were obtained from deceased organ donors (Supplementary Table 1) at the time of organ acquisition for clinical transplantation. In the USA, this was done through an approved protocol and material transfer agreement via LiveOnNY, the organ procurement organization for the New York metropolitan area^3,77. In the UK, tissues were obtained through the Cambridge Biorepository for Translational Medicine (CBTM), REC 15/EE/0152, as previously described⁹. Owing to the different amounts of tissues and some distinct samples (for example, skin, liver and colon) obtained at each location, protocols for processing may differ, as described below.

Tissue processing and CITE-seq at Columbia University

Each tissue was subjected to a tissue-specific protocol to maximize MNC recovery and viability across a diversity of sites⁹. Detailed, step-by-step protocols for immune cell isolation from blood, BM, spleen, LNs, lungs (parenchyma and airway or BAL) and JEJ (JLP and JEL) are presented elsewhere⁷⁸. All single-cell suspensions from each site were centrifuged (400g, 10 min at 4 °C) and washed twice with PBS containing 5% (v/v) FBS and 2 mM EDTA. Cells were counted using the NC-2000 Cell Counter (Chemometec), and 50 million viable cells from each site were treated with TruStain FcX (BioLegend) and FcR Blocking Reagent (Miltenyi). Cells were subsequently labeled for 30 min at 4 °C with biotinylated anti-CD66B, anti-CD235ab and anti-CD326 to remove granulocytes, red blood cells and epithelial cells, respectively, by streptavidin-coated magnetic particles and negative selection (Bangs Laboratories). All single-cell suspensions were subjected to dead cell removal using a Dead Cell Removal Kit (Miltenyi).

Each single-cell suspension was hashtagged to allow pooling of samples for loading on the 10× Genomics Chromium instrument. MNCs from each site (10⁶ per site) were transferred into 4 ml flow cytometry tubes, pelleted by centrifugation as above and resuspended in PBS containing 5% (v/v) FBS and 2 mM EDTA and then incubated with TruStain FcX (BioLegend) and FcR Blocking Reagent (Miltenyi) at 4 °C for 10 min to reduce background labeling. Each hashtag was spun at 14,000g for 10 min, added to each sample (1 µl hashtag per tube), incubated at 4 °C for 30 min, pelleted and washed 3 times with PBS containing 5% (v/v) FBS and 2 mM EDTA. For CITE-seq antibody staining, 200,000 cells from each sample were resuspended in reconstituted TotalSeq-A Universal Cocktail (BioLegend) (donors 496 and 503) and TotalSeq-C Universal Cocktail (BioLegend) (remaining 10 donors) in PBS containing 5% (v/v) FBS and 2 mM EDTA, incubated at 4 °C for 30 min and washed 3 times with PBS containing 5% (v/v) FBS and 2 mM EDTA before resuspension in a final volume of 1 ml. CITE-seq antibody panels are listed in Supplementary Table 1.

Tissue processing and CITE-seq at the University of Cambridge

Each tissue was subjected to a tissue-specific protocol to generate a single-cell suspension of immune cells that has been published in detail elsewhere^79–81. Immune cells were isolated from blood, BM aspirates (sternum), spleen, LNs, lungs, liver, JEJ (JEL and JLP) and skin. Each single-cell suspension was hashtagged to allow pooling of samples for loading on the 10× Genomics Chromium instrument. Approximately 500,000 MNCs per tissue were transferred into 1.5 ml Lo-Bind DNA tubes. Cells were centrifuged at 400g for 5 min, the supernatant removed and resuspended in 50 μl PBS containing 0.04% BSA. Cells were treated with 5 μl TruStain FcX (BioLegend) to reduce background labeling and incubated at 4 °C for 10 min, then each hashtag was added to the sample (1 µl hashtag per tube). Samples were incubated at 4 °C for 30 min, washed three times with PBS containing 0.04% BSA and equal numbers of cells from each tissue were pooled based on the number processed per donor. Cells were incubated with TotalSeq-C Human Universal Cocktail (BioLegend) (Supplementary Table 2) for 30 min at 4 °C and subsequently washed 3 times with PBS containing 0.04% BSA. Cells were resuspended in 500 µl PBS containing 0.04% BSA and passed through a 40 µm Flowmi pipette tip filter to remove any clumps of cells.

Single-cell sequencing

For scRNA-seq experiments, single cells were loaded onto the channels of a Chromium chip (10× Genomics). cDNA synthesis, amplification and sequencing libraries were generated using either the Single Cell 5′ Reagent (v1 and v2) or 3′ Reagent (v3) Kits. TCRαβ and BCR paired VDJ libraries were prepared from samples made with the 5′ Reagent Kit. All libraries were sequenced on either an Illumina HiSeq 4000, NextSeq or NovaSeq 6000 instrument.

Alignment and preprocessing of CITE-seq and scTCR/BCR-seq using Cell Ranger

Alignment was performed using Cell Ranger (v6.0.0) from 10× Genomics⁸² with the appropriate chemistry option (fiveprime or SC3Pv3). We added the cell hashing antibody and the protein antibody fastqs to a single call of cellranger count. Immune receptors (TCR and BCR) were aligned using cellranger vdj. TCR and BCR alignment results from Cell Ranger were used for quality control and filtering of low-quality cells (individual cells with both TCR and BCR detected). In cases where a single cell had both TCR and BCR reads, the immune receptor data were discarded, and the cell was labeled as a multiplet. For all alignments, we used reference genome refdata-gex-GRCh38-2020-A and immune receptor reference refdata-cellranger-vdj-GRCh38-alts-ensembl-5.0.0.

Quality control

Samples were demultiplexed by hashtag expression using hashsolo with default parameters⁸³. Cells that were not uniquely assigned to an individual sample were removed from downstream analysis. Filtering was performed to remove cells with fewer than 50 unique genes detected. Mitochondrial counts were quantified, summing all genes starting with ‘MT-’, and ribosomal counts were quantified using all genes starting with ‘RPS’ and ‘RPL’. For erythrocyte-related counts, all genes starting with ‘HB’ as well as ALAS2 and EPOR (to detect erythrocyte precursors) were quantified. Cells with more than 20% mitochondrial counts were flagged as potentially low-quality for later filtering. Counts for mitochondrial genes and MALAT1 were subsequently removed from the gene expression object and downstream analysis. To exclude contamination from ambient RNA, we processed the data using DecontX⁸⁴. Two samples (one from liver and one from skin) with abnormally high ambient counts were removed, as DecontX could not correct the ambient counts (for example, plasma cell genes like ALB in all immune cells). All downstream analysis was performed on uncorrected counts, as we found few ambient counts in other samples. We used a CellTypist model⁹ (at https://cog.sanger.ac.uk/celltypist/models/Red_Blood_CZI/v1/Red_Blood_CZI.pkl) to detect erythrocytes. Doublets were additionally detected using Scrublet⁸⁵ with a sim_doublet_ratio of ten. For each unique tissue site, we performed an initial integration across all samples by training a single-cell variational inference (scVI) model⁸⁶ on the gene expression with following parameters: 10,000 highly variable genes using the seurat_v3 option⁸⁷ in Scanpy, early stopping enabled and 50 epochs, 10 epochs for n_epochs_kl_warmup, two layers in encoder and decoder, nb gene likelihood and a mini-batch size of 256.

To perform filtering of low-quality events, we used the following quality metrics: the probability of a doublet predicted by Scrublet, the probability of a doublet from HashSolo, the percentage of erythrocyte genes as described above, whether a cell contained both TCR and BCR, whether the CellTypist erythrocyte model predicted a cell to be an erythrocyte, as well as cells with a total count below 2,000 unique molecular identifiers, 1,200 unique genes or 200 protein counts. All scores were added to generate a per-cell quality metric. To perform filtering, we argued that cells that group together and have evidence of low quality should be removed from downstream analysis. We first used Louvain clustering⁸⁸ on the coordinates from scVI latent space using 15 nearest neighbors to cluster the per-tissue integrated data with a resolution of 5.0. Every cluster with a median low-quality score (described above) of at least one was removed from downstream analysis. Although some low-quality events were retained with this filtering, their frequency was drastically reduced. We additionally established tissue-specific cut-offs to remove additional events and removed clusters with a mean low-quality score of 0.3 from all tissues, except for the lung LN and JEJ, for which the threshold was manually increased to recover higher-quality cells. Using a course cell-type annotation based on manual annotation of clusters, we identified cell types that were consistently filtered out, even though their quality did not appear to be spuriously low by manual inspection. We retained mast cells and hematopoietic stem cells from all tissues, all macrophages from LNs and spleen, all erythrocytes and platelets from BM and all monocytes from liver.

We concatenated cells from all tissues and computed the 10,000 top highly variable genes using the seurat_v3 option in Scanpy and used the same parameters as described above, but with a mini-batch size of 1,024 to accelerate the training process. We used this integrated latent space to assign initial cell types and removed all cell types that were not labeled as immune cells. Additionally, we removed all cells for which manual labeling and automatic labeling using MMoCHi (see below) were inconclusive about coarse cell-type identity (for example, B cell, myeloid, T cell). These events were of low quality by manual inspection. We performed post hoc manual removal of these and other clusters of low-quality cells after integrating all cells.

Immune cell subset classification using MMoCHi

To identify canonical immune cell subsets, we used a recently reported, supervised machine learning algorithm, MMoCHi) (v0.2.1)¹². We first normalized the gene expression (GEX) count matrix using log(10,000C_g,i / T_G,i + 1), where C_g,i represents the counts for GEX feature g in cell i, and T_G,i is the total counts for all GEX features in cell i. Similarly, we normalized the antibody-derived tag (ADT) count matrix using log(1,000C_a,i / T_A,i + 1), where C_a,i represents the counts for ADT feature a in cell i and T_A,i is the total counts for all ADT features in cell i. We applied landmark registration (MMoCHi) to batch-correct the ADT expression across experimental batches. In brief, we applied automatic detection of landmarks (peaks) in the expression distributions for each ADT feature in a given sample, applying manual adjustments as needed using the graphical user interface, then performed curve registration and warping to align the positive and negative peaks for each ADT feature across batches.

We provided MMoCHi with a hierarchy of immune cell subsets and their canonical surface protein-level and RNA-level markers (Supplementary Fig. 2) and used the markers to identify high-confidence members (cells) of each subset for training. For each classification level, automatic thresholds for high-confidence ADT or GEX marker-positive and marker-negative cells were manually adjusted as needed using the supplied GUI (Supplementary Table 3). Following MMoCHi’s internal training data refinement, we applied an 80–20 train–test split and trained a random forest classifier, sklearn.ensemble.RandomForestClassifier, on both gene and protein expression¹². For 2 of the 24 organ donors and a subset of samples from a third donor, we did not perform CITE-seq and only had scRNA-seq profiles. Thus, these samples were excluded from the MMoCHi classification described here. However, we used a k-nearest neighbors approach to transfer the classifier labels to individual cells profiled from these two organ donors. Specifically, we used the sklearn.neighbors.KNeighborsClassifier with n_neighbors = 10 to construct a k-nearest neighbors graph in the mrVI embedding of the dataset (see below) and classify the remaining cells. Of the subsets, pDCs were identified using two separate nodes on the hierarchy (Supplementary Fig. 1), as pDCs shared expression with both B cells and myeloid cells. Once classified, the two subsets were merged into a single population of pDCs. The MMoCHi annotation was used at two separate levels throughout the paper, defined as either one of the 34 fine-grained subsets or grouped into CD4⁺ T cells, CD8⁺ and unconventional T cells (including γδ T cells and CD8⁺ MAIT cells), B cells, NK cells and ILCs or myeloid cells (including monocytes, macrophages, cDCs, migratory DCs and pDCs).

Training and application of label transfer models

Owing to the breadth of tissues and human subjects sampled in our dataset and high-resolution annotation of immune subsets, we anticipated that our immune atlas would be useful to the research community as a reference for performing cell-type label transfer. To facilitate this application, we trained a model using popV²², a tool developed for cell-type label transfer that uses several annotation algorithms and consensus voting to determine annotations and evaluate their confidence. popV also calculates joint embeddings of the query and reference datasets, which can be used for visualization of the query data and other analysis tasks. A popV model was trained using the tissues and MMoCHi annotations (Fig. 2) as the reference dataset. Label transfer performance was evaluated using the Human Lung Cell Atlas as a query dataset⁸⁹. To visualize the data, we computed UMAP embeddings as described above on joint scVI embeddings, which were calculated as part of the popV pipeline.

To evaluate the importance of ADT information in the MMoCHi classification performance, we additionally applied a pre-trained CellTypist⁹ model (Immune_All_Low; https://celltypist.cog.sanger.ac.uk/models/Pan_Immune_CellTypist/v2/Immune_All_Low.pkl) using default settings to the tissue immune cells (Fig. 2).

Integration and cell state embedding using MrVI

To integrate scRNA-seq profiles of immune cells in our study, we first used scVI, which did not yield a fully integrated latent space and clustered by site of collection (for example, US or UK). We next leveraged MrVI, which uses a mixture-of-Gaussian as a prior and enforces stronger separation of true cell state and effect of donors on gene expression, as has been recently demonstrated¹⁰. MrVI takes advantage of a prior based on a multimodal variational mixture of posteriors (similar to a VampPrior⁹⁰), which have been shown to outperform Gaussian priors for scRNA-seq integration in benchmarking studies⁹⁰. In brief, MrVI finds a sample-agnostic latent space, U, and computes a sample-specific embedding. A second latent space, Z, is defined by adding an attention-based concatenation between U and the sample embedding space to the original U-space. Another layer of attention is used to incorporate an embedding of 10× Genomics chemistry and experimental site (Cambridge, UK versus Columbia, NY), and this third latent space is decoded using a linear decoder to yield the rate of a negative binomial distribution. We use a cell-type-aware Gaussian mixture prior in U-space. To introduce cell type awareness, we use a bias to the mixture proportions that makes it likely for cells of the same type to be sampled from the same Gaussian.

For the latent embedding highlighted throughout the article and used for manual cell-type curation, we used the donor identities as the sample keys and used the output of MMoCHi classification (see above) as the cell-type prior in MrVI. We used default parameters except n_epochs_kl_warmup of 25, n_latent_u of 20, n_latent in Z-space of 200, dropout in qz as well as pz of 0.03 (adopted from a previous publication¹⁰). To visualize cells (either the total immune component or individual major lineages), we computed nearest neighbors (scanpy.pp.neighbors) on the MrVI U latent space and calculated UMAP embeddings (scanpy.tl.umap) using the 15 nearest neighbors, a minimum distance of 0.4, a spread of 1.0 and initialization with PAGA after running scanpy.tl.paga. To identify additional heterogeneity in cell states within samples in addition to the cell-type annotation provided by MMoCHi, we performed manual annotation. For each MMoCHi annotated population, a new scVI model was trained with donor as the batch key, then Leiden clustering (scanpy.tl.leiden) was performed on the lineage-specific neighbors graph at an appropriate resolution, selected to minimize over-clustering (ranging from 1 to 15). Markers for each cluster were computed by scanpy.tl.rank_genes_groups, and clusters with similar marker expression were merged. To annotate proliferating cells, scanpy.tl.score_genes_cell_cycle was run, and the output was used in combination with the gene expression of MKI67 and TOP2A.

DE and variance decomposition using dreamlet

We focused our DE analysis on immune lineages and cell types with sufficient representation across experimental sites, tissues and donor ages. This included six tissue groups: blood, BM, spleen, gut (JLP and JEL), LNs (ILN, LLN and MLN) and lungs (consisting of BAL and parenchyma); six immune lineages: myeloid, CD4⁺ T cells, CD8⁺ T cells, invariant T cells (that is, γδ T cells and MAIT cells), B cells and ILC/NK cells; and 26 individual cell types within all lineages. Covariates included 10× Genomics chemistry (3′ versus 5′), sex (male versus female), laboratory (Cambridge, UK versus Columbia, NY) and CMV status (positive versus negative). For aging analyses, donors were categorized as being <40 or >40 years of age.

Variance decomposition and pseudobulk DE analysis were performed using LMM through the dreamlet R package (v1.4.1)²³. Depending on the resolution of the analysis, DE was performed separately either for each immune lineage (for example, myeloid cells, B cells, and so on) or for each immune subset (for example, macrophages, naive B cells and so on) using the cluster_id parameter in dreamlet. The raw GEX count matrix was pseudobulked across samples, and each tissue in each donor was treated as a separate sample. Before performing DE, samples and genes with poor representation were filtered using dreamlet::processAssays. Samples with fewer than 50 cells and genes not represented in at least 40% of the samples with at least five counts were excluded. To confirm findings by MrVI counterfactual analysis (see below), these thresholds were reduced to a minimum of ten cells for a sample to be included, and at least 10% of samples with at least five counts. DE for a subset was not performed when fewer than three or four samples (for tissue and age analysis, respectively) met the minimum cell thresholds. Variance decomposition was performed for age analysis for each lineage using dreamlet::fitVarPart with sex, sequencing chemistry, CMV serostatus, age group, processing site and tissue as covariates (Supplementary Table 13). LMM was performed using dreamet::dreamlet with eBayes estimation enabled. Tissue effects (Figs. 3 and 4, Extended Data Figs. 3 and 4, and Supplementary Tables 4, 7 and 10) were modeled by comparing each lineage/subset in one tissue group to the same lineage/subset in the remaining tissue groups, with donor identity encoded as a random effect. Age effects (Figs. 5–7, Extended Data Fig. 5, and Supplementary Tables 15 and 24) were modeled across each tissue-group–age-group combination while controlling for CMV serostatus and sex as fixed effects and with sequencing chemistry and processing site as random effects. Age effects within each tissue group were then measured using the contrasts parameter in dreamlet::dreamlet between old and young for each tissue group (for example, the effect of age in the gut was computed as ‘old-gut − young-gut’). CMV effects (Supplementary Fig. 10 and Supplementary Table 21) were modeled across each tissue-group–CMV serostatus combination while controlling for age and sex as fixed effects and with sequencing chemistry and processing site as random effects. CMV effects within each tissue group were then measured using the contrasts parameter in dreamlet::dreamlet between CMV⁺ and CMV⁻ for each tissue group.

Identification of gene co-expression patterns using consensus scHPF

For identifying cross-tissue and cross-donor gene signatures for each major immune lineage, we constructed probabilistic factor models directly from scRNA-seq count matrices using scHPF. The output of scHPF includes two matrices: an M × K gene score matrix containing weights for each of M genes in each of K factors and a K × N cell score matrix containing weights for each of N cells in each of K factors. In the original report of scHPF, the algorithm required a user-supplied value of K, the number of factors in the model⁹¹. Here, we use a new consensus factorization implementation of scHPF, in which the user specifies a broad range of K values from which many scHPF models are generated²⁷. The gene score matrices for these models are then clustered to identify K recurrent factors, which are combined to seed a final round of training to construct a final consensus model with K factors.

We constructed two types of scHPF models: a tissue-level model (Extended Data Fig. 4), in which the number of cells from each of three tissue groups was balanced by random sub-sampling (gut: JEL and JLP; lung parenchyma; and LNs: MLN and LLN), and a donor-level model (Figs. 5 and 6), in which the number of cells from each organ donor was balanced. We constructed both types of models for the major immune lineages: CD4⁺ T cells, CD8⁺ T cells (including all invariant T cells), NK cells, ILCs, B cells and macrophages. For donor models, donors with fewer than 300 cells for a given lineage were removed. In both models, the count matrices were randomly downsampled such that the average number of transcripts per cell was the same for each donor to avoid coverage bias. scHPF models considered only protein-coding genes (excluding TCR and immunoglobulin cassettes) detected in at least 1% of cells across the final subsampled and downsampled training matrix.

For all consensus scHPF models, we ran scHPF five times for each of 16 values of K (15–30), from which we selected the top three models for each value of K based on convergence criteria for clustering. We applied walktrap clustering to identify recurrent clusters, which we required to form clusters with factors from at least two different models from which we trained the final consensus model²⁷.

Detection of tissue-specific effects

Immune subset composition within each lineage across tissues was visualized by violin plots or box plots (using seaborn). Tissue-specific enrichment of immune subset frequencies in specific tissues was also assayed within each major lineage using scCODA⁹² for Bayesian inference. Significant enrichment of an immune subset in one tissue over the rest was determined using sccoda.util.comp_ana.CompositionalAnalysis to detect credible effects, and was run sequentially, selecting each cell type as the reference. Majority voting was then used to identify cell types that are credibly changing more than half the time with automatic reference-subset selection and the default false discovery rate of 0.05.

To determine tissue-specific gene expression signatures across immune lineages (Fig. 3), significant DEGs were defined as adjusted P < 0.05 and log₂(FC) > 1 by pseudobulk DE across tissues at the lineage level (see above). Mean z-score gene expression was calculated for each pairing of tissue group and lineage. Genes and samples were both hierarchically clustered using scipy.cluster.hierarchy.linkage with Ward’s method and Euclidean distance. Discrete clusters of genes with similar expression patterns were calculated using scipy.cluster.hierarchy.fcluster with the ‘maxclust’ method (Supplementary Table 6).

For each gene cluster, association with specific tissue groups or lineages could arise from DE within one or more specific subsets of that lineage or from compositional shifts in subsets across tissues. To disentangle these possibilities, we first used pre-ranked GSEA to compare the gene clusters identified via lineage-level DE to the effect size (that is, log(FC)) of DE across tissues in the subset-level DE. To visualize potential effects caused by compositional shifts across tissues, we computed the average frequency of the subset (as a proportion of the total cells within that lineage group) within a tissue, the FC of that frequency over the frequency in the remaining tissue groups and the average expression of the gene cluster.

To assess whether differential transcript expression was reflected in the surface protein profiling (Supplementary Tables 11 and 12), we selected ADTs corresponding to DEGs in at least one tissue. To identify enrichment in one tissue group over the other tissue groups, we used scanpy.tl.rank_genes_groups on the normalized expression with Wilcoxon and tie-correction enabled. To minimize the influence of technical staining artifacts or donor covariates, analysis was conducted separately within each donor. Donors with fewer than 50 cells for a particular tissue-group–lineage combination or tissue-group–lineage combinations with fewer than four suitable donors were excluded from analysis. Before DE analysis, the ADT count matrix was subsampled to equalize cell numbers and randomly downsampled such that the average number of transcripts per cell was the same for each group to avoid coverage bias.

We next sought to identify factors from the tissue-level scHPF models of each major immune lineage that were shared across cell types. As described above, we first constructed consensus scHPF models for CD4⁺ T cells, CD8⁺ T cells, macrophages, NK cells, ILCs and B cells with equal representation of cells from each of three major tissue groups (gut, lung and LNs). From each model, we removed probable nuisance factors containing heat shock protein-encoding genes (common dissociation artifact, >1 gene), ribosomal protein-encoding genes (common coverage artifact, >10 genes), genes from the highly inducible metallothionein cluster (>1 gene), hemoglobin transcripts (red blood cell contamination, >0 genes) and genes in a previously published signature of dissociation-induced cell stress in scRNA-seq (>7 genes)⁹³ among the 30 top-weighted genes. Next, we computed the average cell score for each factor in each of the three major tissue groups and identified all factors with an average tissue-group score that was at least 80% higher in one tissue group than the average of the remaining two. Thus, the resulting set of 53 scHPF factors from across all 6 lineage-specific models exhibits some degree of tissue specificity. To compare these factors to each other, we computed the Pearson correlation between the gene score vectors for each pair of factors. We then identified factors with a pairwise correlation that was greater than the 95% confidence threshold with at least two other factors, which yielded 31 scHPF factors from across the six major immune lineages. Finally, we performed hierarchical clustering of the Pearson correlation matrix for these 31 factors (seaborn.clustermap using Euclidean distances) to identify modules containing factors with similar gene signatures that originated from different, lineage-specific scHPF models (Extended Data Fig. 4). Modules of genes were further interrogated by average gene expression and validated in specific immune subsets using pseudobulk GEX DE and ADT DE as described above.

Detection of age-specific and CMV effects on composition and gene expression

To detect shifts in the subset composition of specific lineages across the age groups and CMV serostatus (Supplementary Table 14), we performed generalized linear modeling by fitting a statsmodels.GLM model for each tissue subset, considering sex, sequencing chemistry, CMV serostatus and processing site as additional covariates. Donors with fewer than 50 cells for a particular tissue-group–lineage combination or tissue-group–lineage combinations with fewer than four suitable donors were excluded from analysis. The estimated coefficients were used to calculate a covariate-aware log₂(FC) for visualization. We used the statsmodels.multipletests function to adjust P values for multiple comparisons (Benjamini–Hochberg method), and subset–tissue combinations with adjusted P < 0.05 were considered significantly changing across age.

To depict age-associated effects on the immune system, we visualized the similarity of trending DEGs by age on immune subsets across tissues using t-distributed stochastic neighbor embedding (t-SNE) (Fig. 5d). We first calculated trending DEG (unadjusted P < 0.05; <40 or >40 years old) pairwise similarities by summing the intersection of positively regulated genes (log₂(FC) > 0.1) and negatively regulated genes (log₂(FC) < −0.1), divided by the overall union of both. The similarity or distance (1 − similarity) was applied to cell types containing more than 70 DEGs (unadjusted P < 0.05; mean log-normalized expression, >0.05) and present in at least 3 donors per tissue and age group. The similarity levels of cell types and tissues with more than 200 DEGs were further clustered using the Ward.D2 method and projected into a distance-based t-SNE illustration.

To investigate the effect of age on specific genes within each immune subset, we plotted genes that were significant in at least one tissue (adjusted P < 0.05) and within the top 50 significant genes. Although our power to detect age effects by DE was limited, genes that were significantly DE in one subset were often trending in the same direction across multiple tissue groups. To assess the effect of age on surface protein expression (Supplementary Table 17), we used landmark-registered protein expression data (by MMoCHi, see above) to account for donor-to-donor batch effects in ADT staining quality. Although landmark registration preserves the separation between positive-expressing and negative-expressing cells for thresholding, this non-parametric normalization can obscure changes in overall expression intensity between samples. Therefore, we focused on shifts in percent positivity for a marker in each tissue subset. We performed automatic thresholding by MMoCHi, followed by manual adjustment (as described above), on the landmark-registered expression of all ADTs corresponding to a DEG by age. The percentage of cells with expression of a given ADT above the positive threshold was calculated for each donor–tissue-group–subset combination. Donors with fewer than 50 cells for a particular tissue-group–lineage combination or tissue-group–lineage combinations with fewer than four suitable donors were excluded from analysis. The percent positivity was used as the response variable in the same linear regression model used to detect shifts in composition across age groups. We adjusted P values for multiple comparisons as above, and ADTs with adjusted P < 0.05 were considered significant.

We constructed donor-level scHPF models for each major immune lineage with uniform representation of cells from each donor to identify age-associated gene signatures, as described above. For each scHPF model, we performed LMM to account for covariates and identify age associations. Each LMM contained six categorical covariates as fixed effects: age group, sequencing chemistry, sex, processing site and CMV serostatus. We also considered three tissue types: mucosal (BAL, lung parenchyma, JLP, JEL), LNs (ILN, MLN and LLN) and blood-rich, including blood, BM and spleen), which required us to select one category (blood-rich) as a held-out variable. Thus, we have two categorical variables for tissue, which effectively represent mucosal versus blood-rich and LN versus blood-rich. We encoded donor identity as a random effect. LMM coefficients and P values were computed for each factor in a given scHPF model using the cell scores as response variables by fitting a statsmodels.MixedLM model and using the statsmodels.multipletests function to adjust P values for multiple comparisons (Benjamini–Hochberg method).

To cross-validate age-associated scHPF factors in other datasets, we further analyzed a bone marrow atlas⁵⁴ containing 36 age-annotated donors with good B cell representation for a B cell aging factor, a lung atlas⁴⁷ containing 29 age-annotated donors with good macrophage and CD8⁺ T cell coverage and PBMC data from the Sound Life cohort (age 25–65 years, n = 96) from the Human Immune Health Atlas⁴⁹ (Figs. 5 and 6 and Extended Data Fig. 5). Using the published cell-type annotations from each atlas, we extracted the appropriate scRNA-seq profiles and projected them into the corresponding donor-level scHPF models generated from the data reported here using the scHPF project function. This resulted in cell scores for cells from the external data sets for the same factors that were generated from this data set, allowing us to compare the average cell scores for young versus older donors from the external data. As an orthogonal approach, we also performed pseudobulk DE analysis between older and younger donors (using an age cutoff of 40 years) from the external data sets, ranked the genes by FC and used GSEA to analyze the statistical enrichment of age-associated factors among young versus old donors (Supplementary Table 19). We used the top 200 genes (ranked by scHPF gene score) in each age-associated factor as gene sets for GSEA.

Counterfactual analysis to detect age associations with MrVI

For the DE analysis described in Fig. 7, we subsetted the MrVI sample embeddings to each tissue group and modeled the predicted ε in MrVI by a linear model adjusting for covariates in sex, CMV serostatus and age group as fixed effects, considering sequencing chemistry and processing site as site covariates in MrVI. A ridge regression parameter of 0.1, owing to collinearity of cofactors, was added. This decomposition of ε was performed for every single cell. This yields an estimated effect in Z-space for each covariate. The effect vector was added to the mean cell embedding in Z-space, and DEGs were computed based on the modified and mean embedding for each cell. For downstream analysis, this matrix of estimated log(FC) for each cell and gene was further processed for each immune subset. First, all cells that were represented only in fewer than three samples were filtered out¹⁰. Second, for each cell type, we excluded genes with less than 0.01 raw average expression or an estimated log(FC) across age groups with a 95th percentile below 0.1, retaining only genes that might be affected by age in a group of cells. To dissect predicted gene effects into modules, neighborhood smoothing was performed using 15 nearest neighbors in U-space and multiplying two times the normalized affinity matrix by the predicted gene effects. Spectral co-clustering was performed with four gene clusters and four cell clusters, with mini-batch enabled using sklearn.cluster.SpectralCoclustering. Marker genes for each module were identified by averaging the predicted log(FC) across all cells from the corresponding cell module, and the top 50 genes for each module were identified. We used decoupleR-py to compute a module score of log(FC) scores using weighted means of the signs of those marker genes⁹⁴ (Supplementary Table 23).

For the lung, we isolated a gene module in CD4⁺ T cells that contained T_RM cells. To detect similar cells in other tissues, we computed the best cutoff for the module score to identify cells in a specific cell module based on Youden’s J statistic, computed the module score for all cells from other tissues as described above and applied the same cutoff to all other tissues as the tissue of interest. Given that the gut contained T_RM cells with a T_H17 phenotype and all other tissues had no module-positive cells, we selected all cells with a MrVI predicted negative log₂(FC) of IL17A below −0.05. To confirm our findings on a per-gene level, we selected module-positive cells and used pseudobulk estimates of DE using dreamlet (Supplementary Table 24). Samples with fewer than 5 module-positive cells or 1,000 total counts and genes with fewer than 3 total counts were removed. Aging DE was performed using the contrasts method, as described above. Genes within a shared functional group were manually selected from the MrVI signature for visualization. Pseudobulk DE analysis was performed on the classifier-predicted cells in other tissues using the same settings as above in this cell subset. Enrichment of module or selected marker genes in the pseudobulk DE analysis was performed using GSEA implemented in decoupler.run_gsea (Supplementary Table 23).

TCR and BCR repertoire analysis using Dandelion

Cell Ranger-mapped TCR and BCR contigs contained in ‘all_contigs.fasta’ and ‘all_contig_annotations.csv’ output files were re-annotated using the Dandelion preprocessing pipeline⁹⁵. This pipeline includes the following steps: (1) sample suffix or prefix assignment to each sample barcode; (2) re-annotation of contigs with IgBLAST (v1.19.0)⁹⁶ against IMGT (international ImMunoGeneTics) reference sequences (last downloaded on 24/04/2023); (3) re-annotation of D and J genes separately using blastn to enable the annotation of contigs without the V gene present; and (4) identification and recovery of nonoverlapping individual J gene segments. For BCRs, three additional steps were also performed: (1) additional re-annotation of heavy-chain constant (C) region calls using blastn (v2.13.0) against curated sequences from CH1 regions of respective isotype class; (2) heavy-chain V gene allele correction using TIgGER (v1.0.1)⁹⁷; and (3) BCR mutation calling. Cell-level quality control was performed using Dandelion’s ‘filter_contigs’ function, which only considers productive VDJ contigs, asserts that a single cell should only have one VDJ and one VJ pair or only an orphan VDJ chain and explicitly removes contigs that fail these checks (except for IgM/IgD and TRB/TRD extra pairs). Contigs that did not match any cell barcodes in the gene expression data were also removed at this step. TCRs and BCRs were then grouped into clones or clonotypes. The following default sequential criteria, which apply to both chain contigs, were applied: (1) identical V and J genes usage; (2) identical junctional CDR3 amino acid length; and (3) CDR3 sequence similarity: 100% nucleotide sequence identity at the CDR3 junction for TCRs and 85% amino acid sequence similarity (based on Hamming distance) for BCRs.

TCR or BCR data were then transferred into the corresponding AnnData object. Cells without receptor data or that presented more than one receptor were discarded from further immunoreceptor analysis. For T cell analysis, cells annotated as MAIT cells or ɣδ T cells were also discarded. Clonality of the different populations was calculated as 1 − Pielou’s evenness index, varying from zero (more diverse) to one (less diverse), with the Pielou’s evenness corresponding H_s / H_max, where H_s is the Shannon entropy of sample s and H_max = log₂C, where C is the number of unique clonotypes in s. All clonality scores were calculated on a subsample of 100 cells for each donor, cell type, tissue or cell type and tissue. To detect shifts in the BCR isotype composition of specific B cell lineages across the age groups (Supplementary Table 20), we performed generalized linear modeling by fitting a statsmodels.GLM model for each tissue subset, considering sex, sequencing chemistry, CMV serostatus and processing site as additional covariates. Donors with fewer than 50 cells for a particular tissue-group–lineage combination or tissue-group–lineage combinations with fewer than for suitable donors were excluded from analysis. The estimated coefficients were used to calculate a covariate-aware log₂(FC) for visualization. We used the statsmodels.multipletests function to adjust P values for multiple comparisons (Benjamini–Hochberg method), and subset–tissue combinations with adjusted P < 0.05 were considered significantly changing across age.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41590-025-02241-4.

Supplementary information

Supplementary Information^{(8.8MB, pdf)}

Supplementary Figs. 1–10.

Reporting Summary^{(93KB, pdf)}

Peer Review File^{(126KB, pdf)}

Supplementary Table 1^{(56.9KB, xlsx)}

Donor Metadata. Metadata for each human organ donor sampled in this study, including the processing site, whether the donation was performed after brain death (DBD) or cardiac death (DCD), the age or age range (in years) of the donor, sex, ethnicity/race and EBV/CMV serostatus.

Supplementary Table 2^{(66.6KB, xlsx)}

TotalSeq panels for CITE-seq. A list of antibodies sourced from BioLegend used for CITE-seq surface protein staining. The target antigen and conjugated barcode sequence for all antibodies included in our custom TotalSeq-A (for 3′ 10× chemistry) and TotalSeq-C (for 5′ 10× chemistry) cocktails. Where possible, the clone and associated gene identity are also provided.

Supplementary Table 3^{(17.3KB, xlsx)}

MMoCHi hierarchy. Subset definitions and thresholds used to construct the MMoCHi hierarchy. Hierarchical relationships are defined by subset parents (subsets with no parent represent the first classification level). High-confidence cells for training are selected as events matching the listed marker expression (above the positive threshold for all ‘positive’ markers, and below the negative threshold for all ‘negative’ markers). Thresholds for gene expression markers are defined using log(CP10k + 1). Thresholds are defined for the listed protein markers in landmark-registered expression (A.U.).

Supplementary Table 4^{(51.2MB, csv)}

Lineage-level differential expression across tissues. Pseudobulk differential gene expression was performed using dreamlet linear mixed modeling across tissue groups for each lineage: CD4 T cells (cd4), CD8 T cells (cd8), γδ T and MAIT cells (cd3_unconv), myeloid cells (m), B cells (b) and innate lymphocytes (nk_ilcs). Before differential expression, for each lineage–tissue combination, genes with poor representation were removed (see Methods). Each row represents the differential expression of one gene (column name: “id”) in one lineage (column name: “assay”) in one tissue group (column name: “tissue_group”) against the remaining tissue groups. For each gene tested, we report the estimated log₂(FC) (“logfc”), average expression (“aveexpr”), t-statistic (“t”), P value (“p_value”), FDR-corrected P value (“adj_p_val”), the log-odds that the gene is differentially expressed (“b”) and the P value transformed into a signed z-score (“z-std”), as output by the dreamlet::topTable function for the tissue coefficient.

Supplementary Table 5^{(6KB, csv)}

Lineage-level differential expression across tissues donor table. Pseudobulk differential gene expression was performed using dreamlet linear mixed modeling across tissue groups for each lineage: CD4 T cells (cd4), CD8 T cells (cd8), γδ T and MAIT cells (cd3_unconv), myeloid cells (m), B cells (b) and innate lymphocytes (nk_ilcs). Genes with poor representation in each lineage–tissue combination were removed before analysis (see Methods). Each row corresponds to a tissue-lineage combination (column: “tissue_lineage”) and includes donor IDs with >50 events in that tissue-lineage (column: “donors”), the number of donors (column: “n_donors”), the tissue group (column: “tissue_group”) and the lineage (column: “lineage”).

Supplementary Table 6^{(409.5KB, xlsx)}

Hierarchical clustering of DEGs across tissues and GSEA results. Cluster identity and corresponding cluster color for genes significantly enriched in one tissue over others at the lineage level (P-adj < 0.05, log₂FC > 2) following hierarchical clustering (see Methods). Pre-ranked GSEA was performed to identify enrichment of cluster genes within tissue-related DEGs for specific immune subset–tissue-group combinations. Columns include: cluster identifier (“Term”), enrichment score (“ES”), normalized enrichment score (“NES”), nominal P value (“NOM p-val”), FDR-corrected P value (“FDR q-val”), family-wise error rate-controlled P value (“FWER p-val”), leading-edge genes (“Lead_genes”), the cell subset (“Subset”) and the tissue group (“Tissue_group).

Supplementary Table 7^{(179.3MB, csv)}

Subset-level differential gene expression across tissues. Pseudobulk differential gene expression was performed using dreamlet linear mixed modeling across tissue groups for each immune subset with sufficient representation across tissues (see Methods). Before differential expression, for each subset–tissue combination, genes with poor representation were removed (see Methods). Each row represents the differential expression of one gene (column name: “id”) in one subset (column name: “assay”) in one tissue group (column name: “tissue_group”) against the remaining tissue groups. For each gene tested, we report the estimated log₂(FC) (“logfc”), average expression (“aveexpr”), t-statistic (“t”), P value (“p_value”), FDR-corrected P value (“adj_p_val”), the log-odds that the gene is differentially expressed (“b”) and the P value transformed into a signed z-score (“z-std”), as output by the dreamlet::topTable function for the tissue coefficient.

Supplementary Table 8^{(16.9KB, csv)}

Subset-level differential expression across tissues donor table. Pseudobulk differential gene expression was performed using dreamlet linear mixed modeling across tissue groups for each subset. Genes with poor representation in each lineage–tissue combination were removed before analysis (see Methods). Each row corresponds to a tissue-lineage combination (column: “tissue_subset”) and includes donor IDs with >50 events in that tissue-lineage (column: “donors”), the number of donors (column: “n_donors”), the tissue group (column: “tissue_group”) and the lineage (column: “subset”).

Supplementary Table 9^{(46.2MB, xlsx)}

scHPF cell and gene scores for tissue. Results of factorization analysis for the tissue balanced model, each tab represents either a lineage’s gene scores or cell scores.

Supplementary Table 10^{(84.2MB, csv)}

Subset-level differential gene expression across tissues for scHPF. Same as Supplementary Table 6, except only examining differential expression between the tissue groups included in the scHPF analysis: lung (“resp”), gut (“gut”) and lymph nodes (“ln”).

Supplementary Table 11^{(4.9MB, csv)}

Subset-level differential ADT expression across tissues for scHPF. Results of differential ADT expression by Wilcoxon rank-sum test for proteins whose coding genes displayed differential expression across tissues included in scHPF analysis: lung (“resp”), gut (“gut”) and lymph nodes (“ln”) (see Methods). U₁ test-statistic (“scores”), log₂(FC) (“logfoldchanges”), P value (“pvals”), FDR-corrected P value (“pvals_adj”) are reported for each cell subset in each CITE-seq staining batch, across each tissue-comparison.

Supplementary Table 12^{(3.9MB, csv)}

Subset-level differential ADT expression across tissues. Same as Supplementary Table 8, but for differential expression between all tissue groups: lung (“resp”), gut (“gut”), lymph nodes (“ln”), spleen (“spl”), blood (“blo”) and bone marrow (“bma”).

Supplementary Table 13^{(8.9MB, csv)}

Variance decomposition. Variance decomposition was performed by linear modeling of pseudobulked gene expression using dreamlet. For each gene in each lineage: CD4 T cells (cd4), CD8 T cells (cd8), γδ T and MAIT cells (cd3_unconv), myeloid cells (m), B cells (b), and innate lymphocytes (nk_ilcs), the fraction of gene expression variation that could be explained by sex (“sex”), sequencing chemistry (“chemistry”), CMV serostatus (“cmv”), donor-age (“age_group”; binned into over/under 40 years old), processing site (“site”) or tissue (“tissue”) as well as the remaining (unexplained) residuals (“Residuals”) is reported.

Supplementary Table 14^{(125KB, csv)}

Compositional shifts in immune cell subsets with age. Linear mixed modeling of the tissue-specific composition of immune subsets within each lineage: CD4 T cells (CD4), CD8 T cells (including γδ T and MAIT cells; CD8), myeloid cells (M), B cells (B), and innate lymphocytes (NK/ILC) with fixed effects for sex, age, processing site and CMV serostatus. For each covariate, the estimated coefficient (“coeff”), the standard error (“std_error”), the z-statistic (“z”), the P value, log-likelihood, FDR-corrected P value (“adj_p_val”) and estimated log₂(FC) (“logfc”) are reported. In addition, the number of samples used during linear modeling (and the associated donor identities) are reported (“n_observations” and “donors”, respectively).

Supplementary Table 15^{(150.7MB, csv)}

Subset-level differential gene expression across age. Pseudobulk differential gene expression was performed using dreamlet linear mixed modeling across age groups (over/under 40 years old) for each immune subset–tissue-group combination with sufficient representation (see Methods). Before differential expression, for each subset–tissue combination, genes with poor representation were removed (see Methods). Each row represents the differential expression of one gene (column name: “id”) in one subset (column name: “assay”) in one tissue group (column name: “tissue_group”) across age. For each gene tested, we report the estimated log₂(FC) (“logfc”), average expression (“aveexpr”), t-statistic (“t”), P value (“p_value”), FDR-corrected P value (“adj_p_val”), the log-odds that the gene is differentially expressed (“b”), and the P value transformed into a signed z-score (“z-std”), as output by the dreamlet::topTable function after using contrasts to identify age effects (see Methods).

Supplementary Table 16^{(22.1KB, csv)}

Subset-level differential gene expression across age donor table. Pseudobulk differential gene expression was performed using dreamlet linear mixed modeling across age groups (over/under 40 years old) for each immune subset–tissue-group combination with sufficient representation (see Methods). Before differential expression, for each subset–tissue combination, genes with poor representation were removed (see Methods). Each row corresponds to a tissue-group–subset combination (column: “tissue_groups_subset”) and includes donors with >50 events in that tissue-group subset (column: “donors”), the donors <40 years old (column: “donors_0”) and the donors >40 years old (column: “donors_1”).

Supplementary Table 17^{(2.1MB, csv)}

Subset-level differential ADT expression across age. Linear mixed modeling of the percent positivity of ADT expression for proteins whose coding genes displayed differential expression across tissues for each tissue-group–immune-subset combination (“tissue_subset”). Fixed effects included sex, age, processing site, CMV serostatus and sequencing chemistry. For each covariate, the estimated coefficient (“coeff”), the standard error (“std_error”), the z-statistic (“z”), the P value, log-likelihood, FDR-corrected P value (“adj_p_val”) and estimated log₂(FC) (“logfc”) are reported. In addition, the number of samples used during linear modeling (and the associated donor identities) are reported (“n_observations” and “donors”, respectively). For each protein in each tissue-group–immune-subset combination, the percent positivity within each young donor and each old donor is provided, as a list (“ave_percent_positive_young” and “ave_percent_positive_old”).

Supplementary Table 18^{(88.8MB, xlsx)}

scHPF cell and gene scores for aging. Results of factorization analysis for the donor balanced model to interrogate age effects, each tab represents either a lineage’s gene scores or cell scores.

Supplementary Table 19^{(29.1KB, csv)}

GSEA enrichment of aging factor genes in pseudobulk DE. Pre-ranked GSEA was performed to identify enrichment of the top 200 genes of each age-associated scHPF factor within the age-related DEGs of specific immune-subset–tissue-group combinations. The scHPF factor (“Term”), enrichment and normalized enrichment scores (“ES” and “NES”), nominal P value (“NOM p-val”), FDR-corrected P value (“FDR q-val”) and family-wise error rate-controlled P value (“FWER p-val”) as well as the leading-edge genes (“Lead_genes”). Nominal P values were further FDR-corrected for all comparisons made across a lineage (“adj_p_value”).

Supplementary Table 20^{(36KB, csv)}

Age-related changes in immunoglobulin isotype composition. Linear mixed modeling of the composition of IgM, IgD, IgG, or IgA isotypes in each tissue group for each B cell subset (“tissue_subset”) derived from BCR-sequencing. Fixed effects included sex, age, processing site and CMV serostatus. For each covariate, the estimated coefficient (“coeff”), the standard error (“std_error”), the z-statistic (“z”), the P value, log-likelihood, FDR-corrected P value (“adj_p_val”) and estimated log₂(FC) (“logfc”) are reported. In addition, the number of samples used during linear modeling (and the associated donor identities) are reported (“n_observations” and “donors”, respectively).

Supplementary Table 21^{(31.5MB, csv)}

CD8 T cell differential gene expression across CMV serostatus. Pseudobulk differential gene expression was performed using dreamlet linear mixed modeling across CMV serostatus. For each immune subset–tissue-group combination with sufficient representation within the CD8 T cell lineage (see Methods). Before differential expression, for each subset–tissue combination, genes with poor representation were removed (see Methods). Each row represents the differential expression of one gene (column name: “id”) in one subset (column name: “assay”) in one tissue group (column name: “tissue_group”) across age. For each gene tested, we report the estimated log₂(FC) (“logfc”), average expression (“aveexpr”), P value (“p_value”), and FDR-corrected P value (“adj_p_val”) as output by the dreamlet::topTable function after using contrasts to identify CMV effects (see Methods).

Supplementary Table 22^{(20KB, csv)}

GSEA enrichment of CMV-associated factor genes in pseudobulk DE. Pre-ranked GSEA was performed to identify enrichment of the top 200 genes of each CMV-associated scHPF factor within the CMV-related DEGs of specific immune-subset–tissue-group combinations within the CD8 T cell lineage. The scHPF factor (“Term”), enrichment and normalized enrichment scores (“ES” and “NES”), nominal P value (“NOM p-val”), FDR-corrected P value (“FDR q-val”) and family-wise error rate-controlled P value (“FWER p-val”) as well as the leading-edge genes (“Lead_genes”). Nominal P values were further FDR-corrected for all comparisons made across a lineage (“adj_p_value”).

Supplementary Table 23^{(107.1KB, xlsx)}

MrVI counterfactuals gene modules and gene enrichment. MrVI counterfactual analysis was used to predict age effects on a per-gene and per-cell basis for CD4 T cells in the LNG, LN and JEJ. Following spectral co-clustering, four modules of genes and cells predicted to change similarly with age were obtained. The average predicted effect size for each gene (“gene_symbol” and “gene_ids”) across all cells in these modules is reported (for example, “lfc_mean_module_0” for the first module). The cluster that each gene was associated with and whether the gene was selected as a marker for the module (“coclustering_genes” and “marker_genes”, respectively) are reported. Modules for the LNG, LN and JEJ CD4 T cells are provided on separate sheets. Pre-ranked GSEA was performed to identify enrichment of module genes or selected genes in module-positive cells within pseudobulk differential expression across age (see Methods). The gene sets queried (“Term”) for each tissue group (“tissue”), normalized enrichment scores (“NES”) and significance (whether FDR-corrected P value was below 0.05; column name: “Significance”) are reported in an additional sheet (“GSEA”).

Supplementary Table 24^{(26.2MB, csv)}

Subset-level differential gene expression across age for MrVI. Pseudobulk differential gene expression was performed using dreamlet linear mixed modeling between MrVI module-positive populations (see Methods) across age groups (over/under 40 years old) for each immune subset–tissue-group combination with sufficient representation. Before differential expression, for each subset–tissue combination, genes with poor representation were removed (see Methods). Each row represents the differential expression of one gene (column name: “id”) in one module (column name: “assay”) in one MrVI module-target tissue combination (column name: “model”) across age. For each gene tested, we report the estimated log₂(FC) (“logfc”), average expression (“aveexpr”), t-statistic (“t”), P value (“p_value”), FDR-corrected P value (“adj_p_val”), the log-odds that the gene is differentially expressed (“b”) and the P value transformed into a signed z-score (“z-std”), as output by the dreamlet::topTable function after using contrasts to identify age effects (see Methods).

Source data

Source Data Extended Data Fig. 4^{(17.9KB, xlsx)}

Pearson correlation heatmap for Extended Data Fig. 4a.

Acknowledgements

We thank the deceased organ donors and their families, the extended Cambridge Biorepository for Translational Medicine (CBTM) team, the transplant coordinators at LiveOnNY for providing access to tissue samples, and J. Zamanian of the Lattice: Human Cell Atlas team for data collation and availability on CELL×GENE. This work was supported by a Chan–Zuckerberg Initiative Seed Networks for the Human Cell Atlas grant (CZF2019-002452) to N.Y., D.L.F., P.A. Sims, J.L.J. and S.A.T., along with NIH grants AI106697 and AI128949 awarded to D.L.F. and P.A. Sims and NIHR Cambridge Biomedical Research Centre funding to J.L.J. (BRC-1215-20014). D.P.C. was supported by the Columbia University Graduate Training Program in Microbiology and Immunology (T32AI106711). Flow cytometry analysis was performed in the CCTI Flow Cytometry Core, supported by NIH S10RR027050 and S10OD020056 and the Cambridge NIHR BRC Cell Phenotyping Hub (Department of Medicine, University of Cambridge). This research was funded in part by the Wellcome Trust (203151/Z/16/Z, 203151/A/16/Z) and the UKRI Medical Research Council (MC_PC_17230). For the purpose of open access, the author has applied a CC BY public copyright license to any Author Accepted Manuscript version arising from this submission. The views expressed here are those of the author(s) and not necessarily those of the NIHR, Department of Health and Social Care or National Institutes of Health (NIH).

Extended data

Author contributions

N.Y., D.L.F., P.A. Sims, J.L.J. and S.A.T. designed the study. M.K., R.M., K.M. and K.S.-P. obtained tissues from deceased organ donors. S.B.W., D.B.R., P.A. Szabo, D.P.C., S.K.H., L.B.J., K.L.E. and D.C. performed tissue dissociation and prepared single-cell sequencing libraries. M.M., C.E., S.B.W., D.B.R., P.A. Sims, P.A. Szabo, D.P.C., A.R.M., E.R. and V.V.P.A. analyzed and interpreted the data. D.L.F., N.Y., P.A. Sims and J.L.J. supervised experiments. N.Y. and P.A. Sims coordinated data analysis. D.L.F., N.Y., P.A. Sims, J.L.J., S.A.T., S.B.W., D.B.R., M.M., P.A. Szabo, C.E. and A.R.M. wrote and edited the paper. N.Y., D.L.F., P.A. Sims, J.L.J. and S.A.T. provided funding. All authors read, provided input on and approved the paper.

Peer review

Peer review information

Nature Immunology thanks Dominic Gruen, Claire Gustafson and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available. Primary Handling Editor: Ioana Staicu, in collaboration with the Nature Immunology team.

Data availability

The global, T cell, NK/ILC, B cell and myeloid cell h5ad datasets are available on Lattice: Human Cell Atlas at https://cellxgene.cziscience.com/collections/cc431242-35ea-41e1-a100-41e0dec2665b. Raw sequencing (Fastq) files are available from the Sequence Read Archive (SRA) under accession no. SRP559768 and BioProject accession no. PRJNA1215450. Data are also available on Gene Expression Omnibus under accession number GSE299043. Source data are provided with this paper.

Code availability

The code for data processing and downstream analysis is available on the GitHub repository at https://github.com/YosefLab/CZI-Immuneaging.

Competing interests

In the past 3 years, J.L.J. has consulted for or been a member of a scientific advisory board for Sanofi, Roche and Enhanc3DGenomics, and S.A.T. S.A.T. is a scientific advisory board member of ForeSite Labs, OMass Therapeutics, Qiagen, Xaira Therapeutics, a co-founder and equity holder of TransitionBio and Ensocell Therapeutics, a non-executive director of 10x Genomics, and a part-time employee of GlaxoSmithKline. The other authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Steven B. Wells, Daniel B. Rainbow, Michal Mark, Peter A. Szabo, Can Ergen, Daniel P. Caron, Ana Raquel Maceiras.

These authors jointly supervised this work: Sarah A. Teichmann, Donna L. Farber, Peter A. Sims, Joanne L. Jones, Nir Yosef.

Contributor Information

Sarah A. Teichmann, Email: sat1003@cam.ac.uk

Donna L. Farber, Email: df2396@cumc.columbia.edu

Peter A. Sims, Email: pas2182@cumc.columbia.edu

Joanne L. Jones, Email: jls53@medschl.cam.ac.uk

Nir Yosef, Email: nir.yosef@weizmann.ac.il.

Extended data

is available for this paper at 10.1038/s41590-025-02241-4.

Supplementary information

The online version contains supplementary material available at 10.1038/s41590-025-02241-4.

References

1.Lam, N., Lee, Y. & Farber, D. L. A guide to adaptive immune memory. Nat. Rev. Immunol.24, 810–829 (2024). [DOI] [PubMed] [Google Scholar]
2.Terekhova, M. et al. Single-cell atlas of healthy human blood unveils age-related loss of NKG2C⁺GZMB⁻CD8⁺ memory T cells and accumulation of type 2 memory T cells. Immunity56, 2836–2854.e9 (2023). [DOI] [PubMed] [Google Scholar]
3.Thome, J. J. et al. Spatial map of human T cell compartmentalization and maintenance over decades of life. Cell159, 814–828 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Kumar, B. V., Connors, T. & Farber, D. L. Human T cell development, localization, and function throughout life. Immunity48, 202–213 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Dogra, P. et al. Tissue determinants of human NK cell development, function, and residence. Cell180, 749–763.e13 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Kumar, B. V. et al. Human tissue-resident memory T cells are defined by core transcriptional and functional signatures in lymphoid and mucosal sites. Cell Rep.20, 2921–2934 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Poon, M. M. L. et al. Tissue adaptation and clonal segregation of human memory T cells in barrier sites. Nat. Immunol.24, 309–319 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Yudanin, N. A. et al. Spatial and temporal mapping of human innate lymphoid cells reveals elements of tissue specificity. Immunity50, 505–519.e4 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Dominguez Conde, C. et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science376, eabl5197 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Boyeau, P. et al. Deep generative modeling for quantifying sample-level heterogeneity in single-cell omics. Preprint at bioRxiv10.1101/2022.10.04.510898 (2022).
11.Becht, E. et al. Dimensionality reduction for visualizing single-cell data using UMAP. Nat. Biotechnol.37, 38–44 (2018). [Google Scholar]
12.Caron, D. P. et al. Multimodal hierarchical classification of CITE-seq data delineates immune cell states across lineages and tissues. Cell Rep. Methods5, 100938 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Sallusto, F., Lenig, D., Forster, R., Lipp, M. & Lanzavecchia, A. Two subsets of memory T lymphocytes with distinct homing potentials and effector functions. Nature401, 708–712 (1999). [DOI] [PubMed] [Google Scholar]
14.Zheng, Y. & Rudensky, A. Y. Foxp3 in control of the regulatory T cell lineage. Nat. Immunol.8, 457–462 (2007). [DOI] [PubMed] [Google Scholar]
15.Szabo, P. A. et al. Single-cell transcriptomics of human T cells reveals tissue and activation signatures in health and disease. Nat. Commun.10, 4706 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Provine, N. M. & Klenerman, P. MAIT cells in health and disease. Annu. Rev. Immunol.38, 203–228 (2020). [DOI] [PubMed] [Google Scholar]
17.Miron, M. et al. Maintenance of the human memory T cell repertoire by subset and tissue site. Genome Med.13, 100 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Hashemi, E., McCarthy, C., Rao, S. & Malarkannan, S. Transcriptomic diversity of innate lymphoid cells in human lymph nodes compared to BM and spleen. Commun. Biol.7, 769 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Muramatsu, M. et al. Class switch recombination and hypermutation require activation-induced cytidine deaminase (AID), a potential RNA editing enzyme. Cell102, 553–563 (2000). [DOI] [PubMed] [Google Scholar]
20.Johnson, J. L. et al. The transcription factor T-bet resolves memory B cell subsets with distinct tissue distributions and antibody specificities in mice and humans. Immunity52, 842–855.e6 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Weisel, N. M. et al. Comprehensive analyses of B-cell compartments across the human body reveal novel subsets and a gut-resident memory phenotype. Blood136, 2774–2785 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Ergen, C. et al. Consensus prediction of cell type labels in single-cell data with popV. Nat. Genet.56, 2731–2738 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Hoffman, G. E. et al. Efficient differential expression analysis of large-scale single cell transcriptomics data using dreamlet. Preprint at bioRxiv10.1101/2023.03.17.533005 (2024).
24.Schneider, C. et al. Induction of the nuclear receptor PPAR-γ by the cytokine GM-CSF is critical for the differentiation of fetal monocytes into alveolar macrophages. Nat. Immunol.15, 1026–1037 (2014). [DOI] [PubMed] [Google Scholar]
25.Notta, F. et al. Isolation of single human hematopoietic stem cells capable of long-term multilineage engraftment. Science333, 218–221 (2011). [DOI] [PubMed] [Google Scholar]
26.Choi, Y. S. et al. LEF-1 and TCF-1 orchestrate T_FH differentiation by regulating differentiation circuits upstream of the transcriptional repressor Bcl6. Nat. Immunol.16, 980–990 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Levitin, H. M., Zhao, W., Bruce, J. N., Canoll, P. & Sims, P. A. Consensus scHPF identifies cell type-specific drug responses in glioma by integrating large-scale scRNA-seq. Preprint at bioRxiv10.1101/2023.12.05.570193 (2023).
28.Schuettpelz, L. G. et al. Kruppel-like factor 7 overexpression suppresses hematopoietic stem and progenitor cell function. Blood120, 2981–2989 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Kuwahara, M. et al. The transcription factor Sox4 is a downstream target of signaling by the cytokine TGF-β and suppresses T_H2 differentiation. Nat. Immunol.13, 778–786 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Kunkel, E. J. et al. Lymphocyte CC chemokine receptor 9 and epithelial thymus-expressed chemokine (TECK) expression distinguish the small intestinal immune compartment: epithelial expression of tissue-specific chemokines as an organizing principle in regional immunity. J. Exp. Med.192, 761–768 (2000). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Szabo, P. A. Axes of heterogeneity in human tissue-resident memory T cells. Immunol. Rev.316, 23–37 (2023). [DOI] [PubMed] [Google Scholar]
32.Tellier, J. et al. Unraveling the diversity and functions of tissue-resident plasma cells. Nat. Immunol.25, 330–342 (2024). [DOI] [PubMed] [Google Scholar]
33.Mass, E., Nimmerjahn, F., Kierdorf, K. & Schlitzer, A. Tissue-specific macrophages: how they develop and choreograph tissue biology. Nat. Rev. Immunol.23, 563–579 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Watanabe, K. et al. Requirement for Runx proteins in IgA class switching acting downstream of TGF-β1 and retinoic acid signaling. J. Immunol.184, 2785–2792 (2010). [DOI] [PubMed] [Google Scholar]
35.Gloury, R. et al. Dynamic changes in Id3 and E-protein activity orchestrate germinal center and plasma cell development. J. Exp. Med.213, 1095–1111 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Simmons, D. P. et al. SLAMF7 engagement superactivates macrophages in acute and chronic inflammation. Sci. Immunol.7, eabf2846 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Aegerter, H., Lambrecht, B. N. & Jakubzick, C. V. Biology of lung macrophages in health and disease. Immunity55, 1564–1580 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Kohyama, M. et al. Role for Spi-C in the development of red pulp macrophages and splenic iron homeostasis. Nature457, 318–321 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Hurtado-Navarro, L. et al. NLRP3 inflammasome activation and symptom burden in KRAS-mutated CMML patients is reverted by IL-1 blocking therapy. Cell Rep. Med.4, 101329 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
40.Ainscough, J. S., Gerberick, G. F., Kimber, I. & Dearman, R. J. Interleukin-1β processing is dependent on a calcium-mediated interaction with calmodulin. J. Biol. Chem.290, 31151–31161 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
41.Tokuhisa, M. et al. Expression and localisation of Rab44 in immune-related cells change during cell differentiation and stimulation. Sci. Rep.10, 10728 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Weichhart, T., Hengstschläger, M. & Linke, M. Regulation of innate immune cell function by mTOR. Nat. Rev. Immunol.15, 599–614 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Sica, A. & Mantovani, A. Macrophage plasticity and polarization: in vivo veritas. J. Clin. Invest.122, 787–795 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Roy, S. et al. Macrophage-derived neuropilin-2 exhibits novel tumor-promoting functions. Cancer Res.78, 5600–5617 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Lavin, Y. et al. Tissue-resident macrophage enhancer landscapes are shaped by the local microenvironment. Cell159, 1312–1326 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Bohlen, C. J., Friedman, B. A., Dejanovic, B. & Sheng, M. Microglia in brain development, homeostasis, and neurodegeneration. Annu. Rev. Genet.53, 263–288 (2019). [DOI] [PubMed] [Google Scholar]
47.Natri, H. M. et al. Cell-type-specific and disease-associated expression quantitative trait loci in the human lung. Nat. Genet.56, 595–604 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Pereira, B. I. et al. Sestrins induce natural killer function in senescent-like CD8⁺ T cells. Nat. Immunol.21, 684–694 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Gong, Q. et al. Longitudinal multi-omic immune profiling reveals age-related immune cell dynamics in healthy adults. Preprint at bioRxiv10.1101/2024.09.10.612119 (2024).
50.Pearce, E. L. et al. Control of effector CD8⁺ T cell function by the transcription factor Eomesodermin. Science302, 1041–1043 (2003). [DOI] [PubMed] [Google Scholar]
51.Davis, R. S. Fc receptor-like molecules. Annu. Rev. Immunol.25, 525–560 (2007). [DOI] [PubMed] [Google Scholar]
52.Mogilenko, D. A. et al. Comprehensive profiling of an aging immune system reveals clonal GZMK⁺ CD8⁺ T cells as conserved hallmark of inflammaging. Immunity54, 99–115.e12 (2021). [DOI] [PubMed] [Google Scholar]
53.Oh-hora, M., Johmura, S., Hashimoto, A., Hikida, M. & Kurosaki, T. Requirement for Ras guanine nucleotide releasing protein 3 in coupling phospholipase C-γ2 to Ras in B cell receptor signaling. J. Exp. Med.198, 1841–1851 (2003). [DOI] [PMC free article] [PubMed] [Google Scholar]
54.Lee, N. Y. S., Li, M., Ang, K. S. & Chen, J. Establishing a human bone marrow single cell reference atlas to study ageing and diseases. Front. Immunol.14, 1127879 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
55.Gordon, C. L. et al. Tissue reservoirs of antiviral T cell immunity in persistent human CMV infection. J. Exp. Med.214, 651–667 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
56.Persad, S. et al. SEACells infers transcriptional and epigenomic cellular states from single-cell genomics data. Nat. Biotechnol.41, 1746–1757 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Gaublomme, J. T. et al. Single-cell genomics unveils critical regulators of Th17 cell pathogenicity. Cell163, 1400–1412 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
58.Ivanov, I. I., Zhou, L. & Littman, D. R. Transcriptional regulation of Th17 cell differentiation. Semin. Immunol.19, 409–417 (2007). [DOI] [PMC free article] [PubMed] [Google Scholar]
59.Miron, M. et al. Human lymph nodes maintain TCF-1^hi memory T cells with high functional potential and clonal diversity throughout life. J. Immunol.201, 2132–2140 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
60.Zhou, X. et al. Differentiation and persistence of memory CD8⁺ T cells depend on T cell factor 1. Immunity33, 229–240 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
61.Krasemann, S. et al. The TREM2–APOE pathway drives the transcriptional phenotype of dysfunctional microglia in neurodegenerative diseases. Immunity47, 566–581.e9 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
62.Deczkowska, A., Weiner, A. & Amit, I. The physiology, pathology, and potential therapeutic applications of the TREM2 signaling pathway. Cell181, 1207–1217 (2020). [DOI] [PubMed] [Google Scholar]
63.de Oliveira Matos, A. et al. TREM-2: friend or foe in infectious diseases?. Crit. Rev. Microbiol.50, 1–19 (2024). [DOI] [PubMed] [Google Scholar]
64.Wu, K. et al. TREM-2 promotes macrophage survival and lung disease after respiratory viral infection. J. Exp. Med.212, 681–697 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Schlepckow, K., Morenas-Rodríguez, E., Hong, S. & Haass, C. Stimulation of TREM2 with agonistic antibodies—an emerging therapeutic option for Alzheimer’s disease. Lancet Neurol.22, 1048–1060 (2023). [DOI] [PubMed] [Google Scholar]
66.Lu, J. et al. Heterogeneity and transcriptome changes of human CD8⁺ T cells across nine decades of life. Nat. Commun.13, 5128 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
67.Carrasco, E. et al. The role of T cells in age-related diseases. Nat. Rev. Immunol.22, 97–111 (2022). [DOI] [PubMed] [Google Scholar]
68.Laphanuwat, P., Gomes, D. C. O. & Akbar, A. N. Senescent T cells: beneficial and detrimental roles. Immunol. Rev.316, 160–175 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
69.Rupareliya, M. & Shende, P. Therapeutic potential of stem cells in natural killer-like B cell-associated diseases. Adv. Exp. Med. Biol.1470, 57–72 (2024). [DOI] [PubMed] [Google Scholar]
70.Wang, S. et al. Natural killer-like B cells prime innate lymphocytes against microbial infection. Immunity45, 131–144 (2016). [DOI] [PubMed] [Google Scholar]
71.Zhang, W. et al. SARS-CoV-2 infection results in immune responses in the respiratory tract and peripheral blood that suggest mechanisms of disease severity. Nat. Commun.13, 2774 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
72.Szabo, P. A. et al. Longitudinal profiling of respiratory and systemic immune responses reveals myeloid cell-driven lung inflammation in severe COVID-19. Immunity54, 797–814.e6 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
73.Folorunso, O. S. & Sebolai, O. M. Overview of the development, impacts, and challenges of live-attenuated oral rotavirus vaccines. Vaccines8, 341 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
74.Fraietta, J. A. et al. Determinants of response and resistance to CD19 chimeric antigen receptor (CAR) T cell therapy of chronic lymphocytic leukemia. Nat. Med.24, 563–571 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
75.English, E. P. et al. Engineering CAR-T therapies for autoimmune disease and beyond. Sci. Transl. Med.16, eado2084 (2024). [DOI] [PubMed] [Google Scholar]
76.Guedan, S., Ruella, M. & June, C. H. Emerging cellular therapies for cancer. Annu. Rev. Immunol.37, 145–171 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
77.Carpenter, D. J. et al. Human immunology studies using organ donors: impact of clinical variations on immune parameters in tissues and circulation. Am. J. Transpl.18, 74–88 (2018). [Google Scholar]
78.Jensen, I. J., Wells, S. B., Gras, J. & Farber, D. L. Isolation and characterization of human tissue resident memory T cells. Curr. Protoc.5, e70120 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
79.Rainbow, D., Howlett, S., Jarvis, L. & Jones, J. Multi tissue processing for single cell sequencing of human immune cells v1. protocols.io10.17504/protocols.io.bz4qp8vw (2021).
80.James, K. R. et al. Distinct microbial and immune niches of the human colon. Nat. Immunol.21, 343–353 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
81.Reynolds, G. et al. Developmental cell programs are co-opted in inflammatory skin disease. Science371, eaba6500 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
82.Zheng, G. X. Y. et al. Massively parallel digital transcriptional profiling of single cells. Nat. Commun.8, 14049 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
83.Bernstein, N. J. et al. Solo: doublet identification in single-cell RNA-seq via semi-supervised deep learning. Cell Syst.11, 95–101.e5 (2020). [DOI] [PubMed] [Google Scholar]
84.Yang, S. et al. Decontamination of ambient RNA in single-cell RNA-seq with DecontX. Genome Biol.21, 57 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
85.Wolock, S. L., Lopez, R. & Klein, A. M. Scrublet: computational identification of cell doublets in single-cell transcriptomic data. Cell Syst.8, 281–291.e9 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
86.Lopez, R., Regier, J., Cole, M. B., Jordan, M. I. & Yosef, N. Deep generative modeling for single-cell transcriptomics. Nat. Methods15, 1053–1058 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
87.Hafemeister, C. & Satija, R. Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression. Genome Biol.20, 296 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
88.Blondel, V. D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. Fast unfolding of communities in large networks. J. Stat. Mech.2008, P10008 (2008). [Google Scholar]
89.Sikkema, L. et al. An integrated cell atlas of the lung in health and disease. Nat. Med.29, 1563–1577 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
90.Hrovatin, K. et al. Integrating single-cell RNA-seq datasets with substantial batch effects. Preprint at bioRxiv10.1101/2023.11.03.565463 (2024).
91.Levitin, H. M. et al. De novo gene signature identification from single-cell RNA-seq with hierarchical Poisson factorization. Mol. Syst. Biol.15, e8557 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
92.Büttner, M., Ostner, J., Müller, C. L., Theis, F. J. & Schubert, B. scCODA is a Bayesian model for compositional single-cell data analysis. Nat. Commun.12, 6876 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
93.Denisenko, E. et al. Systematic assessment of tissue dissociation and storage biases in single-cell and single-nucleus RNA-seq workflows. Genome Biol.21, 130 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
94.Badia-I-Mompel, P. et al. decoupleR: ensemble of computational methods to infer biological activities from omics data. Bioinform. Adv.2, vbac016 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
95.Suo, C. et al. Dandelion uses the single-cell adaptive immune receptor repertoire to explore lymphocyte developmental origins. Nat. Biotechnol.42, 40–51 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
96.Ye, J., Ma, N., Madden, T. L. & Ostell, J. M. IgBLAST: an immunoglobulin variable domain sequence analysis tool. Nucleic Acids Res.41, W34–W40 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
97.Gadala-Maria, D., Yaari, G., Uduman, M. & Kleinstein, S. H. Automated analysis of high-throughput B-cell sequencing data reveals a high frequency of novel immunoglobulin V gene segment alleles. Proc. Natl Acad. Sci. USA112, E862–E870 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Information^{(8.8MB, pdf)}

Supplementary Figs. 1–10.

Reporting Summary^{(93KB, pdf)}

Peer Review File^{(126KB, pdf)}

Supplementary Table 1^{(56.9KB, xlsx)}

Supplementary Table 2^{(66.6KB, xlsx)}

Supplementary Table 3^{(17.3KB, xlsx)}

Supplementary Table 4^{(51.2MB, csv)}

Supplementary Table 5^{(6KB, csv)}

Supplementary Table 6^{(409.5KB, xlsx)}

Supplementary Table 7^{(179.3MB, csv)}

Supplementary Table 8^{(16.9KB, csv)}

Supplementary Table 9^{(46.2MB, xlsx)}

scHPF cell and gene scores for tissue. Results of factorization analysis for the tissue balanced model, each tab represents either a lineage’s gene scores or cell scores.

Supplementary Table 10^{(84.2MB, csv)}

Supplementary Table 11^{(4.9MB, csv)}

Supplementary Table 12^{(3.9MB, csv)}

Supplementary Table 13^{(8.9MB, csv)}

Supplementary Table 14^{(125KB, csv)}

Supplementary Table 15^{(150.7MB, csv)}

Supplementary Table 16^{(22.1KB, csv)}

Supplementary Table 17^{(2.1MB, csv)}

Supplementary Table 18^{(88.8MB, xlsx)}

Supplementary Table 19^{(29.1KB, csv)}

Supplementary Table 20^{(36KB, csv)}

Supplementary Table 21^{(31.5MB, csv)}

Supplementary Table 22^{(20KB, csv)}

Supplementary Table 23^{(107.1KB, xlsx)}

Supplementary Table 24^{(26.2MB, csv)}

Source Data Extended Data Fig. 4^{(17.9KB, xlsx)}

Pearson correlation heatmap for Extended Data Fig. 4a.

Data Availability Statement

The code for data processing and downstream analysis is available on the GitHub repository at https://github.com/YosefLab/CZI-Immuneaging.

[CR1] 1.Lam, N., Lee, Y. & Farber, D. L. A guide to adaptive immune memory. Nat. Rev. Immunol.24, 810–829 (2024). [DOI] [PubMed] [Google Scholar]

[CR2] 2.Terekhova, M. et al. Single-cell atlas of healthy human blood unveils age-related loss of NKG2C⁺GZMB⁻CD8⁺ memory T cells and accumulation of type 2 memory T cells. Immunity56, 2836–2854.e9 (2023). [DOI] [PubMed] [Google Scholar]

[CR3] 3.Thome, J. J. et al. Spatial map of human T cell compartmentalization and maintenance over decades of life. Cell159, 814–828 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR4] 4.Kumar, B. V., Connors, T. & Farber, D. L. Human T cell development, localization, and function throughout life. Immunity48, 202–213 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR5] 5.Dogra, P. et al. Tissue determinants of human NK cell development, function, and residence. Cell180, 749–763.e13 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR6] 6.Kumar, B. V. et al. Human tissue-resident memory T cells are defined by core transcriptional and functional signatures in lymphoid and mucosal sites. Cell Rep.20, 2921–2934 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR7] 7.Poon, M. M. L. et al. Tissue adaptation and clonal segregation of human memory T cells in barrier sites. Nat. Immunol.24, 309–319 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR8] 8.Yudanin, N. A. et al. Spatial and temporal mapping of human innate lymphoid cells reveals elements of tissue specificity. Immunity50, 505–519.e4 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR9] 9.Dominguez Conde, C. et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science376, eabl5197 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR10] 10.Boyeau, P. et al. Deep generative modeling for quantifying sample-level heterogeneity in single-cell omics. Preprint at bioRxiv10.1101/2022.10.04.510898 (2022).

[CR11] 11.Becht, E. et al. Dimensionality reduction for visualizing single-cell data using UMAP. Nat. Biotechnol.37, 38–44 (2018). [Google Scholar]

[CR12] 12.Caron, D. P. et al. Multimodal hierarchical classification of CITE-seq data delineates immune cell states across lineages and tissues. Cell Rep. Methods5, 100938 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR13] 13.Sallusto, F., Lenig, D., Forster, R., Lipp, M. & Lanzavecchia, A. Two subsets of memory T lymphocytes with distinct homing potentials and effector functions. Nature401, 708–712 (1999). [DOI] [PubMed] [Google Scholar]

[CR14] 14.Zheng, Y. & Rudensky, A. Y. Foxp3 in control of the regulatory T cell lineage. Nat. Immunol.8, 457–462 (2007). [DOI] [PubMed] [Google Scholar]

[CR15] 15.Szabo, P. A. et al. Single-cell transcriptomics of human T cells reveals tissue and activation signatures in health and disease. Nat. Commun.10, 4706 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR16] 16.Provine, N. M. & Klenerman, P. MAIT cells in health and disease. Annu. Rev. Immunol.38, 203–228 (2020). [DOI] [PubMed] [Google Scholar]

[CR17] 17.Miron, M. et al. Maintenance of the human memory T cell repertoire by subset and tissue site. Genome Med.13, 100 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR18] 18.Hashemi, E., McCarthy, C., Rao, S. & Malarkannan, S. Transcriptomic diversity of innate lymphoid cells in human lymph nodes compared to BM and spleen. Commun. Biol.7, 769 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR19] 19.Muramatsu, M. et al. Class switch recombination and hypermutation require activation-induced cytidine deaminase (AID), a potential RNA editing enzyme. Cell102, 553–563 (2000). [DOI] [PubMed] [Google Scholar]

[CR20] 20.Johnson, J. L. et al. The transcription factor T-bet resolves memory B cell subsets with distinct tissue distributions and antibody specificities in mice and humans. Immunity52, 842–855.e6 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR21] 21.Weisel, N. M. et al. Comprehensive analyses of B-cell compartments across the human body reveal novel subsets and a gut-resident memory phenotype. Blood136, 2774–2785 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR22] 22.Ergen, C. et al. Consensus prediction of cell type labels in single-cell data with popV. Nat. Genet.56, 2731–2738 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR23] 23.Hoffman, G. E. et al. Efficient differential expression analysis of large-scale single cell transcriptomics data using dreamlet. Preprint at bioRxiv10.1101/2023.03.17.533005 (2024).

[CR24] 24.Schneider, C. et al. Induction of the nuclear receptor PPAR-γ by the cytokine GM-CSF is critical for the differentiation of fetal monocytes into alveolar macrophages. Nat. Immunol.15, 1026–1037 (2014). [DOI] [PubMed] [Google Scholar]

[CR25] 25.Notta, F. et al. Isolation of single human hematopoietic stem cells capable of long-term multilineage engraftment. Science333, 218–221 (2011). [DOI] [PubMed] [Google Scholar]

[CR26] 26.Choi, Y. S. et al. LEF-1 and TCF-1 orchestrate T_FH differentiation by regulating differentiation circuits upstream of the transcriptional repressor Bcl6. Nat. Immunol.16, 980–990 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR27] 27.Levitin, H. M., Zhao, W., Bruce, J. N., Canoll, P. & Sims, P. A. Consensus scHPF identifies cell type-specific drug responses in glioma by integrating large-scale scRNA-seq. Preprint at bioRxiv10.1101/2023.12.05.570193 (2023).

[CR28] 28.Schuettpelz, L. G. et al. Kruppel-like factor 7 overexpression suppresses hematopoietic stem and progenitor cell function. Blood120, 2981–2989 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR29] 29.Kuwahara, M. et al. The transcription factor Sox4 is a downstream target of signaling by the cytokine TGF-β and suppresses T_H2 differentiation. Nat. Immunol.13, 778–786 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR30] 30.Kunkel, E. J. et al. Lymphocyte CC chemokine receptor 9 and epithelial thymus-expressed chemokine (TECK) expression distinguish the small intestinal immune compartment: epithelial expression of tissue-specific chemokines as an organizing principle in regional immunity. J. Exp. Med.192, 761–768 (2000). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR31] 31.Szabo, P. A. Axes of heterogeneity in human tissue-resident memory T cells. Immunol. Rev.316, 23–37 (2023). [DOI] [PubMed] [Google Scholar]

[CR32] 32.Tellier, J. et al. Unraveling the diversity and functions of tissue-resident plasma cells. Nat. Immunol.25, 330–342 (2024). [DOI] [PubMed] [Google Scholar]

[CR33] 33.Mass, E., Nimmerjahn, F., Kierdorf, K. & Schlitzer, A. Tissue-specific macrophages: how they develop and choreograph tissue biology. Nat. Rev. Immunol.23, 563–579 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR34] 34.Watanabe, K. et al. Requirement for Runx proteins in IgA class switching acting downstream of TGF-β1 and retinoic acid signaling. J. Immunol.184, 2785–2792 (2010). [DOI] [PubMed] [Google Scholar]

[CR35] 35.Gloury, R. et al. Dynamic changes in Id3 and E-protein activity orchestrate germinal center and plasma cell development. J. Exp. Med.213, 1095–1111 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR36] 36.Simmons, D. P. et al. SLAMF7 engagement superactivates macrophages in acute and chronic inflammation. Sci. Immunol.7, eabf2846 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR37] 37.Aegerter, H., Lambrecht, B. N. & Jakubzick, C. V. Biology of lung macrophages in health and disease. Immunity55, 1564–1580 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR38] 38.Kohyama, M. et al. Role for Spi-C in the development of red pulp macrophages and splenic iron homeostasis. Nature457, 318–321 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR39] 39.Hurtado-Navarro, L. et al. NLRP3 inflammasome activation and symptom burden in KRAS-mutated CMML patients is reverted by IL-1 blocking therapy. Cell Rep. Med.4, 101329 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR40] 40.Ainscough, J. S., Gerberick, G. F., Kimber, I. & Dearman, R. J. Interleukin-1β processing is dependent on a calcium-mediated interaction with calmodulin. J. Biol. Chem.290, 31151–31161 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR41] 41.Tokuhisa, M. et al. Expression and localisation of Rab44 in immune-related cells change during cell differentiation and stimulation. Sci. Rep.10, 10728 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR42] 42.Weichhart, T., Hengstschläger, M. & Linke, M. Regulation of innate immune cell function by mTOR. Nat. Rev. Immunol.15, 599–614 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR43] 43.Sica, A. & Mantovani, A. Macrophage plasticity and polarization: in vivo veritas. J. Clin. Invest.122, 787–795 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR44] 44.Roy, S. et al. Macrophage-derived neuropilin-2 exhibits novel tumor-promoting functions. Cancer Res.78, 5600–5617 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR45] 45.Lavin, Y. et al. Tissue-resident macrophage enhancer landscapes are shaped by the local microenvironment. Cell159, 1312–1326 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR46] 46.Bohlen, C. J., Friedman, B. A., Dejanovic, B. & Sheng, M. Microglia in brain development, homeostasis, and neurodegeneration. Annu. Rev. Genet.53, 263–288 (2019). [DOI] [PubMed] [Google Scholar]

[CR47] 47.Natri, H. M. et al. Cell-type-specific and disease-associated expression quantitative trait loci in the human lung. Nat. Genet.56, 595–604 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR48] 48.Pereira, B. I. et al. Sestrins induce natural killer function in senescent-like CD8⁺ T cells. Nat. Immunol.21, 684–694 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR49] 49.Gong, Q. et al. Longitudinal multi-omic immune profiling reveals age-related immune cell dynamics in healthy adults. Preprint at bioRxiv10.1101/2024.09.10.612119 (2024).

[CR50] 50.Pearce, E. L. et al. Control of effector CD8⁺ T cell function by the transcription factor Eomesodermin. Science302, 1041–1043 (2003). [DOI] [PubMed] [Google Scholar]

[CR51] 51.Davis, R. S. Fc receptor-like molecules. Annu. Rev. Immunol.25, 525–560 (2007). [DOI] [PubMed] [Google Scholar]

[CR52] 52.Mogilenko, D. A. et al. Comprehensive profiling of an aging immune system reveals clonal GZMK⁺ CD8⁺ T cells as conserved hallmark of inflammaging. Immunity54, 99–115.e12 (2021). [DOI] [PubMed] [Google Scholar]

[CR53] 53.Oh-hora, M., Johmura, S., Hashimoto, A., Hikida, M. & Kurosaki, T. Requirement for Ras guanine nucleotide releasing protein 3 in coupling phospholipase C-γ2 to Ras in B cell receptor signaling. J. Exp. Med.198, 1841–1851 (2003). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR54] 54.Lee, N. Y. S., Li, M., Ang, K. S. & Chen, J. Establishing a human bone marrow single cell reference atlas to study ageing and diseases. Front. Immunol.14, 1127879 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR55] 55.Gordon, C. L. et al. Tissue reservoirs of antiviral T cell immunity in persistent human CMV infection. J. Exp. Med.214, 651–667 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR56] 56.Persad, S. et al. SEACells infers transcriptional and epigenomic cellular states from single-cell genomics data. Nat. Biotechnol.41, 1746–1757 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR57] 57.Gaublomme, J. T. et al. Single-cell genomics unveils critical regulators of Th17 cell pathogenicity. Cell163, 1400–1412 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR58] 58.Ivanov, I. I., Zhou, L. & Littman, D. R. Transcriptional regulation of Th17 cell differentiation. Semin. Immunol.19, 409–417 (2007). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR59] 59.Miron, M. et al. Human lymph nodes maintain TCF-1^hi memory T cells with high functional potential and clonal diversity throughout life. J. Immunol.201, 2132–2140 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR60] 60.Zhou, X. et al. Differentiation and persistence of memory CD8⁺ T cells depend on T cell factor 1. Immunity33, 229–240 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR61] 61.Krasemann, S. et al. The TREM2–APOE pathway drives the transcriptional phenotype of dysfunctional microglia in neurodegenerative diseases. Immunity47, 566–581.e9 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR62] 62.Deczkowska, A., Weiner, A. & Amit, I. The physiology, pathology, and potential therapeutic applications of the TREM2 signaling pathway. Cell181, 1207–1217 (2020). [DOI] [PubMed] [Google Scholar]

[CR63] 63.de Oliveira Matos, A. et al. TREM-2: friend or foe in infectious diseases?. Crit. Rev. Microbiol.50, 1–19 (2024). [DOI] [PubMed] [Google Scholar]

[CR64] 64.Wu, K. et al. TREM-2 promotes macrophage survival and lung disease after respiratory viral infection. J. Exp. Med.212, 681–697 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR65] 65.Schlepckow, K., Morenas-Rodríguez, E., Hong, S. & Haass, C. Stimulation of TREM2 with agonistic antibodies—an emerging therapeutic option for Alzheimer’s disease. Lancet Neurol.22, 1048–1060 (2023). [DOI] [PubMed] [Google Scholar]

[CR66] 66.Lu, J. et al. Heterogeneity and transcriptome changes of human CD8⁺ T cells across nine decades of life. Nat. Commun.13, 5128 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR67] 67.Carrasco, E. et al. The role of T cells in age-related diseases. Nat. Rev. Immunol.22, 97–111 (2022). [DOI] [PubMed] [Google Scholar]

[CR68] 68.Laphanuwat, P., Gomes, D. C. O. & Akbar, A. N. Senescent T cells: beneficial and detrimental roles. Immunol. Rev.316, 160–175 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR69] 69.Rupareliya, M. & Shende, P. Therapeutic potential of stem cells in natural killer-like B cell-associated diseases. Adv. Exp. Med. Biol.1470, 57–72 (2024). [DOI] [PubMed] [Google Scholar]

[CR70] 70.Wang, S. et al. Natural killer-like B cells prime innate lymphocytes against microbial infection. Immunity45, 131–144 (2016). [DOI] [PubMed] [Google Scholar]

[CR71] 71.Zhang, W. et al. SARS-CoV-2 infection results in immune responses in the respiratory tract and peripheral blood that suggest mechanisms of disease severity. Nat. Commun.13, 2774 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR72] 72.Szabo, P. A. et al. Longitudinal profiling of respiratory and systemic immune responses reveals myeloid cell-driven lung inflammation in severe COVID-19. Immunity54, 797–814.e6 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR73] 73.Folorunso, O. S. & Sebolai, O. M. Overview of the development, impacts, and challenges of live-attenuated oral rotavirus vaccines. Vaccines8, 341 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR74] 74.Fraietta, J. A. et al. Determinants of response and resistance to CD19 chimeric antigen receptor (CAR) T cell therapy of chronic lymphocytic leukemia. Nat. Med.24, 563–571 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR75] 75.English, E. P. et al. Engineering CAR-T therapies for autoimmune disease and beyond. Sci. Transl. Med.16, eado2084 (2024). [DOI] [PubMed] [Google Scholar]

[CR76] 76.Guedan, S., Ruella, M. & June, C. H. Emerging cellular therapies for cancer. Annu. Rev. Immunol.37, 145–171 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR77] 77.Carpenter, D. J. et al. Human immunology studies using organ donors: impact of clinical variations on immune parameters in tissues and circulation. Am. J. Transpl.18, 74–88 (2018). [Google Scholar]

[CR78] 78.Jensen, I. J., Wells, S. B., Gras, J. & Farber, D. L. Isolation and characterization of human tissue resident memory T cells. Curr. Protoc.5, e70120 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR79] 79.Rainbow, D., Howlett, S., Jarvis, L. & Jones, J. Multi tissue processing for single cell sequencing of human immune cells v1. protocols.io10.17504/protocols.io.bz4qp8vw (2021).

[CR80] 80.James, K. R. et al. Distinct microbial and immune niches of the human colon. Nat. Immunol.21, 343–353 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR81] 81.Reynolds, G. et al. Developmental cell programs are co-opted in inflammatory skin disease. Science371, eaba6500 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR82] 82.Zheng, G. X. Y. et al. Massively parallel digital transcriptional profiling of single cells. Nat. Commun.8, 14049 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR83] 83.Bernstein, N. J. et al. Solo: doublet identification in single-cell RNA-seq via semi-supervised deep learning. Cell Syst.11, 95–101.e5 (2020). [DOI] [PubMed] [Google Scholar]

[CR84] 84.Yang, S. et al. Decontamination of ambient RNA in single-cell RNA-seq with DecontX. Genome Biol.21, 57 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR85] 85.Wolock, S. L., Lopez, R. & Klein, A. M. Scrublet: computational identification of cell doublets in single-cell transcriptomic data. Cell Syst.8, 281–291.e9 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR86] 86.Lopez, R., Regier, J., Cole, M. B., Jordan, M. I. & Yosef, N. Deep generative modeling for single-cell transcriptomics. Nat. Methods15, 1053–1058 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR87] 87.Hafemeister, C. & Satija, R. Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression. Genome Biol.20, 296 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR88] 88.Blondel, V. D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. Fast unfolding of communities in large networks. J. Stat. Mech.2008, P10008 (2008). [Google Scholar]

[CR89] 89.Sikkema, L. et al. An integrated cell atlas of the lung in health and disease. Nat. Med.29, 1563–1577 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR90] 90.Hrovatin, K. et al. Integrating single-cell RNA-seq datasets with substantial batch effects. Preprint at bioRxiv10.1101/2023.11.03.565463 (2024).

[CR91] 91.Levitin, H. M. et al. De novo gene signature identification from single-cell RNA-seq with hierarchical Poisson factorization. Mol. Syst. Biol.15, e8557 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR92] 92.Büttner, M., Ostner, J., Müller, C. L., Theis, F. J. & Schubert, B. scCODA is a Bayesian model for compositional single-cell data analysis. Nat. Commun.12, 6876 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR93] 93.Denisenko, E. et al. Systematic assessment of tissue dissociation and storage biases in single-cell and single-nucleus RNA-seq workflows. Genome Biol.21, 130 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR94] 94.Badia-I-Mompel, P. et al. decoupleR: ensemble of computational methods to infer biological activities from omics data. Bioinform. Adv.2, vbac016 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR95] 95.Suo, C. et al. Dandelion uses the single-cell adaptive immune receptor repertoire to explore lymphocyte developmental origins. Nat. Biotechnol.42, 40–51 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR96] 96.Ye, J., Ma, N., Madden, T. L. & Ostell, J. M. IgBLAST: an immunoglobulin variable domain sequence analysis tool. Nucleic Acids Res.41, W34–W40 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR97] 97.Gadala-Maria, D., Yaari, G., Uduman, M. & Kleinstein, S. H. Automated analysis of high-throughput B-cell sequencing data reveals a high frequency of novel immunoglobulin V gene segment alleles. Proc. Natl Acad. Sci. USA112, E862–E870 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

Multimodal profiling reveals tissue-directed signatures of human immune cells altered with age

Steven B Wells

Daniel B Rainbow

Michal Mark

Peter A Szabo

Can Ergen

Daniel P Caron

Ana Raquel Maceiras

Elior Rahmani

Eli Benuck

Valeh Valiollah Pour Amiri

David Chen

Allon Wagner

Sarah K Howlett

Lorna B Jarvis

Karen L Ellis

Masaru Kubota

Rei Matsumoto

Krishnaa Mahbubani

Kouresh Saeb-Parsy

Cecilia Dominguez Conde

Laura Richardson

Chuan Xu

Shuang Li

Lira Mamanova

Liam Bolt

Alicja Wilk

Sarah A Teichmann

Donna L Farber

Peter A Sims

Joanne L Jones

Nir Yosef

Abstract

Main

Results

Multimodal profiling identifies immune cells across tissues

Fig. 1. A multi-lineage human immune cell atlas encompasses tissues and age.

Extended Data Fig. 1. Additional tissues obtained and annotated covariates in the dataset.

Immune cell subset composition is specific to the tissue

Fig. 2. Multimodal classification reveals immune subset heterogeneity and distribution across sites.

Extended Data Fig. 2. Immune cell subset composition varies by tissue site.

Immune cell phenotype is strongly tissue-associated

Fig. 3. Differential expression analysis identifies tissue localization as a major determinant of immune cell identity across lineages.

Extended Data Fig. 3. Tissue localization shapes gene expression profiles of immune cell subsets.

Extended Data Fig. 4. Factorization reveals shared tissue gene expression programs across immune lineages.

Resident immune cells express tissue-specific signatures

Fig. 4. Cross-tissue analysis of resident immune cell subsets highlights site-specific signatures.

Immune cell profiles change with age across tissues

Fig. 5. Immune aging manifests across lineage, subset and tissue of origin.

T and B lymphocytes express distinct aging signatures

Extended Data Fig. 5. Features of immune aging are shared across lymphocyte subsets and validated in an independent cohort.

Fig. 6. Immune aging in adaptive lymphocytes varies across tissues.

Host factors and subset heterogeneity affect T cell aging

Fig. 7. Integrated analysis reveals tissue-dependent signatures of aging in CD4+ T cells.

Extended Data Fig. 6. MrVI uncovers modules of age-associated changes in CD4+ T cells.

Discussion

Methods

Ethics statement

Statistics and reproducibility

Tissue acquisition from organ donors

Tissue processing and CITE-seq at Columbia University

Tissue processing and CITE-seq at the University of Cambridge

Single-cell sequencing

Alignment and preprocessing of CITE-seq and scTCR/BCR-seq using Cell Ranger

Quality control

Immune cell subset classification using MMoCHi

Training and application of label transfer models

Integration and cell state embedding using MrVI

DE and variance decomposition using dreamlet

Identification of gene co-expression patterns using consensus scHPF

Detection of tissue-specific effects

Detection of age-specific and CMV effects on composition and gene expression

Counterfactual analysis to detect age associations with MrVI

TCR and BCR repertoire analysis using Dandelion

Reporting summary

Online content

Supplementary information

Source data

Acknowledgements

Fig. 7. Integrated analysis reveals tissue-dependent signatures of aging in CD4⁺ T cells.

Extended Data Fig. 6. MrVI uncovers modules of age-associated changes in CD4⁺ T cells.