Unraveling the phenotypic states of human innate-like T cells: Comparative insights with conventional T cells and mouse models

SUMMARY

The “innate-like” T cell compartment, known as T_inn, represents a diverse group of T cells that straddle the boundary between innate and adaptive immunity. We explore the transcriptional landscape of T_inn compared to conventional T cells (T_conv) in the human thymus and blood using single-cell RNA sequencing (scRNA-seq) and flow cytometry. In human blood, the majority of T_inn cells share an effector program driven by specific transcription factors, distinct from those governing T_conv cells. Conversely, only a fraction of thymic T_inn cells displays an effector phenotype, while others share transcriptional features with developing T_conv cells, indicating potential divergent developmental pathways. Unlike the mouse, human T_inn cells do not differentiate into multiple effector subsets but develop a mixed type 1/type 17 effector potential. Cross-species analysis uncovers species-specific distinctions, including the absence of type 2 T_inn cells in humans, which implies distinct immune regulatory mechanisms across species.

In brief

Our study presents a comprehensive atlas of human innate T (T_inn) cells in the thymus and blood, highlighting their blended type 1 and type 17 transcriptional profiles. Loh et al. use single-cell RNA sequencing and flow cytometry to reveal species-specific distinctions and potential divergent developmental pathways of T_inn compared to conventional T cells.

Graphical Abstract

INTRODUCTION

Conventional CD4⁺ and CD8⁺ T cells (T_conv) are essential for adaptive immunity, recognizing peptide antigens presented by HLA class I or II proteins via their T cell antigen receptors (TCRs). Upon antigenic stimulation, these T cells undergo transcriptional and epigenetic changes, secreting pro-inflammatory cytokines, acquiring cytotoxic abilities to clear pathogens, or forming memory T cells for rapid response upon reencounter. Thus, T_conv cells within the circulation are heterogeneous and are commonly classified into naive (T_n), central memory (T_cm), effector memory (T_em), and terminally differentiated effector memory (T_emra) subsets.^1–3

Innate-like T cells (T_inn), such as invariant natural killer T (iNKT) cells, mucosal-associated invariant T (MAIT) cells, and certain γδ T cells, maintain consistent TCRs among individuals and function without prior pathogen exposure. They account for 10%–20% of human T cells.^4,5 T_inn cells, originating from thymic progenitors, do not recognize peptides presented by HLA class I or II. Instead, iNKT cells use semi-invariant TCRs to recognize lipid antigens presented by CD1D.⁶ MAIT cells, also characterized by semi-invariant TCRs,⁷ recognize riboflavin precursor-derived metabolites presented by MR1.^8,9 Some γδ T cells recognize self- and foreign phosphoantigens in conjunction with the transmembrane butyrophilin-family receptors BTN2A1-BTN3A1-BTN3A2 complex.^10–12 The antigens recognized by other human γδ T cell populations remain unclear.¹³ Thus, T_inn cells enhance immune detection of diverse threats.^14–16 Furthermore, the secretion of cytokines at steady state by mouse iNKT cells influences surrounding cells and T_conv development,^17–19 suggesting that they may also function as gatekeepers, ensuring proper T cell development and maturation.

The conservation of T_inn cells throughout mammalian evolution indicates their crucial and non-redundant role in the immune system, which may be attributed to their rapid activation kinetics and their ability for TCR-independent activation.^20–22 In mice, the rapid effector capacity of T_inn cells stems from unique transcriptional programs formed during their development in the thymus.^23,24 Analogous to CD4 T_conv cells, which can be polarized by cytokines into T helper (Th) phenotypes such as Th1, Th2, and Th17 that secrete interferon-γ (IFN-γ), interleukin-4 (IL-4), and IL-17, respectively, mouse T_inn cells diverge into distinct, terminally differentiated subsets that can be readily identified based on the expression of specific transcription factors (TFs) such as PLZF, GATA3, T-bet, and RORγt.¹⁷

In this study, using single-cell genomics and flow cytometry, we assessed the range of phenotypic states T_conv and T_inn cells can adopt in vivo in the human thymus and blood. We uncovered that unlike adult blood T_inn cells most postnatal thymic T_inn cells resemble naive T_conv cells, with few displaying an “effector” state. A distinct transcriptional program is shared among human T_inn cells, and major TFs driving this program are also expressed in mice. However, unlike the mouse, human T_inn cells do not form distinct subsets but develop a mixed type 1/type 17 effector program. We further highlight differences in CD1D and SLAMF6 expression in the thymus between the two species, which could potentially impact the maturation process of iNKT cells in humans. This comprehensive transcriptomic dataset of human T_inn cells provides a valuable resource to explore their fundamental properties and may serve as a guide for their strategic application in immunotherapeutic approaches.

RESULTS

Single-cell RNA-sequencing analysis of T cell maturation and post-maturation stages in humans

We explored the transcriptional profiles of T_inn and T_conv cells through single-cell RNA sequencing (scRNA-seq) of antibody DNA-barcoded (“hashtags”) tetramer-sorted iNKT (PBS57-CD1d tetramer⁺, TRAV10⁺), MAIT (5-OP-RU-MR1 tetramer⁺, TRAV1–2⁺), and total γδ T cells, along with single-positive (SP) CD4 and CD8 T_conv cells from pediatric thymus and adult blood samples (Figure S1). A subset of samples was also subjected to VDJ sequencing (Figure 1A and Table S1). We analyzed 78,355 cells (37,369 pediatric thymus and 40,986 adult blood) after quality control, integrating them into a reference dataset (Figures 1B, 1C, S2A, and S2B). Using unsupervised clustering, we identified 18 distinct and stable clusters (c), primarily separating into thymus (c0–c9) or blood-associated regions (c12–c17, Figures 1C and S2C), with transitioning clusters (c10 and c11) showing cells from both tissues (Figures 1D and 1E), representing naive T cells prepared to leave the thymus or just entering the blood, in agreement with their over-representation of an “egress” gene signature²⁵ (Figure 1F and Table S3).

Figure 1. Integrative view on T_inn and T_conv development and peripheral function.

(A) Experimental setup specifying donor type (postnatal/adult), tissue, and sorted cell types.

(B) Harmony batch-corrected and integrated dataset across donors, tissues, and cell types.

(C–E) (C) Stable Louvain-derived cell clusters distributed across (D) both blood and thymus-derived cells and (E) their respective frequencies in these clusters.

(F) “Egress” score on thymus- and blood-derived cells.

(G) Cells color coded by cluster (as in C) visualized by their hashtag-sorted cell type (columns) and the tissue they originated from (rows).

(H–J) (H) Projection of naive and effector scores and the proportion of (I) thymic and (J) blood cell types per donor, classified on the basis of these scores (bottom row); top row shows analogous proportions by cell cluster (as in C).

(K) Gene-expression programs (GEPs) in thymus and blood identified using cNMF, with color scale representing GEP usage. Sample numbers for all panels as depicted in (A). Score defining genes as described in text.

Cluster identities and transcriptional states were determined using reference genes (Figures S3A and S3B; Table S2), with cell types validated by neighbor voting²⁶ with the human thymus atlas²⁷ (Figure S4). Early T cell development included immature single positive (ISP) cells (c0 and c1), double-positive (DP) thymocytes (c2), and early SP stages characterized by CCR9 expression (c3 for CD4 SP and c9 for CD8 SP). Later stages expressed CCR7 (c11 for CD4 SP, c10 for CD8 SP; Figure S5A). Differentially expressed gene (DEG) analysis comparing CD4 SP cells (c3 and c11) to CD8 SP cells (c9 and c10) further confirms these annotations (Figure S5B). Specialized lineages were also detected, including CD8αα cells (GNG4 and NUCB2), thymic γδ T cells (TRDC, TRGC2), and regulatory T cells (T_regs) with high expression of FOXP3 (c5, c6, and c7, respectively). Other signaling states included “agonist” cells with high levels of TF transcripts associated with TCR signaling (c4; NR4A1, EGR1, EGR3, and NFKBID), cells with high expression of type I IFN signaling genes (IFI6, MX1, and IFI44L in c8), and AP-1 TFs (JUN, FOS, JUNB in c12). We also found cells (c13 through c17) expressing effector-encoding genes (GZMK, GZMH, GZMB, PRF1, and CCL5), suggesting involvement in effector functions of these cells (Figures S3A, S3B, and S5A). Altogether, we observed cells with distinct transcriptional profiles, representing unique cell types (CD8αα, T_regs) and stages of T cell development and maturation.

Identification of the gene-expression programs that characterize T cell populations in thymus and blood

Deciphering scRNA-seq data can be challenging due to the complexity of gene-expression patterns reflecting cell identity and activity. We used consensus non-negative matrix factorization (cNMF)²⁸ to reduce dimensionality, identifying 12 gene-expression programs (GEPs; Figure 1K and Table S4). Each cell type’s contribution to these GEPs was assessed, revealing some GEPs shared across cell types, with others unique to specific clusters (Figures 1G and 1K). GEPs 7–11 were linked to thymic γδ T cells, T_regs, thymic CD8αα T cells, quiescent ISP, and proliferating ISP, respectively (Figure 1K). GEP 12 was excluded due to batch effects (Figure S6). GEPs 1 and 2, marked by CCR9 and CCR7, were active in early and late thymic T cells, respectively. GEP3 was expressed by naive T cells. GEPs 4, 5, and 6, associated with “effectorness,” were active in clusters 12, 13–14, and 15–17, respectively (Figures 1G–1K). Leveraging insights from these GEPs, we conducted an in-depth analysis of thymic and blood T cell populations, providing an integrated understanding of T cell differentiation and function.

Gene-expression similarities between αβ thymocyte lineages

We separated each sample by cell type and tissue, using the identifying tag (Figure 1G). CD4⁺ thymic cells were found in clusters 0–4, 7, 8, and 11, while CD8⁺ thymic cells were in clusters 2 and 5–10. Cell proportions in each cluster were consistent across samples, with about 1% (1.1% ± 0.4%) of the cells showing an effector signature (Figures 1G–1I and Table S3). Thymic iNKT cells were distributed similarly to conventional CD4⁺ SP, and thymic MAIT cells shared clusters with CD8⁺ SP (Figure 1G). Thymic γδ T cells were distinct, mainly in cluster 6 and GEP7 (Figures 1G and 1K).

Global gene-expression analysis between thymocyte lineages revealed three distinct patterns. First, effector stages of iNKT, MAIT, and γδ thymocytes (clusters 12–17) exhibited high gene-expression similarity. Second, non-effector γδ thymocytes showed no expression similarity to other lineages. Third, non-effector iNKT and MAIT thymocytes (clusters 0–11) resembled CD4⁺ and CD8⁺ thymocytes, respectively (Figure 2A). This suggests that specific subpopulations of iNKT thymocytes engage transcriptional programs akin to those of conventional CD4⁺ thymocytes, with a parallel observation between MAIT and CD8⁺ thymocytes. Pseudo-bulk DEG analysis between non-effector iNKT/CD4⁺ and non-effector MAIT/CD8⁺ further demonstrated that non-effector iNKT and CD4⁺ thymocytes share 43 commonly upregulated and 38 commonly downregulated genes (Figure 2B). Key genes such as GATA3, ZBTB7B, and CD40LG were upregulated in non-effector iNKT thymocytes, indicating CD4 lineage differentiation, while CD8A, CD8B, and RUNX3 were elevated in non-effector MAIT thymocytes. Using DEG analysis between CD4⁺/MAIT and CD8⁺/iNKT cells as a negative control revealed only EPHB6, ZFAS1, and EEF1A1 as differentially expressed (Figure 2B). Comparing conventional CD4⁺/CD8⁺ thymocytes with non-effector iNKT/MAIT cells highlighted 11 DEGs (Figure 2C), including ZBTB16 and SLAM family receptors SLAMF5 (CD84) and SLAMF6, underscoring specific pathways crucial for effector differentiation, similar to observations in mice.²³

Figure 2. Gene-expression patterns in thymocyte lineages.

(A) Heatmap showing MetaNeighbor’s AUROC scores between thymocytes split by donor, lineage, and non-effector (c0–11) versus effector (c12–17) clusters. Barplots indicate thymocyte proportions per lineage.

(B) Pseudo-bulk differential expression analysis between CD4⁺/iNKT and CD8⁺/MAIT thymocytes in naive clusters (3, 9, 10, 11). As a negative control, the only three genes that were differentially expressed between CD4⁺/MAIT and CD8⁺/iNKT thymocytes are displayed in the center of the heatmap.

(C) Pseudo-bulk differential expression analysis between CD4⁺/CD8⁺ and iNKT/MAIT thymocytes in naive clusters (3, 9, 10, 11). For both (B) and (C), heatmap displays the expression level of genes (represented with color scale as a Z score of the average normalized expression) that are significantly differentially expressed (p_adj < 0.01).

(D) Clustering of hashtag-separated thymic iNKT cells (top), MAIT cells (middle), and γδ T cells (bottom). Right panel shows the score of type I and type III effector gene signature for the corresponding thymic lineage.

(E) Kernel density estimates of the normalized expression level of genes of interest. The expression level distribution varies between genes and lineage. The range of kernel density estimate values also varies between each panel (from 0 to 0.04 for the smallest range and 0 to 0.4 for the largest range). A unique color scale was represented to indicate the direction of the values.

Unbiased transcriptomic analysis of human T_inn differentiation

To further explore the transcriptional heterogeneity of human T_inn thymocytes, we re-analyzed iNKT and MAIT cell populations individually (Figures 2D–2E and S7). We identified seven clusters for both cell types (Figure 2D), with some variability in their proportion across donors (Figures S7B and S7F). Five major cell signatures were shared across T_conv, iNKT, and MAIT cells (Figures S3C and S3D). First, we observed a distinctive gene signature associated with CD8αα T cells (captured by GEP9; Figure S8; Tables S5 and S6). This signature was characterized by the heightened expression of NUCB2, MINDY2, and HIVEP3 (Figures S3C and S3D), and intriguingly, it was observed in both iNKT and MAIT cells (termed NKT_c0 and MAIT_c1; Figures S7C and S7G). This subset of thymic CD8A-expressing iNKT cells, which also exhibited some PLZF expression while lacking KLRB1 (coding for CD161), EOMES, and GZMK expression (Figure 2E), could be readily identified using flow cytometry (Figure S9). Second, we identified a shared pattern of expression for the CCR9 and CCR7 chemokine receptors across T_conv cells, iNKT cells, and MAIT cells (Figure 2E). Initially, iNKT and MAIT cells exhibited an upregulation of CCR9 in conjunction with TOX and SATB1 (Figures S3C and S3D), resembling the developmental program of early developing CD4 SP and CD8 SP cells, respectively. Subsequently, the elevated expression of CCR7 marked cells at a seemingly more advanced developmental stage (termed NKT_c2 and MAIT_c4). These sequential waves of chemokine receptor expression align with gene modules GEP1 and GEP2 (Figure S8 and Table S4), suggesting their sequential induction during the development of CD4, CD8, iNKT, and MAIT cells. These findings establish that the human thymus harbors iNKT and MAIT cells with a transcriptome resembling that of developing naive T_conv cells. The existence of such naive-like populations (CD161⁻EOMES⁻GZMK⁻) of iNKT and MAIT cells in the human thymus was confirmed by flow cytometry (Figure S9).

Third, we discovered iNKT and MAIT cells characterized by upregulation of genes associated with type I IFN signaling such as MX1 and IFI6, similar to CD4 and CD8 SP cells (NKT_c3 and MAIT_c5; Figures S3C and S3D). Fourth, TCR signaling/AP-1 signatures were found in iNKT cells, with AP-1 TFs FOS and JUN upregulated alongside ZBTB16 and KLRB1. These cells also expressed CD4 transcripts but not CD8A (NKT_c5; Figures 2E and S3C). The TCR signature was more pronounced in MAIT cells, where a small subset showed clear upregulation of genes involved in TCR signaling (NR4A1, NFKBID, REL; MAIT_c3; Figure S3D). Unlike T_conv, a proportion of both thymic iNKT and MAIT cells displayed an “effector” signature. iNKT cells in cluster 6 (NKT_c6) express classically iNKT-associated genes along with upregulation of effector genes usually associated with type 1 or type 17 immunity (EOMES/GZMK and RORA/CCR6, respectively; Figures 2E and S3C). Some of these cells expressed CD8A transcripts, suggesting that CD4⁺ and CD8⁺ iNKT cells might develop into transcriptionally distinct subsets, with CD8⁺ iNKT cells having a more effector-associated signature. We found the same type 1 (EOMES/GZMK) and type 17 (RORA/CCR6) immunity effector co-expression signatures in MAIT cells (MAIT_c6; Figures 2D and S3D), with upregulation of genes encoding for granzymes (GZMA and GZMK) and chemokines (CCL5), as well as genes previously associated with MAIT cells^29,30 (KLRB1, SLC4A10, IL23R; Figure 2E).

Analysis of the thymic γδ T cells identified eight distinct clusters (Figures 2D, S3E, and S7I–S7L; Table S7), confirming previous findings on pediatric γδ thymocytes²⁵ with immature populations (GD_c0, 1, 2), TCR activation profiles (GD_c3), type I IFN response (GD_c6), and effector γδ T cells with egress and mixed type 1/type 17 effector potential (GD_c7). We also observed cells with a cycling gene signature (GD_c4), which was notably absent in iNKT and MAIT cells. Overall, we found a shared utilization of a mixed type 1/type 17 effector programs by iNKT, MAIT, and γδ T cells, with 28.7% ± 22.3% of thymic T_inn cells displaying this effector signature (Figures 1G and 1I, in clusters 12–17).

Lastly, because effector T_inn cells in the thymus might represent cells that initially acquired an effector signature in the blood and subsequently recirculated back to the thymus, we compared expression profiles between thymic and blood effector iNKT and MAIT cells. We found distinctive tissue-specific gene-expression profiles for both cell types (Figure S5C), indicating that thymic effector T_inn cells are unlikely to be derived from recirculating blood cells.

Effect of clonal selection on iNKT, MAIT, and γδ T cells’ effector states

To investigate whether T_inn cells with an effector transcriptome had distinct TCR repertoires compared to naive T_inn cells, we conducted paired VDJ sequencing (Figure 3).

Figure 3. Innate T cell TCR diversity during development.

Cells with VDJ sequencing and their cell-type-specific characteristic chain arrangement for thymic iNKT cells (A), MAIT cells (G), and γδ T cells (L). For each cell type, the respective proportions of gene segment usage in each chain (B and D; H and J; M and N) are shown together with their CDR3 length and sequence logo (C, I, O) and their cluster-specific usage (E, with clusters as in Figure 2A). Shannon index as an estimation of TCR diversity in the naive-like and effector-like iNKT (F) and MAIT (K) cells, based on clusters in Figure 2D. n = 1 human thymus sample for all panels.

For iNKT cells, most VDJ sequenced cells used the TRAV10 gene segment (encoding for the Vα24 chain) rearranged with TRAJ18 (Figures 3A and 3B), resulting in a canonical CDR3α sequence of 14 amino acids (Figure 3C), crucial for antigen recognition.³¹ This invariant TCRα chain paired with diverse TCRβ rearrangements, mainly involving the TBV25 chain (Figure 3D), and was used evenly across all clusters (Figure 3E). The Shannon index, a measure of TCR diversity, revealed that no clonal selection had occurred or was associated with effector transcriptome development (Figure 3F), and no shared TCR clonotypes were identified between naive- and effector-like cells.

Thymic MAIT cells primarily used the TRAV1–2 gene segment (encoding for the Vα7.2 chain) with TRAJ33, TRAJ20, and TRAJ12 (Figures 3G and 3H), maintaining a conserved Y95 residue in the CDR3α loop (Figure 3I) essential for MAIT cell activation.^32,33 These TCRα chains paired with diverse TCRβ chains (Figure 3J) dominated by TRBV6, TRBV20, and TRBV4 gene segments.^34,35 Like iNKT cells, MAIT cells showed no shared TCR clonotypes and no clonal selection in effector transcriptome cells (MAIT_c6) compared to naive-like cells (MAIT_c2–4), based on the Shannon index (Figure 3K).

Effector γδ T cells (GD_c7) were enriched for cells expressing TRDV2 and TRGV9 gene segments, while TRDV1 and TRDV3 segments were excluded from this cluster (Figure 3L). Some TRDV2⁺ or TRGV9⁺ cells were also found in non-effector clusters, suggesting a potential role for these gene segments in the development of effector γδ T cells in the postnatal human thymus. Supporting this hypothesis, we observed that the rearrangements of both the Vδ2 chains and associated Vγ9 chains differed largely between cells in the effector versus non-effector clusters (Figure 3N). Specifically, the Vγ9 chains of effector cells were found to be preferentially rearranged with the TRGJP gene segment and enriched for the public CDR3 sequence typically found among Vδ2Vγ9 γδ T cells in adult blood³⁶ (Figure 3O), whereas Vδ2⁺ cells in the non-effector clusters showed more diverse Vg gene usage and rearrangements (Figure 3N). In summary, the acquisition of effector programs in iNKT and MAIT cells is not associated with changes in TCR diversity, whereas Vδ2 and Vγ9 chain rearrangements in γδ T cells suggest a predisposition toward the effector program.

Gene-expression programs that characterize T cell effector functions

To further characterize the functionality of T_inn and T_conv cells in the blood, we examined the distribution of these cell types across transcriptional clusters. Conventional CD4⁺ T cells were primarily located in cluster 11 (naive CD4 T cells) and cluster 7 (T_regs). CD4⁺ T cells also appeared in effector clusters, particularly clusters 12 and 16, with proportions varying among donors (18.4%–93.7%; Figures 1G and 1J). Blood CD8⁺ T cells were predominantly in cluster 10 (naive CD8 T cells), with varying proportions in effector clusters, reflecting differences in donor immunological history (16.5%–94%; Figure 1J). In contrast, the majority of blood T_inn cells (~94.3% ± 6.7%) were distributed across effector clusters 12–17, regardless of donor (Figure 1J).

We next investigated transcriptional states in blood T cell populations using cell hashtags to reanalyze blood iNKT cells, MAIT cells, γδ T cells, and T_conv CD4⁺ and CD8⁺ T cells individually (Figures 4A and 4B). Each cell type was found within previously identified GEPs (GEP3–6; Figure 1K; Tables S8, S9, S10, S11, and S12), albeit with varying proportions for each cell type (Figures 4C and 4D). To contextualize these GEPs, we computed overlap scores and statistically assessed their enrichment with literature-derived signatures.^37–40 Subsequently, we scored the joint signature-GEP interactions in our dataset (Figure S10). GEP3 was found to be closely associated with signatures of naive T cell characteristics. In contrast, GEP4 displayed similarities with T_cm, T_em, or literature-derived signatures classified as a mix thereof (T_cm/T_em), while GEP6 exhibited characteristics akin to T_emra. GEP5, on the other hand, shared elements with T_em cells and previously identified CD8 MAIT signatures (Figure S10A).

Figure 4. Gene-expression programs in circulating T_inn and T_conv.

(A–C) (A) Clustering of hashtag-separated blood iNKT, MAIT, γδ T, CD4, and CD8 T cells, (B) the respective proportion of cells per cluster and donor, and (C) the effector GEP signature scores (as in Figure 1K) per cell type and cluster.

(D) Top GEP usage for each cell type, based on cNMF-derived usage matrix.

(E) Pseudo-bulk, pairwise differential gene expression between cell types.

(F) Cell-type-specific genes among T_inn cells using GEP5. For both (E) and (F), heatmaps depict the expression level of genes (represented with color scale as a Z score of the average normalized expression) that are significantly differentially expressed (p_adj < 0.01). n = 4, 4, 9, 4, and 9 for iNKT, MAIT, γδ, CD4, and CD8 cells, respectively.

Blood iNKT cells predominantly expressed GEP3 or GEP5, with some expressing GEP4 or GEP6 (Figure 4D). GEP4-expressing iNKT cells had CD4 transcripts, while GEP5 and GEP6 cells lost CD4 expression (Figures S11A and S11B). To validate this observation, we examined the cellular phenotype of blood iNKT cells. Blood CD4⁺ iNKT cells were mostly PLZF⁻CD161⁻EOMES⁻GZMK⁻ but CCR7⁺, indicating a naive GEP3 program (Figure S11C). In contrast, CD8⁺ and double-negative iNKT cells were mostly PLZF⁺CD161⁺ and displayed an effector phenotype (EOMES⁺GZMK⁺CCR7⁻CD62L⁻; Figure S11). Interestingly, the distribution of these programs varied significantly among different donors (Figure 4B). These findings are in line with previous data indicating that CD4-negative iNKT cells become more prevalent in the blood with age, eventually becoming the dominant population in the adult blood iNKT cell compartment.^41,42 This suggests that CD4-negative iNKT cells may originate from CD4⁺ iNKT cells that undergo a loss of CD4 expression as they transition toward a more effector-like state. Conversely, when examining MAIT cells in the blood, we observed that the majority of them exhibited the GEP5 program, with only a minor fraction utilizing the GEP6 program (Figure 4D). This GEP phenotype was further confirmed by flow cytometry, revealing that most MAIT cells were CD8⁺ and, interestingly, that all MAIT cells displayed a uniform effector state (characterized by PLZF⁺CD161⁺EOMES⁺GZMK⁺CCR7⁻CD62L⁻), regardless of CD8 expression (Figure S12). These findings indicate that MAIT cells in the bloodstream primarily exist in an exclusive transcriptional state.

Blood γδ T cells were stratified into five clusters: naive (c0, GEP3), cycling (c4, GEP11), and clusters c1–c3 categorized into GEP5 or GEP6 programs (Figures 4C and 4D). This division in GEP utilization closely mirrored the specific TCR usage among these cells. TRDV2/TRGV9-expressing cells were predominantly GEP5, while TRDV1⁺ or TRDV3⁺ cells were enriched in GEP6 (Figures S13A–S13C). Flow cytometry confirmed Vγ9⁺Vδ2⁺ T cells primarily expressed PLZF and GZMK, while Vδ2⁻ T cells were increased in GZMB⁺ cells (Figures S13D and S13E). Thus, GEP5 is an effector gene module unique to T_inn cells, indicating common transcriptional states in human T_inn cells.

The distribution of CD4 and CD8 T_conv cells in the blood revealed two primary patterns. T_conv cells were found either within clusters containing naive cells (clusters 0 and 1) characterized by high expression of GEP3 or dispersed across clusters of cells displaying a gradient of the GEP6 program, with intermediary cells expressing GEP4 (Figure S14). The proportions of cells in these clusters exhibited variations among donors (Figure 4B). Together, these findings underscore the distinct associations between different T cell types and effector programs. In the blood, CD4⁻ iNKT, MAIT, and Vγ9⁺Vδ2⁺ γδ T cells predominantly employ the GEP5 program, a program also shared by effector T_inn cells in the thymus (Figure S8). Conversely, conventional CD4⁺ and CD8⁺ T cells transition into effector cells along a gradient defined by the GEP6 program. Notably, this GEP6 program is also shared by Vδ3⁺ and Vδ1⁺ γδ T cells.

Next, we used pairwise DEG analyses between T cell lineages in human blood to identify lineage-specific genes. We uncovered a total of 167 genes that exhibited significant differential expression (p_adj < 0.01) in at least two of the comparisons we conducted (Figure 4E). These distinct patterns of DEGs provided insights into changes linked to the transition from a “naive” state to an “effector” state across cell types. Furthermore, we identified genes that were commonly expressed by two distinct cell types when compared to the others. Nevertheless, we did not readily discern any gene-expression patterns specific to a particular cell type. However, intriguingly, among these, a group of 104 genes distinguished γδ, MAIT, and iNKT cells from T_conv CD4 and CD8 T cells, with 63% overlapping with GEP5. Given that only Vγ9⁺/Vδ2⁺ T cells share the GEP5 program with iNKT and MAIT cells, while Vδ2⁻ T cells exhibit greater similarity to T_conv cells as they share the GEP6 program (Figure S13C), we explored whether we could identify cell-type-specific gene signatures specifically among GEP5-expressing cells using the same analytical approach. Surprisingly, the results demonstrated that the only significant DEGs between iNKT, MAIT, and γδ T cells employing the GEP5 program (Figure 4F) were genes encoding the constant regions of the TCR genes (TRGC1, TRAC), the CD8 coreceptor (CD8A, CD8B), and the CD94 receptor (encoded by KLRD1). Thus, human blood T_inn cells, which encompass iNKT, MAIT, and Vγ9⁺/Vδ2⁺ cells, distinguish themselves from T_conv cells by employing a specific gene program, but there is minimal transcriptional difference among T_inn cells themselves.

The effector GEPs exhibit distinct migration, cytokine, chemokine, and integrin characteristics established by distinct gene-regulatory networks

The differentiation states of T cells are intricately linked to their phenotypic, functional, and migratory attributes, making their characterization clinically relevant. Each GEP aligns with distinct sets of chemokine and cytokine receptors as well as molecules related to cytotoxicity, NK receptors, and integrins (Figure 5A). For example, the GEP4 program, shared by T_cm/T_em (Figure S10) and some iNKT cells depending on the donor (Figures 4B–4D), shows high expression of chemokine receptors CXCR3 and CCR4, sphingosine-1-phosphate receptor 4, and oxysterol receptor GPR183, with the latter offering survival and migratory signals to thymocytes and CD4 T follicular helper cells.⁴³ GEP4 also features high levels of IL2RA, IL6R, IL4R, and ITGB1, while conspicuously lacking cytotoxic molecules (Figure 5A). In contrast, the GEP5 program, predominant in T_inn cells across most donors (Figures 4B–4D), shows elevated expression of CCR1, CCR2, CCR5, and CCR6, as well as CXCR6 and cytokine receptors such as IL18R1, IL18RAP, IL12RB1, IL12RB2, IL23R, and IFNGR1 (Figure 5A). This pattern includes GZMA and GZMK but lacks GZMB and GZMH (Figures 5A and S5A) and features the NK receptor KLRB1. On the other hand, the GEP6 program, mainly associated with T_em/T_emra cells (Figure S10) and Vδ1⁺ and Vδ3⁺ γδ T cells across the majority of donors (Figures 4B, 4C, and S13), shows increased CX3CR1 expression (Figure S14), correlating with effector state differentiation in CD4 and CD8 T cells.⁴⁴ GEP6 includes IFNG, CCL4, CCL5, and KLRD1, several integrins (ITGAL, ITGB2, ITGAM), and cytotoxicity genes such as GZMB, GZMH, and granulysin (GNLY), with reduced GZMK compared to GEP5 (Figure 5A). These findings align with studies indicating that GZMK⁺ and GZMB⁺ cells delineate T_cm and T_em/T_emra cell populations.^45,46

Figure 5. Effector gene-expression programs in T_inn and T_conv.

(A) Key genes categorized by function and depicted by their expression level (Z-score color scale) and percentage of expression in cells belonging to the indicated GEPs.

(B–D) (B) Single-cell regulatory network inference and clustering of TFs and enrichment score per cell (as row-scaled Z scores), ordered by cluster (as in Figure 1C), with tissue of origin and GEP assignment (based on cNMF usage) indicated by color bar. Two row clusters are marked, which are preferentially enriched in T_inn (upper bracket) and T_conv (lower bracket). (C and D) TFs with pronounced activity in (C) T_inn and (D) T_conv (corresponding to brackets in B) and their targets. Green dots indicate TFs (y axis) that have other TFs as their target (x axis), where purple labels TFs that can interact in either direction. The marginal bar chart shows the number of TFs per target, color coded by their functional categorization (as in A).

We then used a network inference approach to identify the activity of regulons—TFs and their targets—in single cells.⁴⁷ We identified 149 of such regulons, with 11 regulons more active in T_inn compared to T_conv cells (Figure 5B). These regulons were governed by TFs such as ELK3, MBD2, CREM, NFE2L2, NR1D2, XBP1, MYBL1, RORA, MAF, CEBPD, and FOSL2. Curated analysis of their predicted target genes indicates that these TFs may play a central role in shaping the unique transcriptional profile observed in T_inn cells during steady-state conditions. This role encompasses the regulation of chemokine and cytokine receptors as well as other genes associated with T_inn cells, including ZBTB16 (encoding PLZF), the master regulator of the T_inn cell lineage (Figure 5C). A second group of regulons exhibited enriched activity within effector T_conv cells, although some shared activity with T_inn cells (including EOMES, RUNX3, PRDM1, and FLI1; Figure 5B). As T_conv cells differentiate into T_em/emra cells, there is an increased activity of regulons driven by TBX21, KLF, and NFAT family TFs (Figures 5B and 5D), in agreement with their functions in regulating the cytolytic activity of CD8 T cells.^48–50 Taken together, we discovered novel candidate regulators of T_inn and T_conv effector programs, along with their predicted target genes, which warrant further experimental validation.

Cross-species analysis of thymic T_inn cell development

Only a minority of human T_inn thymocytes display an effector phenotype, contrasting with the predominant effector association of mouse T_inn thymocytes, which develop into distinct effector subsets.^23,51,52 To explore transcriptional similarities between mouse and human T_inn cells in the thymus, we constructed a reference mouse T_inn dataset from nine studies.^51–59 This dataset revealed 13 transcriptionally distinct clusters, with iNKT, MAIT, and γδ T cells coexisting in variable proportions (Figures 6A and S15; Table S13). Lineage-specific clusters included those unique to γδ T cells (c1 and c2, immature Cd24a⁺Gzma⁺ cells^51,58), signaling cells (c3 and c4), cycling cells (c5), type 1 cells (c8 and c9), type 2 cells (c6 and c7), and type 17 cells (c10 and c11). Cluster 0, expressing markers such as Sell (encoding for Cd62l), Klf2, Ccr7, Foxo1, and S1pr1, likely represents T_inn cells positively selected on thymic epithelial cells (TECs), bypassing the “innate” pathway.^23,60 We performed a cross-species comparison of cell identities by assessing the pairwise correspondence between murine T_inn signatures and human iNKT, MAIT, and γδ T cell clusters (Figure 6B). Human iNKT cells in cluster 0 (NKT_c0), exhibiting a CD8αα T cell gene signature, showed the strongest resemblance to signaling cells (Figure 6B), likely due to shared TCR activation genes. Conversely, cells with an effector profile (NKT_c5 and NKT_c6) showed the closest relationship to mouse type 1 and to a lower extent type 17 cells (Figure 6B). Importantly, we did not find human clusters corresponding uniquely to specific mouse subsets, confirming that human iNKT cells do not differentiate into distinct subsets but rather acquire a mixed type 1/type 17 transcriptome. We also did not detect any human iNKT cell clusters that matched with the mouse type 2 subset with a high degree of confidence (area under the receiver operator characteristic curve [AUROC] > 0.65; Figure 6B), suggesting that type 2 iNKT cells are likely absent in the human thymus. Corroborating this finding, we did not detect any expression of IL-4- or IL-13-encoding transcripts in human thymic iNKT cells, which are typically associated with mouse type 2 thymic iNKT cells. Similar patterns were observed for MAIT and γδ T cells in the human thymus, with effector cells resembling more closely mouse type 1 and type 17 effector cells (Figure 6B). This indicates that human T_inn cells follow a distinctive path with mixed effector potential, unlike the mouse model with multiple effector subsets.

Figure 6. Cross-species comparison of mouse and human T_inn development.

(A) Mouse T_inn reference atlas with seven characteristic cell states highlighted, which are found across lineages (as in Figure S15).

(B) MetaNeighbor analyses showing pairwise correspondence (AUROC scores) between murine T_inn (as in A) and human iNKT, MAIT, and γδ T cell clusters (as in Figure 2D). Marginal bar charts indicate number of cells in the corresponding clusters.

(C) Expression of human regulon-driving TFs (as in Figure 5) together with murine TFs of importance in T_inn development (Rorc, Tbx21) projected on mouse T_inn reference atlas (as in A).

We next assessed whether the TFs driving human T_inn cell regulons (Figure 5A) were also expressed in mouse T_inn cells (Figure 6C). Most TFs were indeed expressed in mouse T_inn cells, although their expression varied across clusters. However, exceptions included CEBPD, EOMES, and MYBL1, which were highly expressed in human T_inn cells (Figure S5A) but barely detectable in mouse T_inn cells (Figure 6C). Conversely, mouse type 1 T_inn cells exhibited high T-bet levels (encoded by Tbx21, Figure 6C), while human T_inn cells had low T-bet expression (Figures S5A, S11B, and S12B). These findings highlight some species-specific differences in TF expression that could play a role in modulating T_inn cell development and functions.

CD1D, SLAMF6, and SLAMF1 expression in the mouse and human thymus

The existence of T_inn cells in the human thymus with a transcriptome similar to developing T_conv cells raises questions about their origin. In mice, a subset of MAIT cells is positively selected by radiation-resistant TECs, which do not lead to a memory or effector phenotype acquisition.^52,59 This is because TECs lack SLAM receptors, crucial for T_inn commitment.⁶¹ Although this is more common among MAIT cells, some thymic mouse iNKT cells also exhibit a similar transcriptome.²³

We hypothesized that naive T_inn cells in humans might result from a similar TEC-mediated selection process. Given the challenge of detecting surface MR1 expression at steady state, we investigated CD1D protein expression in the thymus instead. Mouse TECs have been reported to express CD1d on their surface.⁶² scRNA-seq of the mouse thymus²⁷ confirmed Cd1d1 expression across various cell types, including thymocytes and cortical and medullary TECs (Figures 7A and 7B). Flow-cytometry analyses corroborated these findings (Figures 7C and 7D).

Figure 7. CD1D, SLAMF1, and SLAMF6 gene and protein expression in mouse and human thymus.

(A) Clustering of thymic cell populations and (E) their expression of Cd1d1 (mouse)/CD1D (human) derived from the mouse and human thymus cell atlas, respectively.²⁷ (B and F) Normalized expression of Cd1d1/CD1D and Slam/SLAM transcripts across thymic cell populations. Flow cytometry of (C and G) mouse and human TECs and (D and H) thymocyte subsets.

In contrast, human thymus scRNA-seq data²⁷ showed a more limited pattern of CD1D and SLAMF6 expression (Figures 7E and 7F). Human DP thymocytes express CD1D transcripts and surface CD1D molecules, but this expression is lost in mature SP thymocytes. Flow cytometry confirmed that, as in mice,⁶¹ SLAMF1 is expressed by human DP and SP thymocytes, but SLAMF6 is not detectable on the surface of human thymocytes (Figure 7H). Additionally, human cortical TECs (cTECs) express CD1D transcripts and have surface CD1D protein, whereas medullary TECs (mTECs) do not (Figures 7G and 7H). This interspecies difference in CD1D and SLAM family members’ expression might affect iNKT cell development, given the crucial role of mTECs in murine iNKT cell development.^19,63

DISCUSSION

In this study, we employed multi-modal single-cell transcriptomics to explore the diverse phenotypic states of T_inn cells within the human thymus and blood. By comparing these states to those of T_conv cells, we provided insights into human T cell biology and a comprehensive resource for further studies of health and disease. Our work emphasizes T_inn cells as promising candidates for immunotherapies.^64–66

Our study demonstrated that the majority of T_inn cells in adult human blood exhibit a distinct transcriptional program shared by most iNKT, MAIT, and Vδ2Vγ9⁺ T cells under steady-state conditions. This program implies a blended type 1/type 17 transcriptional pattern, driven by specific TFs that enable the expression of distinct chemokine and cytokine receptors, NK receptors, and cytotoxic molecules. This equips T_inn cells to swiftly respond to cytokines such as IL-12, IL-18, and IL-23 independently of TCR signaling.^22,67 Notably, human T_inn cells constitutively express GZMK but lack GZMB while also expressing cathepsins necessary for activating granzymes.⁶⁸ This suggests that T_inn cells are prepared to release active GZMK upon stimulation,⁶⁹ which can induce pro-inflammatory cytokines^45,70 and activate complement,⁷¹ implying a role in immune regulation and inflammatory responses. In contrast, mouse T_inn cells do not express GZMK transcripts but possess pre-formed cytokine-encoding transcripts, allowing for immediate responses.^72,73 Hence, despite their evolutionary conservation, T_inn cells may have evolved species-specific mechanisms to provide early signaling and amplification of the adaptive immune response.

We identified TFs and their predicted target genes with increased transcriptional activity in human T_inn cells compared to naive and effector T_conv cells. Many of these TFs are associated with IFN-γ, cytotoxicity,^48–50,74 and IL-17 production,^75–77 consistent with the type 1/type 17 transcriptional program observed in T_inn cells. In mice, the Th1/Th17 paradigm identifies IL-12 and IL-23 as cytokines that induce IFN-γ and IL-17 production, respectively. By contrast, activating human MAIT cells through their TCR or with IL-12 and IL-18, and stimulating T_inn cells with IL-23, results in IFN-γ production.^67,78 Yet, only a subset of these cells produces IL-17 in the same conditions, a phenomenon thought to be influenced by epigenetic modifications at the IL-17 gene loci.⁷⁸ Underlining the importance of understanding the regulation of IL-17 production in T_inn cells, IL-17 production in human MAIT cells is increased in diseases such as severe asthma, community-acquired pneumonia in children,^79,80 and colorectal cancer patients.⁸¹ Interestingly, NR1D family TFs, associated with Th17 cell regulation,^77,82 drive regulons in human T_inn cells. These factors are regulated by the circadian clock, suggesting that circadian rhythms might affect IL-17 production in T_inn cells, a hypothesis that warrants further investigation.

While many TFs essential for the human T_inn program are also expressed in mouse T_inn cells, there are notable exceptions such as CEBPD, EOMES, and MYBL1, which are highly expressed in human T_inn cells but barely detectable in mouse T_inn. CEBPD regulates CCR6 in human MAIT cells⁸³ and may play a crucial role in the human T_inn program. MYBL1 is expressed in human T_inn cells,⁸⁴ but its function remains to be defined. EOMES, essential for mouse iNKT cell development, shows low expression under steady-state conditions in mice^85,86 but is highly expressed in human T_inn cells. By contrast, T-bet is highly expressed in type 1 mouse T_inn cells and is essential for their development and functions.^86,87 However, human effector T_inn cells, which are most similar to mouse type 1 T_inn cells, express relatively low levels of T-bet. Instead, T-bet’s expression and activity were correlated with the acquisition of the GEP6 program by T_conv cells in humans. These findings suggest the possibility of species-specific transcriptional regulation of T_inn cells, which could be relevant for their future therapeutic applications. Curiously, high-confidence regulons such as PLZF and Rorγt were not identified in the gene-regulatory network of human T_inn cells, possibly due to the relatively low gene detection in this context.

In the postnatal thymus, iNKT and MAIT cells display a transcriptional profile similar to that of developing conventional CD4⁺ and CD8⁺ T cells. Unlike conventional thymocytes selected by cTECs, murine iNKT and MAIT cells undergo selection by DP thymocytes via SLAM family receptors, leading to PLZF expression and effector differentiation.^61,88,89 A minor fraction of murine MAIT cells with a naive-like phenotype is TEC selected.⁵² The corresponding expression patterns of CD1d on cTECs and DP thymocytes in both murine and human thymus suggests a similar selection of human iNKT thymocytes by cTECs or DP thymocytes. The expression of PLZF in non-effector iNKT cells, as observed through both scRNA-seq and flow cytometry, supports the hypothesis of DP selection. However, the initial selection process may not provide all necessary signals for complete maturation, as cells just experiencing positive selection are more akin to CD4⁺ and CD8⁺ T_conv cells. Additional signals may be required for human iNKT and MAIT cells to acquire effector functionalities. This hypothesis aligns with the prevalence of naive iNKT and MAIT cells in human cord blood and the gradual increase in effector Tinn cells with age.^41,42,90 Unlike murine T_inn cells, which use SLAMF6 and SLAMF1 to induce PLZF expression and effector maturation,⁶¹ human thymocytes do not express SLAMF6, possibly affecting T_inn cell maturation. Additionally, human mTECs lack CD1D expression, which could also contribute to the variations in iNKT cell maturation observed in cross-species analyses.^19,63

Our study highlights a distinct path taken by T_inn cells with an effector program in the human postnatal thymus, characterized by a mixed type 1/type 17 effector potential, contrasting with mice where T_inn cells split into multiple effector subsets. We did not observe specific clusters of proliferative human iNKT and MAIT thymocytes. In mice, proliferative thymic iNKT and MAIT cell clusters are identifiable^52,54,55 and reflect the proliferative burst following positive selection,⁹¹ crucial for establishing a substantial T_inn cell pool. While T_inn cells constitute 1%–2% of thymocytes in mice, their proportion is much lower in pediatric humans. Moreover, our analysis did not reveal any type 2 T_inn cells in humans, unlike in mice, where thymus-resident iNKT2 cells significantly impact the thymic environment through IL-4 production.^{17,18,63,92,93} The scarcity of type 2 T_inn cells in the human thymus suggests that these phenomena may be species specific or regulated by different cell types in humans.

Taken together, our findings hold significance in elucidating the diverse functional attributes of human T_inn cells and their potential applications in immunotherapeutic contexts.

Limitations of the study

Our study presents a comprehensive atlas of human T_inn cells in the thymus and blood, revealing a blended type 1 and type 17 transcriptional profile. However, there are several limitations to consider. We noted variations in effector cell percentages among donors, which may be influenced by factors such as age, sex, or immunological history. Additionally, our analysis focused on steady-state transcriptional profiles, leaving unclear how these profiles might change in various disease states. Comparing pediatric thymic T cells with adult blood samples introduces age-related differences that could impact transcriptional profiles. Ideally, matched samples would offer a more accurate comparison, but the limited availability of iNKT and MAIT cells in neonates presents challenges in obtaining sufficient cell quantities. Future research should explore thymus-derived fetal T_inn cells to investigate whether they exhibit specific type 1 and type 17 profiles akin to those described in mice. Recent studies on fetal γδ T cells suggest that fetal T_inn cells might also develop unique transcriptional subsets.²⁵ Given that iNKT and MAIT cells appear from gestational weeks 18–23^94,95 and fetal hematopoietic stem and progenitor cells are predisposed to innate-like lymphocytes,⁹⁶ examining these cells could yield insights into their development and function. Furthermore, the potential for distinct waves of T_inn cells arising from fetal versus adult hematopoietic stem cells in mice⁹⁷ highlights the need for further investigation into their transcriptional differences.

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Laurent Gapin (laurent.gapin@cuanschutz.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• Data that support the findings of this study, including the raw data and Seurat object, were deposited in NCBI GEO with the accession code GEO: GSE249684. The data from this study is also displayed as a ShinyCell application at https://xspeciestcells.cshl.edu/.

• The code used for all analyses presented in this study is publicly available on GitHub: https://github.com/meyer-lab-cshl/xspeciestcells. The code for the browser application can be found at https://github.com/meyer-lab-cshl/xspeciestcells-shiny.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Mice

C57BL/6 mice were purchased from Jackson Laboratories. The Cd1d1d2^−/− mice backcrossed to the C57BL/6 background were previously described.⁹⁹ All mice used were between 6 and 15 weeks and age-matched for each experiment. Mice were raised in a specific pathogen-free environment at the Office of Laboratory Animal Research at the University of Colorado Anschutz Medical Campus or the Animal Core Facility at Cold Spring Harbor Laboratory. Animal procedures were approved by the UCD (00065) Institutional Animal Care and Use Committees and the Cold Spring Harbor Laboratory IACUC (23–1); all procedures were carried out in accordance with the approved guidelines.

Human

Thymus tissues were obtained from anonymous human donors who were undergoing medically necessary surgery where removal of a portion of the thymus was required toa facilitate exposure of the operative field. No tissues were obtained specifically for the purposes of this study. This use of discarded tissue was approved by the Institutional Review Board (IRB) of the University of Colorado Anschutz Medical Campus (IRB-17–2159). Additional samples were collected under the Mount Sinai Biorepository with a Waiver of Consent and under the Northwell Health Biospecimen Repository (IRB 20–0150) with patient consent; the use of these human tissues was reviewed by Cold Spring Harbor Institutional Review Board. Pediatric thymus samples for scRNAseq came from individuals between 10 and 20 weeks old (Table S1), and samples used for flow cytometry experiments came from individuals between 4 days and 5 months old. Plateletpheresis leukoreduction filters (LRS chambers) were purchased from Vitalant Blood Center (Denver, CO, USA). Additional PBMCs were collected under COMIRB #17–2159 at the University of Colorado Clinical and Translation Research Centers (CTRC) which is a part of the Colorado Clinical and Translation Sciences Institute (CCTSI), with all donors having provided written informed consent. Overview of sample metadata is provided in Table S1.

METHOD DETAILS

Murine thymus samples

To isolate thymocytes, thymus tissue was immersed in RPMI 1640 media (Corning, #10–040-CV) and gently pressed through a 40mm cell strainer using the plunger of a 1 mL syringe. For TEC isolation, the thymus tissue was cut into small fragments and submerged in RPMI 1640 media without phenol red (Gibco, #11835030), supplemented with 20mM HEPES, 1.3 U/mL Liberase TH, and 100 U/mL DNase I. These tissue fragments were incubated for 5 min on ice followed by an additional 20 min at 37°C. After digestion, the solution was repeatedly mixed with a micropipette to ensure complete tissue disintegration. To stop the digestion process, cells were suspended in HBSS, 4% heat-inactivated FBS (HI-FBS, FBS preheated for 20 min at 56°C), 20mM HEPES, and 10U/mL DNase I. To remove immune cells, the cell suspension was incubated with rat anti-mouse CD90.2 (clone 53–2.1), anti-mouse CD45 (clone 30-F11), and anti-mouse CD45-BV605 (clone 30-F11) antibodies for 30 min at 4°C. Subsequently, the cell suspension was placed on panning plates coated with goat anti-rat IgG for 20 min at room temperature. Unattached cells were then transferred to new panning plates for a second depletion round. The remaining cell suspension, following this depletion process, was prepared for flow cytometry analysis.

Human thymus samples – Thymocytes

To extract thymocytes for both scRNAseq and flow cytometry, the thymus tissue was placed in complete RPMI 1640 media (Gibco, #22400–071), with 10% HI-FBS, 1% non-essential amino acids, 1% sodium pyruvate, 1X GlutaMAX, 1% Penicillin/Streptomycin, and 1X 2-Mercaptoethanol (Sigma-Aldrich #21985–023). The tissue was cut into small pieces, and gently pressed with the back of a 10 mL syringe to release thymocytes. The resulting suspension was passed through a 70 μm filter. Thymocytes were isolated using a Ficoll-Paque density gradient provided by Cytiva. Tetramer staining for MAIT and iNKT cells, and surface staining for CD4⁺, CD8⁺ and γδ T cells was performed on freshly isolated thymocytes.

Human thymus samples – thymic epithelial cells

To enrich TECs for flow cytometry, thymus tissue was cut into small pieces and placed in RPMI 1640 media without phenol red (Gibco, #11835030), 5% HI-FBS, 1% Penicillin/Streptomycin, 10mM HEPES, and 0.55mM 2-Mercaptoethanol (Gibco, #21985023). The thymus tissue in this media was stirred on a magnetic plate for 40 min. The supernatant was removed and replaced with fresh media every 10 min to remove released thymocytes. The remaining tissue chunks were placed in a digestion buffer consisting of RPMI 1640 media without phenol red (Gibco, #11835030), 2% HI-FBS, 20mM HEPES, 80 U/mL DNase I, 1.6 U/mL Dispase I, and 0.3 mg/mL Collagenase IV, for digestion at 37°C with gentle shaking. This digestion process was conducted in two sessions of 25 min each, in between which the supernatant was extracted and replaced with fresh digestion buffer. At the end of the digestion, the tissue chunks had nearly entirely disintegrated, and the digestion was halted by resuspending cells in the same buffer used for thymocyte release (RPMI 1640 media without phenol red, 5% HI-FBS, 1% Penicillin/Streptomycin, 10mM HEPES, 0.55mM 2-Mercaptoethanol). The combined supernatants were further incubated in TrypLE Express Enzyme, 1mM MgCl₂, 2mM CaCl₂, 100U/mL DNase I for 5 min at 37°C to obtain a single-cell suspension. The digestion was stopped by resuspending cells in the previously described thymocyte release buffer. To remove immune cells and erythrocytes, cells were incubated with mouse anti-human CD3 (clone UCHT1), anti-human CD4 (clone RPA-T4), anti-human CD8ɑ (clone RPA-T8), anti-human CD45 (clone HI30) and anti-human CD235a (clone HI264) antibodies in HBSS, 4%HI-FBS and 20U/mL DNase I, for 30 min at 4°C. Cells were then placed on panning plates coated with goat anti-mouse IgG for 20 min at room temperature, and the unadhered cells were transferred to new panning plates for a second round of depletion. The remaining cells following depletion were then stained for flow cytometry.

Human peripheral blood samples

De-identified peripheral blood samples from the Human Immune Tissue Network Biobank (COMIRB # 17–2159) were collected using sodium heparin tubes. De-identified peripheral blood samples from the Vitalant Blood Center were acquired from plateletpheresis leukoreduction filter chambers (LRS). Peripheral blood mononuclear cells (PBMCs) were isolated from these samples using a Ficoll-Paque density gradient provided by Cytiva. PBMCs were cryopreserved in FBS with 10% DMSO and stored in liquid nitrogen.

Magnetic-bead enrichment of iNKT and MAIT cells

To enrich for thymic MAIT and thymic/blood iNKT cells, up to 2 × 10⁹ cells were incubated with MR1–5-OP-RU-PE or CD1d-PBS57-PE tetramers respectively in MACS buffer (1X PBS, 0.5% BSA, 2mM EDTA), for 25 min at room temperature. Cells were washed twice and incubated with anti-PE microbeads, followed by separation using an autoMACS Pro Separator (Miltenyi) according to manufacturer’s instructions. PE-microbead-labelled cells in the enriched fraction were stained with the specified panel of antibodies listed below.

Fluorescence-activated cell sorting

To sort CD4⁺ T, CD8⁺ T, γδ T, or peripheral blood MAIT cells, 2 × 10⁶ cells unenriched cell suspensions were used. To sort iNKT cells (thymocytes or PBMCs) or thymic MAIT cells, the staining and sorting was performed on magnetic bead-enriched cell suspensions from 0.5 to 10 × 10⁸ cells (above). All single cell suspensions were stained with efluor780 viability dye for 10 min at room temperature and washed once prior to cell surface staining. The cell suspensions were then stained in MACS buffer (1X PBS, 0.5% BSA, 2mM EDTA) at room temperature for 20 min, both with cell surface markers and a unique oligonucleotide-tagged antibody sample tag (Human Single Cell Sample Multiplexing Kit, BD Biosciences) to allow separation in downstream scRNAseq analyses. The following cell surface markers were included in the staining: CD3-AF488 (clone OKT3), CD14-eFluor450 (clone 61D3), CD19-eFluor450 (clone H1B19), Vα7.2-BV785 (clone 3C10), Vα24-PerCP-Cy5.5 (clone C15), CD4-redFluor710 (clone OKT4), CD8α-PE-Cy7 (clone SK1), TCRγδ-BV650 (clone 11F2), FcγR block. Following cell surface staining, cells were washed twice and resuspended in MACS buffer prior to cell sorting on the Aria 3 (BD Biosciences). Purified cell populations were sorted into MACS buffer. Validation of the cell sorting panel was performed on the Cytek Aurora flow cytometry system using SpectroFlo software (v3.0). Overall, from infant thymus and PBMC donors up to 5 populations were sorted after doublet, viability, B cell (CD19⁺CD3⁻) and monocyte (CD14⁺CD3⁻) discrimination: 1. MAIT cells (MR1–5-OP-RU-Tet⁺Vα7.2⁺CD3⁺), 2. iNKT cells (CD1d-PBS57-Tet⁺Vα24⁺CD3⁺), 3. γδ T cells (CD3⁺TCRγδ⁺), 4. CD4⁺ T cells (CD4⁺CD8α⁻CD3⁺) and CD8⁺ T cells (CD8α⁺CD4⁻CD3⁺). Cell subsets sorted for the different donors are listed in Table S1 and the gating strategy is shown in Figure S16.

Flow cytometry — Thymic and peripheral blood T cells

To confirm gene expression from scRNAseq analysis, MAIT and iNKT cells were enriched from the human thymus as described above, as were iNKT cells from human blood. Peripheral blood MAIT and γδ T cells were stained directly without enrichment. Single cell suspensions were stained as above with efluor780 viability dye prior to incubation at 37°C for 10 min with CCR7-APC-Fire810 (clone G043H7) and FcγR block. A combination of the following cell surface markers were subsequently added and cells were stained at room temperature for 15 min: CD3-BUV496 (clone UCHT1), CD14-PE-Cy5 (clone 61D3), CD19-PE Cy5 (clone H1B19), Vα7.2-BV785 (clone 3C10), Vα24-PerCP-Cy5.5 (clone C15), CD4-BV570 (clone RPA-T4), CD8α-BUV395 (clone RPA-T8), TCRγδ-BV650 (clone 11F2), Vδ1-PerCP-Vio700 (clone REA173), Vδ2-FITC (clone 123R3), Vγ9 Vγ9-PE (clone B3), CD161-BUV805 (clone HP-3G10), CD62L-BV650 (clone DREG-56). Cells were then washed twice with MACS buffer, and intracellular staining was performed with BD Transcription Factor Buffer Set according to the manufacturer’s specification. The following antibodies were used to stain for intracellular proteins: PLZF-PE-CF594 (clone R17–809), Eomes-BUV737 (clone X4–83), T-bet-BV605 (clone 4B10), GZMK-eFluor660 (clone G3H69), GZMB-AF700 (clone GB11). Phenotypic analyses were performed on the Cytek Aurora flow cytometry system using SpectroFlo software (v3.0). Data were analyzed using FlowJo software (v10.7.1).

Flow cytometry -— CD1d, SLAMF1, SLAMF6

For murine experiments, thymocytes were resuspended in PBS, 5% FBS, 4mM EDTA and stained for 30 min at 4°C with: Fc blocker anti-CD16/32 (clone 93), CD4-AF488 (clone GK1.5), CD8α-APC (clone 53–6.7), CD1d-PE (clone 1B1). For murine thymus samples which were depleted of immune cells, the single cell suspension was resuspended in HBSS, 4% HI-FBS, 20mM HEPES, 10U/mL DNase I, 2.5mM EDTA, and stained for 30 min at 4°C with: Fc blocker anti-CD16/32 (clone 93), EpCAM-BV421 (clone G8.8), CD45-BV605 (clone 30-F11), UEA1-FITC, Ly-51-AF647 (clone 6C3), and CD1d-PE (clone 1B1). For flow cytometry experiments on human samples, thymocytes were resuspended in PBS, 2% FBS, and stained for 30 min at 4°C with: TruStain FcX, CD45-BV421 (clone HI30), CD4-AF488 (clone OKT4), CD8α-APC (clone RPA-T8), CD1d-PE (clone 51.1). For human samples which were depleted of immune cells and erythrocytes, cells were resuspended in PBS, 2% FBS, and stained for 30 min at 4°C with: TruStain FcX, CD45-AF647 (clone QA17A19), EPCAM-BV421 (clone 9C4), CDR2-AF488 (purified CDR2 antibody kindly provided by Dr. Sheena Pinto, conjugated with the AF488 antibody labeling kit from ThermoFisher Scientific), HLADR-BV711 (clone L243), CD1d-PE (clone 51.1). In all experiments, to measure viability cells were stained with the live/dead Fixable Near-IR dead cell stain kit, simultaneously with cell surface markers. Flow cytometry was performed on a BD LSR Fortessa Cell Analyzer (BD Biosciences).

Single-cell RNA sequencing

Prior to cDNA library preparation for the whole transcriptome (WTA) and VDJ libraries, all cell subsets from the different donors were pooled, with up to 12 unique sample tags combined per library. Single cell WTA and VDJ sequencing libraries were prepared using the BD Rhapsody Single-Cell Analysis System (BD Biosciences) according to the manufacturer’s specifications. Libraries were quantified and pooled according to equivalent molar concentrations and sequenced on the NovaSeq sequencing platform at the University of Colorado Genomics Core with the following read lengths: read 1–150 cycles; read 2–150 cycles; and i7 index - 8 cycles.

Single-cell RNA-seq data analysis

The quality of sequencing reads was evaluated using FastQC and MultiQC. Sequencing reads (FASTQ) were mapped and sample tags were deconvoluted with The BD Rhapsody WTA Analysis Pipeline on the GRCh38 genome sequence. This pipeline produced a gene expression matrix for each sample, which records the number of UMIs for each gene associated with each cell barcode. Aggregated data were then imported into the R environment (version ≥4.0.3) and analyzed with Seurat (v4.3.0). Low-quality cells were filtered using the cutoffs nFeature_RNA ≥ 500 & nFeature_RNA <3000. Cells with a high mitochondrial content were removed using a batch-dependent threshold with the isOutlier function from the scater package (v1.26.1).¹⁰⁰ Genes expressed in less than 20 cells were ignored. This resulted in 78,607 cells with 17,204 genes for downstream analyses. The NormalizeData function of Seurat was performed using default parameters to remove the differences in sequencing depth across cells. Dimensionality reduction was performed prior to integration for visualization purposes (Figure S2A), by selecting 2000 highly variable genes for principal component analysis (PCA) and uniform manifold approximation and projection (UMAP). To integrate the data and remove batch-effects from the PCA subspaces based on the correct cell alignment, we used Harmony¹⁰¹ following PCA to project cells into a shared embedding in which cells group by cell type rather than dataset-specific conditions. We then applied the RunUMAP function on 20 dimensions of the harmony embedding to obtain bidimensional coordinates for each cell. We determined the k-nearest neighbors of each cell using the FindNeighbors function and used this knn graph to cluster cells using the Louvain algorithm from FindClusters based on the same harmony dimensions as the RunUMAP function (20 dimensions, resolution 1.2). For analyses performed on individual lineages (CD4⁺ T, CD8⁺ T, iNKT, MAIT or γδ T cells) in Figures 2 and 4, the dataset was split up based on cell hashing tags. Each lineage from each tissue was re-analyzed individually using the same steps to obtain UMAPs and clusters in Figures 2 and 4. Plots displaying cells on UMAPs were generated using the SCpubR package (v2.0.2).¹⁰²

LISI metric and analysis of cluster stability

The local inverse Simpson’s index (LISI) was used to assess the degree of mixing during batch correction and dataset integration in scRNAseq analysis.¹⁰¹ This approach helps evaluate the effectiveness of data integration methods by quantifying how well datasets are merged without introducing artificial batch effects. To assess the integration process, we employed the “integration LISI” (iLISI) score. iLISI measures the effective number of datasets within a neighborhood and provides an indication of how effectively the individual datasets were harmoniously integrated into a unified whole during the process. In addition, we used the “cell-type LISI” (cLISI) score to evaluate the accuracy of cell-type assignments in the integrated dataset. cLISI is a modified version of the LISI score, but instead of assessing dataset labels, it focuses on the accuracy of cell type assignments within the integrated data. As the specific identities of individual cells were not known beforehand, we assigned mock cell identities based on anticipated gene expression patterns. These mock identities were determined using prior knowledge of gene expression markers associated with distinct cell types. For instance, we identified DN thymocytes as cells expressing PTCRA > 1; B cells as cells expressing CD19 > 1 and IGKC > 1; T_regs as cells expressing FOXP3 > 1; MAIT cells as cells expressing SLC4A10 > 1 and FOXP3 < 1; CD4⁺ T cells as cells expressing CD4 > 1, CD8A < 1, SLC4A10 < 1, FOXP3 < 1, and CCR7 > 1; DP thymocytes as cells expressing RAG1 > 1 and CD1C > 1; and CD8αα thymocytes as cells expressing CD8A > 1 and GNG4 > 1. These mock identities were used as initial cell type assignments and served as the basis for assessing the success of integration, as indicated by increased iLISI scores and the maintenance of a cLISI score of 1 (Figure S2B). Only cells with assigned mock identities were included in the cLISI analysis. To evaluate the stability of clusters, we conducted a bootstrapping procedure in which cells from each predefined cluster were repeatedly sampled and then subjected to re-clustering. Cluster stability was assessed by examining co-assignment probabilities (CP), where higher CP values indicated greater cluster stability (Figure S2C). In essence, a high CP suggests that the cells within a cluster consistently grouped together across multiple iterations, reinforcing the reliability and robustness of that cluster’s identity.

TCR analysis

V(D)J single cell sequencing data were mapped and quantified using the BD Rhapsody WTA Analysis Pipeline and the GRCh38 genome sequence. To connect the VDJ data with transcripts data for each cell, we established links based on cell indexes extracted from the consensus annotation files (VDJ_percell.csv) and MolsPerCell.csv files from each demultiplexed sample. Only TCR paired sequences were retained for subsequent analyses. TCR data from each VDJ-sequenced sample were combined together and added to the metadata of the Seurat object. Clonotypes were defined based on unique TCR VJ usage and complementary-determining region (CDR3) motifs. Basic TCR statistics, such as the distribution of length and counts were computed using the tidyverse package (v1.3.2). The assessment of clonotype diversity was conducted using the mean value of the Shannon index, computed through the diversity function of the vegan R package (v2.6–4) after 100 iterations. Prior to the diversity calculation, the data was subjected to rarefaction to match the lowest sequence count found within the studied groups. Chord diagrams were generated using the circlize package (v0.4.15)¹⁰³ and CDR3 motif logos using the ggseqlogo package (v0.1).¹⁰⁴ The stacked letters’ cumulative height at each position signifies the degree of sequence conservation, portraying the relative abundance of amino acids, which is further depicted by the varying heights of individual letters within the stack.

Identification of differentially expressed genes between clusters

We identified cluster-enriched genes by using the FindAllMarkers function in Seurat with test.use = wilcox. This function identified differentially expressed genes for each cluster by comparing the gene expression for cells belonging to a cluster versus cells belonging to all other clusters. Only those genes that passed an adjusted p value (Benjamini-Hochberg) cutoff of 0.05, log fold change >0.4 and min.pct = 0.3 were included in Figure S3.

Characterizing the replicability of cell types defined by scRNAseq between studies and between species

We assessed the consistency of cell clusters in our integrated thymic data by comparing them with the human thymus atlas from the Park et al. dataset.²⁷ To do this, we focused exclusively on thymocytes, totaling 37,369 cells in our dataset. We also acquired the annotated AnnData object from the Park et al. dataset, which specifically contained T cells. To enable a meaningful comparison, we combined the two raw count matrices, concentrating on the top 2000 highly variable genes shared across both datasets. This resulted in a matrix containing 3,106 genes and 114,363 cells. To evaluate the consistency of cell types between these datasets, we employed the pyMN package to perform unsupervised MetaNeighbor analysis.²⁶ MetaNeighbor assesses the similarity of cell types by constructing a network of cells based on the correlation of their gene expression profiles. It then predicts cell type labels, hiding them from one dataset while using the other. The result is expressed as a mean area under the receiver operator characteristic (AUROC) score, which measures the probability of correctly identifying a cell’s type based on its gene expression profile. We used the ggplot2 package to visualize the AUROC scores obtained from pyMN, comparing our integrated clusters with the thymocyte clusters defined in the Park et al. dataset. For assessing gene expression similarity between thymic lineages in Figure 2A, we considered donors as distinct datasets (study_id parameter); and cell lineages (based on cell hashing) as the cell groups of interest to compare (cell_type parameter), more particularly splitting thymocytes in each lineage into non-effector (clusters 0–11) versus effector states (clusters 12–17). MetaNeighbor was run on the top 2000 highly variable genes chosen on our integrated dataset. For assessing the replicability of cell clusters across species in Figure 6, we utilized the reference scRNAseq murine T_inn dataset (see below) and our human thymic iNKT, MAIT, and γδT individual Seurat objects from Figure 2. To ensure an appropriate comparison, we obtained orthologous genes between mouse and human with the biomaRt package.^105,106 We filtered the murine count matrix to retain only genes with known 1:1 orthologs in humans. Then, we performed unsupervised MetaNeighbor analysis with pyMN on the combined set of highly variable genes from both human and mouse datasets. Finally, we used ggplot2 to create visualizations of the AUROC scores returned by pyMN, including clusters that contained at least 1% of the cells in each species to ensure greater confidence in assessing the replicability of clusters across species.

Identification of gene expression programs

The count matrix was used for conducting consensus non-negative matrix factorization (cNMF).²⁸ This process enabled us to infer both identity and activity programs, along with their respective contributions in each cell. The usage of each program for each cell was added to the metadata of the Seurat object and displayed as a feature plot (Figure 1K). To determine the genes associated with each program, we plotted the gene ranks (ranging from most associated to least associated) against the gene_spectra_score output from the cNMF analysis. The plotted gene ranks were fitted to a sigmoid curve and the slope at the first elbow point was calculated as the minimum threshold for genes to be retained in a given GEP. The same slope was applied to every GEP to prevent bias in ranked gene selection, as the gene rankings between GEPs are not comparable and are relative to each GEP (as depicted in Figure S17). Cells from blood samples were assigned to the GEP with the highest usage (as provided by cNMF), to display an alluvial plot with ggalluvial in Figure 4D (v0.12.5).¹⁰⁷

Scoring of gene signatures

Gene signatures were scored on our Seurat object, or on other dataset’s Seurat or AnnData objects using either the function AddModuleScore in Seurat, or scanpy.tl.score_genes in scanpy. In both cases, the score is computed as the average expression of all genes contained in the gene list, and subtracting the average expression of 100 control genes (randomly chosen to match the expression bins of the gene list). Gene signatures used throughout this manuscript and their source can be found in Table S2.

Gene regulatory network inference

To deduce gene regulatory networks, we employed pySCENIC from a pre-built singularity container, aertslab/pyscenic:0.12.1, a tool utilizing cis-regulatory motif analysis to identify potential transcription factors (TFs) that might govern a cluster of co-expressed genes within individual cells.⁴⁷ pySCENIC was run using the –mask-dropouts flag and a normalized enrichment score threshold of 2 to help mitigate the effects of the varying degrees of sparsity across the datasets we generated. The initial step involved generating modules composed of transcription factors and co-expressed genes using GRNboost2.¹⁰⁸ These modules were pruned to remove indirect targets that lacked significant enrichment for the corresponding TF motif within ±10 kb from the transcription starting site of the putative target (cisTarget). This process yielded a collection of transcription factor regulons. Considering the inherent stochasticity in gene regulatory network inference using GRNBoost2, each run of pySCENIC may yield different quantities of regulons, along with distinct target genes associated with each TF. To mitigate this variability, we performed 100 pySCENIC runs and retained regulons present in 100% of the runs. We also removed regulons that did not have at least 5 target genes defining the regulon activity. Due to the high degree of noise in target genes, we retained target genes that appeared within a regulon in at least 95% of the runs. Furthermore, each target gene also had to overlap with the union of all possible retained ranked gene expression targets across all GEPs generated from cNMF. To identify regulons that were specific to the underlying biology of our cell types and GEPs, we calculated the AUC scores using the R package AUCell, located in the pySCENIC container, for each regulon based on the pruned target gene list. A regulon was deemed specific to a defined cell population if at least 20% of the cells within the annotated population scored in the 90^th percentile of the overall AUC score for all cells.

Comparison of gene expression programs with gene signatures from the literature

To compare the genes characterizing each GEP with known peripheral blood T cell states defined in the literature (Figure S10), we obtained gene signatures identified from (1) differential expression (DE) analysis from bulk RNAseq between sorted naive, T_cm, T_em CD4⁺ and CD8⁺ T cell populations by Rose et al.³⁸; (2) DE genes between cell clusters defined from scRNAseq of naive and memory CD4⁺ T cells isolated from PBMCs by Cano-Gamez et al.³⁷; (3) DE genes between cell clusters defined from scRNAseq of blood immune cells by Terekhova et al.⁴⁰ (4) DE genes between cell clusters defined from scRNAseq of T cells across nine human tissues by Poon et al.³⁹ In the Rose dataset, we kept genes that defined their Figures 2E and 2H (adjusted p-value ≤%0.05). In the Cano-Gamez and Poon datasets, we kept DE genes with a minimum log fold-change of 0.25 (adjusted p-value threshold ≤%0.05 or 0.01, respectively). In the Terekhova dataset, we used the top 100 differentially expressed genes shared in their Table S5. We computed a weighted Jaccard Index (JI) between the gene lists derived from our GEPs and those from the Rose, Cano-Gamez, Terekhova and Poon datasets (see quantification and statistical analysis). For the co-expression analysis of GEPs and gene lists from other datasets, we scored the gene lists on the entire integrated dataset. This was done using functions like Seurat’s AddModuleScore with the blend = TRUE parameter. Additionally, GEP4, GEP5, and GEP6 were scored on the Poon et al. dataset using scanpy’s tl.score_genes function, and their scores in specific cell clusters of interest were displayed in Figure S10B.

Pseudo-bulk differential expression analysis

To investigate for differentially expressed genes between αβ thymocytes lineages in Figure 2, we grouped cells by donor and lineage, restricting our analysis to only donors 2–4 where all αβ lineages were sorted, and also restricting ourselves to thymocytes found in clusters 3, 9, 10 and 11 in the integrated dataset (where CD4⁺/iNKT and CD8⁺/MAIT cells showed high gene expression similarity). We then used DESeq2 (v1.40.2)¹⁰⁹ to perform pseudo-bulk DE analysis between lineages using a Wald test. We restricted our analysis to genes which were significantly upregulated (p_adj < 0.01) in both CD4⁺ and iNKT cells compared to CD8⁺ and MAIT cells, and likewise for other comparisons shown in Figure 2. To investigate for cell lineage-specific gene signatures in PBMCs in Figure 4, we grouped cells by batch, cluster and lineage, restricting our analysis to only batches E, F and I where at least 3 or more cell lineages were sorted and sequenced within the same batch. We then used DESeq2 (v1.34.0)¹⁰⁹ to perform pseudo-bulk DE analysis with a likelihood-ratio test (LRT), where the full model included batch + cluster + lineage, and the reduced model included batch + cluster, in order to detect genes whose expression can be explained by lineage, and not by batch or cell state (i.e., cluster). We used the LRT test by computing pairwise comparisons, contrasting all lineages against each other (CD4vsCD8, CD4vsiNKT, etc.), for each comparison keeping DE genes with an adjusted p-value of 0.01. Then, to extract lineage-specific genes, for each lineage we kept genes that were commonly upregulated in at least 3 or more contrasts. We normalized the raw counts with the rlog function from DESeq2 and batch-corrected them with removeBatchEffect from the limma package (v3.56.2),¹¹⁰ before displaying the final list of DE genes on a heatmap with the pheatmap package (v1.0.12).

Creation of a reference scRNAseq mouse T_inn dataset

ScRNAseq data from mouse thymic iNKT, MAIT and γδ T cells^51–59 were downloaded from Gene Expression Omnibus or the European Bioinformatics Institute (see key resources table). Data were analyzed using the Seurat package. Analyzed cells were selected to express between 800 and 4200 genes per cell, with less than 5% of mitochondrial reads. Datasets were merged and integrated using the FastMNN algorithm,¹¹¹ using 5000 variable features, k = 20 and auto.merge = TRUE. Cell clustering was carried out with a resolution parameter set at 0.5, and potential doublets were detected using scDblFinder¹¹² and subsequently eliminated. To discern differential gene expression between clusters, the FindAllMarkers function was employed, utilizing the MAST algorithm. The analysis considered latent features, specifically the number of genes per cell, and the sample identity, with a log2 fold change threshold of 0.3.

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Human PBS57 CD1d tetramer PE	NIH tetramer core facility	N/A
Human 5-OP-RU MR1 tetramer PE	NIH tetramer core facility	N/A
Anti-human CCR7 - APC-Fire810 (clone G043H7)	Biolegend	Cat#353263; RRID: AB_2894483
Anti-human CD1d - PE (clone 51.1)	Biolegend	Cat#350305; RRID: AB_10642028
Anti-human CD3 purified (clone UCHT1)	Biolegend	Cat#300402; RRID: AB_314056
Anti-human CD3 - AF488 (clone OKT3)	Biolegend	Cat#317310; RRID: AB_571877
Anti-human CD3 - BUV496 (clone UCHT1)	BD Biosciences	Cat#612941; RRID: AB_2916883
Anti-human CD4 purified (clone RPA-T4)	Biolegend	Cat#300570; RRID: AB_2810427
Anti-human CD4 - AF488 (clone OKT4)	Thermo Fisher Scientific	Cat#53-0048-42; RRID: AB_10735503
Anti-human CD4 - redFluor710 (clone OKT4)	Tonbo biosciences	Cat#80-0048; RRID: AB_2621976
Anti-human CD4 - BV570 (clone RPA-T4)	Biolegend	Cat# 300533; RRID: AB_10896788
Anti-human CD8ɑ purified (clone RPA-T8)	Biolegend	Cat#301002; RRID: AB_314120
Anti-human CD8ɑ - APC (clone RPA-T8)	Thermo Fisher Scientific	Cat#17-0088-42; RRID: AB_10669564
Anti-human CD8ɑ - BUV395 (clone RPA-T8)	BD Biosciences	Cat#563795; RRID: AB_2722501
Anti-human CD8ɑ - PE-Cy7 (clone SK1)	Tonbo biosciences	Cat#60-0087; RRID: AB_3106994
Anti-human CD14 - eFluor450 (clone 61D3)	Thermo Fisher Scientific	Cat#48-0149-42; RRID: AB_1272050
Anti-human CD14 - PE-Cy5 (clone 61D3)	Thermo Fisher Scientific	Cat#15-0149-42; RRID: AB_2573058
Anti-human CD19 - eFluor450 (clone H1B19)	Thermo Fisher Scientific	Cat#48-0199-42; RRID: AB_1272053
Anti-human CD19 - PE-Cy5 (clone H1B19)	Thermo Fisher Scientific	Cat#15-0199-42; RRID: AB_10853658
Anti-human CD45 purified (clone HI30)	Biolegend	Cat#304002; RRID: AB_314390
Anti-human CD45 - AF647 (clone QA17A19)	Biolegend	Cat#393406; RRID: AB_2750083
Anti-human CD45 - BV421 (clone HI30)	Biolegend	Cat#304032; RRID: AB_2561357
Anti-human CD62L - BV650 (clone DREG-56)	Biolegend	Cat#304832; RRID: AB_2563821
Anti-human CD161 - BUV805 (clone HP-3G10)	BD Biosciences	Cat#749221; RRID: AB_2873599
Anti-human CD235a (clone HI264)	Biolegend	Cat#349102; RRID: AB_10612565
Anti-human Eomes - BUV737 (clone X4-83)	BD Horizon	Cat# 567170; RRID: AB_2916487
Anti-human EPCAM - BV421 (clone 9C4)	Biolegend	Cat#324220; RRID: AB_2563847
Anti-human GZMB - AF700 (clone GB11)	BD Biosciences	Cat#560213; RRID: AB_1645453
Anti-human GZMK - eFluor660 (clone G3H69)	Thermo Fisher Scientific	Cat#50-8897-42; RRID: AB_2574299
Anti-human HLADR - BV711 (clone L243)	Biolegend	Cat#307643; RRID: AB_11218794
Anti-human PLZF - PE-CF594 (clone R17-809)	BD Horizon	Cat#565738; RRID: AB_2739339
Anti-human T-bet - BV605 (clone 4B10)	Biolegend	Cat#644817; RRID: AB_11219388
Anti-human TCR γδ - BV650 (clone 11F2)	BD Biosciences	Cat#569510; RRID: AB_3106997
Anti-human Vɑ7.2 - BV785 (clone 3C10)	Biolegend	Cat#351721; RRID: AB_2566041
Anti-human Vɑ24 - PerCP-Cy5.5 (clone C15)	Biolegend	Cat#360004; RRID: AB_2562495
Anti-human Vδ1 - PerCP-Vio700 (clone REA173)	Miltenyi Biotec	Cat#130-120-441; RRID: AB_2784469
Anti-human Vδ2 - FITC (clone 123R3)	Miltenyi Biotec	Cat#130-125-095; RRID: AB_2819745
Anti-human Vγ9 - PE (clone B3)	BD Biosciences	Cat#555733; RRID: AB_396076
CDR2 purified98	Dr. Sheena Pinto	N/A
Anti-mouse CD1d - PE (clone 1B1)	Biolegend	Cat#123509; RRID: AB_1236547
Anti-mouse CD4 - AF488 (clone GK1.5)	Biolegend	Cat#100423; RRID: AB_389302
Anti-mouse CD8ɑ - APC (clone 53-6.7)	Biolegend	Cat#100711; RRID: AB_312750
Anti-mouse CD45 (clone 30-F11)	Thermo Fisher Scientific	Cat#14-0451-85; RRID: AB_467252
Anti-mouse CD45 - BV605 (clone 30-F11)	Biolegend	Cat#103139; RRID: AB_2562341
Anti-mouse CD90.2 purified (clone 53-2.1)	Biolegend	Cat#140302; RRID: AB_10641692
Anti-mouse EpCAM - BV421 (clone G8.8)	Biolegend	Cat#118225; RRID: AB_2563983
Goat anti-mouse IgG	Vector Laboratories	Cat#AI-9200-1.5; RRID: AB_3107016
Anti-mouse Ly51 - AF647 (clone 6C3)	Biolegend	Cat#108312; RRID: AB_2099613
Goat anti-rat IgG	Vector Laboratories	Cat#BA-9400-1.5; RRID: AB_3107017
Human Fc receptor blocking FcγR block	Miltenyi Biotec	Cat# 130-059-901; RRID: AB_2892112
Human Fc receptor blocking TruStain FcX	Biolegend	Cat#422302; RRID: AB_2818986
Mouse Fc receptor blocking anti-CD16/32 (clone 93)	Thermo Fisher Scientific	Cat#14-0161-85; RRID: AB_467134
UEA1 - FITC	Vector Laboratories	Cat#FL-1061; RRID: AB_2336767
Human Single Cell Sample Multiplexing Kit	BD Biosciences	Cat# 633781; RRID: AB_2870299
Biological samples
Healthy human whole blood (Human Immune Tissue Network Biobank); COMIRB protocol #17-2159	Colorado Anschutz Medical Campus Clinical and Translational Sciences Institute, Aurora (USA)	N/A
Healthy human blood (from plateletpheresis leukoreduction filter chambers); Vitalant	Vitalant Blood Donation Center, Denver (USA)	N/A
Human pediatric thymic tissue; IRB protocol #20-0150	Children's Hospital Colorado, Aurora (USA)	N/A
Human pediatric thymic tissue; IRB protocol #IRB-23-4	Mount Sinai, New York City (USA)	N/A
Human pediatric thymic tissue; protocol #NHBR2101	Northwell Health Biospecimen Repository, Lake Success (USA)	N/A
Chemicals, peptides, and recombinant proteins
Bovine Serum Albumin (BSA)	ThermoFisher Scientific	Cat#J10857-22
Collagenase type IV	StemCell Technologies	Cat#07427
Dispase I	Sigma-Aldrich	Cat#04942086001
DMSO	Sigma-Aldrich	Cat#D8418-100ML
DNase I	Sigma-Aldrich	Cat#11284932001
Fixable Viability Dye efluor780	ThermoFisher Scientific	Cat#65-0865-14
Fetal Bovine Serum (FBS)	Corning	Cat#35-010-CV
GlutaMAX	Gibco	Cat#35050-061
Liberase TH	Sigma-Aldrich	Cat#5401135001
2-Mercaptoethanol	Sigma-Aldrich	Cat#21985-023
2-Mercaptoethanol	Gibco	Cat#21985023
Penicillin/streptomycin	Gibco	Cat#15140-122
Non-essential amino acids	Sigma-Aldrich	Cat#11140-050
Sodium pyruvate	Sigma-Aldrich	Cat#11360-070
Ficoll-paque Premium	Cytiva	Cat#17544203
Critical commercial assays
AF488 antibody labeling kit	ThermoFisher Scientific	Cat#A20181
Anti-PE microbeads	Miltenyi	Cat#130-048-801
BD Transcription Factor Buffer Set	BD Biosciences	Cat#562574
Live/dead Fixable Near-IR dead cell stain kit	ThermoFisher Scientific	Cat#L10119
BD Rhapsody Cartridge Kit	BD Biosciences	Cat#633773
BD Rhapsody Enhanced Cartridge Kit	BD Biosciences	Cat#664887
BD Rhapsody Cartridge Reagent Kit	BD Biosciences	Cat#633773
BD Rhapsody cDNA Kit	BD Biosciences	Cat#633773
BD Rhapsody TCR/BCR Amplification Kit	BD Biosciences	Cat#665345
BD Rhapsody WTA Amplification Kit	BD Biosciences	Cat#633801
Deposited data
Raw data (fastq files) Preprocessed Seurat object	This study	GEO: GSE249684
Web browser exploration tool (ShinyCell) of human and murine integrated datasets	This study	https://xspeciestcells.cshl.edu/
Code	This study	https://doi.org/10.5281/zenodo.13737916. https://doi.org/10.5281/zenodo.13251142
Experimental models: organisms/strains
C57BL/6J	Jackson Laboratories	Strain #000664; RRID:IMSR_JAX:000664
C57BL/6J CD1d1d2-/-	Bred in house	Chen et al.99
Software and algorithms
Flow cytometry software FlowJo (v10.7.1)	BD Biosciences	RRID: SCR_008520
Flow cytometry software SpectroFlo (v3.0)	Cytek Biosciences	RRID: SCR_019826
FastQC (v0.11.9)	https://github.com/s-andrews/FastQC	RRID: SCR_014583
MultiQC (v1.12)	https://github.com/MultiQC/MultiQC	RRID: SCR_014982
cNMF (v1.5)	https://github.com/dylkot/cNMF	RRID: SCR_025495
Python (version ≥3.8)	https://www.python.org/	RRID: SCR_008394
R (version ≥4.0.3)	https://www.r-project.org/	RRID: SCR_001905
Other
Public dataset - peripheral blood CD4+ T cell gene signatures (Cano-Gamez et al.)	Cano-Gamez et al.37	Supplementary Data 6
Public dataset - peripheral blood T cell gene signatures (Rose et al.)	Rose et al.38	Supplementary Data 2, Supplementary Data 3
Public dataset - peripheral blood T cell gene signatures (Terekhova et al.)	Terekhova et al.40	Table S5
Public dataset - T cell gene signatures across human tissues (Poon et al.)	Poon et al.39	Table S6
Public dataset - scRNAseq of human thymus (Park et al.)	Park et al.27	https://datasets.cellxgene.cziscience.com/de8665a2-0476-4865-b4af-c7b8d3b1b87f.h5ad
Public dataset - scRNAseq of murine iNKT thymocytes (Baranek et al.)	Baranek et al.54	https://datasets.cellxgene.cziscience.com/de8665a2-0476-4865-b4af-c7b8d3b1b87f.h5ad
Public dataset - scRNAseq of murine iNKT thymocytes (Krovi et al.)	Krovi et al.55	GEO: GSE152786
Public dataset - scRNAseq of murine iNKT thymocytes (Maas-Bauer et al.)	Maas-Bauer et al.56	GEO: GSE172169
Public dataset - scRNAseq of murine iNKT thymocytes (Wang et al.)	Wang et al.57	GEO: GSE130184
Public dataset - scRNAseq of murine iNKT, MAIT, γδ thymocytes (Lee et al.)	Lee et al.51	SRA: PRJNA549112
Public dataset - scRNAseq of murine MAIT thymocytes (Chandra et al.)	Chandra et al.59	GEO: GSE189484
Public dataset - scRNAseq of murine MAIT thymocytes (Koay et al.)	Koay et al.53	GEO: GSE137350
Public dataset - scRNAseq of murine MAIT thymocytes (Legoux et al.)	Legoux et al.52	EMBL-EBI: E-MTAB-7704
Public dataset - scRNAseq of murine γδ thymocytes (Li et al.)	Li et al.58	GEO: GSE179422
HBSS	Gibco	Cat#14175079
HEPES (1M)	Gibco	Cat#15630080
RPMI 1640	Corning	Cat#10-040-CV
RPMI 1640 with 25mM HEPES	Gibco	Cat#22400071
RPMI 1640 without phenol red	Gibco	Cat#11835030
TrypLE Express Enzyme	Gibco	Cat#12604013

QUANTIFICATION AND STATISTICAL ANALYSIS

Pseudo-bulk differential expression analysis in thymic samples

Differentially expressed genes in Figures 2B and 2C between αβ thymocyte lineages were identified using a Wald test from the DE-Seq2 package (v1.40.2),¹⁰⁹ with a p_adj < 0.01 threshold. This analysis was restricted only to thymocytes from three sequencing batches (B, C, D; see Table S1) where all αβ lineages were sorted; and to thymocytes in clusters 3, 9, 10 and 11 where naive-like CD4⁺/iNKT and CD8⁺/MAIT had shown transcriptional similarity (based on Figures 1G, 1I, and 2A).

Pseudo-bulk differential expression analysis in PBMC samples

Differential expression analysis in Figure 4 was performed with an LRT analysis, to take into account the heterogeneity of batches and cell states among peripheral blood T cells (see method details). This analysis was restricted only to PBMCs from three sequencing batches (E, F, I; see Table S1), where at least three distinct T cell lineages were sorted and sequenced in the same batch. Only DE genes with p_adj < 0.01 were kept.

Comparison of GEPs with gene signatures from the literature

In Figure S10, we computed the Jaccard Index (JI) between the gene lists derived from our GEPs and those from the Rose, Cano-Gamez, Terekhova and Poon datasets. Since the gene lists varied in length, we weighted the JI to make it comparable across pairwise comparisons. This was achieved by dividing the JI by the maximal theoretical JI for each pairwise comparison, which is the ratio of the length of the smaller list to the length of the larger list. This “weighted Jaccard Index” is shown on the x axis of Figure S10A. To assess the significance of the observed JI, we performed a permutation analysis. We generated 1000 random gene lists A′ and B′, matching in length and expression pattern to the original lists A and B. We computed the weighted JI between these random lists and defined an empirical p-value by counting how many of these weighted JIs were greater than the observed weighted JI divided by the number of permutations. To account for multiple comparisons, we applied a Bonferroni correction to the empirical p-values. Adjusted p-values below 0.01 are shown as text in the figure.

Highlights.

• Comprehensive atlas of human T_inn cells in thymus and blood

• Most postnatal thymic T_inn cells resemble naive T_conv cells

• Adult blood T_inn cells display a mixed type 1/type 17 effector program

• Human T_inn cells do not form distinct subsets like mouse T_inn cells