A series of scRNAseq datasets from ~150 healthy individuals and cancer patients’ samples were downloaded from Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo) integrated and preprocessed as one single reference dataset as depicted [2]. Briefly, the sequencing output bcl2 files were converted to FASTQ format by using cellranger-mkfastq^TM algorithm (10× Genomics), and cellranger-count was applied to align to the GRCh38 reference transcriptome and build the (cell, UMI) expression matrix for each sample. Preprocessing, normalization of UMI, and QC were assessed using the R package Seurat 3.0 [11]. First, genes that correspond to human leukocyte antigens (HLA) and genes located in chromosomes X and Y were removed, and the quality of cells was then determined by the number of detected genes per cell and the proportion of mitochondrial genes. The datasets were normalized together and variable features from each dataset were identified by FinVariable-Data before identification of anchors correspondences across all the datasets, to yield an integrated dataset. Principal components analysis (PCA) was then performed on this integrated dataset before dimensionality reduction by uniform manifold approximation and projection (UMAP) [12]. All its single cells were then scored for various multigene signatures, to further perform the digital extraction of γδ T and of CD8 T lymphocytes by score and gate [2]. An unbiased differentiation trajectory of these γδ T lymphocytes was computed without prior information by a minimum spanning tree (MST) algorithm to determine each γδ T cell’s position, pseudotime, and milestone. The reference γδ T cells matrix (cell, gene, pseudotime, and other hallmarks) is published elsewhere [2]. This reference γδ T cell trajectory was visualized as a pseudotimed plot featuring every single cell by its pseudotime (x-axis) vs. its MST1 and MST2 dimensions projection (y-axis) [2]. For clarity, this reference trajectory is shown as a grey cells background onto which additional γδ T cells, here from the CTCL patient, were mapped. The same pipeline was applied for CD8 T cells whenever mentioned in the text.

In parallel, the γδ T and CD8 T cells were extracted likewise from the CTCL patient MF318 scRNAseq dataset, downloaded from the publicly available GEO GSE173205 [10]. This dataset was separately pre-processed by using the same Seurat 3.0 pipeline, and its single cells were scored for multigene signatures to further digitally extract its γδ T and CD8 T lymphocytes by score and gate as above. Before injection into the reference dataset (see below), the γδ T and CD8 T lymphocytes extracted from this CTCL dataset are referred to as ‘ab extra’ cells.

Since the differentiation trajectory of N cells is computed from their N transcriptomes, the resulting trajectory varies with N. As a consequence, adding any newer dataset of ab extra cells (e.g., ‘M cells’) to a previous ‘N cells’ dataset creates a newer (‘N + M cells’) dataset which trajectory differs from the former one. Here, we aimed at freezing the ‘N cells’ trajectory such as to keep it as a stable reference when adding ab extra cells and identify their hallmarks by their position on this reference.

We incepted and applied ‘single-cell injector’, a two-step procedure based on Seurat’s anchoring integration function [11] (Figure S1). The single-cell injector integrates as described [11] a new dataset (named ‘Query’) into a reference dataset (named ‘reference’) in which cells coordinates are already annotated (cells, cluster, t-SNE, UMAP, trajectory, milestone, pseudotime). The resulting dataset (‘Query 2’) is then used by Seurat’s function MapQuery to anchor Query 2 in the reference as already published [11]. Because pseudotime and milestones are already known in the reference, the N ‘auto’ anchors of the reference’s N cells to themselves allow to precisely determine the pseudotime and milestone attributes of each Query’s single cell, without change to the reference trajectory. The injector procedure was validated using the γδ T cell reference dataset since it comprises cells whose MST1, MST2, milestones, and pseudotime are already known. So, n = 7,449 γδ T cells randomly selected from this reference were used as a Query, and this Query was injected into the reference by the depicted procedure. For each Query cell, the injected position and initial position were compared, using as a metric the Euclidian distance (Δ) between the injected and the initial (MST1, MST2) coordinates. With maximally wrong injections such tests could theoretically lead to Δ _max = 91. However, the distribution of each Query cell’s Δ revealed 87% perfect injection (n = 6,512 cells injected at Δ = 0) and 13% injected close to the expected position (n = 935 cells injected at Δ < 15, all of which for cells from the M5–M8 branching milestones) (Figure S1b). These results validated the ab extra cell injector procedure, so the CTCL derived γδ T lymphocytes were injected likewise in the reference and their respective pseudotimes and milestones were determined accordingly (Tables S1 and S2). When mentioned in the text, the same pipeline was applied to the CTCL-derived CD8 T cells.

Pathway enrichment analysis was performed using Single-Cell Signature Explorer function scorer [13] to score all the signatures from the MSIgDB (http://www.gsea-msigdb.org) collections Kyoto Encyclopedia of Genes and Genomes (KEGG), GO-MF, GO-BP, Reactome, hallmarks, C6, and C7. These genesets were scored in each single γδ T cell of the specified subset from the reference’s healthy donors [2] and from the patch or plaque samples. Each score was compared between the specified CTCL and healthy cell groups by using unpaired Student’s t and Bonferroni correction false discovery rate (FDR). Plots of representative differential pathways were produced by using the R package pathfindR.

Cell classifications of differentiation stage were based on their respective milestones as Tn (M1, M2), Tcm (M3–7), Tem (M8–13), and Temra (M13, 14) (Figure 1). Since γδ T lymphocytes express either the TRGC1-encoded TCRVγ9 or the TRGC2-encoded TCRVγnon9 in a mutually exclusive fashion, subtyping of the TCRVγ9 and TCRVγnon9 γδ T cells was performed by plots of compensated scores of the signatures (‘TRDC, TRGC1’ positive cells) vs. (‘TRDC, TRGC2’ positive cells) [2]. The cell classifications as tissue resident memory T cell (Ttrm)/recirculating (non-Ttrm) and exhausted T cell (Tex)/functional (non-Tex) were performed by ‘at least one binary’ as follows. The single cells were scored for several reference gene signatures (Table S1). For Tex classification, the reference Tex signatures were as published [14–18]. For each signature, a cut-off was defined as the maximal score of the (n = 3,680) control γδ T or CD8 T lymphocytes extracted from healthy adults peripheral blood mononuclear cell (PBMC). This threshold defined the cell’s binary (1 for cell score > threshold, 0 otherwise). The Tex cut-offs were: 3.9 for ‘Chihara_IL27_Coinhib_module’ [14]; 0.52 for ‘Alfei_ d20_tox’ [15]; 0.22 for ‘Khan_Tox_OverExpressed_genes’ [16]; 0.5 for ‘Tosolini_ NHL_IEGS33’ [17]; and 0.16 for ‘Balanca_QP_genes’ [18]. For each tumor-infiltrating lymphocyte (TIL), the five Tex binaries were summed, and any TIL with a non-zero-sum of binaries was classified Tex, or non-Tex otherwise. The Ttrm classification was applied likewise using Ttrm signatures from the literature [19, 20]. The binarizing cutoffs established with the control CD8 T lymphocytes were: 0.46 for ‘Kumar_13g_Ttrm’; 0.3 for ‘Kumar_3g_Ttrm’ [19]; 3.72 for ‘Wu_Tcellcluster4.1_trm’; 2.6 for ‘Wu_Tcellcluster8.3_trm’; 2.8 for ‘Wu_Tcellcluster8.3b_trm’; 4.2 for ‘Wu_ Tcellcluster8.3c_trm’ [20]. These Ttrm binaries were summed, and any TIL with a non-zero-sum of Ttrm binaries was classified Ttrm, or non-Ttrm (recirculating) otherwise.

The γδ T lymphoma cells of CTCL type were sourced from the public scRNAseq dataset of CTCL patient MF318 (GEO GSE173205, Methods), a 55 years old Caucasian female patient diagnosed with stage IIB CD30⁺ γδ MF without large cell transformation [10]. Its γδ T lymphocytes were digitally purified using score and gate by successively discarding the keratinocytes, melanocytes, myeloid, B, natural killer (NK), CD4 T, and finally CD8 T cells from the dataset. A total of 1,220 γδ T cells were obtained. We previously produced a so-called ‘public’ differentiation trajectory of γδ T cells from > 150 individuals’ blood and tissue samples [2]. Using the single-cell injector procedure (Methods, Figure S1), the (n = 1,220) γδ T cells from the CTCL patient MF318 were injected into this reference. With ~20,000 cells distributed in 14 milestones from Tn to Temra cells, this trajectory represents the most exhaustive and resolutive map of γδ T differentiation so far (Figure 1a). Together, these milestones still form a consistent and coherent continuum when the integrated dataset is represented by an alternative computational method such as UMAP (Figure 1b). Also, visualization in this UMAP of the trajectory pseudotime validated the previous sequence of differentiation milestones (Figure 1b). In UMAP, the γδ T cells from the CTCL patient MF318 were embedded in all areas of the dataset, confirming their transcriptomic similarity to the reference γδ T cells. Also, these cells encompassed both TCRVγ9 and TCRVγnon9 γδ T cell subsets (Figure 1d). The expression profile of a series of published marker genes [2, 21] in each milestone showed smooth profile transitions across the trajectory (Figure 1e). Based on expression intensity and percentage of expressing cells, the Tn marker genes typically encompass CCR7, TCF7, LTB, and SELL; Tcm genes include ANXA1, PABPC1, IL7R, H3F3B, CD69, LYAR, CXCR4, KLRB1, NR4A2 and FOS, Tem genes comprise S100A6, KLRD1, NKG7, CST7, PFN1, GZMM, GZMB, DUSP2, FTH1, CCL4, and IL2RG, and Temra genes include FCGR3A, S100A4, PRF1, GZMA, FGFBP2, GZMH, and KLRK1. These profiles confirmed earlier reports [2, 21] depicting a progressive differentiation of type-1 cytotoxic functions along Tem and Temra stages, with the Temra cells detected here in both M13 and M14 instead of M14 only [2]. Few innate-like γδ T cells were present in this integrated reference dataset, as KLRB1 (encoding for CD161) was expressed by most of the late Tcm (M7) cells but KLRC1 (encoding for NKG2A) and TBX1 (encoding for PLZF1) by very few Tem (M13) cells. Fetal differentiation-derived type-3 effector γδ T cells revealed by their CCR6, RORC, IL23R, and DPP4 marker genes [21] were not abundant in the integrated reference, though clearly represented in most milestones (not shown).

Given the correspondence of the differentiation trajectory and UMAP visualizations of the integrated dataset, we investigated further the γδ T cells from the CTCL patient MF318 with the higher resolution of pseudotimed trajectory maps.

Among the variants of CTCL, MF is depicted as a malignancy of skin-resident Tem cells, in contrast to leukemic CTCL which is composed of skin-recirculating malignant Tcm cells [9]. So to determine precisely the differentiation of the skin γδ T cells from patient MF318, the (n = 1,220) γδ T cells were analyzed by injection in the reference γδ T pseudotimed trajectory. In parallel controls, (n = 238) CD8 T lymphocytes extracted from the same CTCL dataset were injected into their respective CD8 T reference differentiation trajectory [2]. These CD8 T cells encompassed a few Tn, Tcm, and early Tem cells, but no Temra (Figure 2). In contrast, the γδ T cells displayed all stages of differentiation. When cells were stratified by patch or plaque origin, the CD8 T cells differentiation was similar in both cases. In contrast, the γδ T cells from the patch were only Tn and early Tcm cells, while in the plaque, they comprised a few late Tcm and mostly Tem cells (Figure 2).

These analyses unveiled a γδ T cell-selective differentiation from Tcm to Tem upon the patch-to-plaque progression of patient MF318’s skin lesions.

By scoring various gene signatures and applying a classifier based on these scores (Methods), the above γδ T cells were characterized for TCRVγ9 or TCRVγnon9 subset, tissue-residency or recirculating, and exhaustion or functional cell states. These distinct features were then visualized simultaneously on the mapped γδ T cells (Figure 3a). This showed that the TCRVγnon9 subset (n = 1,189 TCRVγnon9 cells vs. n = 31 TCRvγ9 cells) was largely prominent among the skin γδ T cells (Figure 3). Unexpectedly however, these TCRVγnon9 lymphocytes were in majority recirculating (n = 723 recirculating, n = 466 Ttrm), and their (n = 167) Tex cells were more often Ttrm (n = 116 Tex; 25% of the Ttrm) than recirculating lymphocytes (n = 51 Tex; 7% of the recirculating). The fewer TCRvγ9 cells comprised (n = 19) recirculating cells including (n = 2) Tex and (n = 12) functional Ttrm.

There were evident changes during the patch-to-plaque progression concerning cell counts: the TCRVγ9 cells remained unchanged and scarce in both samples (n = 13 cells in patch, n = 18 in plaque), while the TCRVγnon9 cells strongly increased (n = 471 cells in patch, n = 718 cells in plaque). This was a doubling of the cell counts for recirculating cells (patch: n = 247 Ttrm and n = 224 recirculating, plaque: n = 219 Ttrm and n = 499 recirculating). In the patch, the recirculating compartment of TCRVγnon9 lymphocytes comprised 18% Tn, 62% Tcm (n = 139 cells), 20% Tem, and 0% Temra cells, while in the plaque it comprised 0% Tn, 7% Tcm, 86% Tem (n = 431 cells), and 7% Temra (Figure 3b).

Thus the patch-to-plaque progression of the γδ CTCL patient MF318 involved a Tcm proliferation and Tem differentiation of the TCRVγnon9 subset of skin γδ T lymphocytes.

To further determine cellular hallmarks of the plaque γδ T lymphocytes, we compared various pathway signatures between the circulating TCRVγnon9 γδ T Tem cells (milestones M8–M12) and in their corresponding control cells (n = 1,431) from reference’s healthy donors [1].

Several signatures appeared significantly enriched in the plaque Tem cells (Figure 4). Their most upregulated encompassed T cell activation signatures (graft-versus-host disease, type 1 diabetes, allograft rejection, autoimmune thyroid disease) and immune escape to cell cytotoxicity (protection from NK cell lysis, protein tyrosine kinase inhibition). Asthma and immunoglobulin E (IgE) binding corresponding to an allergic context were enriched in the plaque γδ T cells, suggesting that some contact dermatitis could underlie this progression of lesions [22]. The ‘viral ribonucleoprotein (RNP) complexes in the host cell nucleus’ signature were also significantly upregulated which, together with significantly downregulated cytosolic DNA sensing, toll-like receptors (TLR), cytokine, and chemokine signalings, supported the hypothesis of a viral infection in CTCL [23]. By contrast, activities of several transcription factors [targets of early 2 factor (E2F), MYC, and the polycomb repressive complex 2-enhancer of zeste homolog 2 (PRC2-EZH2) and PRC1-BMI1] and all the main energetic pathways [glycolysis, oxidative phosphorylation, fatty acid (FA) metabolism] were significantly upregulated in the Tem TCRVγnon9 γδ T cells from the plaque (Figure 4 right panel). Further comparisons of the same lymphocytes from patches to their respective controls unveiled that circulating Tem TCRVγnon9 γδ T cells with the above signatures were less abundant but actually present in the earlier patches (Figure 4 left panel).

So the TCRVγnon9 γδ T cell proliferation and differentiation-associated with the patch-plaque switch of patient MF318 involved a global evolution of their transcriptome profile.

During the differentiation of control γδ T lymphocytes from healthy individuals, there is no flagrant change in the expression level of the main energetic pathways (Figure S2). So here, we aimed at characterizing more precisely those skin recirculating Tem TCRVγnon9 γδ T cells displaying increased levels of the above signatures. Plotting all the patient’s γδ T cells for both energetic pathways (sum of glycolysis, OXPHOS, and FA metabolism scores) vs. transcription factor activity (E2F targets) confirmed that some cells, more numerous in the plaque, had simultaneously up-regulated both signatures (Figure 5a). When selected for a focused analysis, these (n = 342) cells comprised 98% of TCRVγnon9 lymphocytes (Figure 5b). Their positioning on trajectories characterized a majority of Tcm cells in the patch and of effector cells in the plaque (Figure 5c). However, rare (n = 8) Tn cells from the patch also displayed this activated phenotype (Figure 5c). Finally, mapping the oxidative phosphorylation levels in ungated γδ T cells from healthy controls and MF318 patch and plaque (not shown) confirmed that this activated phenotype of plaque TCRVγnon9 cells was not the mere consequence of normal γδ T differentiation, consistent with their malignant state.

Hence, the MF318 patient’s skin lesions encompass a sizeable subset of recirculating skin TCRVγnon9 γδ T cells whose energetic and transcriptional profiles were above those of normal cells all along with their Tn, Tcm, and Tem differentiation, consistent with the malignant status of these MF skin-derived γδ T cells [10].