Circulating Tumor Reactive KIR+CD8+ T cells Suppress Anti-Tumor Immunity in Patients with Melanoma

Effective anti-tumor immunity is largely driven by cytotoxic CD8+ T cells that can specifically recognize tumor antigens. However, the factors which ultimately dictate successful tumor rejection remain poorly understood. Here we identify a subpopulation of CD8+ T cells which are tumor antigen-specific in patients with melanoma but resemble KIR+CD8+ T cells with a regulatory function (Tregs). These tumor antigen-specific KIR+CD8+ T cells are detectable in both the tumor and the blood, and higher levels of this population are associated with worse overall survival. Our findings therefore suggest that KIR+CD8+ Tregs are tumor antigen-specific but uniquely suppress anti-tumor immunity in patients with melanoma.

and/or stockholders of Repertoire Immune Medicines.SH, FJO, and JK are employees of Repertoire Immune Medicines (Switzerland) AG, formerly Tepthera Ltd.All other authors report no competing interests.

Introduction
Over the past decade, therapies modulating the immune system's ability to identify and lyse cancer cells have revolutionized cancer care.Monoclonal antibodies targeting negative signaling receptors on immune cells -termed checkpoint inhibitors -to break tolerance on functionally exhausted tumor antigen-speci c T cells have improved survival across multiple cancer types, as exempli ed by the durable clinical responses seen in > 50% of patients with advanced melanoma treated with combination anti-PD-1 and anti-CTLA-4 therapy 1 .However, the cellular and environmental elements dictating the success of anti-tumor immunity are complex and remain poorly understood, which limits the development of novel therapies.The presence of certain cell types, such as tumor antigen-speci c CD8 + T cells and CD4 + regulatory T cells, are critical cellular elements to sustain or impair anti-tumor immune responses, respectively.However, how features of tumor antigen-speci c T cells relate to circulating T cells and the clinical relevance of either population remain poorly characterized.
Here, we used single-cell RNA sequencing paired with T cell receptor (TCR) sequencing, multi-parametric ow cytometry, and high-throughput peptide screens to characterize tumor antigen-speci c T cells in resected tumors and blood from patients with melanoma.Combining a transcriptional-based approach to predict tumor antigen reactivity with experimental con rmation of antigen speci city, we identify a subpopulation of tumor antigen-speci c CD8 + T cells which are shared between tumor and blood and resemble KIR + CD8 + regulatory T cells (Tregs) recently described to have a role in regulating autoin ammatory responses 2,3 .Moreover, the frequencies of circulating KIR + CD8 + Tregs in patients with melanoma correlate with worse overall survival.These results provide new evidence for a role of CD8 + Tregs in regulating human immune responses that may have clinical importance as an alternative mechanism of tumor immune evasion.

Circulating tumor antigen reactive T cells are predominantly clonally expanded, cytotoxic CD8 + T cells
To understand the transcriptional and clonal relatedness of tumor antigen reactive T cells in the tumor and blood, we performed droplet-based, single-cell RNA sequencing with paired TCR sequencing in sorted immune cells from resected tumors and blood of 17 patients with immunotherapy-naive, cutaneous melanoma (Fig. 1A, table S1).In total, 78,840 tumor and 124,329 blood T cells were analyzed, of which 42,094 (53.4%) and 66,114 (53.2%) had paired TCR α-and β-chains in the tumor and blood, respectively.
We used a stringent de nition for clonal relatedness, whereby clonally related T cells were de ned as those with identical single CDR3α-and single CDR3β-chain pairs based on amino acid sequences.
Using previously described, functionally validated gene signatures (neoTCR8 and neoTCR4 for CD8 + and CD4 + T cells, respectively) 4 , we then performed single-cell gene set enrichment analysis to predict the tumor antigen reactivity of T cells in the tumor (predicted-reactive T cells).We validated the accuracy of this transcriptional-based classi cation using external datasets annotated with experimentally-observed tumor antigen reactivity ( g. S1) 5,6 .Using TCR sequences to link clonally related T cells, we identi ed 7,183 CD8 + T cells (17.8% of total CD8 + ) and 145 CD4 + T cells (0.71% of total CD4 + ) in the blood which were clonally related to predicted-reactive T cells in the tumor (Fig. 1B).Consistent with prior work, we noted a wide range in the frequency of circulating predicted-reactive T cells across patients (CD8 + T cells 0-50.6%,CD4 + T cells 0-4.28%) ( g. S2) 4,[7][8][9][10][11] .We additionally identi ed 16,387 circulating T cells clonally related to predicted-unreactive tumor T cells.
To quantify the degree of clonal expansion, we calculated the proportion of the total TCR repertoire occupied by each clonotype (Expansion Score) and observed that predicted-reactive CD8 + T cells in the blood were more largely expanded than predicted-unreactive T cells (Fig. 1F).Consistent with a higher degree of clonal expansion, circulating predicted-reactive T cells also had more restricted TCR repertoires (Fig. 1G, g.S4).Using a publicly available database of TCR sequences annotated with antigen reactivity (VDJdb 13 ), we observed that 1.5% of predicted-reactive and no predicted-unreactive clonotypes in the blood matched TCR αβ amino acid sequences previously reported to be tumor antigen-speci c (Fig. 1H).Our ndings are thereby consistent with prior work showing that tumor antigen-speci c T cells in the blood are predominantly a clonally expanded, cytotoxic CD8 + T cell population 7,[10][11][12]14 .
A subpopulation of circulating, predicted-reactive CD8 + T cells resemble KIR + CD8 + Tregs To derive a transcriptional signature with greater speci city for predicted-reactive T cells in the blood, we focused our analyses on CD8 + T cells which could be found in both the tumor and blood (tumor matched).We performed local two-sample testing by random-walk distributions to identify subpopulations of cells differentially abundant for tumor matched, predicted-reactive or predictedunreactive CD8 + T cells (Fig. 2A) 15 .We identi ed a subpopulation of 3,051 cells enriched for predictedreactive CD8 + T cells.Consistent with previous studies 7 , circulating CD8 + T cells clonally related to tumor-in ltrating T cells had high expression of genes related to cytotoxicity (GNLY, GZMB, FGFBP2) and cell migration (ITGB1) regardless of predicted reactivity (Fig. 2B, table S5-6).However, the subpopulation of predicted-reactive CD8 + T cells had high expression of several natural killer cell-associated genes (NKG7, KLRD1, KLRC2, KLRC3, FCG3A), including genes encoding the inhibitory subset of killer cell immunoglobulin-like receptor (KIR) family (e.g.KIR2DL3).While KIR expression is known to be upregulated in terminally differentiated CD8 + T cells 16,17 , human KIR + CD8 + T cells were also recently demonstrated to have regulatory-like function (KIR + CD8 + Tregs) in the context of autoimmune diseases and viral infections 2,18,19 .Similar to the better characterized CD4 + Tregs, KIR + CD8 + Tregs limit the effector function of pro-in ammatory immune cells, though do so through contact-dependent cytolysis 2 .
Previous studies have also reported the presence of KIR + CD8 + Tregs in the context of tumor immune surveillance 3,20 , though the antigen speci city and clinical signi cance remain unclear.
To interrogate whether the identi ed subpopulation of circulating predicted-reactive KIR + CD8 + T cells may represent KIR + CD8 + Tregs, we examined the expression pattern of several hallmark genes (Fig. 2C).In line with KIR + CD8 + Tregs in patients with autoimmune diseases, the predicted-reactive KIR + CD8 + T cell population we observed highly expressed genes encoding several members of inhibitory KIRs (KIR2DL1, KIR2DL3, KIR3DL1, KIR3DL2) in addition to cytotoxic molecules (GZMB, PRF1, GNLY) and cell tra cking proteins (CX3CR1, CCL4, CCL5, ITGB1) hypothesized to be involved in their regulatory function.Similar to previously described CD8 + Tregs, the predicted-reactive KIR + CD8 + T cells lacked expression of costimulatory receptors (CD27, CD28, ICOS) and lacked expression of the CD4 + Treg-de ning transcription factor FOXP3 2,21 .There was also high expression of IKZF2, which encodes the transcription factor Helios highly expressed by KIR + CD8 + Tregs in humans and is an essential element for the suppressive function of Tregs, including the analogous Ly49 + CD8 + Treg population in mice 2,[22][23][24] .Notably, the aggregate pattern of expression was only present in the subpopulation of predicted-reactive KIR + CD8 + T cells in comparison with other transcriptional clusters of blood CD8 + T cells ( g. S5).Gene set enrichment analysis also showed signi cant enrichment of the top 200 upregulated genes in KIR + CD8 + Tregs from patients with multiple sclerosis (Fig. 2D; normalized enrichment score = 3.045, false discovery rate = 0) 2 .

Hallmark transcriptional features of KIR + CD8 + Tregs are conserved in blood and tumor
To understand how the differentiation state of predicted-reactive KIR + CD8 + T cells relates to other CD8 + T cells in the blood, we performed pseudotime trajectory analysis on the subpopulation of predictedreactive T cells and clonally related cells.De ning clonally related cells with a naive differentiation state as the start of the differentiation trajectory, we observed three branched lineages (Fig. 3A).The subpopulation of predicted-reactive KIR + CD8 + T cells were enriched along lineage 2 while cells transcriptionally resembling predicted-unreactive CD8 + T cells were enriched in lineage 1. Lineage 3 was enriched with a minor population of proliferative cells.
To understand which genes are most associated with differentiation across lineages, we calculated the association between gene expression and pseudotime along each trajectory then identi ed genes which were differentially associated between lineages (Fig. 3B, table S7).IKZF2 was among the most differentially associated genes along lineage 2 in comparison with other lineages (Wald statistic = 622.9,p = 0), which is consistent with the hypothesized role in de ning the suppressive function of KIR + CD8 + Tregs (Fig. 3C).In comparison, ZNF683, which encodes the transcription factor Hobit, instructs tissue residency, and regulates cytotoxicity in lymphocytes [25][26][27][28] , was the most highly differentially associated gene in lineage 1 (Wald statistic = 3184.8,p = 0).
To understand whether the transcriptional state of predicted-reactive KIR + CD8 + T cells was also present in the tumor microenvironment, we examined the gene expression pro le of 3,037 clonally related cells in the tumor (Fig. 3D).We again observed a high expression of hallmark KIR + CD8 + Treg genes and an absence of FOXP3 (Fig. 3E).As inhibitory KIR proteins are known to be expressed in terminally differentiated CD8 + T cells and contain a cytoplasmic immunoreceptor tyrosine-based inhibitory motif (ITIM) domain, we examined whether KIR + CD8 + T cells in the tumor expressed features of T cell exhaustion (PDCD1, HAVCR2, LAG3, TIGIT, CTLA4, ENTPD1, ITGAE, CXCL13, LAYN, CXCL13, IFNG).
Compared with other tumor T cells, KIR + CD8 + T cells had lower exhaustion gene module scores (p < 0.0001 vs.Other neoTCR8 reactive, p < 0.0001 vs. Unreactive; g.S7), whereas other neoTCR8 predictedreactive T cells had the highest features of exhaustion.These data suggest that CD8 + T cells that are clonally related to circulating predicted-reactive KIR + CD8 + T cells share hallmark features of KIR + CD8 + Tregs and lack features of T cell exhaustion induced by chronic antigen stimulation.
We next identi ed genes that distinguish the subpopulation of predicted-reactive KIR + CD8 + T cells in the blood and clonally related T cells in the tumor from all other CD8 + T cells in the blood and tumor, respectively.Genes that were differentially expressed by both predicted-reactive KIR + CD8 + T cells in the blood and clonally related cells in the tumor were considered to have conserved expression across tissue compartments (Fig. 3F; table S8).The conserved gene signature was comprised of 135 genes and again included genes encoding the KIR protein (KIR2DL3), cytotoxic molecules (GNLY, PRF1, GZMB), transcription factors (TBX21, IKZF2), NK cell-associated proteins (KLRC2, NKG7, KLRD1, KLRG1), and cell tra cking molecules (CX3CR1, ITGB2, S1PR5) in addition to the immunosuppressive cytokine transforming growth factor-β (TGFB1).Using the conserved gene signature, we next calculated gene expression module scores in all CD8 + T cells in the tumor and blood.Expression of the conserved signature in the tumor were positively correlated with those in the blood (R 2 = 0.394, p = 0.005; Fig. 3G).In aggregate, these data suggest that KIR + CD8 + T cells represent a distinct differentiation state, and that the core transcriptional program is preserved across tissue compartments.Furthermore, the conserved gene signature in the blood is correlated with the expression in the tumor, providing support for using blood sampling to assess the presence of this population in the tumor.

Functional con rmation of tumor antigen reactivity in KIR + CD8 + T cells
To functionally con rm the tumor antigen speci city of predicted-reactive KIR + CD8 + T cells, we performed an unbiased epitope screen in 12 patients with melanoma for whom we have previously performed paired single-cell RNA and T cell receptor sequencing (Fig. 4A, table S1) 7 .We selected candidate TCRs from a broad representation of transcriptional and clonal expansion states, including T cells with high expression of the neoTCR8 and the conserved KIR + CD8 + gene signatures, which were then cloned into a T cell line.Epitope screening was performed by selecting peptides related to known tumor associated antigens, melanoma driver mutations, human endogenous retroviral elements, and non-tumor protein as controls.For six patients, we also included personalized tumor neoantigens identi ed through wholeexome sequencing data of tumor from the same patient.Ten amino acid peptide sequences (10-mer) were cloned into single-chain trimers linked to an intracellular domain of TCR zeta chain (SCTz) to construct peptide and major histocompatibility complex hybrid molecules (MCR) 29,30 .In total, over 60,000 10-mers were included in the MCR library.The MCR were then transfected into 16.2X cell line with an NFAT-low uorescent timer (sFT) reporter and co-cultured with the T cell line expressing selected TCRs.
Of 96 KIR + CD8 + clonotypes screened, three clonotypes (TCR2328, TCR2409, TCR2647) from separate patients were con rmed to be reactive to tumor antigens (Fig. 4B-E).TCR2328 and TCR2647 were reactive to known tumor associated antigens (MAGEA1 31 and MELOE-1 32 , respectively), while TCR2409 was reactive to a neoantigen derived from a personalized mutation in the SORL1 gene.Screened clonotypes from KIR -CD8 + T cells that were speci c to tumor antigens were all predicted to be reactive based on neoTCR8 signatures and are the subject of further studies (manuscript under preparation).
Notably, of all clonotypes speci c to tumor antigens, two of the three KIR + CD8 + clonotypes (TCR2328, TCR2409) were the only TCRs identi ed in both tumor and blood in the single-cell sequencing data (Fig. 4E).While the demonstrated reactivity was limited to a minority of clonotypes, which is in line with work by others 5,6,10,29,30,[33][34][35] , these data suggest that at least a subpopulation of KIR + CD8 + T cells are tumor antigen-speci c.
Tumor antigen reactive KIR + CD8 + T cells are associated with worse clinical outcome To assess the potential clinical relevance of KIR + CD8 + T cells, we examined overall survival and its association with the frequency of KIR + CD8 + T cells in our single-cell sequencing cohort.We adopted a Lasso Logistic model to construct a classi er to prospectively classify cells based on clinical variables and select gene expression.We focused on genes encoding cell surface proteins where gene-level expression positively correlated with protein-level expression based on three blood samples for which we performed paired single-cell RNA and DNA oligonucleotide antibody labeling (table S1, S9), which may allow for future validation using clinically-available techniques, such as ow cytometry ( g. S8).Training and testing datasets were generated from the immunotherapy-naive single-cell RNA sequencing cohort, which demonstrated a high prediction performance (Training: AUC 0.864, Testing: AUC 0.867; Fig. 5A).Application to patients who had disease progression following immunotherapy (immunotherapy resistant cohort) also demonstrated acceptable prediction performance (AUC 0.753), suggesting the potential generalizability of the model regardless of past treatment history.We then applied this model to the combined single-cell sequencing cohorts.Using a median cut-point to de ne patients with high (> median) or low (≤ median) frequencies of predicted-reactive KIR + CD8 + T cells, we observed that patients with a higher proportion of predicted-reactive KIR + CD8 + T cells in the blood had shorter three-year overall survival (p = 0.037; Fig. 5B, table S11).This inverse association was particularly pronounced in the immunotherapy resistant cohort (p = 0.029; Fig. 5C, table S11).
To validate the association of KIR + CD8 + T cells and worse overall survival, we next examined the prognostic signi cance of KIR + CD8 + T cells in our ow cytometry (n = 47, table S10).Using a median cutpoint to de ne patients with high versus low frequencies, we again observed that patients with a higher proportion of KIR + CD8 + T cells in the blood had shorter three-year overall survival (p = 0.047; Fig. 5D, table S10).Overall, the association with worse clinical outcome is consistent with the potential regulatory nature of KIR + CD8 + T cells on anti-tumor immunity.

DISCUSSION
In this study, we employed high-resolution single-cell techniques to pro le tumor antigen-speci c T cells in patients with advanced melanoma.By using TCR sequences as a molecular barcode to link clonally related T cells in the tumor and blood, we identi ed a subpopulation of circulating predicted-reactive CD8 + T cells resembling KIR + CD8 + Tregs, which have been shown to speci cally kill pro-in ammatory, autoreactive T cells in patients with autoimmune diseases and viral infections.Using an unbiased epitope screen, we experimentally con rmed the tumor antigen speci city of KIR + CD8 + T cells.To our knowledge, our study is the rst to demonstrate the tumor antigen speci city of KIR + CD8 + Tregs in patients with cancer.
While counterintuitive in the context of anti-tumor immunity, as tumor antigen-speci c CD8 + T cells are the predominant effector immune cells necessary for tumor clearance, it is possible that the suppressive function in KIR + CD8 + Tregs may represent a differentiation state induced by tumor cells as a mechanism of immune evasion.Similar mechanisms have been described with CD4 + Tregs, whereby tumor cells induce the proliferation of tumor antigen-speci c, suppressive CD4 + Tregs through the direct and indirect presentation of human leukocyte antigen (HLA) class II tumor antigens 6 .An analogous induction of CD8 + Tregs through the presentation of HLA class I tumor antigens, as observed in our data, is possible, and prior ex vivo studies suggest that TCR signaling in KIR + CD8 + Tregs is required for maximal suppressive function 2 .A recent study additionally described the peripheral induction of Ly49 + CD8 + Tregs from thymic extruded T cells after encountering self-antigen in mouse models, which impaired anti-tumor immunity 36 , lending further support for the antigen-speci c nature of CD8 + Tregs.
Our analyses also link KIR + CD8 + T cells in the tumor and blood and demonstrate the potential relevance as a clinical, blood-based biomarker.Importantly, we observed that the expression of a conserved gene signature in KIR + CD8 + T cells is correlated between tumor and blood, providing rationale for using the blood to detect the presence of KIR + CD8 + T cells in the tumor.Lastly, KIR + CD8 + T cells are associated with worse overall survival in patients with advanced melanoma, which is consistent with the immunosuppressive function of CD8 + Tregs.
Our study has several limitations.While our epitope screen and analyses mapping clonotypes to publicly available references were able to identify tumor antigen speci city in a subset of KIR + CD8 + T cells, the epitope speci city of the remaining of KIR + CD8 + T cells is unknown.As KIR + CD8 + T cells have been described at homeostasis and increase with progressive in ammation as an alternative mechanism to preserve peripheral tolerance, it is unclear whether tumor antigen speci city is required to impair antitumor immunity or whether the presence of KIR + CD8 + T cells is instead representative of host immunity more broadly.Further studies are also required to elucidate the speci c interactions and contexts needed to induce KIR + CD8 + Tregs.Additionally, the speci c mechanism by which KIR + CD8 + Tregs impair antitumor immunity remain unclear.While KIR + CD8 + Tregs have been described to target antigen-speci c CD4 + T cells through interactions between both classical and non-classical major histocompatibility complex interactions in autoimmunity 2,19 and one prior study reported that the interactions between FOXP3 -CD8 + CD25 + KIR + CD127-and CD4 + Tregs were important in in uencing anti-tumor immunity 3 , we observed no association between the frequency of KIR + CD8 + T cells and CD4 + T cell subsets in the tumor microenvironment (data not shown).We did observe a trend towards an inverse correlation between the frequency of KIR + CD8 + T cells and other neoTCR8 predicted-reactive CD8 + T cells in the tumor, raising the possibility that KIR + CD8 + Tregs may target other neoantigen-speci c CD8 + T cells.
Overall, our study demonstrates the potential clinical relevance of KIR + CD8 + Tregs in patients with advanced melanoma.Importantly, the presence of KIR + CD8 + Tregs in the blood was indicative of the tumor immune microenvironment, providing rationale for further exploration of this population as a circulating biomarker and potential therapeutic target.The prognostic signi cance of KIR + CD8 + Tregs regardless of treatment history further raises the possibility that this population is not impacted by current immune checkpoint inhibitor therapies and may be a relevant mechanism of therapeutic resistance.Further insight into the function of KIR + CD8 + Tregs in the tumor microenvironment may lead to the development of future therapies.

Patient samples
Freshly resected, histologically-con rmed tumor tissue and/or blood were collected from patients with cutaneous melanoma.Tumor samples were obtained immediately following surgery and processed, as outlined below.Heparinized blood samples were collected prior to surgery or treatment and immediately processed, as outlined below.Patient sample details, including the tissue site sampled, processing details, sex, age, treatment history, and clinical outcome, are available in table S1.

Tumor dissociation and peripheral mononuclear cell isolation
Single-cell tumor and blood suspensions were prepared, as previously described 7 .In brief, freshly resected tumor samples were mechanically dissociated and digested in HBSS medium with collagenase IV (2.5 mg/ml; Roche) and DNase I (0.2 mg/ml; Worthington Biochemical Corporation) at 37°C for 30 min.Dissociated tumor suspensions were then isolated using Lymphoprep gradient centrifugation.Peripheral blood mononuclear cells were isolated from whole blood using Lymphoprep gradient centrifugation.For samples Mel-T-09, Mel-T-11, Mel-UT-03, Mel-UT-05, Mel-UT-08-13, and Mel-UT-17-18, single-cell suspensions were cryopreserved in GemCell human AB serum (Gemini) with 20% DMSO in liquid nitrogen.The single-cell suspensions were then thawed per 10X Genomics protocols CG000233 and CG000447 for tumor and blood samples, respectively, prior to further processing.

Cell sorting
Single-cell suspensions were stained for live cells (Live/Dead Cell Viability Assay; Life Technologies) followed by uorophore-conjugated anti-human antibodies, as summarized in table S2.Samples were sorted on a BD FACS Aria II with the gating strategies outlined in table S3.At least 30,000 live cells were sorted per sample.

CITEseq antibody staining
Cryopreserved peripheral blood mononuclear cell suspensions were thawed, as above.Lyophilized TotalSeq™-C Human Universal Cocktail V1. 0 (BioLegend) were reconstituted and used for staining, per the manufacturer's protocols.Each sample was reconstituted to a concentration of 20 million cells/mL and incubated with Human TruStain FcX™ Fc Blocking reagent (BioLegend) for 10 minutes at 4°C.The blocked cells were then stained with the reconstituted TotalSeq-C antibody cocktail and uorophore conjugated surface antibodies for cell staining (table S2) for 30 minutes at 4°C as washed with Cell Staining Buffer (BioLegend) prior to cell sorting (table S3).

10X library preparation and sequencing
Sorted samples were prepared for single-cell sequencing by the Keck Microarray Shared Resource (KMSR) and sequenced by the Yale Center for Genome Analysis (YCGA) at Yale University.Cells were processed following the recommended protocol with the Chromium Single Cell 5' Library Construction Kit and Chromium Single Cell V(D)J Enrichment Kit (Human T Cell; Single Cell 5' Chemistry).Libraries were sequenced on an Illumina NovaSeq S4 instrument.scRNAseq libraries were sequenced at a read length of 26 × 8 × 91 bp and a depth of 300 million reads per sample.CITEseq libraries were sequenced at a read length of 26 × 8 × 91 bp and a depth of 100 million reads per sample.scTCRseq libraries were sequenced at a read length of 150 x 150 bp and a depth of 20 million reads per sample.FASTQ les were generated and analyzed with CellRanger (version 5.0.1) using the GRCh38 human reference genome for alignment.

Whole exome sequencing
Melanoma tumors and adjacent normal tissue were analyzed with whole-exome sequencing, as previously described in-detail 37 .Brie y, genomic DNA was extracted from snap-frozen tumors using DNeasy puri cation kits (Qiagen Inc.).Genomic DNA was sheared, end repaired, ligated with custom adapters (Integrated DNA Technologies), ampli ed, and size selected.Sample concentrations were normalized to 2nM and sequenced on an Illumina NovaSeq 6000 using 101 bp paired-end sequencing reads, per Illumina protocols.Raw sequencing les were aligned to the GRCh38 human genome.

Bulk RNA sequencing
Bulk RNA sequencing was performed on melanoma tumors, as previously described 37,38 .In brief, RNA was extracted from tumors using the RNeasy PowerLyzer Tissue & Cells Kit (QIAGEN) and quality was assessed with a 2100 Bioanalyzer System (Agilent Technologies).Ribosomal RNA was depleted Kapa RNA HyperPrep Kit with RiboErase (Kapa Biosystems, Inc., Cape Town, South Africa).Sequencing libraries were generated using TruSeq RNA sample prep kits (Illumina) and sequenced on an Illumina HiSeq 2500 using 2 × 150 bp paired-end reads with a target of 20-25 million reads per sample.Raw sequencing les were aligned to the GRCh38 human genome.

Immunophenotyping human peripheral blood mononuclear cells
Cryopreserved peripheral blood mononuclear cells were thawed, as above, and rested at 37°C for 16 hours.Cells were then stained for viability with Fixable Near-IR Dead Cell Stain (Life Technologies) at room temperature for 10 minutes in the dark.Samples were then washed with PBS with 2% FBS, prior to staining with a master mix of surface antibodies (table S2) at room temperature for 20 minutes.Cells were then washed with PBS with 2% FBS.Samples undergoing intracellular staining were xed for 45 minutes with Fixation Buffer, per the manufacturer's protocol (eBioscience).Cells were washed with Permeabilization Buffer (eBioscience), stained with a master mix of intracellular antibodies (table S2) at 4°C for 16 hours, and resuspended in PBS + 10% FBS + 0.2% EDTA for analysis on a BD LSR Fortessa in the Yale Flow Cytometry Core.

MCR construction
HLA typing was computationally inferred from the single-cell RNA sequencing data from blood cells using the python-based package, arcasHLA 39 .MCR libraries were generated by cloning oligonucleotide libraries (Twist Bioscience) into MCR-SCTz retroviral expression vectors (having an IRES-hCD4tr (1-420aa) downstream) for each patient HLA, as described previously 30 .For cloning of single peptide-MCR constructs, oligos were purchased from Microsynth AG (Balgach, Switzerland).The libraries cover the full length of selected antigens, as detailed below, by sliding 10-mers shifted by one amino acid.After retroviral production, 16.2X cells were transduced with the MCR libraries.After 2-3 days, transduced cells were sorted based on hCD4 expression (BioLegend) on a BD FACSAria Fusion.

MCR screening
MCR screening was performed as previously described 29,30 , with the adaption of pooling of up to ve different TCRs per co-culture.TCR-cells and MCR-cells were co-cultured with a 5-fold excess of each TCR compared to MCR-cells.Up to 7 rounds of enrichment were done.When a NFAT-signal was detected, the enriched MCR cells were co-cultured with the individual TCR-expressing cell lines to identify the TCR recognizing the MCR cells.Sorting was performed on BD FACSAria Fusions, ow-cytometry for acquisition only was done on a BD LSRFortessa or Luminex Guava easyCyte.When multiple peptides (due to multiple transduction) were detected in the sequencing of positive single cell clones, the peptides were subcloned and re-expressed as single peptides-MCR-SCTz (e.g.TCR2328).Expanded single cells were harvested, and DNA was isolated (KAPA Express Extract), followed by Sanger sequencing of the linked peptide and the HLA class I heavy chain.

Data and Statistics
Single-cell RNA sequencing processing and analysis CellRanger output matrices containing gene expression values for each cell were analyzed using the open-source toolkit, Seurat v4.2 40 .Cell level quality control was performed for each sample separately.
Thresholds for removing low-quality cells were selected based on the distribution of the number of unique genes per cell, the number of genes per unique molecular barcode (novelty score), the percentage of genes that map to the mitochondrial genome, and the expression of housekeeping genes.Library size normalization and variance stabilization was performed using SCTransform to apply a regularized negative binomial regression based on the 3,000 most variable genes 41 .Genes encoded by the Y chromosome were removed for downstream analysis.Normalized datasets from tumor and blood samples for each patient were then combined using the FindIntegrationAnchors and IntegrateData functions in Seurat with default parameters 42 .Scaled z-scores for each gene were calculated and used for principal component analysis (PCA) based on the integrated dataset.
Cell clustering was performed by rst using the shared nearest neighbor (SNN) method based on statistically signi cant principal components, as applied by the FindNeighbors function.Differentially expressed genes between clusters of interest were then identi ed using the Wilcoxon Rank Sum test with Bonferroni correction (q < 0.05) 43 .For subsequent T cell subset analyses, T cell receptor genes were removed to mitigate the effects of clustering driven strictly by highly expanded clones.Cluster annotation was manually assigned based on the expression of canonical marker genes, differentially expressed genes, overlap coe cients with previously published cell phenotype gene signatures 5,25,44,45 , automated reference-based annotations 46 , and the proportion of cells with paired T cell receptor CDR3α and CDR3β chains.Clusters with T cells were selected for downstream T cell phenotyping (table S4-S5).CD8A, CD8B, and CD4 expression following Adaptively-thresholded Low Rank Approximation imputation (ALRA, version 1.0.0) 47were also used to assign CD8 + and CD4 + phenotypes.
T cell receptor repertoire analysis T cell receptor contigs called with high con dence were extracted from CellRanger VDJ outputs and retained for downstream analyses.Nucleotide and amino acid CDR3α and CDR3β sequences and count matrices of gene expression were matched for each cell based on barcode identities using custom R scripts.Only cells with a single CDR3α and single CDR3β sequence were retained for downstream analyses.Cells with multiple sequences or those with missing sequences were not considered for clonal analyses.Clonal relatedness was de ned by cells with exact matching single CDR3α and single CDR3β sequence per patient.
To quantify the degree of clonal expansion, we calculated the proportion of T cell clones occupied by a given clonotype per patient.To predict T cell receptor speci cities, we matched clonotype sequences in our dataset with VDJdb, a publicly available database of T cell receptor sequences annotated with experimentally validated speci cities 13,48 .Human species clonotypes were extracted from VDJdb and matched based on exact amino acid sequences of the CDR3 from the paired a-chain and b-chains.We then calculated the number of tumor antigen-speci c and viral-speci c clonotypes in each reactivity group.
Additional single-cell T cell receptor repertoire analysis was performed using the scRepertoire package (version 1.12.0) 49.Filtered annotated contig les generated by the CellRanger VDJ pipeline were used as input les.Cells with paired single alpha-and beta-chains were selected for downstream analysis.Clonotypes were de ned by using the CDR3 nucleotide and amino acid sequences.Indices of diversity (Shannon entropy, Inverted Simpson) and species richness (Chao, ACE) were calculated using clonalDiversity function.Between group differences were calculated with Kruskal-Wallis one-way analysis of variance testing and Dunn's multiple comparison testing.Visualization of repertoire metrics were generated using GraphPad Prism (version 10.1.1).droplets were identi ed by examining the protein library size distribution.The normalized ADT matrices were then added to Seurat objects generated from the RNA count matrices and analyzed, as described above.

Surface protein-encoding gene selection
To identify candidate surface markers where gene expression correlated with protein-level expression, we calculated Spearman's Rank correlation coe cients between normalized expression in the ADT and RNA assays.Of 137 markers analyzed, 40 genes positively correlated (r > 0.1) with protein-level expression and were therefore included in the surface marker classi er models (table S9).

Cell classi cation analysis
To further investigate the relationship between surface markers and tumor reactivity, we adopted a Lasso Logistic model [66][67][68] to analyze normalized count matrices from CD8 + T cells (glmnet version 4.1-4).
Training and testing datasets were constructed by randomly splitting cells from the immunotherapy-naive cohort into equal sample sizes.In the training phase, the Lasso Logistic model was tted to the count matrices for the 40-surface protein-encoded genes, select clinical variables (sex, age, clinical stage, lesion location, technical batch), and predicted reactivity.The classi er was then applied to the testing datasets and the immunotherapy resistant cohorts.Receiving operating characteristics statistics were calculated based on the differential abundance reactivity labels for the immunotherapy naive cohort and neoTCR8 reactivity labels for the immunotherapy resistant cohort and visualized using the R-based package pROC (version 1.18.0).
To facilitate the application of the Lasso Logistic model classi er to clinically applicable techniques such as ow cytometry, we adopted a Decision Tree model to assign hierarchical combinations of markers using samples where simultaneous RNA sequencing and CITE-seq were performed.Testing and training datasets were constructed by randomly splitting cells from the CITEseq cohort.Thresholds to dichotomize protein-level expression of the markers selected by the Lasso Logistic model classi er were determined by examining the pattern of expression in the ADT assay.A Decision Tree was constructed based on the binary protein expression levels to predict tumor reactivity in the training set using R-based package rpart (version 4.1.19).

Flow cytometry data analysis
Flow data was analyzed using FlowJo™ (version 10.9).Gating was determined by comparison to Fluorescence Minus One (FMO) controls using healthy donor human peripheral blood mononuclear cells.

Survival analysis
The association between the frequency of reactive KIR + CD8 + T cells (high, or greater than or equal to the median frequency for that cohort, vs. low, or less than the median frequency for the cohort) as classi ed by ow cytometry staining (table S10), the Lasso logistic regression classi er (table S11), or the decision

Figures
Figures

Figure 3 Relationship
Figure 3

Figure 4 High
Figure 4