Two cohorts of patients participated in this observational study. The first cohort consisted in 15 NMIBC and 9 MIBC patients recruited at the Lausanne University Hospital (Table 1). All NMIBC patients were treated by TURBT and 9 (BCG-naïve patients) of them received a subsequent BCG treatment. All MIBC patients were treated by cystectomy and 2 of them received neoadjuvant chemotherapy. The second cohort consisted in 4 patients diagnosed with NMIBC, recruited between February and July 2020 at the department of urology of the Cliniques universitaires Saint-Luc in Brussels. They were surgically treated and received BCG treatment. The study protocol was approved by the ethics committee of the hospital (Belgian registration n°B403201938826) and of Canton de Vaud (Switzerland; #2019-00546). All patients provided written informed consent before enrollment in the study.

For the patients of cohort 1, peripheral blood was collected before TURBT or cystectomy and peripheral blood mononuclear cells (PBMCs) were isolated by density gradient centrifugation and cryopreserved, as described (16, 17). Finely minced fresh bladder tumors were cultured in RPMI with 8% of human serum (HS) and recombinant human (rh) IL-2 at 6,000 U/ml for three weeks. From the established tumor infiltrating lymphocyte (TIL) cell lines, a fraction of the cells was further amplified over 14 days with a rapid expansion protocol (REP) (17). For NMIBC patients undergoing BCG treatment, urine samples were obtained after each instillation to expand urinary T cells with REP, as described (16).

For the patients of cohort 2, bladder tumor samples, healthy bladder tissues and PBMCs were collected during TURBT (pre-BCG) and after BCG treatment (i.e. 6 weeks after the sixth BCG instillation; post-BCG). For tissue samples, histological analyses confirmed that tumor cells were present in all tumoral pre-BCG samples and absent from the non-tumoral pre-BCG samples and from the post-BCG samples. Tissue samples were finely minced and mechanically dissociated with a gentleMACS dissociator (Miltenyi), then passed through a 40 µM filter and washed in PBS 1% HS 2 mM EDTA. All samples were then labelled for 30 min at 4°C with an antibody panel (Supplementary Methods) prepared in 10 µl Brilliant Stain buffer plus (BD Biosciences, #566385). Samples were acquired on a FACSAria III Cell Sorter (BD Biosciences) and single CD8⁺ T cells were sorted in 384-well plates containing 1 µl per well of capture mix, composed of RNase-free H₂0, 2.5 mM dNTP (ThermoFisher Scientific, #R0192), 0.5% Triton X100 (Sigma, #T8787), 125U recombinant RNase Inhibitor (Takara, #2313A) and 2.5 µM reverse-transcription oligonucleotide, and frozen at -80°C. Single cell RNA-seq libraries were prepared from thawed sorted single cells using the Smartseq2 protocol, in order to extract TCR sequences. Libraries were sequenced on an Illumina HiSeqX (Macrogen Europe). Reads were aligned to human reference genome hg38 using STAR and sorted and indexed with SAMtools. Gene expression quantification was obtained with htseq-count. TRA and TRB sequences were extracted using Mixcr3 (18). Data analysis was performed in R/Bioconductor as described elsewhere (19).

Whole-exome-sequencing (WES) libraries were prepared using the xGen Dual Index UMI Adapters (IDT), and genomic DNA (gDNA) extracted (DNeasy kit, Qiagen) from formalin-fixed paraffin-embedded (FFPE) or snap-frozen tumor tissue, as well as autologous PBMCs or TILs. Libraries were subjected to paired-end sequencing using Illumina HiSeq4000 with a coverage depth of 250X. The exome reads were aligned to the human reference genome hg38, followed by somatic variant calling using a bioinformatics pipeline consisting of MicMap (https://github.com/sib-swiss/micmap ) and the Variant Effect Predictor (VEP) tool (https://www.ensembl.org/info/docs/tools/vep/index.html). Library preparation was carried out with the AllType FASTplex NGS 11 Loci Flex Kit (One Lambda), and sequencing was conducted on the Illumina MiSeq platform. The binding affinity of 8- to 11-mer candidate mutated peptides to their corresponding HLA class I molecules was assessed using PRIME v1.031814 (20). For each patient, up to 150 peptides with the lowest PRIME %Rank scores were synthesized at the Protein and Peptide Chemistry Facility of the University of Lausanne.

For neoepitope antigenicity evaluation in TILs, IFN-γ-secreting TILs were assessed using an IFN-γ ELISpot assay (Diaclone) upon stimulation with a pool of neoepitope candidates. Neoepitope was considered antigenic if it induced an average number of IFN-γ-secreting cells exceeding the negative control by more than three times the standard deviation and was confirmed in ≥2 independent experiments. For urinary T cells, CD8⁺ cells were magnetically isolated, underwent a second REP protocol before a challenge with a pool of predicted neoepitope peptides followed by IFN-γ ELISpot. For neoepitope antigenicity evaluation in PBMCs, CD8⁺ T cells were magnetically isolated and stimulated for 14 days with autologous irradiated CD4^negCD8^neg fractions and a pool of predicted neoepitope peptides in RPMI supplemented with 8% HS and rec-hu-IL-2 (20 IU/mL for the first 2 days, followed by 150 IU/mL). The cells were then restimulated overnight with the same pools of peptides and irradiated autologous CD4⁺ blast cells (21), and their reactivity against neoepitope peptides was evaluated by IFN-γ ELISpot assays. The recognition of wild-type (WT) peptides by neoepitope-specific T cells was assessed by IFN-γ ELISpot assays in the presence of neoepitope or WT peptide dilutions.

To test neoepitope-reactive T-cell polyfunctionality, TILs or PBMCs (5x10⁵ cells/well) were cultured in RPMI supplemented with 10% FCS, with autologous CD4⁺ blasts (5x10⁵ cells/well) and neoepitope (1 µM) for 6h. After 1h, Protein Transport Inhibitor Cocktail (eBioscience) was added with anti-CD107a-BV605 (H4A3, Biolegend). Surface antigen staining was performed by incubating stimulated cells with a panel of monoclonal antibodies (mAbs) for 20 minutes at 4°C in staining buffer (PBS, 0.2 % bovine serum albumin, 2 mM EDTA), together with the Aqua live/dead stain kit (Thermo Fisher Scientific) and Fc-receptor blocking reagent (Miltenyi). mAbs were used at optimal dilutions and included: anti-CD4-BUV661 (M-T477), anti-CD8-BUV395 (G42-8), anti-CD45-BUV737 (HI30) (BD Biosciences), and anti-CD3-PE/Cy7 (UCHT1) (BioLegend). For intracellular cytokine staining, cells were fixed for 30 minutes at room temperature using the Intracellular Fixation and Permeabilization Buffer Foxp3 set (eBioscience), followed by staining with anti-IFN-γ-BV421 (4S.B3), anti-TNF-α-AF647 (Mab11), and anti-IL-2-PE (MQ1-17H12) (BioLegend). Sample acquisition was performed using the CytoFLEX-LX2 flow cytometer (Beckman Coulter), and data analysis was conducted with FlowJo software.

DNA was extracted from PBMCs or frozen tissue samples. For patient UC4, DNA was extracted from FFPE tissue. For WES, DNA fragment libraries were prepared from gDNA extracted either from frozen tumor fragments or PBMC and were amplified for exon DNA using the SureSelect V7 method (Agilent) for patients UC1, 2 and 3. For patient UC4, the library was prepared from gDNA extracted from frozen or FFPE tumor tissue and PBMC using the Human Core Exome kit (Twist Bioscience). The libraries were paired-ended sequenced on a Novaseq (Illumina), generating 151 bp-long sequences with a coverage depth of 50X (for PBMC) and 200X (for tumor samples) in paired-end mode. The sequenced reads were trimmed and filtered as above. The resulting reads were aligned to the GRCh38 reference human genome sequence using the align function from the Rsubread Bioconductor package (22) in DNA mode and built-in genome annotations.

For bulk RNAseq, poly-A RNA was purified from total RNA extracted from frozen tumor fragments (patients UC1, 2 and 3) and used as template to prepare a cDNA library according to the TruSeq Stranded mRNA method (Illumina). The cDNA libraries were paired-end sequenced on the Novaseq platform (Illumina). The sequenced RNA-seq reads were trimmed with the ShortRead Bioconductor package (23) to remove adapter sequences as well as leading and trailing bases with Phred quality scores <20 and were further filtered to remove reads shorter than 50 bases and unpaired reads. The resulting reads were aligned to the GRCh38 reference human genome sequence using the splice-aware subjunc function from the Rsubread Bioconductor package and the Homo_sapiens.GRCh38.105.gtf genome annotations (Ensembl). Raw read counts per gene were calculated with featureCounts (Rsubread package, with multi-mapping reads reported as fractionated counts) and normalized in TPM.

Somatic variants were identified using Mutect2 (24), Strelka2 (25) and Varscan2 (26) and annotated with Funcotator (GATK4). Only missense variants identified by at least 2 of the 3 variant caller tools were selected. For each variant amino acid and its wild-type counterpart, a 25 amino acid-long sequence spanning the centered position was generated. For each of these sequences, 8 to 11-amino acid-long peptides able to bind to the HLA-A, HLA-B and HLA-C types of the patient were predicted with NetMHCpan4.0 (27). Only peptides derived from a gene expressed at a level >1 TPM as determined by RNA-seq on the same tumor sample and associated with a strong predicted binding affinity (%Rank <0.5%) were selected for synthesis.

The other methods for the analyses of the patients of cohort 2 are further described in Supplementary Methods.

We first analyzed a cohort of 24 patients with non-metastatic bladder cancer (Table 1). The tumor mutational burden was established from surgical samples. The median of non-synonymous single nucleotide substitutions per tumor was 172.5 with a range of 20-567, which is in line with previous reports (10, 28). No difference was observed when stratifying our cohort according to tumor stage or grade (data not shown). Mutant peptides of 8 to 11 amino acids predicted to bind to autologous HLA class I molecules were obtained with PRIME v1.0318 (20) and the 105–150 best-ranked peptides were synthesized. We used IFN-y ELISPOT assays to screen bulk TILs and PBMCs against pools of 10 to 20 mutant peptides (3400 peptides in total). Out of 400 such recognition assays, 14 were positive (Figure 1A). It led to the identification of 10 different neoepitopes (Figure 1B, Supplementary Table 1). Only 3 patients had T cells against neoepitopes in both TILs and PBMCs. In two patients, URO671 and URO878, T cells from both TILs and PBMCs recognized the same neoepitope. Three patients, URO291, URO845 and URO878, had T cells against two neoepitopes. Overall, 9 out of the 24 patients had detectable neoepitope-specific T cells: 6 NMIBC and 3 MIBC patients (Figure 1B). Of note, in two patients (URO475 and URO845), we observed recognition of a pool of peptides, but failed to deconvolute to the single peptide, owing to the scarcity of biological materials. Among the identified mutated genes (Supplementary Table 1), only IQGAP3 (29) and RXRA (30) mutations have been already observed in cancers. Notably, mutations of the RXRA gene can be found in 9% of MIBC and 58% of them consist in S427F/Y mutation (30). For most of all the identified peptides, we verified that the corresponding WT peptide was not or weakly recognized (Figure 1C). In addition, upon stimulation by their cognate neoepitope, TILs from 6 patients were shown by flow cytometry to upregulate expression of CD107a, IFN-γ, TNF-α and IL-2, indicating polyfunctionality (Figure 1D and Supplementary Figure 1). However, we were unable to generate autologous tumor cell lines, so we could not confirm that the anti-neoepitope CD8⁺ T cells recognize autologous tumor.

As expected, the median tumor mutational burden was higher for patients with detectable anti-neoepitope T cells as compared to those without detectable anti-neoepitope T cells (Figure 1E and Table 1). Furthermore, in contrast to NMIBC patients, MIBC patients harboring detectable spontaneous neoepitope CD8⁺ T-cell responses exhibited improved recurrence-free survival compared to patients without neoantigen responses (Table 2, Supplementary Figure 2). Of note, no difference in recurrence-free survival was observed when considering only NMIBC patients (n=9; 4 with and 5 without detectable neoantigen-specific T cells) who underwent a BCG therapy (data not shown). Larger cohorts are needed to confirm this result.

Finally, we established 10 T-cell lines from urines collected during BCG treatment from 5 out of 9 patients with NMIBC. They were screened with IFN-γ ELISPOT for recognition of the autologous mutant peptides, with negative results. However, no neoepitope-specific T cells were detected in the TILs or PBMCs of these 5 patients (data not shown).

We conclude from this cohort of 24 patients with non-metastatic bladder cancer screened for the presence of anti-neoepitope T cells, the largest series thus far analyzed in this cancer type, that about one third of patients mounted a spontaneous anti-neoepitope T-cell response detectable in blood or tumor, which might be associated with tumor control in MIBC patients.

Since we did not detect urinary neoantigen specific-CD8⁺ T cells from the BCG-treated patients from cohort 1, we focused on 4 new NMIBC patients with high grade tumors (Table 3) to gain more insights on bladder neoantigen-specific T-cell features upon BCG therapy. Thus single CD8⁺ T cells from tumor collected during the TURBT (pre-BCG) and non-tumor tissues collected at the TURBT and after the BCG therapy (post-BCG) were sorted for RNAseq using Smart-Seq2 (31) (Table 3).

t-SNE dimensionality reduction analysis of CD8⁺ T cells from pre-BCG tumors and corresponding post-BCG bladder biopsies followed by graph-based clustering revealed 6 clusters (Figure 2A). First, cells qualified as senescent that express KLRG1 but not IL7R (32) and strongly express cytotoxicity-associated genes such as GZMB, GZMA and PRF1. Second, a population of cells strongly expressing marker genes associated with T regulatory cells (Tregs) such as FOXP3, IL2RA, CCR8 and CTLA4. Third, a population of exhausted T cells expressing HAVCR2, ENTPD1, IFNG and ZNF683, which encodes Hobit, a canonical transcription factor associated with tissue-residency. Next, two populations of resident cells expressing ZNF683, CD69 and ITGAE (CD103), distinguished by the expression of IL7R. Finally, a population of memory cells that express EOMES, encoding the transcription factor eomesodermin, associated with T-cell memory differentiation (Figure 2B) (32).

Comparing these CD8⁺ populations in the tumors of the four patients revealed major differences. Exhausted cells were present only in the tumors of patients UC1 and UC2. The CD8⁺ Treg population was present only in the tumor of patient UC2. In the tumor of patient UC3, most of the cells were of the resident IL7R⁺ phenotype (Figure 2C). These results clearly show that each tumor contained functionally distinct populations of CD8⁺ T cells. Moreover, several of these populations were present in some tumors but absent from others. The causes of these important differences between tumors are unknown. Next, we compared these populations of pre-BCG CD8⁺ TILs with those present in the corresponding tumor-free bladder regions post-BCG (Figure 2C). The populations of exhausted and Treg CD8⁺ cells present in tumors of patients UC1 and UC2 were not detected anymore after BCG. In contrast, no clear difference between pre- and post-BCG samples was observed for the other CD8⁺ populations (Figure 2C). The results suggest that no new population of CD8⁺ T cells appeared in the bladder of these BCG-treated patients.

Anti-neoepitope CD8⁺ T cells infiltrating metastatic melanomas, breast, colon and rectal tumors have been shown to be present mostly among exhausted T cell populations, presumably as a consequence of their chronic stimulation by tumor antigens (33, 34). As exhausted CD8⁺ T cells were clearly present in the tumor of patient UC1, we aimed to identify the antigens recognized by some of these cells (Supplementary Figure 3).

The TCR repertoire of the 222 exhausted CD8⁺ cells extracted from the tumor of patient UC1 consisted of 48 clonotypes (Figure 3A). Six of them, all enriched in tumor tissue as compared to non-tumoral adjacent pre-BCG bladder tissue and blood (Figure 3A), were screened for antigen recognition. Reconstructed TCR sequences were transduced into Jurkat 2D3 cells, which have lost endogenous TCR expression and express CD8ß and a NFAT reporter gene (35). TCR-transduced Jurkat 2D3 were screened for recognition of UC1 EBV-B cells incubated with a set of 125 synthetic mutant peptides predicted to bind to UC1 HLA class I molecules. Four TCRs recognized a mutant peptide (Figure 3A). TCR #10 and #8 recognized the same peptide. The results of one screening are shown on Figure 3B. TCRs #49 and #31 recognized none of the mutant peptides present in the tested set and were screened on a cDNA library prepared with mRNA extracted from UC1 tumoral tissue. One screening result is shown in Figure 3C. The two TCRs proved to recognize also a mutant peptide (Figure 3A). One of them, DCHELWSNKR presented by HLA-A*68:01 to TCR #31, was encoded by gene CDK12. CDK12 mutations are frequently found in human tumors, including bladder cancer (36).

The CD8⁺ TILs of patient UC2 were functionally more diverse than those of patient UC1. We selected 6 TCRs from clusters ‘exhausted’, ‘Tregs’ and ‘memory’, none of which scored positive against a set of 50 mutant peptides containing epitopes predicted to bind to autologous HLA class I molecules. One clonotype, TCR #52 from the ‘memory’ cluster, recognized autologous EBV-B cells. Previous reports have identified such bystander T cells in human tumors (37). These negative results were surprising for the TCRs of the 5 ‘exhausted’ T cells and left open the possibility that they could recognize other tumor-specific antigen(s) than mutant peptides resulting from a single nonsynonymous nucleotide polymorphism. All 5 remaining TCRs were therefore screened against transfectants of a tumoral cDNA library (Supplementary Table 2). These TCRs did not appear to recognize tumor-specific antigens. For patient UC3, two clonotypes (#65 and #68) were expanded in the tumor of the patient. The TCRs of these clonotypes showed no reactivity against a set of 106 mutant peptides. No TCRs were selected for patient UC4, as no clonotype was enriched in the tumor of the patient compared to blood and adjacent normal tissue.

Concerning global TCR repertoire, we did not find clear patterns differentiating pre-and post-BCG PBMC samples, in terms of clonality index, V-gene usage, CDR3 length (data not shown). In addition, comparing pre-and post-BCG repertoires for each patient did not show global shifts in TCR frequencies (data not shown).

HLA class I expression is necessary for presentation of epitopes to CD8⁺ T cells. We performed beta-2-microglobulin (B2M) staining by immunohistochemistry, as a proxy for HLA class I expression, on the tumors of patients UC1, UC2 and UC3. Unfortunately, there was not enough tissue available to perform B2M staining for UC4 patient. We found that tumoral B2M expression was high in patient UC1, intermediate in patient UC2 and absent in patient UC3 (Supplementary Figure 4). These results led us to evaluate the status of the HLA and B2M loci, as genetic modifications of these genes are a frequent cause of altered HLA class I expression. No genetic anomaly was found for patient UC1 and UC3. A clear loss of heterozygosity at the HLA locus was detected in the tumor of patient UC2, providing an explanation for the reduced intensity of the B2M staining. We did not find a genetic alteration that could explain the undetectable HLA class I expression in the tumor cells of patient UC3.

In conclusion, from these analyses of CD8⁺ T cells present in NMIBC of 4 patients at the TURBT, we could demonstrate the presence of a spontaneous tumor-specific T cell response only in the tumor of patient UC1. All the tumor-specific CD8⁺ T cells that we analyzed (n=6) proved to recognize mutant peptides, which is in line with a higher tumor mutational burden in this tumor than in those of patients UC2, UC3 and UC4 (Table 3), and an intact HLA class I expression in the tumor. It is remarkable that by analyzing only 6 out of the 48 TCRs expressed by 222 CD8⁺ cells with an ‘exhausted’ phenotype we identified 5 different tumor-specific antigens, suggesting that many more tumor-specific antigens were immunogenic to CD8⁺ T cells in that tumor. These positive results are in sharp contrast with what we observed in the other two tumors, for which only negative results were obtained when screening sets of candidate mutant peptides as well as tumoral cDNA libraries.

The clinical benefit of adjuvant BCG instillations in the bladder could result from a local inflammatory response, which could contribute to either increase the frequency of existing tumor-specific CD8⁺ cytolytic T cells or prime new ones. We first addressed the potential boosting effect of the BCG therapy in patient UC1, who had mounted a spontaneous tumor-specific response detected in TILs prior to BCG treatment. In the bladder tissue collected 6 weeks after the last BCG instillation, the frequencies of the six tumor-specific TCRs had decreased at least 10-fold as compared to the tumor at the time of TURBT, even though the difference was statistically different only for clonotype #19 (Figure 4A). Considering that after the treatment there was no tumoral tissue left detected in the bladder biopsies, the absence of tumor-specific CD8⁺ TILs was expected. In addition, these tumor-specific T cells did not appear to persist in the bladder as tissue-resident memory T cells, which could participate in a long-term immunosurveillance of the primary tumor site. We also looked for an increase in the frequencies of tumor-specific CD8⁺ T cells in the blood of patient UC1, concomitant with BCG instillations. As shown in Figure 4B, 3 of the 6 clonotypes (#8, #10 and #19) were detected in pre-BCG blood but not anymore after BCG. Two other clonotypes had decreased blood frequencies after BCG. One tumor-specific clonotype (#24) was never detected in blood. Thus, for the tumor-specific CD8⁺ TILs of patient UC1, we have no element in favor of a local or systemic frequency increase concomitant with BCG instillations. Next, we examined whether in patient UC1 new CD8⁺ T cell clonotypes appeared after BCG. In the post-BCG bladder, the frequencies of 4 clonotypes increased significantly compared to the pre-BCG tumor. Two of these clonotypes (#28 and #47) were negative in our screenings for tumor specificity. We did not explore the specificity of the two other TCRs. For patients UC2, UC3 and UC4, some clonotypes had frequencies that increased significantly in the post-BCG bladder as compared to pre-BCG tumor. Three of them, #36 and #56 for patient UC2 and #90 for patient UC4, were tested for tumor specificity with negative results (Supplementary Table 2).

In conclusion, similarly to the cohort 1, our results demonstrate that tumor-specific CD8⁺ TILs in NMIBC can be spontaneously present with multiple specificities at the time of TURBT. However, this is not the case in all patients, as we detected no tumor-specific CD8⁺ T cells in three out of four examined tumors. In addition, in the patient with tumor-specific T cells, the frequencies of these T cells did not increase in blood or the bladder after the BCG instillations.