Recurrent Immunogenic Neoantigens and Their Cognate T-cell Receptors in Treatment-Resistant Metastatic Prostate Cancer

Recurrent AR^H875Y-derived immunogenic neoantigens and their cognate T-cell receptors are identified in treatment-resistant prostate cancer, revealing shared, targetable immune vulnerabilities for T-cell receptor immunotherapies.

Abstract

New approaches that generate long-lasting therapeutic responses in patients with therapy-resistant metastatic cancer are urgently needed. To address this challenge, we developed Spot Neoantigens in Metastases (SpotNeoMet), a novel data-driven pipeline that systematically identifies recurrently presented neopeptides in treatment-resistant patients. We identified seven therapy resistance mutations predicted to produce neopeptides presented by common HLAs. Using HLA immunopeptidomics, we discovered three novel neopeptides derived from androgen receptor (AR) H875Y, a common metastatic castration-resistant prostate cancer (mCRPC) mutation. We validated these neoantigens as highly immunogenic and then isolated and characterized cognate T-cell receptors (TCR) from healthy donor peripheral blood mononuclear cells. We demonstrated that AR H875Y–specific TCRs are highly specific and kill prostate cancer cells presenting AR neopeptides in vitro and in vivo. Our new pipeline identifies novel immunotherapy targets and potential treatment options for patients with mCRPC. Moreover, SpotNeoMet offers a systematic route to identify “HLA–peptide” pairs and their cognate TCRs across treatment-resistant cancers.

Significance:

As the emergence of resistance to targeted treatments in patients with metastatic cancer, there is an urgent need for innovative therapeutic approaches for this population. Our study provides a new analytic framework to identify neoantigens from treatment-resistant mutations and a proof-of-concept T cell–based immunotherapy treatment for mCRPC.

Introduction

Prostate cancer is one of the most common malignancies in men, with global estimates of 1,466,680 new cases and 396,792 deaths in 2025 (1). Most patients with metastatic prostate cancer (mPC) initially receive androgen-targeting therapies that provide temporary disease control, but nearly all patients develop resistance and progress to metastatic castration-resistant prostate cancer (mCRPC). Although new therapies such as lutetium-177, radium-223, and PARP inhibitors (PARPi) have improved the quality of life and overall survival, the 5-year survival rate remains 30% (2).

The advent of immune checkpoint blockade has marked a paradigm shift in cancer treatment, with long-lasting successes in melanoma and non–small cell lung cancer, but with limited efficacy in treating other cancer types such as prostate cancer and breast cancer (3).

Neoantigens are degradation products of somatically mutated proteins in cancerous cells (4). Neoantigen-based immunotherapies, including neoantigen vaccines, cell transfer therapy, and bispecific antibodies, have recently garnered significant attention owing to their potential to specifically target tumor cells.

Despite remarkable achievements in the field, most neoantigens identified stem from private subclonal somatic mutations, confining their effectiveness to the individual patient and potentially only to the subclonal portion of the tumor cells. To develop “off-the-shelf” immunotherapies for a broader patient population, there is a need to identify hotspot neoantigens, especially from driver mutations, which are expected to be clonal and critical for cell function. However, only a few hotspot neoantigens have been identified to date (5, 6).

Our previous study used a data-driven approach to identify recurrent neoantigens and their corresponding T-cell receptors (TCR) in melanoma (7). Here, we established a novel analysis pipeline, Spot Neoantigens in Metastases (SpotNeoMet), that systematically identifies treatment resistance mutations and their predicted derived neoantigens and validates the neoantigen presentation and immunogenicity. We successfully identified three immunogenic neoantigens derived from the prevalent androgen receptor (AR) mutation (AR H875Y), a major driver of resistance to antiandrogen-targeted therapies in mCRPC. We then identified three cognate and highly specific TCRs that reacted to AR neoantigens. Our work presents new immunotherapy options for patients with mCRPC and a pipeline applicable to identifying neoantigens and their cognate receptors across diverse cancers.

Results

Data-Driven Identification of Recurrent Candidate Neoepitopes from Treatment Resistance–Associated Hotspot Mutations: SpotNeoMet

We aimed to systematically identify potential recurrent neopeptides specifically presented in treatment-resistant tumors. We reasoned that for a candidate to be considered a “hotspot neoantigen” with significant therapeutic potential, the combined frequency of the specific binding HLA allele and the associated mutation should be highly prevalent among patients with these cancers. To this end, we developed a pipeline named SpotNeoMet to identify recurrent neoantigens, further described in the Methods section (“Identification of Recurrent Neoantigens”). We first identified 1,138 recurrent (>3 samples) somatic single-nucleotide variant hotspots in protein-coding sequences (8) in the Hartwig Medical Foundation (HMF) dataset, a metastatic cancer cohort. This dataset comprises 3,235 metastatic tumors with clinical annotations on the treatments received for their primary tumors (Methods; refs. 8, 9). As a control, we surveyed 6,062 hotspots in a cohort of 9,644 primary tumors from The Cancer Genome Atlas (TCGA; ref. 10)–a primary cancer cohort. We predicted HLA class I (HLA-I) binding among peptides derived from missense hotspots in both cohorts, resulting in 54,478 unique peptide/HLA-binding pairs (Fig. 1A; “Methods”). We aimed to identify recurrent neoantigens generated from driver mutations (11) suspected to confer treatment resistance in primary tumors of metastatic patients with one or more cancer types with limited prospective therapeutic options. Thus, we selected clonal driver (11) neoantigens derived from hotspots enriched in the metastatic cohort, which were found to generate potential neoepitopes predicted to bind the most frequent HLA-I alleles, as described in the “Methods” section (Fig. 1A–C), to maximize their potential as T cell–based immunotherapy targets for treatment-resistant tumors.

Figure 1.

Pipeline for targeted therapy-resistant neoantigen prediction in the cancer genome. A, Schematic overview of the pipeline for identifying and prioritizing recurrent hotspot neoantigens associated with therapy-resistance genes. Single-nucleotide variant (SNV) hotspots were first identified in TCGA or across metastatic cancer samples from the HMF cohort using HotspotFinder. Hotspots resulting in missense variants were used for further analysis. The same strategy was applied to primary tumors from the TCGA database. For each hotspot, we generated all possible mutated and WT peptides 8–14 amino acids long that overlapped with the mutation and predicted neoantigen presentation to HLA-I alleles using MHCflurry, MHCnuggets, and NetMHCpan. Finally, we prioritized hotspot neoantigens for experimental validation using different criteria (“Methods”), sorted by the total number of mutated samples in IntOGen and median variant allele frequency, followed by artifact removal and prediction of driver mutations using BoostDM. B, Frequency of candidate hotspot mutations across metastatic and primary pan-cancer cohorts to prioritize treatment-resistant mutations in patients with metastasis in TCGA (primary tumors) and Hartwig (metastases). C, Top-ranked hotspot mutations of interest were exclusively found in metastases, suggesting that they likely arise from resistance to targeted therapies. Note that for representation purposes, all other neoantigens found in B were excluded from this plot. D, Table depicting selected neoantigens from C, showing the number of patients with mutations bearing the mutation, clonality status, and treatments received.

Among the hotspot neoantigens fulfilling these criteria, we identified seven recurrent mutational hotspots in four genes suspected to confer resistance to the treatment of primary tumors, including the AR in prostate cancer, epidermal growth factor receptor (EGFR) in non–small cell lung cancer, estrogen receptor 1 (ESR1) in breast cancer, and SMAD family member 4 (SMAD4) in colorectal cancer (Fig. 1A–D; Supplementary Table S1). The lack of effective treatments and immunotherapeutic options for mCRPC led us to focus on the three hotspots affecting AR, a key driver of disease progression and drug resistance in advanced prostate cancer. These hotspots were AR L702H, H875Y, and T878A, which were present in 4.25%, 3.5%, and 2% of patients with metastatic prostate adenocarcinoma in the HMF cohort, respectively (Fig. 1D).

Analysis of AR H875Y Candidate Neoantigens

To validate the presentation of these potential neopeptides, we applied HLA-immunopeptidomics to B-LCL 721.221 cells coexpressing a minigene (MNG) and one HLA allele (Fig. 2A). Mass spectrometry (MS) successfully identified three neopeptides of the AR H875Y mutation restricted to HLA-A*01:01 and HLA-B*15:01 (Fig. 2B–D) and three wild-type (WT) AR peptides restricted to HLA-A*02:01 (Supplementary Fig. S1A–S1C). Synthetic light peptides bearing the same amino acid sequence were used to validate the MS results. The MS/MS spectra of the synthetic peptides showed a high Pearson correlation coefficient relative to that of the neopeptides (Fig. 2B–D; Supplementary Fig. S1A–S1C). Importantly, HLA-immunopeptidomics of the prostate cancer cell line LNCaP, which endogenously expresses HLA-A*01:01, transduced with AR H875Y MNG and HLA-B*15:01 (LNCaP^{H875Y/B*15:01}), demonstrated robust presentation of all three AR-derived neopeptides (Supplementary Fig. S1D–S1F).

Figure 2.

From spectra to structure: comprehensive analysis of resistance-derived neopeptides. A, Diagram illustrating the identification of neopeptides presented by monoallelic B cells. B-LCL 721.221 HLA-I–null cells were transduced with a single HLA allele to generate a library of monoallelic B-cell lines. These monoallelic B cells were then transduced with one of the MNGs to create all combinations predicted to bind HLA. This was followed by HLA-immunopeptidomics, in which monoallelic B cells were analyzed using MS to identify the neopeptides presented by each specific HLA allele. A schematic was created using BioRender.com. The head-to-tail spectrum of the identified neopeptides by HLA-immunopeptidomics in monoallelic B cells of (B) VQPIARELY neopeptide, harboring the B*15:01/AR H875Y combination; (C) SVQPIARELY neopeptide, harboring the B*15:01/AR H875Y combination; (D) LLDSVQPIARELY neopeptide, harboring the A*01:01/AR H875Y combination; and (E) T2-A*01:01 (left) and T2-B*15:01 (right) cells were incubated overnight with the indicated peptides. HLA surface stabilization was quantified by flow cytometry and is shown as fold change in normalized mean fluorescence intensity relative to the DMSO control. F, PyMOL-based structural model of the peptide SVQPIARELY depicted as sticks presented by the HLA-B*15:01 molecule, depicted as a gray ribbon, viewed from the top (left) and side (right). Peptide amino acids are labeled with one-letter symbols and positions. G, PyMOL-based structural model of the peptide VQPIARELY depicted as sticks presented by the HLA-B*15:01 molecule, depicted as a gray ribbon, viewed from the top (left) and side (right). Peptide labeled as in E. H, PyMOL-based structural model of peptide LLDSVQPIARELY depicted as sticks presented by the HLA-B*01:01 molecule, depicted as a gray ribbon, viewed from the top (left) and side (right). Peptide labeled as in E.

We collected published HLA-binding peptides from the IEDB database (12) and our HLA-immunopeptidomics to validate HLA preferences and analyzed their sequence motifs using GibbsCluster (13). The resulting motifs (Supplementary Fig. S1G and S1H) demonstrated that our AR H875Y–derived neopeptides closely matched the anchor residue preferences for their respective HLAs (14). We predicted by NetMHCpan that for HLA-A*01:01, several mutant peptides were predicted to bind (9- to 13-mer), but for HLA-B*15:01, binding was prominent for the 9- and 10-mer peptides (Supplementary Tables S2 and S3; ref. 15).

To functionally validate the binding capacity of the identified AR H875Y–derived neopeptides, we performed a T2 stabilization assay. T2 cells were transduced with one of the HLA alleles (A*01:01, B*15:01), incubated overnight with peptide and β-2-microglobulin, and evaluated for surface HLA using flow cytometry. The identified neopeptides bind to their respective HLA molecules similarly to or greater than known canonical viral peptides and peptides derived from NRAS Q61K, NRAS Q61R, NRAS WT, and HLA-E sequences (Fig. 2E; refs. 7, 16). We conclude that VQPIARELY, SVQPIARELY, and LLDSVQPIARELY are naturally processed AR H875Y–derived neopeptides that are robustly presented in the context of HLA-A*01:01 and HLA-B*15:01.

The Mutational Residue Tyrosine in AR H875Y Mutant Neoantigens Is Essential for HLA-A*01:01 and HLA-B*15:01 Presentation and CD8⁺ T-cell Recognition

Differences in amino acid profiles between immunogenic and nonimmunogenic peptides highlight the importance of specific residues in eliciting T-cell responses (17). As the tyrosine residue is at the C-terminal position of the AR H875Y peptides (9-, 10-, and 13-mer), we used PyMOL (“Methods”) to visualize how the mutation position in the neopeptides affects antigen presentation and T-cell recognition. Structural modeling revealed that the tyrosine residue in the neopeptides acts as an anchor for both HLAs (HLA-A*01:01 and HLA-B*15:01) and is not presented to CD8⁺ T cells (Fig. 2F–H). Interestingly, in contrast to neopeptides, structural modeling indicated that these HLAs could not present the corresponding WT peptides. We validated these predictions using three independent methods. First, we used NetMHCpan, which predicted that the WT peptides were nonbinders to HLA alleles. Next, we analyzed the HLA atlas data using MaxQuant (18) and did not identify any of these peptides. Lastly, we experimentally validated their presentation by HLA-immunopeptidomics of monoallelic B cells overexpressing the WT MNG and found that these peptides were not present. We conclude that although the mutational residue position does not participate in the TCR interaction, it might create a neopeptide that would be perceived as completely foreign by CD8⁺ T cells because the WT peptide cannot bind to these HLA alleles, which could promote its immunogenicity.

AR H875Y Neopeptides Are Immunogenic in Peripheral Blood Mononuclear Cells from Healthy Donors

To evaluate the immunogenicity of AR H875Y neoantigens in healthy donors’ peripheral blood mononuclear cells (PBMC), we used a previously published protocol with some modifications (19). Following three in vitro stimulation rounds with the corresponding peptides, HLA-B*15:01/HLA-A*01:01 monoallelic B cells were pulsed with 1 µg/mL AR H875Y or WT peptides for 2 hours, followed by 16 hours of coincubation with T cells to evaluate reactivity using flow cytometry with 4-1BB, IFNγ, and TNFα staining (Fig. 3A). All three neopeptides elicited a significantly effective immune response in PBMCs from healthy donors compared with the WT peptides (Fig. 3B; Supplementary Fig. S2A–S2C). We quantified the number of patients who could be treated with agents against these neopeptides utilizing publicly available mCRPC datasets (20–22). In silico HLA typing (Supplementary Tables S4–S7), integrated with our original Hartwig Medical Database (Hartwig) dataset, showed that AR H875Y was present in ∼5.86% (∼17,580 annually) of mCRPC cases (range: 2.36%–9.09%), with 13.93% of them (range: 0%–35.71%) carrying at least one of the relevant HLA alleles (∼2,450 annually; ref. 1). Altogether, our findings point to applying these neopeptides as immunotherapeutic targets that could be implemented in patients with mCRPC.

Figure 3.

Immunogenicity assessment and TCR identification for AR H875Y–derived neopeptides. A, T cells derived from healthy donor PBMCs were stimulated and expanded for 26 days. Following coculture with human B-LCL 721.221 at an E:T ratio of 2:1 of targets expressing HLA-A*01:01 or HLA-B*15:01 alleles that were pulsed with 1 μg/mL AR H875Y neopeptides (VQPIARELY, SVQPIARELY, LLDSVQPIARELY) or 1 μg/mL AR WT peptide (VQPIARELH, SVQPIARELH, LLDSVQPIARELH). A schematic was created using BioRender.com. B, Flow cytometry analysis evaluating T-cell activation based on IFNγ secretion, TNFα secretion, and 4-1BB expression. The normality of the data distribution was confirmed using the Shapiro–Wilk test (P > 0.05). Statistical significance was determined using an unpaired two-tailed Student t test. Specific P values are indicated: *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001. Images are representative of ≥2 replicates. C, Healthy donor T cells were expanded and stimulated for 26 days. Following expansion and stimulation, T cells were stained with AR H875Y double fluorophore-labeled dextramers to evaluate the percentage of the AR H875Y–specific subpopulation. D, Double-positive and double-negative dextramer cells were sorted by FACS for single-cell TCR sequencing using 10x Genomics. TCR clonal expansions are shown for double-positive dextramer cells (dark blue) and negative cells (light gray). Circle sizes represent the number of single cells sharing the same TCR sequence. E, The table includes the TCR sequences of the major TCR clones identified in D as sequenced by 10x TCR single-cell sequencing.

Identification of AR H875Y Neopeptide-Specific TCRs

To characterize the antigen-specific TCR repertoire within the in vitro stimulated T cells described above, we performed single-cell TCR sequencing on dextramer double-positive (0.31%, 0.5%, and 2.44% of stimulated cells of the 9-, 10-, and 13-mer neopeptides, respectively) and double-negative sorted populations on day 26 (Fig. 3C; Supplementary Fig. S2D). The dextramer-positive population comprised 45.37% of cells belonging to a single TCR clone for the 9-mer peptide, 48.45% and 45.41% of cells belonging to two TCR clones for the 10-mer peptide, and 44.27% of cells belonging to a single TCR clone for the 13-mer peptide. These clones were absent in the dextramer-negative population (Fig. 3D; Supplementary Tables S8–S10). To assess whether the identified TCR clones were reactive toward their respective neoantigens, we first cloned their TCRα and TCRβ chains into an MSGV1 vector designed with a mouse constant region (Fig. 3E; ref. 23).

To test their reactivity and the potential amount of peptide needed for a TCR response, we retrovirally infected Jurkat TCRαβ knockout cells (24) with each TCR individually (T157.1, T157.2, T157.3, T112.1) and validated their infection rate by mouse constant TCR (mTCR) expression and binding to the AR H875Y dextramer (Supplementary Fig. S3A–S3C). Next, the reactivity of TCR-transduced Jurkat T cells to neoantigens was assessed by coculturing them with antigen-presenting cells loaded with titrating concentrations of the relevant neoantigen peptides. Activation of the NFAT reporter gene was measured using the Bio-Glo Luciferase Assay. Strikingly, we found that concentrations as low as 0.54, 1.9, 6.6, and 6.6 nmol/L for T157.1, T157.2, T157.3, and T112.1, respectively, were sufficient to trigger a response above baseline (Fig. 4A).

Figure 4.

Avidity assessment of H875Y cognate TCRs. (A) TCRαβ-knockout (CD8⁺) Jurkat cells were transduced with the TCR-MSGV1 plasmid using retroviruses. One week after transduction, the cells were cocultured at a 1:1 E:T ratio with monoallelic B cells (HLA-B*15:01 or HLA-A*01:01) and titrated with AR H875Y peptides (VQPIARELY, SVQPIARELY, and LLDSVQPIARELY) or AR WT peptides (VQPIARELH, SVQPIARELH, and LLDSVQPIARELH) for 6 hours. The luminescence assay measured the TCR activation for T157.1, T157.2, T157.3, and T112.1. The plots represent ≥2 biological replicates. Healthy donors’ PBMCs were transduced with the TCR-MSGV1 (T157.1, T157.3, and T112.1) plasmid using retroviruses. One week after transduction, the cells were cocultured at a 1:1 E:T ratio with monoallelic B cells (HLA-B*15:01 or HLA-A*01:01) and titrated with AR H875Y peptides (VQPIARELY, SVQPIARELY, and LLDSVQPIARELY) or AR WT peptides (VQPIARELH, SVQPIARELH, and LLDSVQPIARELH). The activation of T cells was evaluated based on (B) 4-1BB expression by flow cytometry staining and (C) IFNγ secretion by ELISA. Images are representative of two replicates. The normality of the data distribution was confirmed using the Shapiro–Wilk test (P > 0.05). Levene’s test for homogeneity of variance was not violated (P > 0.05), and statistical significance was determined using an unpaired two-tailed Student t test. Specific P values are indicated: *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001.

AR H875Y–Specific Reactivity of the Identified TCRs

Having established that all our cloned TCRs react to their respective neoantigens, we sought to further characterize their functional reactivity, sensitivity, and specificity toward their target. To this end, we retrovirally infected T cells from two healthy donors with each of the TCRs and evaluated their mTCR expression and binding to AR H875Y dextramer by FACS staining (Supplementary Fig. S4A–S4C). Next, we cocultured transduced T cells with monoallelic B cells pulsed with mutant or WT peptides. We assessed their responses by measuring 4-1BB expression via flow cytometry and IFNγ secretion by ELISA across a range of peptide concentrations. The AR H875Y–derived neopeptides elicited robust immune responses, with a significant upregulation of 4-1BB and IFNγ production at low concentrations. Specifically, T157.1 had a substantial difference in 4-1BB and IFNγ responses at 10⁻² μmol/L, whereas T157.3 and T112.1 demonstrated significant differences at 10⁻¹ μmol/L for both markers (Fig. 4B and C; Supplementary Fig. S4D and S4E). Notably, T157.1-transduced T cells displayed a small background elevation in 4-1BB expression, regardless of whether DMSO or the WT peptide control was pulsed. This elevation was not apparent in IFNγ production or NFAT activity in Jurkat cells, and the response to the neopeptide remained significantly higher than in controls, confirming specificity and heightened sensitivity.

To rigorously evaluate the specificity and potential cross-reactivity of our TCR candidates, we performed coculture assays using monoallelic B cells expressing either HLA-B*15:01 or HLA-A*01:01. Monoallelic B cells were pulsed with each of the three neopeptides, and TCR reactivity was measured by flow cytometry of 4-1BB (Supplementary Fig. S5A). T112.1 recognized only the 13-mer peptide presented by HLA-A*01:01. In contrast, T157.1 responded to both the 10- and 13-mer peptides, but only in the context of HLA-B*15:01, whereas T157.3 was primarily specific for the 9-mer peptide, with modest reactivity to the 10- and 13-mer peptides, also restricted to HLA-B*15:01. No cross-reactivity was observed for mismatched HLA alleles. These findings confirm that although some flexibility in peptide length exists, each TCR maintains strict HLA restriction and limited cross-recognition.

To confirm the specificity and sensitivity of our initial functional assays, we validated these findings using FACS staining for three different markers: IFNγ, TNFα, and 4-1BB. T157.1, T157.3, and T112.1 exhibited significant increases in IFNγ, TNFα, and 4-1BB levels (Fig. 5A; Supplementary Figs. S4C–S4E and S5B–S5D), whereas T157.2 only exhibited significant increases in IFNγ and TNFα levels (Supplementary Fig. S6A).

Figure 5.

Reactivity of specific TCRs to AR H875Y using healthy donors’ PBMCs. A, Healthy donors’ PBMCs were transduced with TCR-MSGV1 (T157.1, T157.3, and T112.1) plasmids using retroviruses. One week after transduction, the cells were cocultured with monoallelic B cells (HLA-A*01:01 or HLA-B*15:01) pulsed with 1 μg/mL AR H875Y peptides (VQPIARELY, SVQPIARELY, or LLDSVQPIARELY) or AR WT peptides (VQPIARELH, SVQPIARELH, or LLDSVQPIARELH) at a 1.5:1 E:T ratio. Flow cytometry analysis was used to evaluate the activation of T cells based on IFNγ secretion, TNFα secretion, and 4-1BB expression. B, The LNCaP cancer cell line harboring the HLA-A*01:01 allele was stably transduced with the AR H875Y MNG and HLA-B*15:01 alleles (LNCaP^{AR H875Y}, LNCaP^{AR H875Y/B*15:01}). The cells were cocultured at a 1:1 E:T ratio with T cells stably expressing the AR H875Y–specific TCRs, followed by flow cytometry analysis of activation markers, including IFNγ secretion, TNFα secretion, and 4-1BB expression. C, LNCaP cells were stably transduced with AR H875Y full-length cDNA and the HLA-B*15:01 allele (LNCaP^{AR H875Y}, LNCaP^{AR H875Y/B*15:01}). The cells were cocultured at a 1:1 E:T ratio with T cells stably expressing the AR H875Y–specific TCRs, followed by flow cytometry analysis of 4-1BB expression. D, Diagram illustrating the CRISPR knock-in. LNCaP cells were cotransfected with a plasmid encoding Cas9-GFP, single-guide RNA (sgRNA) targeting the AR gene, and an oligonucleotide repair template for homology-directed repair. After 48 hours, GFP-expressing cells were isolated as single cells in individual wells of a 96-well plate using FACS. Single cells were expanded into clonal populations. Genomic DNA was extracted from each clone, and Sanger sequencing was performed to verify the incorporation of the intended knock-in mutation. A schematic was created using BioRender.com. E, CRISPR KI AR H875Y LNCaP cells (SCC18) and control cells (SCC8) were stably transduced with the HLA-B*15:01 allele. The cells were cocultured at a 1:1 E:T ratio with T cells stably expressing the AR H875Y–specific TCRs, followed by flow cytometry analysis of 4-1BB expression. F, C4-2 cells were stably transduced with AR H875Y full-length cDNA and the HLA-B*15:01 allele (C4-2^{AR H875Y}, C4-2^{AR H875Y/B*15:01}). The cells were cocultured at a 1:1 E:T ratio with T cells stably expressing the AR H875Y–specific TCRs, followed by flow cytometry analysis of 4-1BB expression. G, The 22Rv1 cancer cell line harboring the AR H875Y mutation was stably transduced with the HLA-B*15:01 allele. The cells were cocultured at a 1:1 E:T ratio with T cells stably expressing the AR H875Y–specific TCRs, followed by flow cytometry analysis of activation markers, including IFNγ secretion, TNFα secretion, and 4-1BB expression. The plots represent ≥2 biological replicates. The normality of the data distribution was confirmed using the Shapiro–Wilk test (P > 0.05). Levene’s test for homogeneity of variance was not violated (P > 0.05), and statistical significance was determined using an unpaired two-tailed Student t test. Specific P values are indicated: *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001.

To further evaluate our TCR candidates’ functional reactivity and cytotoxic potential, we investigated their ability to recognize and respond to prostate cancer cell lines that endogenously process and present relevant neoantigens. We coincubated TCR-transduced T cells, specifically T157.1, T157.3, and T112.1, with LNCaP cells expressing the AR H875Y mutation using an MNG construct and overexpressing HLA-B*15:01. In these assays, all three TCRs (T157.1, T157.3, and T112.1) induced robust activation, as indicated by significant increases in IFNγ, TNFα, and 4-1BB expression (Fig. 5B). In contrast, T157.2 did not show appreciable stimulation of these markers under the same conditions (Supplementary Fig. S6B). Although T112.1 exhibited statistically significant activation, the magnitudes of IFNγ and TNFα responses were consistently lower than those of T157.1 and T157.3. This reduced functional potency led us to focus subsequent validation efforts on T157.1 and T157.3, as these TCRs demonstrated the strongest reactivity across all tested parameters.

Subsequently, we assessed T-cell reactivity in several additional model systems to confirm and extend our findings; however, as prostate cancer has limited existing models, we pursued several approaches. To validate the processing and presentation of the AR H875Y peptide from the full-length protein, we overexpressed either the full-length AR H875Y or the WT gene together with HLA-B*15:01. We observed a consistent upregulation of 4-1BB in T157.1 and T157.3 only in cells expressing the mutated version of the gene (Fig. 5C). We then employed an additional system, with physiologically relevant AR transcription levels, by CRISPR knock-in of the AR H875Y mutation (SCC18) or AR WT (SCC8) to LNCaP cells, each with or without HLA-B*15:01 transduction. Notably, only the combination of mutant AR and HLA-B*15:01 resulted in significant 4-1BB upregulation (Fig. 5D and E).

We further validated our results in the C4-2 prostate cancer cell line, a castration-resistant LNCaP derivative, expressing full-length AR H875Y or WT AR and HLA-B*15:01 positive. In this system, T157.1 and T157.3 induced 4-1BB responses only to the mutated AR, indicating their capacity to recognize antigens in a model relevant to advanced prostate cancer (Fig. 5F). Finally, we extended our analysis to the 22Rv1 cell line, which endogenously expresses the AR H875Y mutation and was transduced with HLA-B*15:01. The TCR-transduced T cells again demonstrated robust activation of all three markers: IFNγ, TNFα, and 4-1BB. The robust recognition and activation observed with T157.1 and T157.3 TCRs highlight their potential utility for targeting AR H875Y–expressing prostate cancer cells (Fig. 5G).

AR H875Y–Specific TCRs Kill Endogenously Processed AR Mutant Cancer Cells, In Vitro and In Vivo

To examine the cytotoxicity of these TCRs toward LNCaP cells, we used an active cleaved caspase-3 staining assay. T157.1 and T157.3 demonstrated a significant cytotoxic capacity toward LNCaP^{H875Y/B*15:01} cells overexpressing the MNG and endogenously mutated 22Rv1^B*15:01 castration-resistant cells (Supplementary Fig. S7A–S7C).

To further assess the cytotoxic capacity of TCRs toward neoantigen-expressing prostate cancer cells, we utilized an S3 Incucyte live-cell imaging assay, which enables continuous, longitudinal, real-time monitoring of the killing of prostate cancer cells by T157.1- and T157.3-transduced T cells at varying effector-to-target (E:T) ratios. In the LNCaP^{H875Y/B*15:01/GFP} MNG model, both TCRs induced significant cytotoxicity, as demonstrated by the marked reduction in viable cancer cells (Supplementary Fig. S7D and S7E). In addition, we assessed T157.1 and T157.3 cytotoxicity in the C4-2 prostate cancer cell line expressing HLA-B*15:01 and the full-length mutated or WT AR. Both TCRs mediated specific target cell killing compared with the WT transduced cells although the reduction in viable C4-2 cells was less pronounced than that in the LNCaP MNG model (Supplementary Fig. S7F and S7G). Thus, we evaluated the cytotoxic activity of T157.1 and T157.3 in additional prostate cancer cell lines that produce the AR H875Y peptides at physiologic levels, including 22Rv1 (Fig. 6A and B), LNCaP cells overexpressing full-length AR (mutant and WT; Fig. 6C and D), and LNCaP CRISPR knock-in lines (SCC18 AR H875Y and SCC8 AR WT; Fig. 6E and F), all expressing HLA-B*15:01. In these models, both TCRs mediated potent dose-dependent cytotoxicity, with a pronounced reduction in viable target cells observed within 72 hours, even at low E:T ratios. Minimal killing was observed in WT and HLA-mismatched controls, confirming the specificity of the response. This finding highlights the potent and specific cytotoxic potential of the identified TCRs in AR H875Y prostate cancer cells.

Figure 6.

Cytotoxicity of specific TCRs to AR H875Y using healthy donors’ PBMCs. Transduced GFP cells were seeded in triplicate in 96-well plates, with each well containing 3,000–5,000 cells. The next day, T cells expressing the AR H875Y–specific TCR were added at different E:T ratios with 300 IU/mL IL-2. Cellular interactions were monitored using an Incucyte SX3 system over 4 days, with images captured every 4 hours. A, 22Rv1 cells (5,000 cells/well ± HLA-B*15:01) with T157.1. B, 22Rv1 cells (5,000 cells/well ± HLA-B*15:01) with T157.3. C, AR full-length LNCaP cells (3,000 cells/well, HLA-B*15:01, H875Y, or WT) with T157.1. D, AR full-length LNCaP cells (3,000 cells/well, HLA-B*15:01, H875Y, or WT) with T157.3. E, LNCaP CRISPR KI (3,000 cells/well, SCC18 AR H875Y ± HLA-B*15:01, SCC8 AR WT + HLA-B*15:01) with T157.1. F, LNCaP CRISPR KI cells (3,000 cells/well, SCC18 AR H875Y ± HLA-B*15:01, SCC8 AR WT + HLA-B*15:01) with T157.3. G, Experimental timeline for NSG mouse studies. NSG mice were subcutaneously inoculated with 5 × 10⁶ 22Rv1-B*15:01 tumor cells, randomized, and treated as indicated. Tumor growth was monitored for 24 days. H, NSG mice bearing 22Rv1-B*15:01 tumors were treated with T cells expressing T157.1, T157.3, or an irrelevant TCR (N = 6 for each group). Images are representative of ≥2 biological replicates.

Finally, adoptive cell transfer experiments were performed in NOD.Cg⁻ Prkdc^scid Il2rg^tm1Wjl/SzJ (NSG) mice to assess the in vivo efficacy of the TCR candidates. NSG mice were subcutaneously inoculated in the right flank with 22Rv1 prostate cancer cells expressing HLA-B*15:01. Four days after tumor implantation, T cells transduced with either T157.1 or T157.3 TCRs were administered intravenously, and IL-2 was provided by s.c. injection on days 4, 7, and 9 (Fig. 6G). The control arm included mice treated with an irrelevant TCR transduced in parallel. Mice treated with T157.1 or T157.3 TCR-transduced T cells showed a marked reduction in tumor size compared with those receiving T cells expressing an irrelevant TCR (Fig. 6H). These results highlight the ability of T157.1 and T157.3 TCRs to selectively identify and eliminate target cells in vivo, further supporting their potential as targeted immunotherapies for mCRPC.

Discussion

In this study, we established a new data-driven approach, SpotNeoMet, to systematically identify recurrent clonal neoantigens derived from treatment-resistant driver mutations. We demonstrated the power of our approach by identifying and validating numerous treatment-resistant neoantigens, including those derived from the common AR H875Y mutation, which is associated with resistance in prostate cancer. We further identified cognate TCRs for AR H875Y neoantigens, demonstrating that they are highly sensitive and specific to their targets and cytotoxic in vitro and in vivo.

Methods for identifying recurrent neopeptides have evolved significantly in recent years. Early approaches focused on specific recurrent mutations or HLA allele frequencies and provided few successful targets. However, several challenges and limitations remain. First, improved methods are needed to prioritize and predict the identified neoantigens. Indeed, current prediction algorithms are not optimized to identify all neopeptides (7). A second challenge is that some data-driven approaches do not consider neoantigen mutation clonality, resulting in a lack of applicability for immunotherapy (25). Third, not all neopeptides are derived from driver genes. These mutations may not directly contribute to oncogenesis, and the expression levels of nondriver genes may vary across different cancer cells. Finally, no pipeline has yet been established to identify therapy-resistant neopeptides specifically.

To address these challenges, this study established a SpotNeoMet pipeline. Our systematic and unbiased in silico approach for neoantigen prioritization combines three prediction methods. It leverages the power of a large dataset and comprehensive genomic coverage, including whole-genome and whole-transcriptome sequencing. This approach improves prediction accuracy and allows high-quality annotation of genomic variants, including variant clonality and mutation-specific expression. Analyzing treatment-naïve patients facilitates the identification of treatment resistance–associated variants that can be selectively targeted. This allowed us to identify targets derived from recurrent mutations in ESR1, SMAD4, EGFR, and AR in breast cancer, colorectal cancer, non–small cell lung cancer, and metastatic prostate adenocarcinoma, respectively. These mutations collectively encompassed ∼8% of all patients with these cancer types in the HMF cohort. Among these targets, we focused on AR mutations in prostate cancer because of their critical role in disease progression and therapy resistance. AR H875Y is a mutation in the ligand-binding domain of AR that enables ligand-independent, constitutive activation of AR; maintains oncogenic signaling even without ligand–receptor interaction; and drives resistance to AR inhibitors. Although it is infrequently identified in treatment-naïve patients, it is commonly found in mPCs upon progression to AR pathway inhibitors (26) when patients with mCRPC have limited therapeutic options.

Importantly, although our computational pipeline predicted 34 candidate HLA-neoantigen pairs derived from AR mutations, HLA-I immunopeptidomic validation identified only three mutant and three WT peptides. This discrepancy highlights both the limitations of current neoantigen prediction tools and the critical value of integrating MS-based validation. By directly identifying peptides present on tumor cells, immunopeptidomics provide a more accurate framework for target discovery than computational predictions alone. Enhancing both the predictive accuracy of peptide–HLA binding models and the sensitivity of HLA-immunopeptidomics could significantly improve the neoantigen identification workflow (27). All three validated peptides in our study conformed to the canonical motifs of HLA-A*01:01 or HLA-B*15:01, exhibiting expected anchor residues and optimal lengths of 9 to 13 amino acids. Moreover, the identified peptides demonstrated stronger binding than previously reported canonical binders, as our T2 binding assay shows. However, stable presentation of HLA is insufficient to ensure immunogenicity; a matching TCR must exist and engage the peptide–MHC complex with sufficient affinity, an unpredictable step. Current models of TCR–peptide–MHC recognition perform near-randomly when challenged with novel epitopes (28), underscoring the necessity of empirical screening. Notably, TCR recognition of neoantigens is typically private and distinct for each donor (29). Accordingly, we used in vitro stimulation of PBMCs from healthy donors to functionally isolate potent mutation-specific TCRs. This approach enabled the identification of rare, high-affinity TCRs that can recognize and kill cancer cells harboring endogenous mutations.

Our results have strong clinical implications, particularly given that mCRPC remains the second leading cause of cancer death in men worldwide (1). mCRPC is often considered an immunologically cold tumor, marked by a low mutational burden and minimal infiltration by effector T cells (30). Various immunotherapy approaches have been explored, including sipuleucel-T (31), PD-1 and CTLA-4 immune checkpoint inhibitors (32), chimeric antigen receptor T-cell therapies (33), and bispecific T-cell engagers (34). However, each demonstrated limited and inconsistent clinical benefits. To date, only TCR-based therapies evaluated for prostate cancer have focused on nonmutated proteins that are highly expressed in tumors but are absent or barely detectable in normal tissues (35). However, because these targets are not derived from tumor-specific mutations, immune tolerance may limit the strength and durability of the response. In contrast to previous strategies, our approach targets a distinct class of molecules, recurrent gene mutations, that drive resistance to therapy. These alterations are not only prevalent across patients but are also critical for tumor survival, particularly in advanced treatment-refractory disease. Because tumors depend on these mutations to evade standard therapies, they remain consistently expressed under therapeutic pressure, making them promising and stable targets for TCR-based immunotherapy. This approach enables the development of TCR therapies that exploit tumor escape mechanisms, transforming essential resistance pathways into selective and durable points of immune intervention in mCRPC. Interestingly, recent findings have demonstrated that AR activity in CD8⁺ T cells directly inhibits T-cell function, particularly by repressing IFNγ production (36). Combining our identified TCRs with AR inhibition in future therapeutic approaches may potentially yield synergistic effects. This hypothesis is supported by the fact that AR blockade improves responsiveness to PD-1–targeted therapy (36). Thus, combining our identified TCRs, AR inhibitors, and checkpoint inhibitors may be a promising strategy for future studies, potentially overcoming multiple immune evasion mechanisms simultaneously.

In conclusion, our identification of the AR H875Y neoantigens in mCRPC demonstrates the robustness of our novel SpotNeoMet pipeline. Our data further suggest that resistance drivers can be turned into therapeutic opportunities. Indeed, SpotNeoMet can be extended to other mutations and cancer types with similar therapeutic resistance. Such innovative approaches are critical for addressing the challenges posed by advanced-stage cancers.

Methods

Cell Lines and PBMCs

Human Epstein–Barr virus (EBV)-transformed B721.221 (IHW00001; RRID:CVCL_6263) was purchased from the Fred Hutch International Histocompatibility Working Group. The human cell line LNCaP (RRID:CVCL_0395) was a generous gift from Michael Elkin. The human cell lines 22Rv1 (RRID: CVCL_4Y35) and C4-2 (RRID:CVCL_4782) were a generous gift from Joaquin Mateo. All the cell lines were cultured in RPMI 1640 (Gibco) supplemented with 10% heat-inactivated FBS (Gibco), 2 mmol/L L-glutamine, 2.5% HEPES, 100 IU/mL penicillin, 0.1 mg/mL streptomycin, and 1 mmol/L sodium pyruvate. Jurkat TCRαβ-knockout CD8⁺ cell line (cat. #GA1162) was purchased from Promega and maintained in RPMI 1640 (Gibco) supplemented with 10% FBS (Gibco), 1% sodium pyruvate, 28 mmol/L HEPES, 1% minimum essential medium nonessential amino acids, 400 μg/mL hygromycin B, and 10 μg/mL blasticidin. The human embryonic kidney HEK293T cell line (CRL-3216, RRID:CVCL_0063) was obtained from the ATCC and maintained in DMEM (Gibco) supplemented with 10% FBS (Gibco, RRID:SCR_013780), 1% L-glutamine, 1% penicillin–streptomycin, and 1% sodium pyruvate. The 293GP packaging cell line (RRID:CVCL_H716), a generous gift from Michal Lotem, was maintained in DMEM (Gibco, Thermo Fisher Scientific, RRID:SCR_021173) supplemented with 10% FBS (Gibco, RRID:SCR_013780), 1% L-glutamine, 1% penicillin–streptomycin, and 1% sodium pyruvate. The human TAP-deficient T2 lymphoblast cell line (CRL-1992, RRID:CVCL_2214) was obtained from ATCC and maintained in RPMI 1640 medium (Gibco) supplemented with 10% FBS (Gibco, RRID:SCR_013780), 2.5% HEPES, 1% L-glutamine, 1% penicillin–streptomycin, and 1% sodium pyruvate. All cells were cultured at 37°C in humidified 5% CO₂ and tested routinely for Mycoplasma using a Mycoplasma EZ-PCR test kit (20-700-20, Biological Industries).

Human PBMCs were isolated from blood donations obtained after written informed consent from all donors, in accordance with Weizmann Institute of Science ethical guidelines and with approval from the Institutional Review Board (2274-2).

Human PBMCs were extracted using Ficoll-Paque PLUS (GE HealthCare) via density gradient centrifugation and were freshly used or cryopreserved in aliquots in human serum with 12% DMSO and thawed as needed per experiment.

HLA Typing of Healthy Donors’ PBMCs and Cell Lines

High-resolution HLA typing was performed on genomic DNA extracted from PBMCs of healthy donors, prostate cancer cell lines, and EBV-transformed B721.221 cells to determine the HLA-A, HLA-B, and HLA-C alleles. Typing was performed using the MX6-1 next-generation sequencing kit (GenDx) following the protocols provided by the manufacturer. Sequencing libraries were prepared and run as paired-end reads (2 × 150 bp) on the Illumina iSeq 100 platform. FASTQ files were analyzed using NGSengine software version 2.13 (GenDx), with reference to the IPD-IMGT/HLA database version 3.33.0. HLA allele assignments were finalized following manual curation and confirmation of next-generation sequencing–based genotype calls. A summary of HLA-I genotypes for the cell lines used in this study is provided in Supplementary Table S11.

CRISPR-Cas9–Mediated Knock-In

A single-guide RNA (5′- CCATGTGTGACTTGATTAGC -3′) targeting the human AR gene was designed to knock in the H875Y mutation cloned into the pSpCas9(BB)-2A-GFP (PX458) vector, using standard BbsI cloning sites. pSpCas9(BB)-2A-GFP (PX458) was gifted by Feng Zhang (Addgene, plasmid #48138; http://n2t.net/addgene:48138; RRID:Addgene_48138). A 128-nucleotide single-stranded DNA oligonucleotide (ssODN) repair template was synthesized (IDT) containing the desired H875Y knock-in mutation flanked by ∼64-nucleotide homology arms on either side of the Cas9 cut site. To prevent recutting by Cas9 after homology-directed repair, ssODNs were designed to introduce silent mutations that disrupt the protospacer adjacent motif at three separate sites within the region of interest (5′- CTGTTTTTCTCCCTCTTATTGTTCCCTACAGATTGCGAGAGAGCTGCATCAGTTCACTTTTGACCTGCTAATCAAGTCACACATGGTGAGCGTGGACTTTCCGGAAATGATGGCAGAGA -3′). LNCaP cells were cotransfected with 2 μg of PX458 plasmid and 40 μmol/L of ssODN using Lipofectamine 2000 (Thermo Fisher Scientific), following the protocol described by Ran and colleagues (37). Forty-eight hours after transfection, GFP-positive cells were isolated by FACS using a CytoFLEX SRT Cell Sorter (Beckman Coulter Life Science) and distributed into 96-well plates to obtain single-cell clones. Genomic DNA from expanded clones was extracted using the DNeasy Blood & Tissue Kit (QIAGEN), and PCR amplification of the target region spanning the knock-in site was performed with gene-specific primers (forward: 5′-GAGGACCAAGGAAGTACGGG-3′, reverse: 5′-GGGTGTGGAAATAGATGGGC-3′) to generate a 248-bp amplicon. The PCR products were purified using the Wizard Genomic DNA Purification Kit (Promega) and subjected to Sanger sequencing to verify the precise knock-in of the mutation.

Tumor Somatic Mutations

Cancer somatic mutations were obtained from (i) primary tumor whole-exome sequences from TCGA (10) and (ii) metastatic whole-genome sequences from the HMF (38), comprising a total of 9,644 and 3,235 samples, respectively, after discarding hypermutators within each cancer cohort (39). Somatic mutations from different cancer types were merged into two pan-cancer files representing primary and metastatic tumors. Somatic mutations were liftovered from hg19 to hg38 using pyliftover version 0.3 (https://pypi.org/project/pyliftover), and those failing to liftover were discarded. A total of 1,547,627 and 65,139,581 mutations in primary and metastatic tumors, respectively, were analyzed.

Identification of Recurrent Neoantigens

We first identified all recurrent mutations (or hotspots) in primary and metastatic pan-cancer cohorts (10, 38) using our in-house developed HotspotFinder (version 1.0; ref. 8). We used the somatic mutation calls provided by the original sources as inputs for hotspot identification using HotspotFinder. HotspotFinder identifies unique genomic positions that are recurrently mutated to the same alternate (e.g., 2 C>T transitions) across tumors and annotates the overlap of these positions with genomic data, including genome mappability, population variants, and genomic elements. Hotspots were identified as recurrent mutations of equal alternates (split_alternates = True) in three or more samples. All other parameters were maintained at their default values. To remove potential artifacts, hotspots overlapping low-mappability regions, population variants, or those falling outside protein-coding sequences of Ensembl canonical transcripts were discarded (8). To identify which of the resulting hotspots generated missense mutations, we ran Variant Effect Predictor version 92 through bgvep (https://bitbucket.org/bgframework/bgvep/src/master/) and retained those hotspots creating missense variants in canonical transcripts.

Hotspots were then prioritized for their potential to encode neoepitopes using the following criteria:

1. HLA-I presentation: A list of all possible neoantigens with lengths between 8 and 14 amino acids overlapping each hotspot position was derived using the protein sequence of the canonical transcript. A matched WT sequence was generated for each mutant peptide. For each neoantigen and WT peptide, HLA-I presentation was validated across all HLA-I alleles of the matched mutated patients using three independent tools: MHCflurry version 2.0 (40), NetMHCpan version 4.1 (15), and MHCnuggets version 2.3 (41). Candidate neoantigens were defined as those peptides and HLA-I allele pairs classified as binders by MHCflurry 2.0 (40) and an additional prediction algorithm. Binders were pairs with predicted binding affinities lower than 500 nmol/L. Neoantigens for which the WT version was predicted to bind to the same HLA-I allele were discarded. HLA-I germline allotypes have been reported in LILAC (42). Neoantigens binding to at least one of the following high-frequency HLA-I alleles in the Hartwig medical dataset population were kept: HLA-A*01:01, HLA-A*02:01, HLA-A*03:01, HLA-A*11:01, HLA-A*24:02, HLA-B*07:02, HLA-B*08:01, HLA-B*15:01, HLA-B*35:01, HLA-B*40:01, HLA-C*03:03, HLA-C*03:04, HLA-C*04:01, HLA-C*05:01, HLA-C*06:02, HLA-C*07:01, HLA-C*07:02, and HLA-C*12:03.

2. Hotspot prevalence in the baseline cohort of prebiopsy treated patients with cancer: Candidate neoantigens passing the previous steps were then ranked based on their source alteration prevalence matched with the predicted HLA-I allotypes across the cohort of metastatic treated patients. The prevalence of grouped alterations was estimated across all patients with HLA-I alleles predicted to present the candidate neoantigen.

3. Clonality: Mutation clonality, which is an essential factor in triggering strong and durable immune responses, was considered. Subclonal single-nucleotide variants were filtered based on estimates from the PURPLE tool (9).

4. Potential role in tumorigenesis: Driver predictions (i.e., driver or passenger) were retrieved from BoostDM (11, 43). Hotspots were annotated as drivers if the BoostDM score was ≥0.5 in a cancer type. Passenger mutations were excluded.

5. Neoantigen-specific RNA expression: For each selected hotspot mutation, RNA expression of the corresponding mutated gene in the given sample (if available) was annotated using the Isofox tool (https://github.com/hartwigmedical/hmftools/blob/master/isofox/README.md). This information was then filtered to exclude neoantigens derived from nonexpressed genes in the corresponding mutated sample hotspot mutations.

Study of AR H875Y Frequency and HLA Presentation in mCRPC

We retrieved data from three publicly available datasets of mCRPC: Abida and colleagues (SU2C/PFC cohort; ref. 20), Quigley and colleagues (21), and Herberts and colleagues (22). AR H875Y somatic mutations and clinical data from the SU2C/PFC dataset were downloaded from cBioPortal (www.cbioportal.org, accessed on March 27, 2025). For the Quigley and Herberts datasets, AR H875Y somatic mutations and clinical data were retrieved from the supplementary materials in the original publications (21, 22). The Hartwig prostate cancer cohort (38) was also included in this analysis.

For the Quigley and colleagues, Herberts and colleagues, and Hartwig cohorts, HLA-I typing was performed using LILAC version 1.5. For the SU2C/PFC cohort, HLA-I types were kindly provided by the authors and obtained using OptiType, with DNA as the input.

To estimate the frequency of AR H875Y mutations in each cohort, only one sample per donor was analyzed (i.e., duplicate samples per donor were removed). Only donors with available HLA types were used in this analysis (Supplementary Tables S4–S7). Sample numbers, tumor type distributions, and metastatic biopsy locations for the Hartwig cohort, TCGA, and the three publicly available datasets of mCRPC are detailed in Supplementary Tables S12–S15.

Assessment of the Expression of Genes in Normal Tissues

We extracted the expression levels (transcripts per million) of AR, ESR1, EGFR, and SMAD4 from the Genotype-Tissue Expression project (44).

Lentivirus Production

To generate lentiviruses, HEK293T cells were seeded in 10 cm² dishes and cultured to approximately 70% to 80% confluency. Cells were then cotransfected with 4.4 µg of pCDH-CMV-MCS-EF1α-Puro-GFP (System Biosciences, CD513B-1) or pCDH-CMV-MCS-EF1α-Neo (System Biosciences, CD514B-1) transfer plasmid, 2.2 µg of pMD2.G envelope plasmid (pMD2.G was a gift from Didier Trono, Addgene, plasmid # 12259; http://n2t.net/addgene: 12259; RRID:Addgene_12259), and 3.4 µg of psPAX2 packaging plasmid (psPAX2 was a gift from Didier Trono, Addgene, plasmid # 12260; http://n2t.net/addgene:12260; RRID:Addgene_12260) using the TurboFect Transfection Reagent (Thermo Fisher Scientific, RRID:SCR_015569), according to the manufacturer’s protocol. After 72 hours, the supernatant containing lentiviral particles was harvested, clarified by filtration through a 0.22 µm filter, aliquoted, and stored at −80°C until use.

Generation of B-cell Lines Coexpressing HLA Alleles and MNG

Synthetic optimized recombinant DNA sequences of HLA-I alleles were purchased from Twist Bioscience and cloned into a lentiviral system pCDH-CMV-MCS-EF1α-Neo (CD514B-1, System Biosciences) to generate stable monoallelic overexpression in the HLA-I–deficient B-cell line B721.221, as previously described (7). Following selection with 800 µg/mL G418 Geneticin (11811-031, Thermo Fisher Scientific), transduced B721.221 cells were labeled with anti–HLA-I Allophycocyanin(APC)-conjugated antibodies and evaluated for HLA-I surface expression using the W6/32 antibody (BioLegend, cat. #311410, RRID:AB_314879). Data were analyzed using Kaluza software (RRID:SCR_016182). MNGs were designed as 47-mer peptides encompassing hotspot driver mutations (AR H875Y, AR T878A, AR L702H, EGFR T790M, ESR1 D538G, ESR1 Y537S, and SMAD4 P356L). These fragments were purchased from Twist Bioscience and cloned into the lentiviral system pCDH-CMV-MCS-EF1α-GFP-Puro (CD513B-1, SBI). Stable B721.221 monoallelic cells were transduced with lentiviruses containing one of the MNGs. The cells were then selected using 1 µg/mL puromycin for 7 days and subsequently verified for GFP expression by flow cytometry.

Generation of Cell Lines Coexpressing HLA-B*15:01 and AR H875Y

Synthetic codon-optimized full-length cDNA sequences of AR WT and AR H875Y were synthesized using GenScript and cloned into the lentiviral vector pCDH-CMV-MCS-EF1α-Puro-GFP (CD513B-1, System Biosciences). Both constructs were used to transduce LNCaP and C4-2 prostate cancer cells, followed by selection with 1.5 μg/mL puromycin for 7 days. HLA-B*15:01 was used to transduce 22Rv1 cells and all AR-overexpressing LNCaP and C4-2 cell lines. Transduced cells were selected with 1,500 to 2,000 μg/mL Geneticin (G418, Thermo Fisher Scientific) and verified by flow cytometry using an APC-conjugated anti–HLA-I antibody (W6/32, BioLegend, cat. #311410, RRID:AB_314879). AR H875Y MNG was introduced into parental LNCaP cells, and the cells were transduced with HLA-B*15:01. LNCaP SCC18 CRISPR AR H875Y knock-in cells and LNCaP SCC8 CRISPR control cells were transduced with HLA-B*15:01. The cells were subjected to puromycin selection (1.5–2 μg/mL for 7 days) and evaluated for GFP positivity by flow cytometry. A summary of all cell lines, their transgenes, and the corresponding antibiotic resistance markers (puromycin or neomycin) is provided in Supplementary Table S11.

Purification and Isolation of Membrane HLA Peptides by Immunopeptidomics

Cell pellets consisting of 2 × 10⁸ cells were homogenized and lysed on ice with lysis buffer containing 1% octyl-β-D-glucopyranoside, 0.25% sodium deoxycholate, 1 mmol/L EDTA, 0.2 mmol/L iodoacetamide, 1:200 Protease Inhibitor Cocktail (P8340, Sigma-Aldrich), and 1 mmol/L phenylmethylsulfonyl fluoride in PBS, as previously described (7). Lysates were cleared by centrifugation at 48,000 × g for 45 minutes at 4°C and passed through a preclearing column containing Protein-A Sepharose beads (GenScript). HLA–peptide complexes were then immunoaffinity-purified from the cleared lysate using a W6/32 pan–HLA-I antibody (purified from HB95 hybridoma cells). The HLA–peptide complexes were then eluted with 1% trifluoroacetic acid (TFA), followed by separation of the peptides from the HLAs using a Sep-Pak tC18 100 mg sorbent 96-well plate (Waters). The peptide fraction was eluted with 28% acetonitrile (ACN) in 0.1% TFA. The HLA-eluted peptides were dried and resolubilized using vacuum centrifugation with 0.1% formic acid.

Identification of Eluted HLA Peptides by MS

For Orbitrap MS/MS experiments, the peptides were separated by reversed-phase chromatography using the nanoAcquity system (Waters), with a Symmetry trap column (180 × 20 mm) and HSS T3 analytical column, 0.75 to 250 mm (Waters), mobile phase A: H₂O + 0.1% formic acid, and B: ACN + 0.1% formic acid. The peptides were separated with a linear gradient over 2 hours from 5% to 28% B, 28% to 35% in 15 minutes, 35% to 95% in 15 minutes, maintained at 95% for 10 minutes, and back to initial conditions at a flow rate of 0.35 μL minute⁻¹.

The LC was connected online via a nano-electrospray ionization source (Flex Ion, Thermo Fisher Scientific) using an emitter (Fossil) to either a quadrupole orbitrap MS (Q Exactive HF, Thermo Fisher Scientific) or a tribrid MS (Fusion Lumos, Thermo Fisher Scientific). Data were acquired using a data-dependent method, and the peptides were fragmented using higher-energy collisional dissociation. On the Q Exactive HF, full-scan MS spectra were acquired at a resolution of 120,000 at 200 m/z with an automated gain control (AGC) value of 3 × 10⁶ ions, a mass range of 300 to 1,800 Th, and a maximum injection time of 100 ms. MS/MS scans were acquired with an AGC target value of 10⁵ with a maximum injection time of 150 ms, isolation of 1.7 Th, normalized collision energy of 30%, and MS/MS resolution of 15,000 at 200 m/z. The fragmented m/z values were dynamically excluded from further selection for 20 seconds.

Full-scan MS spectra were acquired at a resolution of 120,000 at 200 m/z with an AGC value of 200%, a mass range of 300 to 1,800 Th, and a maximum injection time set to auto. MS/MS scans were acquired with an AGC target value of 100%, a maximum injection time of 150 ms, an isolation of 1.7 Th, a normalized collision energy of 27%, and an MS/MS resolution of 15,000 at 200 m/z. The fragmented m/z values were dynamically excluded from further selection for 20 seconds.

For the trapped ion mobility (TIMS) time-of-flight (TOF) mass spectrometry (timsTOF Pro 2; TTP, BRUKER) experiments, the peptides were resolubilized with 0.1% TFA and 5 mmol/L Tris (2-carboxy-ethyl)-phosphin-HCl before LC/MS-MS analysis. Five microliters of each sample were loaded using nanoElute 2 liquid chromatography (Bruker). Mobile phase A contained 0.1% formic acid in water, whereas mobile phase B contained 0.1% formic acid in ACN. Peptides were separated using an Aurora Ultimate C18 nano column (0.075 × 250 mm; IonOpticks), with a gradient of 2% B to 29% B in 80 minutes, then to 95% B in 0.5 minutes, and maintained at 95% B for 2.9 minutes at a flow rate of 300 nL/minute. The column was placed in a column toaster and connected to a CaptiveSpray electrospray ionization source. The column was maintained at 50°C. Data were acquired with a timsTOF Pro (Bruker) in data-dependent acquisition–parallel accumulation-serial fragmentation mode with the following parameters: capillary voltage of 1,600 V, temperature of 180°C, mass range of 100 to 1,700 Th, ion mobility of 0.6 to 1.57 1/K0, tims ramp time of 300 ms, number of parallel accumulation-serial fragmentation MS/MS scans 10, target intensity of 20,000 with a threshold of 2,500, charge range of 0 to 5, and collision energy of 20 at 0.6 1/K0 and 59 at 1.6 1/K0.

MS raw files were analyzed using MaxQuant software (version 2.1.3.0; ref. 45) with a 5% FDR. Peptides were searched against the UniProt human proteome database ID UP000005640 (June 20, 2021), to which we added the MNG peptide sequence MGSRRFYQLTKLLDSVQPIARELYQFTFDLLIKSHMVSVDFPEMMAE.

Validation Using Synthetic Peptides

For spectral validation, light synthetic peptides (high-purity high-performance liquid chromatography grade ≥85, GeneScript) were analyzed using the same LC/MS-MS system as endogenous peptides, with minor modifications to the gradient and collision energy settings. Specifically, the ACN gradient was adjusted from 4% to 30% over 20 minutes, and the normalized collision energy was set to 27 or 30 for HF2 and FS1, respectively. Data processing was conducted using MaxQuant software, with FDR set to 1% and the individual peptide mass tolerance disabled. The OrgMassSpecR R package (https://rdrr.io/rforge/OrgMassSpecR/man/SpectrumSimilarity.html) compared endogenous peptides with their synthetic counterparts. The synthetic spectrum with the highest MaxQuant score was selected for each peptide, whereas all endogenous spectra scoring above 60 in MaxQuant were included in head-to-tail comparisons. Only endogenous-synthetic peptide pairs with the same charge were used in our analysis. The analysis utilized default software parameters, and both the normalized dot product and Pearson correlation were evaluated. Statistical significance was determined using a P value of 0.05.

Peptide Stabilization Assay on HLA-Transduced T2 Cells

T2 cells were transduced with either HLA-B*15:01 or HLA-A*01:01 constructs and selected with 3,500 μg/mL Geneticin (G418; Thermo Fisher Scientific). Peptide stabilization assays were performed as previously described (46). T2-B*15:01 or T2-A*01:01 cells (1 × 10⁵) were incubated overnight at 37°C in RPMI 1640 medium supplemented with 50 μg/mL peptide (Supplementary Table S16), 0.5% DMSO (MP Biomedicals, cat. #0219605580), and 10 μg/mL human β2-microglobulin (ProSpec, cat. #PRO-553). The following day, the cells were washed with PBS, stained with LIVE/DEAD Fixable Green Viability Dye (Invitrogen, Thermo Fisher Scientific), and then incubated with anti–HLA-A/HLA-B/HLA-C antibody (W6/32, APC-conjugated, BioLegend, cat. #311410, RRID:AB_314879). The cells were then analyzed using flow cytometry.

Peptide–HLA Model Generation

To generate VQP/SVQ-HLA-B*15:01 models, we compiled peptide–HLA-B*15:01 (p-HLA-B*15:01) complex coordinates from the Protein Data Bank (PDB). The model consists of a nanometer peptide derived from the SARS-CoV-2 spike protein presented by HLA-B*15:01 (PDB: 8ELH; ref. 47). The spike peptide was manually mutated to a VQPIARELY neopeptide using PyMOL (Schrödinger, LLC, RRID:SCR_000305) mutagenesis. The rotamer orientation of each amino acid was selected to avoid steric hindrance by HLA-B*15:01. To accurately represent the decamer peptide conformation within the HLA-binding groove, the PDB coordinate (PDB: 3DX7), consisting of a peptide-derived Epstein-Barr virus nuclear antigen 6 presented by HLA-B*44:03 (48), was superimposed onto the p-HLA-B*15:01 coordinate. The Epstein-Barr virus nuclear antigen 6 decamer peptide was mutated to SVQPIARELY for model generation. To obtain the LLD-HLA-A*01:01 model, we superimposed MAGE-A1–HLA-A*01:01 (PDB: 3BO8; ref. 49) onto the BZLF1–HLA-B*35:08 (PDB: 1ZHL; ref. 50) complex to precisely model the putative superbulged conformation of the 13-mer peptide. BXLF1 peptide residues were then manually mutated to LLDSVQPIARELY for LLD-HLA-A*01:01 model generation.

T-cell Reactivity toward AR H875Y Neopeptides

The priming of naïve CD8 T cells from healthy donors with neopeptides was performed as previously described (19) with several modifications. Written informed consent was obtained from healthy blood donors under a protocol approved by the Institutional Review Board Ethics Committee (2274-2). Briefly, PBMCs were isolated from healthy donors’ blood using Ficoll gradient separation (Cytiva). On day −4, monocytes were isolated from PBMCs using CD14-reactive MicroBeads (Miltenyi Biotec) and cultured for 3 days in CellGro GMP DC medium (CellGenix) supplemented with 1% human serum (Valley Biomedical), 1% penicillin–streptomycin (dendritic cell-T cell medium) containing 10 ng/mL IL-4 (PeproTech), and 800 IU/mL GM-CSF (PeproTech). On day −1, monocyte-derived dendritic cells (DC) were matured for 16 hours with 800 IU/mL GM-CSF, 10 ng/mL IL-4, 10 ng/mL lipopolysaccharide (Sigma-Aldrich), and 5 ng/mL IFNγ (PeproTech) supplemented into the cultures. Autologous naïve T cells were isolated using Pan T Cell MicroBead (Miltenyi Biotec) and cultured overnight in DC-T medium containing 5 ng/mL IL-7 (PeproTech). On day 0, monocyte-derived DC were pulsed for 2 hours with 1 μg/mL AR H875Y neopeptides (VQPIARELY, SVQPIARELY, LLDSVQPIARELY) or 1 μg/mL AR WT peptide (VQPIARELH, SVQPIARELH, LLDSVQPIARELH). Subsequently, monocyte-derived DCs were cocultured with isolated naïve T cells in DC-T medium supplemented with 30 ng/mL IL-21 (PeproTech) at a DC/T-cell ratio of 1:2. On days 3, 5, and 7, half of the medium was replaced with fresh medium supplemented with 10 ng/mL of both IL-7 and IL-15 (PeproTech). On day 10, 25 IU/mL IL-2 (PeproTech) was added to the cytokine cocktail. On day 12, T cells were restimulated with irradiated (35 Gy) human B-LCL 721.221 cells expressing HLA-B*15:01 or HLA-A*01:01. Irradiated monoallelic B721.221 was pulsed for 2 hours with neopeptides or WT peptides and then cocultured with T cells in DC-T medium at a DC/T-cell ratio of 1:2. On days 14 and 17, half of the medium was removed and replenished with fresh medium supplemented with 10 ng/mL of both IL-7 and IL-15 and 25 IU/mL IL-2 or 50 IU/mL IL-2 (PeproTech), respectively. On day 19, the same restimulation and replenishment of cytokines were repeated until day 26. On day 26, T cells were collected and cocultured with pulsed B-LCL 721.221 cells expressing the corresponding HLA allele. Brefeldin and monensin were added to evaluate IFNγ and TNFα secretions. Following 16 hours of stimulation, the surface expression of 4-1BB and the secretion of IFNγ and TNFα were examined via flow cytometry. Cells were stained with the LIVE-DEAD Fixable Green Dead Cell Stain Kit (Invitrogen), CD3 (BioLegend, cat. #300308, RRID:AB_314044; BioLegend, cat. #300318, RRID:AB_314054), CD8 (BioLegend, cat. #300912, RRID:AB_314116), 4-1BB (BioLegend, cat. #309820, RRID:AB_2563830), IFNγ (BioLegend, cat. #502509, RRID:AB_315234), and TNFα (BioLegend, cat. #502932, RRID:AB_10960738). All experiments were performed using the CytoFLEX LX Flow Cytometer (Beckman Coulter). Data were analyzed using Kaluza software (RRID:SCR_016182).

Defining Peptide-Specific CD8 T-cell Subpopulation via Dextramer Staining

Following the stimulation and expansion of T cells derived from healthy donors’ PBMCs, T cells were collected and stained with AR H875Y 9-, 10-, and 13-mer loaded dextramers, according to the manufacturer’s protocol (Immudex). Briefly, naïve T cells that were previously exposed to mutant or WT peptides were stained using the LIVE-DEAD Fixable Green Dead Cell Stain Kit (Invitrogen). T cells were labeled with double AR H875Y dextramers (APC and PE), stained with anti-CD8 (BioLegend, cat. #301036, RRID:AB_10960142), and sorted by BD FACSAria Special Order Research Product (SORP) cell sorter for 10x Genomics. Cells were freshly forwarded to a 10x library preparation immediately after sorting.

Single-Cell TCR Sequencing

Sorted double and negative dextramer cells were processed immediately after sorting as described above for single-cell library construction. Briefly, sorted cells were washed, resuspended in PBS containing 0.04% BSA, and counted using trypan blue staining. Single cells were captured in droplets by loading onto a Chromium Controller, with a targeted cell recovery of 10,000 cells per sample. According to the manufacturer’s protocol, single-cell gene expression (GEX) and TCR-enriched libraries were prepared using the Chromium Single Cell 5’ V(D)J version 2 dual index kit (10x Genomics). Samples were sequenced on an Illumina NovaSeq 6000 sequencer with 26 cycles read1, 10 cycles i7 index, 10 cycles i5 index, and 90 cycles read2 (GEX and TCR were sequenced together on the same sequencing kit). The 800M reads were obtained in a 4:1 GEX library to TCR sequencing library ratio.

Data Processing of single-cell RNA and TCR Sequencing Libraries

Reads from 10x scRNA expression libraries were aligned to the human genome assembly GRCh38 (hg38) and quantified using the Cell Ranger count (10x Genomics, version 3.1.0). Filtered feature barcode matrices containing only cellular barcodes were used for further analysis. Single-cell GEX matrices were imported into R (version 3.6.1) and analyzed using Seurat (version 3.1.1). Cells with less than 4,000 UMIs or more than 20,000 UMIs were excluded. Cells with mitochondrial RNA reads greater than 10% were excluded from subsequent analyses. Single-cell TCR reads were aligned to the human genome assembly GRCh38 (hg38) and assembled into reconstructed TCR consensus sequences using Cell Ranger V(D)J (version 3.1.0, 10x Genomics). Only productive TCRα and TCRβ sequences were considered for further analysis. Overall, TCR sequences were annotated for 95.2% of cells, with paired TCRβ sequences detected in 37,934 of 41,542 cells (91.3%). Only cells with conventional paired TCR chain combinations of αβ were used for downstream analyses. Cells sharing identical CDR3αβ amino acid sequences were considered to belong to the same TCR clone.

Identification of Neoantigen-Reactive Clones

TCR clonal frequencies were compared between the sorted double-positive and double-negative dextramer populations for each healthy donor. Clones were considered double dextramer-specific if (i) a ≥100-fold higher clone frequency was identified in the dextramer-positive population than in the dextramer-negative populations and (ii) each TCR chain was ≥100-fold enriched in the double positive dextramer than in the double negative dextramer populations. Based on the above criteria, four neoantigen-reactive clones were defined as dextramer-specific for the two healthy donors.

TCR Construction

TCR sequences (T157.1, T157.2, T157.3, and T112.1) were codon-optimized for human expression and synthesized using GenScript. The constructs were cloned into the MSGV1 retroviral vector (RRID:Addgene_11174) in a TCRβ–TCRα orientation. Human constant regions were replaced with murine cysteine-modified constant regions to minimize mispairing with endogenous TCRs and to enable detection with antibodies against the murine TCRβ constant region, as previously described (23). TCR chains were linked using a furin-P2A sequence. The final constructs were cloned using NcoI and EcoRI restriction sites and confirmed using Sanger sequencing before use in retroviral production.

Retroviral Transduction of Human T Cells

293GP cells were seeded at 1.2 × 10⁶ cells per well onto poly-D-lysine–coated six-well plates (CELLCOAT, Greiner Bio-One, RRID:SCR_013782) and incubated overnight. Transfection was performed by mixing 2 μg of MSGV1 plasmid, 1.4 μg of RD114 envelope plasmid (RRID:Addgene_17576), and Lipofectamine 2000 (Thermo Fisher Scientific, cat. #11668-019, RRID:SCR_015527) in OptiMEM (Gibco). PBMCs from healthy donors were resuspended in 50/50 AIM V (Thermo Fisher Scientific, cat. #12055091, RRID:SCR_021179) and RPMI 1640 (Gibco) supplemented with 10% human AB serum (HP1022HI, BioIVT), 1% L-glutamine, and 1% penicillin–streptomycin. Cells were stimulated at 3.75 × 10⁶ cells/mL with 300 IU/mL IL-2 and 50 ng/mL OKT3 antibody (BioLegend, cat. #317326, RRID:AB_2562971) and plated at 7.5 × 10⁶ cells per well in 24-well plates. For retroviral infection, non-tissue culture–treated plates were coated overnight at 4°C with 20 μg/mL retronectin (Takara Bio, cat. #T100A, RRID:SCR_013825) in PBS and then blocked with 2% BSA for 30 minutes at room temperature. The viral supernatant was collected from transfected 293GP cultures, centrifuged at 1,000 rpm for 10 minutes, and plated onto retronectin-coated plates, followed by spinoculation at 2,000 × g for 2 hours at 32°C. PBMCs were washed thoroughly and resuspended in 50/50 AIM V/RPMI medium with 300 IU/mL IL-2 at 2 × 10⁵ cells/mL. After spinoculations, the viral supernatant was aspirated, and PBMCs were added to each well, followed by centrifugation at 1,500 rpm with acceleration and deceleration settings set to 1 for 10 minutes, followed by overnight incubation at 37°C. Jurkat TCRαβ-knockout CD8⁺ cells were infected in parallel using the same retronectin-based protocol. Transduction efficacy was assessed on days 4 to 7 after transduction by mTCR (BioLegend, cat. #109208, RRID:AB_313431; BioLegend, cat. #109220, RRID:AB_893624) surface staining using FACS.

Rapid Expansion

Following transduction, T cells were transferred to tissue culture plates and expanded in a 50/50 medium supplemented with 300 IU/mL recombinant human IL-2 and 50 ng/mL anti-human CD3 antibody (OKT3; BioLegend, cat. #317326, RRID:AB_2562971). The cells were cocultured with irradiated PBMC feeders at a responder-to-feeder ratio of 1:50. Feeder cells were irradiated at 35 Gy and prepared in RPMI 1640 supplemented with 10% FBS (Gibco, RRID:SCR_013780), 1% L-glutamine, 1% penicillin–streptomycin, and 1% sodium pyruvate. Half of the medium was replaced every 2 to 3 days with fresh cytokine-supplemented 50/50 medium, or the cells were split as needed.

In Vitro TCR Immunogenicity Assessment toward Pulsed B Cells

HLA-matched monoallelic B cells were pulsed with 1 µg/mL AR H875Y or AR WT peptides and coincubated with TCR-transduced T cells. The surface expression of 4-1BB and the secretion of IFNγ and TNFα were examined by flow cytometry.

Peptide Titration Assays

Titration assays were performed using TCR-transduced Jurkat reporter cells (TCRα/β knockout) and PBMCs from healthy donors. For reporter-based assays, monoallelic HLA-expressing B721.221 cells were pulsed for 2 hours at 37°C with decreasing concentrations of mutant or WT peptides (ranging from 10 μg/mL to 0.1 ng/mL), washed, and cocultured with TCR-expressing Jurkat cells. Following the manufacturer’s protocol, TCR activation was measured after 6 hours using a Bio-Glo-NL Luciferase Assay System (Promega).

PBMCs from healthy donors transduced with the same TCR constructs were cocultured overnight with peptide-pulsed monoallelic B721.221 cells. Activation was assessed by measuring the surface expression of 4-1BB on CD8⁺ T cells using flow cytometry. Supernatants from these titration assays were collected and analyzed for IFNγ secretion using an ELISA kit (R&D Systems, DY285B, RRID:AB_2928044). All peptide titration experiments were performed using at least two independent donors.

In Vitro TCR Immunogenicity Assessment of Prostate Cancer Cell Lines

TCR immunogenicity was evaluated by coculturing TCR-transduced healthy donor PBMCs with various prostate cancer cell lines modified to present the AR H875Y neoantigen in the context of HLA-B*15:01 or HLA-A*01:01. The cell line combinations included LNCaP expressing the AR H875Y MNG and HLA-B*15:01, LNCaP overexpressing full-length AR H875Y or WT with HLA-B*15:01, LNCaP with CRISPR/Cas9-mediated AR H875Y knock-in or control, HLA-B*15:01, C4-2 overexpressing full-length AR H875Y or WT with HLA-B*15:01, and 22Rv1 transduced with HLA-B*15:01. T-cell activation was evaluated using flow cytometry to measure the cell surface expression of 4-1BB, IFNγ, and TNFα. Apoptosis was independently analyzed in selected cell lines (LNCaP with AR H875Y MNG + HLA-B*15:01 and 22Rv1 + HLA-B*15:01) using flow cytometry for active caspase-3 (BD Biosciences, cat. #550914, RRID:AB_393957), according to the manufacturer’s instructions.

Incucyte Fluorescent Killing Assay

TCR-mediated cytotoxicity against prostate cancer cell lines was measured using live-cell imaging on an Incucyte SX3 system (Sartorius). The same panel of transduced prostate cancer cell lines was used for TCR immunogenicity assays. Each cell line was further modified to express GFP. The cell lines were seeded in triplicate in a 96-well plate, with each well containing 3,000 to 5,000 cells. The following day, T cells stably expressing AR H875Y–specific TCR (T157.1, T157.3, and T112.1) were added to the cancer cells at various E:T ratios. These ratios were calculated based on the percentage of mTCR-positive cells as determined by FACS analysis conducted on the same day. All conditions included the addition of IL-2 at a concentration of 300 U/mL in the T-cell medium. Control cancer cells received IL-2, but not T cells. Cellular interactions were monitored using an Incucyte SX3 system over 4 days, with images captured every 4 hours. Each well was photographed four times per imaging session. The area covered by the GFP-positive cells was analyzed.

In-Vivo Evaluation of TCR-Engineered T Cell

NSG (The Jackson Laboratory, RRID:IMSR_JAX:005557) were housed under specific pathogen-free conditions at the Weizmann Institute of Science. All procedures were approved by the Institutional Animal Care and Use Committee under the protocol number 00730124-1. On day 0, 5 × 10⁶ 22Rv1 prostate cancer cells expressing HLA-B*15:01 in PBS were subcutaneously injected into the right flank. On day 4, mice were randomized into treatment groups and intravenously injected with 20 × 10⁶ healthy donor PBMCs transduced with either AR H875Y–specific TCRs (T157.1 or T157.3) or an irrelevant control TCR (T104; KRAS-specific, unpublished data). All the mice received s.c. injections of recombinant human IL-2 (200,000 IU per mouse) on days 4, 7, and 9. Tumor size was measured using calipers three times per week for 24 days, and the volume was calculated using the following formula: length × width² × 0.5.

Supplementary Material

Supplementary Table S1

MS/MS spectra for resistant gene-derived peptides and the corresponding head-to-tail analysis.

Supplementary Table S2

HLA-A*01:01 Binding Predictions for AR H875Y by NetMHCpan 8 to 14 flanking the mutation

Supplementary Table S3

HLA-B*15:01 Binding Predictions for AR H875Y by NetMHCpan 8 to 14 flanking the mutation

Supplementary Table S4

Frequency of AR H875Y mutations and HLA allotypes in the Hartwig Cohort

Supplementary Table S5

Frequency of AR H875Y mutations and HLA allotypes in the Abida and colleagues, 2019 Cohort

Supplementary Table S6

Frequency of AR H875Y mutations and HLA allotypes in the Quigley and colleagues, 2018 Cohort

Supplementary Table S7

Frequency of AR H875Y mutations and HLA allotypes in the Herberts and colleagues, 2022 Cohort

Supplementary Table S8

Single Cell TCRseq of donor 157 with AR H875Y 9-mer dextramer staining sorted T-cells

Supplementary Table S9

Single Cell TCRseq of donor 157 with AR H875Y 10-mer dextramer staining sorted T-cells

Supplementary Table S10

Single Cell TCRseq of donor 112 with AR H875Y 13-mer dextramer staining sorted T-cells

Supplementary Table S11

Cell lines HLA allotypes

Supplementary Table S12

Cancer types and sample numbers in the Hartwig cohort

Supplementary Table S13

Distribution of cancer types and metastatic locations in the Hartwig cohort

Supplementary Table S14

Cancer types and sample numbers in the TCGA

Supplementary Table S15

Cancer types and sample numbers in three metastatic prostate cancer cohorts

Supplementary Table S16

Synthetic peptides used in this study

Supplementary Figure S1

MS/MS spectra for resistant gene-derived peptides and the corresponding head-to-tail analysis.

Supplementary Figure S2

Immunogenicity gating strategy of reactive CD8+ T-cells derived from healthy donor PBMCs

Supplementary Figure S3

Retroviral transduction of Jurkat (TCR alpha beta KO) CD8+ T-cells with AR H875Y specific TCRs

Supplementary Figure S4

Retroviral transduction of healthy donor PBMCs with AR H875Y-specific TCRs

Supplementary Figure S5

Immunogenicity and cross reactivity of CD8+ T transduced with specific AR H875Y TCRs

Supplementary Figure S6

Specific AR H875Y T157.2 TCR reactivity in healthy donor PBMCs

Supplementary Figure S7

Specific AR H875Y TCR cytotoxicity in healthy donor PBMCs

Data Availability

All single-cell RNA and single-cell TCR sequencing datasets generated in this study were deposited in the NIH Sequence Read Archive (RRID:SCR_004891) under BioProject ID PRJNA1336784. All raw MS files were deposited in the ProteomeXchange Consortium via the PRIDE Partner Repository (accession number: PXD055544, RRID:SCR_004055). All data are publicly available on the date of publication. The source code is now available on our GitHub repository: https://github.com/bbglab/SpotNeoMet. All raw data, supplementary files, and any additional information required to reanalyze the data reported in this article are available from the lead contact upon request. All statistical test results that are not shown (e.g., nonparametric tests) are available from the corresponding author upon request. The credentials for GitHub are provided at https://github.com/bbglab/SpotNeoMet. Code availability for data processing and figure generation will be shared upon request by the corresponding author.

Authors’ Disclosures

C. Arnedo-Pac reports grants from “la Caixa” Foundation (ID 100010434) fellowship (LCF/BQ/ES18/11670011) during the conduct of the study. O.G. Troyanskaya reports that she is a scientific advisory board member at Caris Life Sciences. A.W. Wyatt reports personal fees from AstraZeneca, Merck, Janssen, and Pfizer and grants from Tyra Biosciences, ESSA Pharma, and Promontory Therapeutics outside the submitted work. E.M. Van Allen reports personal fees from Tango Therapeutics, Genome Medical, Genomic Life, Monte Rosa Therapeutics, Manifold Bio, Enara Bio, Serinus Bio, Foaley & Hoag, TracerDx, and Riva Therapeutics; grants and personal fees from Novartis; and grants from Bristol Myers Squibb, Janssen, Sanofi, and NextPoint outside the submitted work, and reports institutional patents filed on chromatin mutations and immunotherapy response and on methods for clinical interpretation, pending and issued. N. Anandasabapathy reports other support from Panther Life Sciences, Network Bio, and Shennon Bio and personal fees from Genmab outside the submitted work. J. Mateo reports grants, personal fees, and nonfinancial support from AstraZeneca and Pfizer Oncology; grants from Amgen; personal fees from MSD, Bayer, and Roche; and nonfinancial support from Guardant Health outside the submitted work. No disclosures were reported by the other authors.

Authors’ Contributions

N. Gumpert: Validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. S. Sagie: Validation, methodology, writing–original draft, writing–review and editing. C. Arnedo-Pac: Methodology. T. Babu: Methodology. C. Weller: Methodology. A. Gonzalez-Perez: Methodology. Y. Wang: Methodology. L. Michel Todó: Methodology. R. Levy: Methodology. X. Chen: Methodology. P. Greenberg: Methodology. M. Dayan-Rubinov: Methodology. E. Yakubovich: Methodology. T. Wasserman-Bartov: Methodology. M. Zerbib: Methodology. J. Gong: Methodology. R.J. Rebernick: Methodology. A. Oliveira Tercero: Methodology. L. Agundez Muriel: Methodology. G. Benedek: Methodology. M. Kedmi: Methodology. R. Oren: Methodology. S. Ben-Dor: Methodology. Y. Levin: Methodology. O.G. Troyanskaya: Methodology. A.D. Munzur: Methodology. A.W. Wyatt: Methodology. M.P. Cieslik: Supervision, methodology. D.A. Quigley: Methodology. E.M. Van Allen: Methodology. N. Anandasabapathy: Methodology. J. Mateo: Methodology. X. Yang: Methodology. F. Martínez-Jiménez: Methodology. N. Lopez-Bigas: Supervision, methodology. Y. Samuels: Supervision, writing–original draft, writing–review and editing.