Ketogenic Diet Alters the Epigenetic and Immune Landscape of Prostate Cancer to Overcome Resistance to Immune Checkpoint Blockade Therapy Free

Prostate cancer is the second most commonly diagnosed malignancy and the fifth leading cause of cancer-related mortality for men worldwide (1). A handful of targeted therapy drugs have been approved to treat advanced prostate cancer, but only offer modest survival benefits (1). Therefore, new therapeutics are highly desired. Immune checkpoint blockade (ICB) using antibodies against cytotoxic-T-lymphocyte–associated protein 4 (CTLA4) or programmed cell death 1/ programmed cell death 1 ligand 1 (PD1/PD-L1) generates durable therapeutic responses in a significant subset of patients across various cancer types (2). However, advanced prostate cancer shows overwhelming de novo resistance to ICB (3). Prostate cancer features a moderate mutation burden and a suppressed tumor immune microenvironment (TIME), posing unique challenges for ICB therapy.

One promising immunotherapy sensitization approach is targeting histone deacetylases (HDAC) with HDAC inhibitors (HDACi; refs. 4, 5). HDACi have diverse immune-modulatory activities. An intact immune system is required for the robust anticancer effects of class I/II HDACi vorinostat and panobinostat in murine cancer models (6). Both class I-specific and pan-HDACi increase the cancer cell immunogenicity through upregulating tumor-associated antigens and the components of the antigen processing machinery and MHC molecules (7). HDACi also has a profound impact on immune populations in the TIME, including dampening regulatory T cells (Treg) and myeloid-derived suppressor cells (MDSC; ref. 8) and biasing the effector functions of T-helper (Th) cells and dendritic cells (DC; refs. 9, 10). Translation of these findings to early-phase clinical trials of combining entinostat and high-dose IL2 in treating renal cell carcinoma showed promising clinical activity (11). Vorinostat is an HDACi approved by FDA to treat cutaneous T-cell lymphoma (12). Enhanced antitumor effects by combining vorinostat and various types of immunotherapy were reported in mammary, renal, and colorectal tumor models (13). Whether vorinostat sensitizes prostate cancer to ICB therapy remains to be determined.

Dietary interventions, including the ketogenic diet (KD), are emerging metabolic approaches to enhancing cancer therapy (14). KD consists of high fat, moderate to low protein, and very low carbohydrates, forcing the body to burn fat rather than glucose for energy (15). KD may be a potential dietary treatment for cancer by exploiting inherent oxidative metabolic differences between cancer and normal cells (15). The effect of KD in animal tumor models is varied, with an overall moderate antitumor activity (16). Fatty acid oxidation of ingested KD in the liver produces ketone bodies, with β-hydroxybutyrate (BHB) being the most abundant (17). Once taken up by a target tissue, BHB is converted back to acetyl-CoA to feed the tricarboxylic acid (TCA) cycle for oxidation and ATP production. In addition to being a carrier of energy, BHB functions as a signaling metabolite through several mechanisms (17). Importantly, BHB is an endogenous pan-HDACi and modulates gene expression and function by inhibiting HDAC1 & 3 (class I) and HDAC4 (class II; ref. 18). The immune-modulatory activity of BHB started to be appreciated recently (19).

This study generated an isogeneic murine prostate cancer cell series showing distinct sensitivity to anti-PD1 and anti-CTLA4. Using this model, we demonstrated combinatorial efficacy by combining ICB with HDACi vorinostat, a CKD, or BHB supplementation in eradicating ICB-resistant tumors. Through functional experiments, mass cytometry (CyTOF) and single-cell RNA-sequencing (scRNA-seq), we found that both cancer cell–intrinsic and -extrinsic mechanisms underlie the combinatorial efficacy. The optimized KD regimens, especially the 1,3-butanediol-supplemented diet (BD), may pave the road for applying these strategies to enhance patient outcomes.

All animal work performed in this study was approved by the Institutional Animal Care and Use Committee (IACUC) at University of Notre Dame. All animals were maintained under pathogen-free conditions and cared for per the International Association for Assessment and Accreditation of Laboratory Animal Care policies and certification. C57BL/6J (RRID:IMSR_JAX:000664), FVB/NJ (RRID:IMSR_JAX:001800), Batf3^−/− (RRID:IMSR_JAX:013755), Rag1^−/−(RRID:IMSR_JAX:002216) mice were purchased from the Jackson Laboratory and bred in house. Bdh1^−/− allele was a generous gift from Daniel P. Kelly at University of Pennsylvania. Only male mice were used for experiments, given the gender association of prostate cancer.

PPS-6239, PPS-PD1R, PPS-CTLA4, PPS-ICBR, MyC-CaP (ATCC, CRL-3255, RRID: CVCL_J703), and MC38 (Millipore Sigma, SCC172, RRID: CVCL_B288) were cultured in DMEM (GE Healthcare, SH30243.FS) supplemented with 10% FBS (GE Healthcare, SH30396.03) and 100 U/mL penicillin–streptomycin (Cytiva, SV30010). RM9 (ATCC, CRL-3312, RRID: CVCL_B461) was cultured in DMEM/F12 (VWR, 45000–344) supplemented with 10% FBS and 100 U/mL penicillin–streptomycin. All the cell lines were cultured at 37°C in a humidified incubator with 5% CO₂. Ectopic Hdac1 overexpression in PPS-6239 cells was conducted through a well-established lentiviral packaging and transduction procedure (20) with Hdac1-encoded lentiviral vector TFORF3338 (Addgene, 144814, RRID:Addgene_144814). All cells were cultured for no more than five passages and tested monthly for Mycoplasma-free status using a Mycoplasma Assay Kit (Agilent Technologies, 302109). No authentication was performed for mouse cell lines.

Syngeneic tumors were generated by injecting a total of 5×10⁵ tumor cells subcutaneously into both flanks of 6 to 8-weeks-old male mice. When tumors reached 50 to 100 mm³, mice were randomized to receive therapeutics, including anti-PD1 (BioLegend, 114116, RRID: AB_2566280) and anti-CTLA4 (BioLegend, 106207, RRID: AB_2616631) at 10 mg/kg each intraperitoneally twice a week, or vorinostat (MedChemExpress, HY-10221) at 25 mg/kg i.p. daily. For CD8⁺ T-cell depletion, tumor-bearing mice received intraperitoneal injection of three doses of 200-μg anti-CD8 (BioXCell, BE0061, RRID:AB_1125541) at the indicated time points. For blocking Cxcr3, anti-mouse Cxcr3 antibody (BioLegend, 126538) was dosed at 0.5 mg per intraperitoneal injection, twice per week. For B-cell depletion, anti-CD20 antibody (Genentech, clone 5D2, RRID: AB_2715460) was dosed at 100 μg per intraperitoneal injection, weekly. For dietary interventions, the randomized mice were fed ad libitum with standard chow (Teklad, 2918), or KD (Teklad, 160153) replenished daily, or cyclic KD (CKD) that followed a weekly cycles with five days on daily KD (Teklad, 160153) then two days on standard chow (Teklad, 2918), or BD that was prepared by mixing 80 mL of 1,3-butanediol (Sigma-Aldrich, 309443), 120 mL of water, 198 g of standard chow (Teklad, 2918), and 2 g of Saccharine (Sigma-Aldrich, 109185), replenished every other day. Mice on the KD were given extra water due to the metabolic requirements of fatty acid degradation. All treatments were continued until the specified experimental endpoints were reached.

PPS-ICBR cells grown in the regular medium were treated with 5 μmol/L vorinostat (MedChemExpress, HY-10221), 1:2,000 diluted DMSO (Sigma-Aldrich, D4540), or 10 ng/mL mouse IFNγ (Sigma-Aldrich, I4777) for the indicated durations. For BHB treatment, the medium was removed once cells in the regular medium reached approximately 70% confluence. Cells were washed once with warm PBS and changed to a medium of DMEM without glucose (VWR, 97060–876), 10% FBS, and 1% Pen/Strep. Glucose was restored to 5 mmol/L to mimic the human blood glucose level by diluting from a 500 mmol/L stock of glucose (Sigma-Aldrich, G8270). BHB (Cayman Chemical, 14148) was added to 3 mmol/L, and the medium was titrated to pH 7.2 using HEPES (VWR, 16777–032). The treatments were maintained for indicated durations before cells were harvested for subsequent experiments.

Cells and fresh tissue samples were lysed in RIPA supplemented with protease inhibitors (Bimake, B14012) and phosphatase inhibitors (Bimake, B15002). The western blot procedure has been described previously (20). The primary antibodies were used to detect β-actin (Santa Cruz Biotechnology, sc-47778, RRID:AB_626632), α-tubulin (Santa Cruz Biotechnology, sc-5286, RRID:AB_626632), acetyl-Histone H3 Lys9 (Cell Signaling Technology, 9649, RRID:AB_823528), Histone H3 (Cell Signaling Technology, 4499, RRID:AB_10544537), acetyl–α-tubulin Lys40 (Cell Signaling Technology, 5335, RRID:AB_10544694), BDH1 (ProteinTech, 15417–1-AP, RRID: AB_2274683), and HDAC1 (Cell Signaling Technology, 34589, RRID:AB_2756821). The secondary antibodies include horseradish peroxidase (HRP)–conjugated goat anti-mouse (Cell Signaling Technology, 7076, RRID:AB_330924) and HRP-conjugated goat anti-rabbit (Cell Signaling Technology, 7074, RRID:AB_2099233). Signals were detected with Clarity Max ECL Substrate (Bio-Rad, 1705062).

Animal tissues were fixed overnight in 10% formalin and embedded in paraffin. Antigen retrieval was performed by heating in a pressure cooker at 95°C for 30 minutes, followed by 115°C for 1 minute in citrate-unmasking buffer (pH 6.0). IHC staining with CD8α antibody (Cell Signaling Technology, 98941, RRID:AB_2756376) was performed using the VECTASTAIN Elite ABC-HRP Kit (Vector Laboratories, PK-6101) with signals detected with DAB Substrate Kit (Vector Laboratories, SK-4100). Counterstain was performed using hematoxylin (VWR, 10143–608). The IHC slides were scanned using an Aperio ScanScope (Leica, RRID:SCR_022191). Signals were quantified using functions in ImageJ Fiji (RRID:SCR_002285).

RNA from formalin-fixed, paraffin-embedded (FFPE) samples was isolated using the FFPE RNA Purification Kit (Norgen Biotek, 25360). cDNA was synthesized using the SuperScript III First-Strand Synthesis System for RT-PCR (Invitrogen, 18080–051). qRT-PCR was performed using SYBR Green qPCR Master Mix (Bimake, B21203) on the CFX Connect Real-Time PCR Detection System (Bio-Rad, 1855201). Primers used: Pdl1 Forward GACCAGCTTTTGAAGGGAAATG, Reverse CTGGTTGATTTTGCGGTATGG, Ctla4 Forward TTTTGTAGCCCTGCTCACTCT Reverse CTGAAGGTTGGGTCACCTGTA, Gzmb Forward TCATGCTGCTAAAGCTGAAGAG Reverse CCCGCACATATCTGATTGGTTT, Ifng Forward ATGAACGCTACACACTGCATC Reverse CCATCCTTTTGCCAGTTCCTC, H2-K1 Forward TCCACTGTCTCCAACATGGC Reverse CCCTCCTTTTCCACCTGTGT, Gapdh Forward AGGTCGGTGTGAACGGATTTG Reverse TGTAGACCATGTAGTTGAGGTCA.

Mouse blood (100 μL) was collected with submandibular bleeding. Serum BHB concentration was measured following the manual instructions of the BHB colorimetric detector kit (Cayman Chemical, 700194).

Flow cytometry samples were prepared as described previously (20) and run on CytoFLEX (Beckman Coulter). Data were analyzed using FlowJo v10.8 (FlowJo, RRID: SCR_008520). Fluorochrome-conjugated antibodies included H2-K^b-APC (BioLegend, 116518, RRID: AB_10564404), H2-D^b-PE/Cyanine7 (BioLegend, 111516, RRID: AB_2565864), B220-PE (BioLegend, 561878, RRID: AB_394619), CD45-FITC (Cytek, 35–0451, RRID:AB_2621689), CD19-APC (BioLegend, 115511, RRID: AB_313646), CD3-APC/Cy7 (Cytek, 25–0032, RRID:AB_2621619), CD4-PE (Cytek, 50–0041, RRID:AB_2621736), CD8a-APC (Cytek, 20–0081, RRID:AB_2621550), and CD45-PE/Cy7 (BD Biosciences 552848 RRID:AB_394489).

PPS-ICBR cells treated with DMSO or vorinostat for 12 hours were used to extract total RNA using the RNeasy Kit (Qiagen). RNA samples were profiled on the Mouse Genome 430 2.0 Array (Affymetrix) at the Genomics Core Facility at the MD Anderson Cancer Center. The data were analyzed using Transcriptome Analysis Console software (Thermo Fisher Scientific) to generate a list of differentially expressed genes with cutoff FDR <0.05, absolute fold change ≥3. Pathway enrichment was performed using Enrichr (RRID: SCR_001575).

Four cell lines, PPS-6239 (control), PPS-PD1R (anti-PD1 resistant), PPS-CTLA4R (anti-CTLA4 resistant), and PPS-ICBR (dual resistant), were profiled with RNA-seq. The total RNA was harvested using the Rneasy mini plus kit (Qiagen). The quantity and quality of each sample were measured using an Agilent 2100 Bioanalyzer. RNA libraries were prepared for sequencing using standard Illumina protocols. RNA-seq was performed at Sequencing and Microarray Facility at MD Anderson Cancer Center using standard stranded RNA-seq protocol on an Illumina HiSeq 4000. The raw reads were aligned to the mouse reference genome build mm10, using Tophat RNA-seq alignment software. The aligned reads were verified for quality using FASTQC software. HTseq software was used to summarize the gene expression counts from Tophat alignment data after sorting the BAM files. The raw counts were normalized, and differential expression analysis was performed on contrasts of interest using the DEseq2 package. The raw and normalized data files were deposited to Gene Expression Omnibus (GEO) with accession number GSE252986. To plot heat map showing genes commonly and most significantly upregulated or downregulated in PPS-PD1R, PPS-CTLA4R, and PPS-ICBR compared with PPS-6239, Up genes (271 genes) were filtered by log₂ (fold change) ≥1 for PPS-PD1R/PPS-6239, PPS-CTLA4R/PPS-6239 and PPS-ICBR/PPS-6239, and read count for PPS-6239 ≥ 50; Down genes (246 genes) were filtered by log₂ (fold change) ≤ −1 for PPS-PD1R/PPS-6239, PPS-CTLA4R/PPS-6239, and PPS-ICBR/PPS-6239, and read count for PPS-6239 ≥ 200. The read count threshold helped remove genes with generally very low expression levels. Pathway enrichment was performed by inputting the gene lists into the Enrichr (RRID: SCR_001575).

Tumors were minced into homogenate and rotated at 37°C in dissociation media, DMEM with 10% FBS and 1 mg/mL collagenase IV (STEMCELL Technologies, 07427) for 1 hour, followed by passing through 40-μm strainers. Erythrocytes were depleted via hypotonic lysis. The CyTOF procedure has been described previously (20). The samples were run with Helios CyTOF mass cytometer (Fluidigm, RRID:SCR_019916) in the Flow Cytometry and Cellular Imaging Core Facility at the MD Anderson Cancer Center. For the first CyTOF experiment, data analysis was performed with Cytobank (Beckman Coulter, RRID:SCR_014043). Samples were concatenated on the basis of treatment types and then uploaded to Cytobank. tSNE (RRID:SCR_024305) plots of single viable CD45⁺ cells were generated using optSNE with 5,000 iterations and a KL divergence of 4.329796. FlowSOM (RRID:SCR_016899) was used with a pre-determined number to create 14 cell clusters manually annotated on the basis of marker expression patterns. For the second myeloid-focused CyTOF experiment, data analysis was performed with FlowJo v10.8 (FlowJo, RRID: SCR_008520). After gating cell populations manually, CD45⁺ cells were used for dimension reduction using tSNE with iterations = 1,000, perplexity = 30, and learning rate = 2,800. The Approximate (random projection forest—ANNOY) KNN algorithm was used with the FFT interpolation gradient algorithm. Cell clustering was performed using the FlowJo plugin tool PhenoGraph (RRID:SCR_016919) with parameters K = 30 and Run ID = auto, with clusters manually annotated on the basis of marker expression patterns. We used the plugin tool ClusterExplorer to generate the intensity heat map table of proteins for each condition.

Tumors were minced and dissociated with a mouse dissociation kit (Miltenyi Biote, 130–096–730) at 37°C for 1 hour. Dissociated tumor single-cell suspension was passed through 40-μm strainers followed by red blood cell lysis. Then the dead cell was removed using EasySep Dead Cell Removal (Annexin V) Kit (STEMCELL Technologies, 17899) followed by multiplexing with hashtag labeling of TotalSeq-A0303 (BioLegend, 155805 RRID: AB_2750035), TotalSeq-A0304 (BioLegend, 155807 RRID: AB_2750035), TotalSeq-A0305 (BioLegend, 155809 RRID: AB_2750036), TotalSeq-A0306 (BioLegend, 155811 RRID: AB_2750037), TotalSeq-A0307 (BioLegend, 155813 RRID: AB_2750039), and TotalSeq-A0308 (BioLegend, 155815 RRID: AB_2750040). At this step, the samples were transferred to Genomics and Bioinformatics Core Facility at University of Notre Dame, where cells were counted, evenly combined, and loaded onto the Chromium Controller (10X Genomics RRID:SCR_019326). The samples were then processed following the manuals of the Chromium Single Cell 3′ Reagent Kits. Next, the samples were sequenced with Illumina NovaSeq 6000 (RRID:SCR_016387) platform at the Medical Genomics Center at Indiana University.

Barcode processing, alignment, filtering, and UMI counting were performed using the Cell Ranger analysis pipeline (v. 6.1 RRID:SCR_017344). Sample demultiplexing was performed using the CITE-seq-Count pipeline (v. 1.4). Cells with larger than 500 UMI counts in the demultiplexing step were used for downstream analysis in R (v. 4.1). Data pre-processing, normalization, and clustering was performed using the R package Seurat (v. 4.1 RRID:SCR_007322). Single-cell transcriptomes were initially filtered for quality control. Cells with fewer than 200 genes detected and genes expressed in fewer than three cells were removed. We also filtered cells with the percentage of mitochondrial counts larger than 5%. The resulting expression matrix contained 57,49 cells by 18,848 genes, with 1,215 cells from the standard diet (SD) treatment, 490 cells from the BD treatment, 2,314 cells from the SD+ICB treatment, and 1,730 cells from the BD+ICB treatment. Data normalization was performed using the function NormalizeData with normalization.method = “LogNormalize” and scale.factor = 10,000. 2,000 variable genes were chosen using the function FindVariableFeatures with selection.method = “vst.” Data scaling was performed using the ScaleData function. Dimensional reduction was accomplished by performing principal component analysis and then using the first 50 principal components for Uniform Manifold Approximation and Projection (UMAP) using default parameters associated with the RunUMAP function. Unsupervised clustering was finished by constructing a shared nearest neighbor (SNN) graph using the FindNeighbors function and then performing graph-based clustering using the “Louvain” algorithm with resolution = 1.8, resulting in 22 clusters. We merged clusters based on their marker genes, leading to 19 cell types in our analysis. Differential expression analysis between clusters was performed using a Wilcoxon rank sum test by the FindAllMarkers function. The dot plots and feature plots for chosen marker genes were obtained by the DotPlot and FeaturePlot functions, respectively. The expression heat map was generated using the function DoHeatmap.

Signaling pathway gene signatures were curated from the hallmark gene sets, C2, C6, C7, and C8 collections within MSigDB (RRID:SCR_016863). Normalized enrichment scores were calculated using the fgsea package v. 1.20, with parameters minSize = 15 and maxSize = 500. Cell–cell communication analysis for each treatment group was conducted using the CellChat (RRID:SCR_021946) package v. 1.4. The analysis followed the standard pipelines provided by the CellChat package, using the ligand–receptor database curated for mouse tissues. We computed the communication probability for each treatment condition using the computeCommunProb function and filtered out the cell–cell communications if there are less than 10 cells in certain cell groups, using the filterCommunication function with parameter min.cells = 10. We then inferred the cellular communication network and aggregated cell–cell communication, using computeCommunProbPathway and aggregateNet functions, respectively. The netVisual_circle function generated the circle plots for each condition. The bubble plots with multiple functions were created with the netVisual_bubble function. The circle plots for individual ligand–receptor pairs (LRP) were created with the netVisual_individual function. The violin plot for chosen genes under different cell types and conditions was generated by the plotGeneExpression function. Visualizations of the dominant senders and receivers for all the conditions were generated with the netAnalysis signalingRole scatter function. The comparison of the overall information flow of signaling pathways among all the conditions was made by the rankNet function.

The pseudotime analysis was conducted using the monocle3 R package v. 1.3.1. (RRID:SCR_018685) We followed the recommended pipelines from the monocle tutorials. Ten cell types from the BD+ICB condition were analyzed. The dimension reduction was performed using the preprocess cds and reduce dimension functions. The trajectory was learned using the learn graph function with parameter close_loop = FALSE. We further inferred the pseudotime using the order cells function and chose the monocyte cell population as the root point.

In vitro and in vivo experiments were performed three or more times and conclusions were drawn only when the results were reproducible. One representative result among the replicates was shown in the figures. For non-omics data, statistical analyses were performed using GraphPad Prism v9.3 (RRID: SCR_002798). All data are presented as mean ± SEM (standard error of the mean). We followed this workflow for statistical testing as previously described (21): The Shapiro–Wilk test was performed to assess for normality of data distribution: (i) In case of normality, when only two conditions were to test, we performed unpaired t test; when more than two conditions were to compare, we performed a parametric one-way or two-way ANOVA followed by post hoc test with recommended correction for multiple comparisons to assess the significance among pairs of conditions. (ii) In case of non-normality, when only two conditions were to test, we performed a Mann–Whitney U test; when more than two conditions were to compare, we performed a nonparametric one-way ANOVA followed by a recommended test to assess the significance among pairs of conditions. For survival data, log-rank test was used. Sample sizes, error bars, P values, and statistical methods are noted in the figure legends. Statistical significance was defined as a P value of <0.05.

Microarray data of DMSO or vorinostat-treated PPS-ICBR cells are available in the GEO (RRID: SCR_005012) with accession number GSE205469. RNA-seq data of the four cell lines are available at GEO with accession number GSE252986. scRNA-seq data are available at GEO with accession number GSE206561. All other raw data and the materials generated in this study are available from the corresponding author upon request.

Previously, we developed murine prostate cancer cell lines from PB-Cre⁺ Pten^L/L p53^L/L Smad4^L/L mouse model of metastatic prostate cancer (22). These PPS cell lines grew in C57BL/6 mice after subcutaneous inoculation. We focused on one line, PPS-6239, and observed that PPS-6239 tumors (∼100 mm³) showed partial response to anti-PD1 or anti-CTLA4 alone (10% cure rate in each case) but substantial response to the combination (80% cure rate, Fig. 1A). The mice whose tumors were either resected surgically or eradicated by treatments were challenged with PPS-6239 or an irrelevant syngeneic prostate cancer cell line RM9. RM9 grew in all cases, but PPS-6239 showed a distinct pattern dependent on the treatment history: Mice treated with isotype IgG displayed no rejection, mice treated with anti-PD1 or anti-CTLA4 displayed moderate rejection, and mice treated with combination therapy rejected all inoculation attempts (Fig. 1B). This result demonstrates that effective immunotherapy targeting PPS-6239 generated enduring and specific memory immunity.

To generate ICB-resistant variants of PPS-6239, we digested the rare tumors resistant to anti-PD1 or anti-CTLA4 single therapy and developed cell lines PPS-PD1R and PPS-CTLA4R, respectively (Fig. 1C). Both lines gained resistance to anti-PD1, anti-CTLA4 and the combination (0% cure rate in all groups, Fig. 1D). From one of the PPS-CTLA4R-bearing mice treated with the combination therapy, we further developed a subline PPS-ICBR (Fig. 1C), which exhibited resistance to anti-PD1 plus anti-CTLA4 therapy (0% cure rate, Fig. 1E). Consistent with the response pattern, PPS-6239 tumors had massive infiltration of CD8⁺ T cells under the dual ICB, whereas PPS-ICBR tumors lacked CD8⁺ T cells under the same therapy (Supplementary Fig. S1A). The fact that PPS-PD1R, PPS-CTLA4R and PPS-ICBR were all resistant to dual ICB suggests certain shared resistance mechanisms. Genes commonly upregulated or downregulated in PPS-PD1R, PPS-CTLA4R and PPS-ICBR compared with PPS-6239 were identified through RNA-seq (Supplementary Fig. S1B). The commonly upregulated genes enriched for pathways such as epithelial–mesenchymal transition, KRAS signaling up, hypoxia, IL2/STAT5 signaling, TNFα signaling via NF-κB, inflammatory response, and IFNγ response, whereas the commonly downregulated genes enriched for some unique pathways such as UV response down and Hedgehog signaling (Supplementary Fig. S1C). Many of these commonly enriched pathways are immune-related and may contribute to the gained resistance to the dual ICB therapy by all three sublines. Overall, the PPS-6239, PPS-PD1R, PPS-CTLA4R, and PPS-ICBR series of isogenic cell lines provide a valuable model to study the mechanism of ICB resistance and the strategy to overcome it.

Consistent with the critical role of antigen presentation by MHC-I in tumor immunity, H2-K^b and H2-D^b were significantly downregulated in the resistant sublines (Fig. 2A). At the tumor level, PPS-ICBR tumors expressed several CTL-related genes (H2-K1, Pdcd1, Ctla4, Ifng and Gzmb) at lower levels compared with PPS-6239 tumors, but expressed Pdl1(Cd274) at a higher level (Supplementary Fig. S1D). To assess whether HDACi could revert gene expression associated with ICB resistance, we profiled the transcriptome of PPS-ICBR treated with DMSO or 5 μmol/L vorinostat for 12 hours in vitro. Among the differentially expressed genes, MHC-encoded genes were significantly upregulated by vorinostat, including H2-K1 and H2-D1 (Fig. 2B; Supplementary Table S1). H2-K^b protein level was restored by 5μmol/L vorinostat (Fig. 2C), but PD-L1 (Cd274) was not altered by vorinostat (Supplementary Fig. S2A and S2B). Chromatin immunoprecipitation followed by quantitative PCR (ChIP-qPCR) using Histone H3 acetyl Lys9 (H3K9Ac) antibody showed augmented histone acetylation of the H2-K1 locus in vorinostat-treated cells compared with DMSO-treated cells (Fig. 2D). Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis showed that vorinostat-upregulated genes enriched for immune-related pathways, such as TNF signaling, IL17 signaling, and cytokine–cytokine receptor interaction (Fig. 2E; Supplementary Table S2), whereas vorinostat-downregulated genes enriched for proliferation-related pathways such as DNA replication and cell cycle (Fig. 2E; Supplementary Table S3). Transient Hdac1 knockdown with shRNA upregulated H2-K^b and H2-D^b in PPS-ICBR cells (Fig. 2F). Conversely, ectopic Hdac1 overexpression in PPS-6239 reduced H3K9Ac and H2-K^b (Supplementary Fig. S2C and S2D).

Despite the in vitro activity, vorinostat monotherapy in vivo failed to affect PPS-ICBR tumors (Fig. 2G). Because α-tubulin acetyl Lys40 (α-tub-K40Ac) is deacetylated by HDAC6 (a class IIb HDAC inhibited by vorinostat), α-tub-K40Ac is a valid pharmacodynamic marker for vorinostat. Vorinostat indeed augmented α-tub–K40Ac levels in vitro and in vivo, yet the signals in vivo declined rapidly and returned to the baseline 6 hours after dosing (Fig. 2H). This result is consistent with the short half-life of vorinostat in patients (23), presumably accounting for its limited efficacy as monotherapy in our model.

Strikingly, vorinostat dramatically enhanced the efficacy of ICB (anti-PD1+anti-CTLA4), with complete response for 9 of 24 combo-treated tumors (Fig. 2G). Vorinostat enhanced H3K9Ac levels in vorinostat-treated and combo-treated tumors (Fig. 2I). When H2-K^b level on tumor cells from the treated cohorts was measured, the moderate increase by vorinostat was consistent with the corresponding HDACi activity in vitro, whereas the higher H2-K^b by ICB treatment was likely due to the induction by elevated intratumoral IFNγ (Fig. 2J; Supplementary Fig. S2E), consistent with the reported activity of IFNγ to induce MHC-I expression in human cancer cells in preclinical (24) and clinical studies (25). The complementary activities from HDACi and IFN-γ might contribute to the highest H2-K^b level by ICB+Vo (Fig. 2J).

At the histological level, whereas the other three groups showed poor CD8⁺ T-cell infiltration, the ICB+Vo group was infiltrated with myriad CD8⁺ T cells (Fig. 2K; Supplementary Fig. S2F). Depletion of CD8⁺ cytotoxic T lymphocytes (CTL) abrogated the antitumor effect of ICB+Vo (Fig. 2K and L; Supplementary Fig. S2F–S2G). The efficacy from ICB+Vo was also abrogated when Batf3^null mice were used as the hosts (Supplementary Fig. S2H). Because Batf3 deficiency depletes DCs capable of cross-presenting tumor antigens to CD8⁺ T cells (26), this result further proves the essential role of CTLs in ICB+Vo efficacy. Finally, we re-challenged three mice fully cured by ICB+Vo with PPS-ICBR cells 55 days after the initial start of treatment. Having detected no tumor growth 40 days later, we challenged one of the three mice with MC38 colorectal cancer cells and readily observed engraftment (Fig. 2M). In summary, vorinostat upregulated MHC-I expression and various immune-related pathways in PPS-ICBR cells and markedly restored the sensitivity of PPS-ICBR tumors to ICB through CTL activities.

We immunophenotyped the tumors 12 days after the start of the therapy using CyTOF (Fig. 3A and B). The CyTOF antibody panel targeted 27 cell surface markers and 10 intracellular markers (Supplementary Table S4). Cytobank was used to analyze the CyTOF data, where tSNE was used for dimension reduction, and cell clusters were annotated on the basis of marker expression patterns (Fig. 3C; Supplementary Fig. S3A). Both cell cluster percentages and densities showed visually recognizable changes associated with the ICB+Vo condition, including the increase of NK cells and CD8⁺ T cells (purple) and conventional type 1 DCs (cDC1; dark blue) as well as the disappearance of cancer cells (brown) and M2-polarized macrophages (cyan; Fig. 3D and E). CD8⁺ CTLs and the central memory (CD44⁺ CD62L⁺) and effector memory (CD44⁺ CD62L⁻) subsets increased substantially only in the ICB+Vo group, CD4⁺ Th subset was upregulated in both ICB and ICB+Vo groups, and Treg (CD25⁺ Foxp3⁺) showed no change (Fig. 3F). As a result, CD8⁺ effector/Treg ratio peaked in the ICB+Vo condition. M1-macrophages (CD11b⁺ F4/80⁺ MHC-II⁺) and M2-macrophages (CD11b⁺ F4/80⁺ MHC-II⁻ CD206⁺) followed the opposite trend—ICB+Vo group had the highest M1 and lowest M2 populations (Fig. 3G). DCs (CD11c⁺ MHC-II⁺) also showed the highest level in ICB+Vo, but Mo-MDSCs (CD11b⁺ Ly6G⁻ Ly6C^high) or PMN-MDSCs (CD11b⁺ Ly6G⁺ Ly6C^low) did not change much (Fig. 3G). In summary, ICB+Vo caused the most dramatic changes compared with vehicle, whereas ICB caused moderate changes in some immunocytes, and vorinostat alone had negligible effect.

To further elucidate the shift in myeloid cells, a second CyTOF experiment with a myeloid-focused antibody panel (Supplementary Table S4) was conducted for a similar four-group therapeutic experiment (Supplementary Fig. S3B). Among the 20 cell clusters identified by Phenograph (Supplementary Fig. S3C), 5 clusters had their fractions in the ICB+Vo therapy over 50% among all conditions, including CD8⁺ T, BATF^hi cDC1, BATF^lo cDC1, iNOS⁺ DCs, and an unspecified CD11b⁺ population; by contrast, the 2 clusters with fractions in the ICB+Vo therapy lower than 10% among all conditions were both M2-macrophages (CX3CR1^high and CX3CR1^low; Supplementary Fig. S3D and S3E). Among the changes, the dramatic increase of iNOS⁺ DCs (CD11b⁺ CD11c⁺ MHC-II⁺) by the ICB+Vo therapy (Supplementary Fig. S3F) echoes the previous report on this population being indispensable for effective antitumor CD8⁺ T-cell therapy (27). Overall, the lymphocyte and myeloid landscapes reshaped by the ICB+Vo therapy favor an antitumor T-cell immunity, explaining the superior tumor-eradiating activity of this combination.

On the basis of BHB being an endogenous HDACi (18), we postulated that KD or BHB could replace vorinostat to cooperate with ICB in treating PPS-ICBR tumors. First, the HDACi activity of BHB was confirmed because PPS-ICBR cells treated with 3 mmol/L BHB showed higher H3K9Ac and MHC-I, comparable with the effect from vorinostat (Fig. 4A and B). When C57BL/6 mice were fed with SD and KD (Fig. 4C), KD generated 10-fold increase in serum BHB compared with SD (Supplementary Fig. S4A). The basal BHB level (0.2−0.3 mmol/L) and KD-induced BHB level (2−3 mmol/L) in mice were consistent with the corresponding values in humans (28). Continuous KD feeding for 2 weeks delayed tumor growth (Supplementary Fig. S4B), yet also led to over 20% weight loss (Supplementary Fig. S4C). To alleviate the adverse effect of continuous KD, we designed a CKD regimen where tumor-bearing mice were fed 5 days on KD and 2 days on SD every week (Fig. 4D). For a 7-day cycle, serum BHB reached 3 mmol/L at day-5 of the KD period and returned to basal level at day-7 (Fig. 4E). CKD not only allowed the mice to gain weight at a comparable pace as those fed with SD (Supplementary Fig. S4D), but also delayed tumor growth more effectively than continuous KD (Fig. 4F). CKD augmented H3K9Ac signals of PPS-ICBR tumors (Fig. 4G), and prostate cancer cells of CKD-fed mice exhibited higher level of H2-K^b and H2-D^b than prostate cancer cells of SD-fed mice (Fig. 4H). The efficacy of CKD was abolished in both Batf3^−/− mice and Rag1^−/− mice (Fig. 4I). In brief, CKD had single-agent tumor-decelerating activity, but vorinostat did not, and this activity is dependent on the adaptive immunity.

To elucidate the role of BHB in the antitumor effect of CKD, we compared CKD-influenced PPS-ICBR tumor growth in wild-type (WT) and Bdh1-deficient mice (29). Bdh1 (3-hydroxybutyrate dehydrogenase 1) is the enzyme converting acetoacetate to BHB during hepatic ketogenesis (Fig. 4J). Bdh1^−/− mice lacked Bdh1 expression and serum BHB accumulation despite CKD feeding (Fig. 4K and L). Remarkably, the effect of CKD on tumor growth was abrogated in Bdh1^−/− mice (Fig. 4M), proving the essential role of BHB in ketogenesis-induced tumor retardation.

To determine whether ketogenesis enhances ICB, we treated PPS-ICBR–bearing mice with SD, CKD, SD+ICB (anti-PD1+anti-CTLA4), or CKD+ICB. CKD (but not ICB) extended survival as a single therapy compared with SD (Fig. 4N). CKD+ICB further improved survival (Fig. 4N) and reached 14.3% tumor curing (Supplementary Fig. S4E). To test the direct effect of BHB on tumor response, we supplemented SD with the ketone diester 1,3-butanediol, a precursor for BHB that is readily converted to BHB in vivo through hydrolysis (30). Gratifyingly, this BD showed an even more potent effect than CKD to delay tumor growth and cured 23.1% of the tumors when combined with ICB (Fig. 4O; Supplementary Fig. S4E). BD-fed mice reached 1.5 mmol/L serum BHB (Fig. 4P), a level lower than the 3 mmol/L reached by CKD but comparable with the 0.75–1 mmol/L level previously reported by BD-feeding (30). The increased response of BD over CKD may be partly a result of constant BHB presence as opposed to the intermittent presence of BHB by a CKD. H2-K^b level on prostate cancer cells was higher in BD-fed mice than in SD-fed mice and was further enhanced in BD+ICB mice (Fig. 4Q). H3K9Ac level was enhanced in tumors treated with CKD, BD, CKD+ICB, and BD+ICB (Fig. 4R). A direct comparison of ICB+Vo and ICB+CKD confirmed that the two combination regimens reached similar levels of antitumor efficacy (Supplementary Fig. S4F and S4G). In summary, CKD or BD raised BHB level, delayed PPS-ICBR tumor growth as a single agent, and sensitized the tumors to ICB therapy with prolonged survival and 10% to 20% complete response.

We dissociated PPS-ICBR tumors treated with SD, BD, ICB and BD+ICB at day 13 (three biological replicates/condition) and profiled viable cells with 10X Genomics scRNA-seq. Single-cell transcriptomes from 5,749 cells passed the quality control with the biological replicates well represented for each condition (Supplementary Fig. S5A). The cells were clustered into 19 clusters encompassing lymphocytes (CD8⁺ T, CD4⁺ T, Treg, NK) and myeloid cells (2 monocyte subsets, 3 DC subsets, 4 macrophage subsets, 3 neutrophil subsets) based on distinct marker expressions (Fig. 5A–C; Supplementary Fig. S5B–S5C). BD, SD+ICB and BD+ICB exhibited distinct shifts of cell fractions compared with SD (Fig. 5D). BD+ICB showed the most dramatic changes, including 4 upregulated populations (CD8⁺ T, Vcan⁺ monocytes, M1-macrophages and Nos2⁺ DCs) and 4 downregulated populations (Trem2⁺ M2-macrophages, Hilpda⁺ neutrophils, Cxcr2⁺ neutrophils, mesenchymal cells; Fig. 5E). Nos2⁺ DCs expressed Tnf (Supplementary Fig. S5C), defining them as TNF/iNOS-producing (Tip)-DCs with reported essential role in CTL therapies (27). The immune population changes in the BD+ICB condition reinforced an overall anticancer activity reminiscent of the changes caused by ICB+Vo. Nonetheless, the reduction of neutrophils (Hilpda⁺ and Cxcr2⁺) by BD+ICB and the reduction of Cxcr2⁺ neutrophils by BD alone (Fig. 5E) were not observed by either Vo or ICB+Vo, suggesting the existence of HDACi-independent activity of BHB to restrict tumor-associated neutrophils/PMN-MDSCs.

Besides cell fraction changes, we also examined the differentially expressed genes for the cell clusters between SD and BD+ICB (Fig. 5F; Supplementary Table S5). CD8⁺ T cells from BD+ICB upregulated several genes essential for cytotoxic functions, including Id2, Hif1a, Cxcr6 and Il2ra (Cd25). For myeloid cells, the genes simultaneously upregulated in monocytes, macrophages and neutrophils are involved in IFNγ response and antigen presentation (e.g., B2m, Tap1, Irf1, Stat1, Cxcl10, Cd274). Several MHC-I genes (H2-K1, H2-D1, H2-Q7) were upregulated in macrophages and neutrophils, suggesting the stimulation of MHC expression by BHB extends beyond cancer cells. The significant downregulation of Apoe in Vcan⁺ monocytes, M1-macrophages and Nos2⁺ DCs likely reduced immunosuppression and contributed to ICB responsiveness (31).

We next conducted gene set enrichment analysis of MSigDB hallmark and canonical pathways (Fig. 5G). For CD8⁺ T cells, both SD+ICB and BD+ICB enriched for IFNα and IFNγ responses, T-cell receptor signaling, IL12-STAT4 pathway, IL2-STAT5 pathway, CD28 costimulation pathway, and mTORC1 signaling. For Vcan⁺ monocytes, both SD+ICB and BD+ICB enriched for various IL pathways (IL2, IL6, IL7, IL20, and IL27), type I IFNα/IRF3 signaling, Toll-like receptor signaling, STING pathway, and CD28 costimulation pathway. Hence, SD+ICB and BD+ICB shared marked similarities in pathway activations for CD8⁺ T cells and Vcan⁺ monocytes, whereas BD alone had little effect. Therefore, the significance of BD in the BD+ICB combination is more likely through mitigating immunosuppression (e.g., reducing Cxcr2⁺ neutrophils). The comparable effect of SD+ICB and BD+ICB on pathway enrichment was also observed for Trem2⁺ M2-macrophages (Fig. 5G), consistent with these cells being reduced in proportion by both treatments (Fig. 5E). Nos2⁺ DCs (Tip-DCs) relied on the BD+ICB combination to enrich for pathways relevant to proliferation, activation, and antigen-presentation (Fig. 5G), suggesting the active participation of these cells to stimulate CTLs (27). M1-macrophages also relied on the BD+ICB combination to activate pro-inflammatory pathways, and the combination appeared capable of reversing the negative influence of BD on these pathways (Fig. 5G). Finally, BD and BD+ICB weakened glycolysis and gluconeogenesis pathways in Hilpda⁺ and Cxcr2⁺ neutrophils (Fig. 5G), suggesting BHB-induced metabolic debilitation of these neutrophils. Overall, scRNA-seq reveals that, although BD and ICB each generated a moderate impact on the immune landscape, the combination drastically fostered a CTL-favoring microenvironment and enabled cancer cell elimination.

To explore intercellular communications, we used CellChat to quantitatively infer cell–cell communication networks based on ligand–receptor expression patterns from scRNA-seq data (32). First, interactomes for each condition were depicted with circle plots for CD8⁺ T, CD4⁺ T (Treg excluded), monocytes (Vcan⁺), DCs (Nos2⁺), macrophages (M1, Trem2⁺ M2, Mmp2⁺ M2), neutrophils (Hilpda⁺, Cxcr2⁺), and mesenchymal cells (Fig. 6A). SD+ICB and BD+ICB conditions featured visually more extensive connections among cell clusters, especially feeding in and out of CD8⁺ and CD4⁺ T cells. When displayed on scatter plots to compare the outgoing and incoming interaction strengths for all the cell clusters, SD+ICB and BD+ICB also showed dramatically augmented counts of sent or received signals for CD8⁺ T, CD4⁺ T, monocytes (Vcan⁺, Ifit⁺), macrophages (M1, Trem2⁺ M2, Mmp⁺ M2) and DCs (Nos2⁺), yet Treg and Batf3⁺ cDCs remained weak (Supplementary Fig. S6A). The information flow of each signal molecule charted in CellChat also demonstrated the augmented intercellular communications under SD+ICB and BD+ICB conditions, with BD+ICB selectively enriched for a few signals, including IL2, IL4, VEGF, TRAIL, LIGHT, and ACTIVIN (Supplementary Fig. S6B).

To gain insight into how myeloid cells regulate T cells, we focused on the LRPs with myeloid cells (Vcan⁺ monocytes, Nos2⁺ DCs, M1-macrophages) as the ligand producers (sender) and CD8⁺ or CD4⁺ T cells as the receptor owners (receivers). The significant LRPs for the BD+ICB condition support various functions, including MHC interaction, T-cell stimulatory/inhibitory signaling, chemokine/cytokines signaling, and adhesion (Fig. 6B). Many of these LRPs were also significantly represented in the SD+ICB condition but weaker communication probabilities. A few LRPs were only present in BD+ICB, suggesting their specific contribution to the BD+ICB therapy, with the most obvious example being the several LRPs with Cxcr3 as the receptor (Cxcl4-Cxcr3, Cxcl9-Cxcr3, and Cxcl10-Cxcr3; Fig. 6B). Circle plots for the Cxcr3-oriented LRPs under BD+ICB condition depicted the strong and directional connections from myeloid and mesenchymal cells toward CD8⁺ and CD4⁺ T cells (Fig. 6C). Using anti-Cxcr3 antibody to block Cxcr3-mediated chemoattraction, we confirmed that Cxcr3 was critical for the effect of BD+ICB on efficient CD8⁺ T-cell infiltration (Fig. 6D) and antitumor activity (Fig. 6E). Interestingly, although the Cxcr3 chemokine Cxcl9 originated mainly from Nos2⁺ DCs and M1-macrophages, another Cxcr3 chemokine Cxcl10 predominantly emerged from Vcan⁺ monocytes (Fig. 6C), suggesting myeloid subtype-specific chemokine production.

The simultaneous increase of Vcan⁺ monocytes, Nos2⁺ DCs, and M1-macrophages by BD+ICB suggests relationships among these cells. We focused on the myeloid island in the middle of the UMAP and applied Monocle 3 to construct the potential transitional trajectories among myeloid cells under BD+ICB condition (Fig. 6F). Pseudotime suggested a trajectory from monocytes to both DC and macrophage branches (Fig. 6G), consistent with the ability of monocytes to differentiate into macrophage and monocyte-derived DCs in the TIME (33).

Although the tumor modeling was based on the subcutaneous site instead of the bone, the effect of the treatments may still have relevance to the bone microenvironment. Jiao and colleagues (34) reported that the bone metastatic castration-resistant prostate cancer (mCRPC) microenvironment restrained Th1 development and hampered the therapeutic response to ICB therapy; blocking immunosuppressive signaling in the bone (especially TGFβ) along with ICB increased Th1 and decreased Treg cells, promoting CD8⁺ T-cell expansion and regression of bone mCRPC. BD+ICB combination strongly upregulated markers for CD8⁺ T and Th1 cells but downregulated Foxp3 for Treg (Supplementary Fig. S6C). Both Th1/Treg ratio and CD8⁺ T/Treg ratio were highest in the BD+ICB condition (Supplementary Fig. S6D). These results suggest that the T-cell immunity shift promoted by BD+ICB shares a similar trend to the shift promoted by the anti-TGF-β+ICB therapy in bone mCRPC (34).

Lastly, we examined the effect of CKD and BD on the MyC-CaP model of CRPC, which showed complete resistance to ICB (20). MyC-CaP tumors progressed to CRPC after castration, and CKD or BD alone decelerated tumor growth (Supplementary Fig. S6E–S6H). When CKD or BD were combined with ICB (anti-PD1+anti-CTLA4), no additional efficacy was achieved (Supplementary Fig. S6F–S6H), suggesting that MyC-CaP CRPC may harbor additional immunosuppressive forces absent in the PPS-ICBR tumors. PPS-ICBR tumors did not contain noticeable level of B cells (Fig. 5D); by contrast, MyC-CaP CRPC tumors contain a population of immunosuppressive B cells capable of promoting CRPC and impeding CTL-dependent immunogenic chemotherapy (35, 36). We depleted B cells with anti-CD20 (Supplementary Fig. S6I) and compared its effect with BD or BD plus anti-CD20. Both BD and anti-CD20 retarded MyC-CaP CRPC, and the combination caused the most significant delay (Supplementary Fig. S6J). This result highlights that BD may enhance different types of immunotherapies.

Our study connects two promising anticancer regimens, HDACi and KD, demonstrating that HDACi plus ICB and KD plus ICB generate combinatorial anti-tumor efficacy in a newly developed ICB-refractory prostate cancer model. Multiple early-phase clinical trials are testing the combination of vorinostat and anti-PD1 in malignancies such as glioblastoma (NCT03426891), urothelial and renal carcinomas (NCT02619253), non–small cell lung cancer (NCT02638090), and lymphomas (NCT03150329). Early reports of several trials showed that pembrolizumab plus vorinostat was well tolerated and had promising antitumor activities (37–39). Vorinostat used in our study (25 mg/kg daily in mice) is equivalent to 2 mg/kg for human patients with a reference body weight of 60 kg based on the mouse–human conversion factor 12.3 (FDA guidance; ref. 40). This dose is significantly lower than the recommended vorinostat dose to treat adult cutaneous T-cell lymphoma (400 mg/d) and falls below or within the lower spectrum for the clinical trials that combine vorinostat and anti-PD1 (e.g., NCT02638090 used vorinostat at 200 or 400 mg/d to combine with pembrolizumab; ref. 37). Therefore, the vorinostat dose based on our preclinical results should be tolerated if translated to clinical trials. One clinical trial was initiated recently to test the combination of KD (continuous or discontinuous) or BHB supplementation with nivolumab plus ipilimumab to treat metastatic renal cell carcinoma (NCT05119010). The result from this trial is eagerly awaited to provide evidence for ketogenesis-enhanced immunotherapy.

Two recent studies reported the effect of KD to enhance ICB therapy in preclinical models (41, 42). Although our study agrees with the overall conclusion of these two publications, a few unique strengths exist in our study. (i) We are the first to demonstrate ketogenesis-enhanced immunotherapy in prostate cancer models. Compared with other cancer types, prostate cancer exhibits unique challenges in ICB resistance, accentuated by its low response to dual anti-PD1 and anti-CTLA4 inhibitions (43). (ii) scRNA-seq was applied in our study to explore the TIME reprogramed by BD and ICB, which revealed the dramatic upregulation of CD8⁺ T cells and three myeloid populations (Vcan⁺ monocytes, Nos2⁺ DCs, M1-macrophages) only under the BD+ICB treatment. (iii) Although Dai and colleagues (42) discovered that KD-triggered energy deprivation stimulated AMPK activation, causing PD-L1 degradation and IFN upregulation, Ferrere and colleagues (41) associated the KD-enhanced immunotherapy with compositional changes in gut microbiota. By contrast, our result emphasizes both cancer cell–intrinsic and -extrinsic mechanisms elicited by BD+ICB. At the cancer cell–intrinsic level, the HDACi activity of BHB augmented MHC-I expression in prostate cancer cells, rendering them more visible to the immune system. At the cancer cell–extrinsic level, BD alone had a relatively moderate effect on the TIME (except for reducing Cxcr2⁺ neutrophils). However, the BD+ICB combination drastically reshaped the TIME to favor tumor-killing T-cell immunity. We believe these intrinsic and extrinsic mechanisms at least complement the mechanisms in Dai and Ferrere's studies and illustrate how an optimized KD regimen can take diverse routes to enable effective immunotherapy.

Our experiments using Bdh1^−/−, Batf3^−/− and Rag1^−/− mice demonstrated the essential role of both BHB and adaptive immunity in establishing the antitumor activity of CKD. This result echoes recent studies that elucidated various immune-modulatory activities of BHB in various physiological contexts, which can be classified as T-cell–related and myeloid cell-related. For T cells, KD or BHB can expand γδ T cells (44), fueling oxidative phosphorylation and restoring the function of CD4⁺ T cells (45), enhance memory T cells (46), and bolster the effector function of CTLs (47). Surprisingly, BHB was even preferred over glucose by CTLs for acetyl-CoA production (47). For myeloid cells, BHB often plays an anti-inflammatory role (30, 48) and increases M2-macrophages (49, 50). We observed that BD alone increased Mmp⁺ M2-macrophages and decreased Cxcr2⁺ neutrophils (Fig. 5E). Intriguingly, the addition of ICB curbed BD's effect on Mmp⁺ M2-macrophages yet amplified BD's effect on Cxcr2⁺ neutrophils (Fig. 5E), with the net result supporting antitumor immunity. Future work will further resolve why the interplay between BD and ICB generates an immune landscape that maximizes CTL activity and minimizes immunosuppression.

In summary, we developed a syngeneic prostate cancer cell line series with PPS-ICBR as a robust model for dual ICB resistance. Using this model, we demonstrated that epigenetic modulation (in a broad sense) with pharmacological HDACi vorinostat or endogenous HDACi BHB (via CKD or BD) sensitized ICB-refractory prostate cancer to PD1/CTLA4 dual ICB therapy. Although these two combination therapies cause some similar changes in the tumors, such as cancer cell MHC-I upregulation and innate and adaptive immune landscape remodeling favoring antitumor activities, they also have apparent differences in the influences on specific immune populations (e.g., Cxcr2⁺ neutrophils) and possibly other unexamined tumor hallmarks. Among the strategies, BD may represent the most attractive option to enhance ICB therapy in patients due to its high chance of patient compliance, low cost, and minimal toxicity.

Z.T. Schafer reports grants from NIH during the conduct of the study as well as grants from CureFA Foundation outside the submitted work. No disclosures were reported by the other authors.

S. Murphy: Conceptualization, data curation, formal analysis, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. S. Rahmy: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft. D. Gan: Data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft. G. Liu: Investigation, methodology. Y. Zhu: Investigation, Methodology. M. Manyak: Investigation, methodology. L. Duong: Investigation, methodology. J. He: Methodology. J.H. Schofield: Methodology. Z.T. Schafer: Supervision. J. Li: Supervision. X. Lu: Supervision. X. Lu: Conceptualization, resources, formal analysis, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing.

We thank the Lu laboratory members for their essential comments and suggestions during this work. We thank Daniel P. Kelly from University of Pennsylvania for sharing Bdh1^−/− mice. We appreciate the technical assistance on scRNA-seq from Qingfei Wang and Siyuan Zhang. We are grateful for the support from core facilities used in this study, especially Freimann Life Science Center (Teri Highbaugh) and Genomics and Bioinformatics Core Facility (Michael Pfrender, Melissa Stephens, Jacqueline Lopez, Brent Harker). This work was supported by a research grant from American Institute for Cancer Research (to X. Lu), National Institutes of Health grant 5F99CA274694 (to S. Murphy), and a core facility grant from Indiana Clinical and Translational Sciences Institute and National Institutes of Health grant UL1TR002529 (to X. Lu). Other support included National Institutes of Health grants R01CA248033 and R01CA280097, Department of Defense grants W81XWH2010312, W81XWH2010332, HT94252310010, and HT94252310613, and Boler Family Foundation (to X. Lu) at University of Notre Dame.