Melanoma MM200, ME4405, MEL-RM and B16/F10 were purchased from the American Type Culture Collection (ATCC, Manassas, VA, USA). The obtained cell lines were authenticated by short tandem repeat (STR) profiling through the Victorian Centre for Functional Genomics. The most recent authentication was on March 1, 2017. The cells were maintained in a recommended medium with 10% fetal bovine serum (Gibco, Grand Island, NY, USA) and antibiotics (50 mg/mL-1penicillin/streptomycin) and routinely tested for Mycoplasma species prior to performing related experiments.

The plasmids encoding human or mouse SOX2 (pmCherry-Flag-hSOX2 and pCDH-Flag-mSOX2), plasmid knockdown (KD) human SOX2 (pLKO.1-hSOX2) and plasmid knockout mouse SOX2 (pLenti-V2-mSOX2) were constructed, and siRNAs were obtained from GenePharma (Shanghai, CHN). Transfection of plasmids or siRNAs to melanoma cells was performed using the Lipofectamine 2000 (No. 11668019, Invitrogen, Carlsbad, CA, USA) or Lipofectamine RNAiMAX transfection reagent (No. 13 778–150, Invitrogen, Carlsbad, CA, USA) according to the manufacturer’s instructions. Information for primer sequences are as listed: siSOX2#1/shSOX2#1, human, 5′-AGAAATATCCTCCTTACTC-3′; siSOX2#2/shSOX2#2, human, 5′-TGACGTCAATGCTGCCATA-3′, siSOX2#3, human, 5′CGAGGUGCUGAGCAAGAAA-3′, sgSOX2#1, mouse, 5′-GCAGGGCGCTGACGTCGTAG-3′, sgSOX2#2, mouse, 5′-CATGTATAACATGATGGAGA-3′.

The antibodies used for western blotting: anti-SOX2 (No. ab92494, abcom, Cambridge, MA, USA), anti-p-JAK1 (Tyr1022/1023) (No. GTX50207, GeneTex, San Antonio, TX, USA), anti-JAK2 (No. GTX31943, GeneTex), anti- SOCS3 (No. A0694, Abclonal, Wuhan, CHN), anti-PTPN1 (No. A1590, Abclonal), anti-GAPDH (No. 60 004–1, Proteintech, Chicago, IL, USA). Anti-JAK1 (No. 3344S), anti-p-JAK2 (Tyr1007/1008) (No. 3776S), anti-STAT1 (No. 9172S), anti-p-STAT1 (Tyr701) (No. 7649S), anti-STAT3 (No. 9139S), anti-p-STAT3 (Tyr705) (No. 9145S), anti-Acetylated-Lysine (No. 9681S) and anti-β-Tubulin (No. 2146S) were purchased from Cell Signaling Technology (Danvers, MA, USA).

Flow cytometry antibodies: Alexa Fluor 488 antimouse CD3ε (No. 100321), Alexa Fluor 647 antihuman/mouse GranzymeB (No. 515406) and PerCP/Cyanine5.5 antimouse IFNγ (No. 505821) were purchased from Biolegend (San Diego, CA, USA). Ms CD4 PE-Cy7 RM4-5 (No. 552775), Ms CD8a PerCP-Cy5.5 53–6.7 (No. 551162), FVS700 (No. 564997), Ms CD45.1 PerCP-Cy5.5 A20 (No. 560580), Ms Foxp3 Alexa 647 R16-715 (No. 563486) and PE antimouse NK cells (NK1.1) (No. 557391) were purchased from BD Pharmingen (San Diego, CA, USA).

The epigenetic compounds library, inhibitors and cytokines are as listed: epigenetic compounds library (No. L1900, Selleck Chemicals, Houston, TX, USA), SAHA (No. S1047, Selleck Chemicals), Etidronate (No. S1857, Selleck Chemicals), PTP1B inhibitor (No. T4256, TargetMol, Boston, MA, USA), protein tyrosine phosphatase (PTP) inhibitor I (No. T7084, TargetMol), SPI 112 (No. T4340, TargetMol), murine IFNγ (No. 315-05-100, PeproTech, Rocky Hill, NJ, USA), human IFNγ (No. 300-02-1000, PeproTech), human IL2 (No. 200-02-100, PeproTech), antimouse PD-1 (CD279) (No. BE0146, BioXCell, West Lebanon, NH, USA), Rat IgG2a isotype control (No. BE0089, BioXCell).

Six-week-old C57BL/6 and nude mice were procured from the Laboratory Animal Center of Guangdong province (Guangzhou, CHN). These animals were maintained in accordance with the institutional guidelines and the experiments were approved by the Animal Care and Use Committee of Sun Yat-sen University Cancer Center.

5×10⁶ human MEL-RM cells were subcutaneously injected into the flank of nude mice or 0.5×10⁶ B16/F10 cells were injected into the back flank C57BL/6 mice. Four days following the injection, the mice were randomized into the following four groups: (1) vehicle; (2) treatment with SAHA (25 mg/kg, intraperitoneally, every other days); (3) treatment with anti-PD-1 (100 µg per mice, intraperitoneally, every 3 days) or IFNγ (2.5 µg per mice, intraperitoneally, every 3 days); (4) treatment with a combination of SAHA and anti-PD-1 or IFNγ. The tumors were measured every third-day for nude mice or second-day for CB7BL/6 mice and tumor volume was calculated as (length x width²)/2. Tumor growth was monitored over a period of 8 days and the mice were sacrificed.

The tumors were cut into small pieces, mechanically disrupted and filtered through a 70 µm mesh to generate a single-cell suspension. Dissociated tumor cells were lysed with Red Blood Cell Lysis Solution (No. 420301, Biolegend) and incubated with phorbol 12-myristate 13-acetate (PMA) (No. P1585, Sigma-Aldrich, St. Louis, MO, USA), ionomycin (No. 73724, Stemcell, Vancouver, BC, CAN) and Golgi stop (No. 554715, BD Pharmingen) at 37°C. For cell surface staining, cell suspensions were stained with indicated fluorescent labeled antibodies for 30 min on ice. For intracellular staining, the cells were sorted for fixation and permeabilization using Cytofix/CytoPerm BUF KIT (No. 554714, BD Pharmingen) according to the manufacturer’s guidelines and incubated with primary antibodies. Flow cytometry acquisition was carried out on LSRFortessa (BD Biosciences, San Jose, CA, USA) or a Navios and Gallios (Beckman Coulter, Miami, FL, USA). Compensation beads were used to evaluate spectral overlap. The data were analyzed using FlowJo and CytExpert software according to manufacturers’ instructions.

CD3+ T cells were isolated from human peripheral blood mononuclear cells and activated with 10 ug/mL LEAF Purified anti-CD3 (No. 300314, Biolegend), 2 ug/mL LEAF Purified anti-CD28 (No. 302913, Biolegend) and 100 IU/mL human IL2 (No. 200-02-100, PeproTech) as previously reported.15 Pretreated melanoma cells were cocultured with the activated T cell in a ratio of 10:1 or 20:1 for 6–8 hours. The cells were sorted, stained with FITC Anti-Active Caspase-3 (No. 559341, BD Pharmingen) and used for flow cytometric analysis.

For colony formation assay, a total of 300 melanoam cells were seeded in a 6-well plate and cultured in complete DMEM medium for 10 days. Colonies were fixed with 4% polymethanol and dyed with 0.1% crystal violet (1 mg/mL), and the number of colonies with over 50 cells was counted. For CCK-8 assay, a total of 300 melanoam cells were seeded in triplicate in each well of a 96-well plate, and the cell numbers were counted every day by CCK-8 (No. CK04-500, DOJINDO, Kumamoto, Japan) for 7 days.

The cells were lysed in 1x Cell Lysis Buffer (No. 9873, CST) supplemented with protease inhibitor phenylmethylsulphonyl fluoride. Protein was quantitated using the BCA Protein Assay Kit (No. 23227, Thermo Fisher Scientific, Waltham, MA, USA). Equal amounts of protein were boiled by adding 6x SDS sample buffer for 10 min at 100°C and resolved using SDS-PAGE. Membranes were blocked in 5% milk/TBS-T (Boston Bioproducts) and incubated with their respective antibodies.

Total RNA was extracted using TRIzol reagent (No. 15596018, Invitrogen) according to the manufacturer’s instructions. The RNA concentration was determined using a Nanodrop (Thermo Fisher Scientific). cDNA was reversed from 1 µg of total RNA using the HiScript II Q Reverse Transcription kit (No. R233-01, Vazyme, Nanjing, CHN). QPCR amplification was performed using the SYBR master mix plus (No. Q311-03, Vazyme) and run on Bio-Rad CFX96 qPCR system. All reactions were performed in duplicate. Relative gene expression was quantified and normalized to GAPDH. Both the forward and reverse primers used are designed form Primerbank https://pga.mgh.harvard.edu/primerbank.

HEK293T cells were cotransfected with luciferase-reporter plasmids and PGL3 or PGL3-SOCS3/PGL3-PTPN1 plasmids using the Lipofectamine 2000 transfection reagent. Lysis buffer (25 mM of Tris phosphate, 1% Triton X-100, 1 mM of dithiothreitol, 2 mM of ethylenediaminetetraacetic acid, 10% glycerol, PH=7.8) was added at 48 hours post-transfection. Luciferase assay was performed using a luciferase assay kit (No. E1910, Promega, Madison, WI, USA). An enzyme-linked immunosorbent-assay plate reader (Bio-Rad) was applied to measure O-nitrophenol at a wavelength of 490 nm to evaluate β-galactosidase activity.

Epigenetic compound library was purchased from Selleck. Five hundred thousand MM200 cells were seeded overnight in 12-well plates. Following IFNγ stimulation, the inhibitors were added to the culture medium at a concentration of 10 µM. After 24 hours, the cells were harvested and lysed for western blot.

Normalized gene expression data and paired clinical feature data of patients without immunotherapy were download from the The Cancer Genome Atlas (TCGA) database (https://tcga-data.nci.nih.gov). Data of patients treated with nivolumab (accession number GSE 9106116) or pembrolizumab (accession number GSE7822017) were obtained from Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo). Overall survival (OS) was defined as the time from randomization to death from any cause and progression-free survival (PFS) was defined as the time from randomization to documented disease progression according to Response Evaluation Criteria in Solid Tumors, version 1.1 (RECIST V.1.1) or death from any cause. Objective response (ORR) was the proportion of patients with confirmed complete response or partial response according to RECIST V.1.1.18

The optimal cut-off for PD-L1 was the median, meaning that mRNA level ranking at the top 50% of the patients was defined as PD-L1 high and the others were PD-L1 low.

The optimal cut-offs for SOX2 were determined using receiver operating characteristic (ROC) analysis. In the prenivolumab cohort (GSE96061), the cut-off of SOX2 was 0.054, the sensitivity, specificity and area under the curve (AUC) were 77.8, 87.5% and 79.7%, respectively (p=0.017). In the onnivolumab cohort (GSE96061), the cut-off of SOX2 was 0.204, the sensitivity, specificity and AUC were 80.0, 80.0% and 80.0%, respectively (p=0.013). In the perpembrolizumab cohort (GSE96061), the cut-off of SOX2 was 0.925, the sensitivity and specificity were 83.3% and 57.1%, respectively.

Kaplan-Meier plots for OS and PFS were compared using the log-rank test. The comparison of clinicopathological characteristics was performed with the use of a log-rank test stratified at a two-sided alpha level. A Cox proportional-hazards model was used to estimate the HR and its corresponding CI for death. The R programming language (http://cran.r-project.org) was used to classify the different gene signature patterns of the patients. The DAVID database (Database for Annotation, Visualization and Integration Discovery, http://david.abcc.ncifcrf.gov) was used to conduct a functional enrichment analysis.19 Statistical analyses were performed using GraphPad Prism (V.5.0, http://www.graphpad.com) and SPSS V.24.0. All data were presented as the mean±SD or were otherwise noted in the legends. All reported p values were two-tailed. For all analyses, p<0.05 was considered as statistically significant unless specified otherwise. All data shown are representative of two or more independent experiments, unless indicated otherwise.

To establish relationship between SOX2 and host antitumor immune response in vivo, vector and SOX2 OE B16/F10 cell lines were generated and verified by western blot (online supplemental figure 1A) and transplanted into the syngeneic C57BL/6 mice (figure 1A). SOX2 OE promoted melanoma growth as confirmed by the growth curve of the xenografts volume (figure 1B). Immune cell subsets pooled from tumor infiltrating lymphocytes were analyzed by multicolor flow cytometry analysis (online supplemental figure 1B). The results showed that SOX2 OE suppressed CD8+ T cells (figure 1C, online supplemental figure 1C), while had no influence on CD4+ T/effector T cells, natural killer (NK) cells and Treg cells (figure 1F–J). Importantly, CD8+ T cells infiltration were accompanied by a decreasing in cytotoxic function, which was confirmed by GranzymeB and IFNγ staining (figure 1D and E, online supplemental figure 1C). Next, we investigated whether SOX2 could impair CD8+ T-cell killing in vitro. We generated SOX2 KD melanoma cell lines (ME4405 and MM200) by indicated shRNA (online supplemental figure 2A,B) and performed a T-cell-mediated cytotoxicity assays. SOX2 KD enhanced the sensitivity of melanoma cells to T-cell killing (figure 1K and L). We next validated whether SOX2 exhibited direct impact on tumor growth in vitro. As assessed by colony formation and CCK8 assay, we found SOX2 OE (online supplemental figure 2C) or KD had mild effect on colony formation (online supplemental figure 2D,E) and tumor growth (online supplemental figure 2F–I). These finding suggested that SOX2-related tumor growth was caused by immune evasion of CD8+ T-cell killing.

To explore the impact of SOX2 on treatment-naive patients OS, 420 melanoma patients data download from TCGA database was analyzed. The result showed that SOX2 high status indicated a poor OS (p value, 0.031) in CD8+ T-cell proficient patients while it could not stratify survival in CD8+ T-cell absent patients (online supplemental figure 2J). IFNγ and GranzymeB were both effectors of cytotoxic T cell and had a critical role in antitumor immunity, we also observed that SOX2 predicted poor OS in IFNγ or GranzymeB high expression patients (p value, IFNγ: 0.016, GranzymeB: 0.031). Besides, IFNγ or GranzymeB low expression could not contribute to the apparent discrepancy of OS between different SOX2 statuses patients (online supplemental figure 2K,L). Together, we concluded that SOX2 prompted tumor progression through causing CD8+ T-cell tolerance.

Since the duration of tumor IFNγ signaling and overexpression of ISG.RS accounted for the resistant to ICIs,20 21 we hypothesized that SOX2 impaired CD8+ T-cell killing through regulating the IFNγ signaling. We detected ISG.RS mRNA level in negative control or SOX2 KD melanoma cell lines and observed that SOX2 KD decreased the ISG.RS expression (IDO1, PDL1, IFI27, and USP16) (figure 2A,B). ME4405 and MM200 cells were pretreated with IFNγ to induce ISG.RS expression and then cocultured with activated human peripheral blood T cells. We found IFNγ exposures caused melanoma cells escaping from T-cell killing. When SOX2 was knocked down, the sensitivity was recovered without significantly difference between the IFNγ treated and untreated group (figure 2C,D), speculating that SOX2 KD regained the sensitivity to T-cell killing depending on attenuating IFNγ-induced ISG.RS expression.

Next, we further explored how the transcriptional repressor SOX2 regulated ISG.RS expression. We found SOX2 KD suppressed the JAK-STAT pathway activity (figure 2E). Previously reported that suppressor of cytokine signaling family and protein tyrosine phosphatase family coud counter the activities of JAK-STATs pathway,22 23 we proposed that SOX2 inhibition might increase phosphatase expression. In fact, we found that SOX2 knockout (online supplemental figure 2M) increased SOCS3 and PTPN1 mRNA level (figure 2F). Further, luciferase report assay confirmed that SOCS3 and PTPN1 were transcriptionally induced by SOX2 KD (figure 2G). Together, SOX2 caused immune evasion to CD8+ T-cell killing through inhibiting SOCS3 and PTPN1 transcription, further inducing the hyperactivation of the JAK-STAT pathway and OE of ISG.RS.

Since SOX2 caused immune evasion of CD8+ T-cell killing, we assumed that suppressing SOX2 could argument antitumor immunity. As methylation and acetylation promoted SOX2 ubiquitination and proteasome degradation,24 25 we performed western blot-based screen using an epigenetic compound library to identify epigenetic inhibitors decreasing SOX2. A collection of 96 agents were tested and 20 histone deacetylases could decrease the SOX2 level (figure 3A, online supplemental figure 3A), of which SAHA was chosen for the following experiments since SAHA has been approved for the treatment of cutaneous T-cell lymphoma.26 We confirmed that SAHA decreased SOX2 protein levels but not mRNA levels (figure 3B, online supplemental figure 3B). Then, we introduced CHX to inhibit translation of SOX2. The protein level of SOX2 was decreased significantly when cells were treated with SAHA for 12 hours (online supplemental figure 3C), indicated that SAHA prompted SOX2 degradation. Furthermore, the SAHA-related inhibition of SOX2 could mostly be blocked by the proteasome inhibitor MG132 and partially by the autophagy inhibitor bafilomycin A1 (online supplemental figure 3D). Besides, SAHA increased SOX2 acetylation and ubiquitylation levels (online supplemental figure 3E,F). These data demonstrated that SAHA facilitated SOX2 acetylation and proteasome degradation.

To test whether SAHA could reverse the SOX2-dependent CD8+ T-cell tolerance, vector or SOX2 OE melanoma cells were treated with SAHA before cocultured with activated peripheral blood T cells. As expected, SAHA sensitized the melanoma cells to T-cell killing and partially reversed SOX2-induced CD8+ T-cell tolerance, indicating that the function of SAHA might be achieved through targeting SOX2 (figure 3C,D).

Next, vector or SOX2 OE B16/F10 cells were injected into C57BL/6 mice and treated with SAHA (figure 3E). The result showed that SAHA could reduce the tumor burden and partially reversed SOX2-induce tumor promotion effect (figure 3F and G). Tumor infiltrating immune cells analysis displayed that SAHA increased CD8+ T cells and CD8+ cytotoxic T cells infiltration, and partially eliminated SOX2-related CD8+ T cells and CD8+ cytotoxic T cells inhibition (figure 3H–J, online supplemental figure 4A–C). Additionally, SAHA had a mild influence on the CD4+ T/effector T cells, NK cells and Treg cells (figure 3K–O). Based on these findings, we concluded that SAHA could overcome SOX2-related CD8+ T cells tolerance in vivo.

Mechanistically, SAHA stimulated PTPN1 and SOCS3 transcription (figure 4A and B, (online supplemental figure 5A) and declined the JAK-STAT pathway activity (figure 4C, online supplemental figure 5B). Besides, the p-JAK1/2 and p-STAT1/3 levels were also reduced by SAHA in the SOX2 OE group but were still slightly higher than the vector group (figure 4D, online supplemental figure 5C), confirming that SAHA inhibited JAK-STAT activity through targeting SOX2. To test whether PTPN1 and SOCS3 were required, we treated melanoma cells with SAHA and phosphatase inhibitors for 24 hours. We found that phosphatase inhibitors increased the p-JAK1/2, p-STAT1/3 levels, corroborating that PTPN1 and SOCS3 inhibition could augment the JAK-STAT pathway activity. On PTPN1 and SOCS3 inhibition, SAHA could not fully decrease p-JAK1/2, p-STAT1/3 (figure 4E), suggesting that SAHA blocked SOX2-mediated JAK-STAT pathway excessive activation through stimulating PTPN1 and SOCS3 expression.

Next, we tested whether SAHA could enhance the antitumor effect of IFNγ. SAHA and IFNγ pretreated melanoma cells were cocultured with activated human peripheral blood T cells. We found that SAHA could sensitize the melanoma cells to T-cell killing in IFNγ pretreated group (figure 4F, online supplemental figure 5D). Then, we detected the antitumor effect of SAHA and IFNγ in immune-competent (C57BL/6) mice and immune-compromised (nude) mice. SAHA or IFNγ alone showed mild antitumor response but remarkable tumor regression was achieved in the combination group (figure 4G,H, (online supplemental figure 5E). Besides, the tumors in immunocompromised mice failed to show response to both therapies (online supplemental figure 5F–H), signifying that SAHA could significantly augment the antitumor effect of IFNγ in an immunocompetent context with intact T-cell immunity.

Given that SAHA decreased SOX2 expression and recovered the sensitivity of melanoma cells to T-cell killing, we hypothesized that SAHA could enhance the effect of anti-PD-1 in vivo. For this purpose, C57BL/6 mice were inoculated with B16/F10 cells and then treated with SAHA and anti-PD-1 (figure 5A). SAHA demonstrated slight antitumor activity and anti-PD-1 had no antitumor effect, while the combined treatment could significantly decrease tumor growth (figure 5B,C). Then, tumors infiltrating immune cells were detected. CD8+ T cells and CD8+ IFNγ+ T cells or CD8+ GranzymeB+ T cells were slightly increased in SAHA or anti-PD-1 group, but were significantly infiltrated in the combination group (figure 5D–F, online supplemental figure 6A–C). Additionally, neither treatment increased the CD4+ T/effector T-cells, NK cells and Treg cells population (figure 5G–K). Thus, we concluded that SAHA could reinforce the effect of anti-PD1 through overcoming SOX2-ralated CD8+ T-cell tolerance.

As our results uncovered that SOX2 caused immune evasion to CD8+ T cells killing through regulating IFNγ pathway, we hypothesized that SOX2 might contribute to the apparent discrepancy of response and OS to anti-PD-1 in CD8+ T-cell inflamed patients. To elucidate it, 84 metastatic melanoma patients with transcriptional profiles and clinical data from the GEO repository were used. 58 patients received nivolumab, 43 had paired prenivolumab (before treatment) and onnivolumab (4 weeks after the initiation of treatment) data, seven had only prenivolumab data and eight had only on-nivolumab data. 26 patients treated with pembrolizumab. Of all the patients treated with anti-PD-1 antibodies, the median patient overall OS was 94.6 (95% CI: 61.7 to 127.5) weeks and consisted of 2 M0, 12 M1a, 13 M1b, 47 M1c and 10 patients with unknown stage. As expected, we found that SOX2 high status indicated a poor OS (p value, prenivolumab: 0.042; onnivolumab: 0.007; prepembrolizumab: 0.025) and PFS (p value, prepembrolizumab: 0.064) in CD8+ T-cell inflamed patients while it could not stratify survival in CD8+ T-cell absent patients (online supplemental figure 7A,C. Similarly, SOX2 also had prognostic effect in IFNγ proficient patients (online supplemental figure 7B).

Since PD-L1 is widely used to predict clinical outcomes of anti-PD-1 and was strongly associated with CD8+ T infiltration,27 we evaluated the prognostic impact of SOX2 across different PD-L1 status. No significant difference between the clinicopathological characteristics of the PD-L1^low SOX2^low, PD-L1^low SOX2^high, PD-L1^high SOX2^low and PD-L1^high SOX2^high patients were observed (online supplemental tables 1,2). SOX2 high status (both prenivolumab and onnivolumab) corresponded to a poor OS in PD-L1 high patients (p value, prenivolumab: 0.005; onnivolumab: 0.022) (figure 6A) but it could not stratify the survival of PD-L1 low patients (p value, prenivolumab: 0.826; onnivolumab: 0.863) (online supplemental figure 8A). These results indicated that the baseline and evolution status of SOX2 had a prognostic value for high PD-L1 patients regardless of the intratumor heterogeneity evolution during the nivolumab treatment.16 Similarly, for patients treated with pembroluzimab, SOX2 was also correlated with poor OS and PFS (p value, OS: 0.039; PFS: 0.047) in PD-L1 high patients (figure 6B) while it could not stratify the survival of PD-L1 low patients (p value, OS: 0.589; PFS: 0.854) (online supplemental figure 8B).

Besides, ORR was higher in PD-L1 high patients with SOX2 low expression than SOX2 high expression (SOX2 low vs SOX2 high, prenivolumab: 36.4% (4 of 11) vs 7.7% (1 of 13); onnivolumab: 27.3% (3 of 11) vs 15.4% (2 of 13) (figure 6C, online supplemental figure 8C); prepembroluzimab: 80.0% (4 of 5) vs 25.0% (2 of 8) (figure 6D). Multivariate cox regression analysis using BRAF/NRAS/NF1 mutation status, tumor stage and previous ipilimumab treatment as adjusting factors identified SOX2 as a negative independent predictor for OS in PD-L1 high patients (prenivolumab, HR: 5.67, 95% CI: 1.68 to 19.17, p value: 0.005; onnivolumab, HR: 3.69, 95% CI: 1.02 to 13.40, p value: 0.047) (online supplemental tables 3,4). According to the AUC of the ROC model, we found that the SOX2 status was a better predictor in high PD-L1 patients than low PD-L1 patients (AUC, PD-L1 high vs PD-L1 low, prenivolumab: 0.799 vs 0.521; onnivolumab: 0.800 vs 0.531) (figure 6E and F). The patients treated with pembroluzimab were not included in the analysis due to small sample size (26 patients). We found that SOX2 was an independent predictor for poor survival and resistance to anti-PD-1 therapy in melanoma with high PD-L1 expression.