Verteporfin Inhibits PD-L1 through Autophagy and the STAT1–IRF1–TRIM28 Signaling Axis, Exerting Antitumor Efficacy

Discovered first as a B7 family ligand of the programmed cell death 1 (PD1) encoded by the CD274 gene (1), PD-L1 is a checkpoint molecule that initiates an inhibitory pathway, suppressing CTLs upon PD-1 engagement. This mechanism that normally serves to prevent over activation of immune responses is frequently coopted by cancer cells to evade immune surveillance (2). Subsequent studies have provided compelling evidence that the PD-L1/PD1 cascade is a highly effective therapeutic target for immune checkpoint blockade therapy that yields durable anticancer efficacy and prolongs patient survival (3).

PD-L1 is expressed in a variety of cancer types in either a constitutive (or intrinsic) or IFN-induced manner, and the mechanisms controlling the expression of PD-L1 mRNA and protein have been the subject of numerous studies. The STAT1–IRF1 axis plays a central role in mediating IFN-induced PD-L1 transcription in both cancer and noncancer cells (4). In addition to transcription regulation, epigenetic mechanisms, 3′ UTR variations, and miRNAs have roles in fine-tuning the relative expression of PD-L1 in context-dependent settings (5). Other mechanisms control PD-L1 degradation, ubiquitination, autophagy, glycosylation, and recycling from plasma membrane and intracellular compartments (6, 7). In contrast to IFN-dependent PD-L1 transcription, the mechanisms underlying intrinsic PD-L1 expression in cancer are poorly understood.

Toxicity and cost are two key issues, among others, that limit the use of antibody approaches to immune checkpoints (8, 9). The complexity of the mechanisms controlling PD-L1 expression poses another major challenge for the development of small-molecule PD-L1 inhibitors. Targeting one of the mechanisms controlling either intrinsic or IFN-induced PD-L1 expression may not be sufficient because this pathway controls the expression of >7,000 genes. Many of these genes are essential for immune responses, especially those along the CIITA-MHC cascade, which are crucial for cancer immunogenicity (10). Thus, it is imperative to preserve the CIITA-MHC cascade and immune responses while still targeting PD-L1 expression.

In mammalian cells, the endoplasmic reticulum (ER)-Golgi network plays an important role in autophagy and is a major source of membrane structures contributing to formation of autophagosomes (11, 12). PD-L1 is one of a large number of glycosylated proteins, including most of the MHC antigens, that are processed through the Golgi apparatus (13). The conserved oligomeric Golgi proteins (COG) in the Golgi apparatus play an important role in posttranslational processing and transport of secreted peptides and proteins that are glycosylated and targeted to the plasma membranes (14). The oligomeric Golgi complex includes 8 members (COG1-8), with the central COG1 surrounded by a COG2-4 forming lobe A and COG5-8 forming lobe B (15–19). PD-L1 associates with 7 of the 8 COGs (20), whereas the MHC molecules associate with the coatomers (20, 21), indicating distinct routes of processing. It is suggested that an essential role for the COGs is maintaining posttranslational homeostasis of either protein glycosylation or glycosylated proteins (22). However, it remains unclear whether PD-L1 expression can be targeted selectively because the central mechanisms controlling PD-L1 expression, for example, IRF1-dependent transcription and posttranslational processing of PD-L1 through the ER–Golgi network, remain insufficiently understood. Indeed, despite the essential role of IRF1 in PD-L1 mRNA expression, little is known regarding how this mechanism is regulated at the PD-L1 gene promoter and whether PD-L1 transactivation can be dissected from other IRF1 downstream genes. As such, the goal of this study was to identify potential therapeutic agents that could inhibit PD-L1 while maintaining antitumor immune responses.

All cell lines obtained from the ATCC were curated by and redistributed through The University of Texas MD Anderson Cancer Center (MD Anderson, Houston, TX) Characterized Cell Line Core from 2015 to 2019. CRF-SB, EFE184, ETN1, KLE, HOC7, UNP251, OVCAR3, OVCAR8, and MCF7 cells were maintained in RPMI1640 medium supplemented with 5% FBS and incubated at 37°C in a humidified incubator in an atmosphere of 95% air/5% CO₂. Cell line identities were reauthenticated by short tandem repeat (STR) DNA fingerprinting using the AmpF STR identifier kit according to the manufacturer's instructions (Applied Biosystems, catalog no. 4322288). The STR profiles were compared with known ATCC fingerprints; with the Cell Line Integrated Molecular Authentication database, version 0.1.200808 (23); and with the MD Anderson fingerprint database. The STR profiles matched known DNA fingerprints or were unique. The ID8 mouse ovarian surface epithelial cells, obtained from Vahid Afshar-Kharghan (MD Anderson, Houston, TX), were maintained in DMEM (high-glucose, Cellgro) supplemented with 4% FBS, 100 U/mL penicillin, 100 μg/mL streptomycin, 5 μg/mL insulin, 5 μg/mL transferrin, and 5 ng/mL sodium selenite. All lines were tested for Mycoplasma every 6 months. MDA-MB-231 demonstrates reliable cell surface expression of PD-L1 for flow cytometery (24).

PD-L1 (E1L3N-#13684), CIITA (#3793), IDH1 (#8137), mTOR (7C10-#2983), Ub (P4D1-#3936), p62, AKT (40D4-2920), LDHB, LC3 (D11-#3868), ATG5 (2630), GM130 (D6B1-#12480), HSP60, VPS34 (D9A5-4263), β-actin (8H10D10-#3700), ATG16L1 (D6D5-8089), BECN1 (#3738), IFIH1 (#5321), IRF1 (D5E4-#8478), IRF2 (#4943), IRF3 (#11904), STAT1 (D1K9Y-#14994), YAP1 (#15117), TRIM28 (#4123), and protein A-HRP (#12991) were obtained from Cell Signaling Technology; p62/SQSTM1 (610832) from BD Biosciences, HLA-Ds (sc-53302), and ERK2 from Santa Cruz Biotechnology; COG3 (11130-1-AP) from ProteinTech group; and COG7 (EPR9942-ab168362) from Abcam.

Cell lysis was performed by lysing 5 × 10⁶ cells with NP-40 lysis buffer containing 50 mmol/L Tris (pH 7.4), 150 mmol/L NaCl, 0.5% NP-40, protease inhibitors (1873580001, Millipore Sigma), and phosphatase inhibitors (4906837001, Millipore Sigma) on ice for 15 minutes followed by microcentrifugation at 14,000 × g for 15 minutes at 4°C to remove debris. Immunoblotting was performed using 50 μg of protein lysates resolved in SDS-PAGE, transferred to polyvinylidene difluoride membranes (88520, Thermo Fisher Scientific), probed with primary and secondary antibodies (A9917 and AP307P, Sigma), followed by signal detection using the ECL reagents (GERPN2019) from Sigma as instructed. Equal protein loading was verified by blotting ERK2 (tumor lines), LDHB, β-actin, mTOR, or pan AKT.

Immnoprecipitation was performed per the manufacturer's protocol (Cell Signaling Technology). Briefly, 1 mg of protein lysate was precleared with protein A magnetic beads (#73778) and was used for the COG3 and IRF1 immunoprecipitation at 1:50 antibody to lysate dilution (v/v). Following overnight incubation at 4°C, the immunocomplexes were incubated with magnetic protein A beads for 20 minutes at room temperature and then washed five times followed by Western blotting.

Chromatin immunoprecipitation (ChIP)-PCR assay was performed as described previously (25). Briefly, cells were crosslinked with 1% formaldehyde for 10 minutes and the reaction was stopped by adding 125 mmol/L glycine. Cells were then lysed and chromatin DNA was fragmented using a Bioruptor Sonicator (Diagenode). The samples were immunoprecipitated with 4 μg of the IRF1 (#8478, Cell Signaling Technology) antibody overnight at 4°C and a concentration-matched rabbit IgG isotype control (#3900, Cell Signaling Technology) was used for mock precipitation. The protein–DNA cross-links were then reversed and the DNA purified and analyzed by quantitative real-time PCR on the ABI ViiA 7 Real-Time PCR system using the Power SYBR Green PCR Master Mix (4368577, Thermo Fisher Scientific). The primers used for qPCR are as follows. Pair 1: -666 forward-TCCTTAGGGTGGCAGAATATCAG, -605 reverse-CCCATCCCGAGCTAC-ATCTTT; Pair 2: -103 forward-TGAAAGCTTCCGCCGATTT, -53 reverse-TGCCGGG-CGTTGGA; Pair 3: +120 forward-TGCCCACGGCCCAGTAT, +179 reverse-GTAGAGACCCTCCGTCCTAAAGTG; Pair4: +146 forward-GCTCGCTGGGCACTTTAGG; +205 reverse-TCCTCTCTCCATCCCAAAGAAA.

Tissues were first fixed in in 10% neutral buffered formalin for 24 hours, preserved in 70% ethanol, processed, and embedded in paraffin. Tissue blocks were then cut into 5-μm sections on positively charged slides, deparaffinized, and rehydrated. Mouse specific PD-L1 (#64988, 1:200) and CD8a (#98941, 1:200) antibodies from Cell Signaling Technology, and CD3e (sc-20047, 1:100) from Santa Cruz Biotechnology were used for IHC using SignalStain reagents (#8112, #14746, and #8114 or #8125) per Cell Signaling Technology protocol. Quantitative data were obtained with ImageJ analysis of positive cells per random area.

Cell surface PD-L1 was stained with PE-conjugated PD-L1 antibody (clone MIH1; catalog no. 12-5983-42) from Thermo Fisher Scientific for 30 minutes at 4°C. Control staining was performed using manufacturer recommended mouse IgG1 isotype and treatment-matched samples. Data were acquired using the FACSCelesta Cell Analyzer (BD Biosciences) and analyzed using the Flowjo software (Flowjo, LLC).

ON-TARGETplus siRNAs targeting human COG2 (L-019487-01-0005), COG3 (L-013499-02-0010), ATG5 (LQ-004374-00-0005), YAP/TAZ (LQ-012200-00-0005/L-009608-00-0010), IRF3 (L-009608-00-0010), IRF1 (LQ-011704-00-0005), and IRF2 (L-011705-02-0010) were from Dharmacon. RNAi silencing was performed using Lipofectamine RNAiMAX Transfection Reagent (Thermo Fisher Scientific) according to the manufacturer's protocol. Briefly, 20 nmol/L siRNA was transfected and control cells were transfected with 20 nmol/L nontargeting siRNA (Dharmacon).

Total RNA was isolated using the RNeasy Plus Mini Kit (Qiagen, #74136) according the manufacturer's instructions. Briely, 5 × 10⁶ cells were lysed and homogenized. RNA concentration was determined using a NanoDrop2000c Spectrophotometer (Thermo Fisher Scientific) and 2 μg RNA was used for first strand cDNA synthesis using the High Capacity cDNA Reverse Transcription kit (Invitrogen, #4368814). Real-time PCR was performed in a total volume of 20 μL per reaction in duplicates in twin.tec real-time PCR Plates (Eppendorf, #0030132718), covered with Masterclear Real-Time PCR Film (Eppendorf, #0030132947). PCR probes CD274 (Hs01125301_m1) and ACTB (Hs01060665_g1) from TaqMan Gene Expression Assays (Invitrogen) were used as instructed. Each reaction contained 40 ng of cDNA, 10 μL of TaqMan Fast Universal PCR Master Mix (2×), no AmpErase UNG (Invitrogen, #4366073). TaqMan reactions were run on Mastercycler ep realplex (Eppendorf) with the following thermal cycling protocol: 95°C for 2 minutes followed by 40 cycles of 95°C for 15 seconds, 55°C for 15 seconds, and 68°C for 20 seconds. CD274 gene expression was quantified using the comparative C_t (2^−ΔΔC_t) method with C_t values normalized to the housekeeping gene (ACTB).

Subcellular fractionation was performed using the FractionPREP cell fractionation kit (K270) from Biovision according to the manufacturer's instructions.

Cell viability was assayed using the CellTiter-Blue Cell Viability Assay kit (G8081) from Promega according to the manufacturer's protocol.

Reverse-phase protein microarray analysis (RPPA) assays were performed in the MD Anderson CCSG core as described at http://www.mdanderson.org/education-and-research/resources-for-professionals/scientific-resources/core-facilities-and-services/functional-proteomics-rppa-core/index.html. Briefly, serially diluted lysates were spotted onto FAST slides (Schleicher & Schuell BioSciences) using a robotic GeneTAC arrayer (Genomic Solutions, Inc.). After printing, the slides were blotted sequentially with Re-Blot (Chemicon), I-Block, and a biotin blocking system (Dako), probed with primary antibodies, and incubated with biotin-conjugated secondary antibodies. The signals were then amplified using a Catalyzed Signal Amplification kit (DakoCytomation) according to the manufacturer's instructions. The processed slides were scanned and quantitated using the Microvigene software (VigeneTech Inc.) and the quantitative values of five consistently expressed proteins (p38, JNK, ERK, mTor, GSK) as internal controls. To screen for the effects on PD-L1 protein levels, a total of 40 drugs or drug combinations targeting cancer pathways that are entering or likely to enter clinical trials were surveyed across established cancer cell lines and RPPA datasets (Supplementary Table S1).

All animal work and protocols were supervised and approved by the Institutional Animal Care and Use Committee of MD Anderson (Houston, TX). ID8 cells (1 × 10⁶ in 100 μL PBS) were intraperitoneally injected into C57BL/6 mice (female, 6–8 weeks old, CRL/NCI). After transplantation, cells were allowed to grow for 1 week, and then mice with established tumors were randomly sorted into different treatment groups with 5 mice/group. ID8-bearing mice were then treated with isotype control IgG or PD-L1 antibody (200 μg/mouse in 100 μL PBS, B7-H1, clone 10F.9G2, BioXCell; i.p. injection) every three days, verteporfin 60 mg/kg daily in 8% DMSO (100 μL; i.p.), BMN 673 0.33 mg/kg daily in 2.64% ethanol (100 μL; oral gavage), or the combination of BMN 673 with either PD-L1 antibody or verteporfin. Tumor progression was monitored once a week using a Xenogen IVIS Spectrum in vivo bioluminescence imaging system (PerkinElmer). Tumor volume was determined on the basis of the total flux (photons per second). Mice reaching a humane endpoint or weighing more than 35 g as a result of tumor growth and/or ascites were euthanized.

Lewis lung carcinoma (LLC) cells (5 × 10⁵ in 100 μL PBS) were inoculated subcutaneously into C57BL/6 mice or nude (NU/J, The Jackson Laboratory; 8–12 weeks) in both flanks. Five days after inoculation, mice were randomly divided into 4 groups (n = 6/group) and treated with vehicle, BMN 673 0.33 mg/kg daily (oral gavage), verteporfin 30 mg/kg daily (i.p. injection), or the combination of BMN 673 with verteporfin for 16 days. Tumors size were measured every 2 days by digital calipers to determine tumor volume using the formula [length/2] × [width²].

All statistical analyses were done in GraphPad Prism 8 software. Correlations between TRIM28 and CIITA-MHC II gene expression was analyzed using the linear regression test. On the basis of pilot studies of the anti–PD-1 experiment, 5 mice per group was sufficient to identify the expected effects with 90% power. Overall survival of various treatment groups was analyzed using the Cox regression model, and the log-rank test was used to determine the P values. Otherwise, unpaired t tests were used to generate two-tailed P values and P < 0.05 was considered statistically siginificant. The Cancer Genome Atlas (TCGA) data were analysed using cBioPortal (cbioportal.org) and PanCancer Atlas datasets for each cancer types were used.

Using downregulation of PD-L1 as a read-out by RPPAs to assess drugs in use or likely to enter trials with well-characterized molecular targets, we performed a candidate drug screen (n = 40) for inhibition of PD-L1 expression discarding low-potency, narrow-spectrum, and upregulating candidates, which identified a single drug candidate, verteporfin. Verteporfin suppressed PD-L1 expression effectively in all six cell lines (T-cell leukemia; B-cell leukemia; ovarian; endometrium n = 3; Fig. 1A; Supplementary Table S1; Supplementary Fig. S1). In an additional panel of eight human cancer cell lines (ovarian, n = 5; osteoblastoma, n = 1; and lung cancers, n = 2) and two murine cancer cell lines (ovarian and lung), verteporfin abolished basal PD-L1 protein expression, including differential glycosylated states as reflected by the double bands on Western blots (7), regardless of genetic background, lineage specificity, and basal (intrinsic) PD-L1 expression (Fig. 1A–D). Cell fractionation revealed that verteporfin decreased membrane-associated PD-L1 (functionally relevant PD-L1) in EFE184 cells (endometrial cancer; Fig. 1E) and flow cytometry showed that verteporfin reduced PD-L1 expression on both the surface of cancer cells (Fig. 1F) and on antigen-presenting cells (Supplementary Fig. S1D). Verteporfin suppressed both IFN-induced PD-L1 protein expression (Supplementary Fig. S1B–S1D) and mRNA expression (Fig. 1G). However, in contrast to the marked loss of PD-L1 protein, verteporfin had little effect on intrinsic PD-L1 mRNA expression in the absence of IFNγ (Fig. 1H). Thus, verteporfin engaged at least two independent mechanisms to downregulate PD-L1 expression.

As verteporfin inhibits autophagy (26), we evaluated whether autophagy was required for verteporfin-induced loss of PD-L1. Although verteporfin exhibited a modest dose-dependent growth-inhibitory effect on EFE184 cells, cotreatment with chloroquine, which inhibits the final step of autophagy, led to an approximately 16-fold increase in verteporfin-induced growth suppression and a marked loss of cell viability (Fig. 2A); these data were consistent with a crucial role of autophagy in maintenance of viability of cells treated with verteporfin. Transmission electron microscopy (TEM) revealed a marked increase in autophagosomes in the cells treated with verteporfin, and altered Golgi apparatus with swollen and disrupted structures (Fig. 2B). The morphologically disrupted Golgi networks were in proximity to autophagosomes. Consistent with the TEM data, verteporfin treatment led to progressive loss of high molecular weight ubiquitinated proteins in a dose-dependent manner, increased LC3 lipidation (LC3 1 and II), and loss of the selective autophagy substrate and adaptor p62/SQSTM1 (Fig. 2C and D), indicating active autophagy consistent with a prior report (27).

In addition to loss of p62 and PD-L1, verteporfin induced a marked decrease in the Golgi proteins COG3, COG7, and the Golgi matrix protein GM130 (Fig. 2D). Chloroquine treatment abrogated verteporfin-induced loss of PD-L1, p62/SQSTM1, and Golgi proteins (Fig. 2D). These data suggested that verteporfin induced autophagy-mediated degradation of the Golgi apparatus, which was likely a consequence of verteporfin-induced organelle damage. Indeed, we found COG3 physically associated with VPS34 (a crucial membrane component of the autophagy machinery), the autophagy adaptor molecule p62, and GM130 (Fig. 2E). Next, we interrogated whether the lobe A subunits might have a role in autophagy-dependent PD-L1 removal in cancer cells. RNAi-mediated gene silencing was performed for COG2 and COG3. The COG2 RNAi led to cross depletion of COG3, consistent with the finding that both COG3 and COG2 are required for stabilizing lobe A complexes (28). Nonetheless, RNAi of either COG2 or COG3 increased intrinsic PD-L1 expression, with COG2 depletion having a stronger effect (Fig. 2F), consistent with a role for the oligomeric Golgi complex in regulating PD-L1 expression as a quality control mechanism. However, depletion of COG3 but not COG2 attenuated verteporfin-induced loss of PD-L1 (Fig. 2F), suggesting a crucial role for COG3, but not COG2, in verteporfin-induced PD-L1 removal.

Supporting a role for autophagy in mediating verteporfin-induced loss of PD-L1, RNAi-mediated knockdown of ATG5, a gene that is essential for canonical autophagy (29), mitigated the effect of verteporfin on PD-L1 downregulation (Fig. 3A). Despite a role for autophagy in downregulation of PD-L1 protein expression, verteporfin was sufficient to decrease PD-L1 expression in ATG5-depleted cells at increased concentrations (1 μmol/L vs. 0.5 μmol/L). At higher drug concentrations, verteporfin abolished IFN-induced PD-L1 expression in the presence of chloroquine (Fig. 3B), suggesting that chloroquine-sensitive autophagy was dispensable for loss of PD-L1 in this setting. Because verteporfin treatment also resulted in significant downregulation of PD-L1 mRNA expression (Fig. 1G), we explored mechanisms controlling PD-L1 gene expression.

Verteporfin inhibits YAP/TAZ function and YAP1 inhibits IRF3 signaling (30). Thus, RNAi was performed to interrogate the roles of the YAP1, TAZ (WWTR1), IFIH1, and IRF3 cascade in association with PD-L1 expression. Knockdown of these genes had minimal effects on both the basal expression of PD-L1 or IFN-induced PD-L1 expression, with depletion of YAP1 and TAZ (WWTR1) leading to increased rather than decreased PD-L1 in the presence of IFNγ (Supplementary Fig. S2). Thus, the effect of verteporfin on PD-L1 expression is unlikely to be attributed to the YAP/TAZ signaling cascade.

Given the ability of verteporfin to abolish IFN-induced PD-L1 expression, we next sought to determine whether verteporfin acted to disrupt PD-L1 gene transcription. Both the signal transducer and activator of transcription 1 (STAT1) and the main transcription factor IRF1, downstream of the IFNγ receptor, are essential for IFN-induced PD-L1 gene expression (31). Indeed, IFN increased IRF1 expression in a panel of cancer cell lines, and IRF1 protein expression directly correlated with PD-L1 levels (Supplementary Fig. S3). IRF1 depletion abolished IFN-induced expression of PD-L1 (Fig. 3C), whereas IRF2 knockdown had no effect despite a possible role of IRF2 in either opposing or complimenting that of IRF1 (32). Thus, we examined the effect of verteporfin on PD-L1 mRNA in the context of IRF1 depletion. Verteporfin markedly decreased PD-L1 mRNA expression, and IRF1 depletion had a stronger effect, whereas verteporfin did not cause decreases in PD-L1 mRNA expression in the setting of IRF1 depletion (Fig. 3D), suggesting that verteporfin acted via inhibition of IRF1-dependent PD-L1 transcription. However, IFN-induced STAT1 and IRF1, including their nuclear localization, were not affected by verteporfin (Fig. 3E), suggesting intact IFN-STAT1 signaling upstream of IRF1.

To determine where in the STAT1 pathway verteporfin might be acting, we performed IRF1 immunoprecipitation and found that whereas IFN-γ induced a marked increase in IRF1–STAT1 association, verteporfin treatment led to complete disruption of the IRF1–STAT1 complex (Fig. 3F). We then performed ChIP assays to determine the effect of verteporfin on IFN-induced IRF1-binding to PD-L1 gene promoter regions (Fig. 3G). Indeed, IRF1-binding could not be detected in unstimulated cells, but IFNγ induced marked increases in the binding of IRF1 to the PD-L1 promoter (Fig. 3H). Strikingly, despite inhibiting PD-L1 expression, verteporfin treatment led to a modest increase in IRF1 binding in the absence of IFN, and a marked increase was detected in the presence of IFN (Fig. 3F). The paradoxical increase in IRF1 binding to the PD-L1 promoter after verteporfin treatment appeared to suggest futile recruitment of IRF1, probably as a result of disruption of the IRF1–STAT1 protein complex. Indeed, additional ChIP assays showed that verteporfin decreased the recruitment of STAT1 to the PD-L1 promoter in either the absence or presence of IFN (Supplementary Fig. S4). Taken together, our data suggested that verteporfin disrupts IFN-induced IRF1–STAT1 interaction, leading to suppression of PD-L1 transcription with nonproductive trapping of IRF1 to the PD-L1 promoter.

In addition to loss of PD-L1 expression, IRF1 knockdown also led to a marked decrease in the CIITA transcription factor and its downstream CD74 protein (Fig. 4A). The MHC II cascade branch of IRF1 signaling was not inhibited in verteporfin-treated cells in the presence of IFN (Fig. 4B), suggesting that verteporfin blocked the IFN–IRF1–PD-L1 axis specifically but had little effect on IRF1-CIITA. Such pathway specification might allow the PD-L1 immune checkpoint pathway to be blocked without collateral suppression of CIITA-dependent tumor immunogenicity. We next sought to understand the underlying mechanism behind this. TRIM28 is a cofactor acting either to activate or repress target gene transcription (33). Although previously reported in association with IRF1 (34), the role of TRIM28 in regulating IRF1 signaling remains unclear. We found in parallel to STAT1, that the TRIM28–IRF1 interaction was markedly increased in response to IFN and that the interaction was abrogated by verteporfin (Fig. 3F). Strikingly, shRNA-mediated TRIM28 depletion led to marked increases in CIITA and HLA-D (Fig. 4C), suggesting TRIM28 normally acted to inhibit the CIITA-MHC II cascade in a feed-forward manner that is responsive to IFN.

Indeed, analyses of TCGA data revealed that TRIM28 expression inversely correlated with CIITA-MHC II gene expression and CD74 in the majority of human cancer types that span immunologic reactivity (Supplementary Fig. S5A and S5B). Gene expression profiling revealed two distinct tumor clusters in lung cancer, conforming to high and low CIITA signature classes of gene expression, respectively (Supplementary Fig. S5C), which in turn exhibited a very strong (P = 9.27 × 10⁻¹⁶) inverse association with TRIM28 mRNA expression (Supplementary Fig. S5D–S5F). Thus, the role of TRIM28 in suppressing the IRF1–CIITA pathway appeared to be generalizable.

Importantly, higher TRIM28 expression was associated with poorer patient outcomes in melanoma and lung cancer (Fig. 4D and E), which are considered immunologically reactive, with a demonstrated benefit to immune checkpoint inhibition (35). To determine whether the statistical association between TRIM28 and patient outcomes might simply reflect a noncausal correlation, we analyzed the effect of a downstream event, that is, the expression of CIITA. We found that the mRNA levels of CIITA were indeed positively associated with more favorable patient outcomes, with high statistical power across a broad spectrum of immunologically reactive human cancers albeit not necessarily in immunologically nonreactive tumors such as GBM (Fig. 4F–K), suggesting that the effects of TRIM28 and CIITA on patient outcomes might be mechanistically linked in specific cancer lineages.

We next hypothesized that therapeutic interventions that activate the CIITA-MHC immunogenicity cascade, which usually also drive PD-L1 expression, would synergize with verteporfin. Because PARP inhibitors synergize with immune checkpoint blockade (36), we sought to determine whether PARP inhibition might affect the CIITA-MHC cascade. We found that PARP1 copy number increases in over 85% of lung cancers (adenocarcinomas), which correlated with PARP1 mRNA expression (Fig. 5A; Pearson r = 0.511). Higher PARP1/2 mRNA expression was associated with significantly poorer overall and disease-free survival in patients with lung cancer and with decreased mRNA expression of CIITA-MHC and PD-L1 genes (Fig. 5B–G).

PARP inhibitors have activity across a spectrum of cancers and are approved for first-line treatment of BRCA-mutant ovarian cancers (37). Notably, murine ID8 ovarian carcinoma cells exhibited little sensitivity to verteporfin in cell culture and no synergism with the BMN 673 PARP inhibitor (Fig. 6A). Given these properties, we tested the in vivo therapeutic effects of verteporfin monotherapy or in combination with BMN 673 in established immune competent mice bearing ID8 cells intraperitoneally in direct comparison with PD-L1 antibody (Fig. 6B). Monotherapy with either verteporfin or PD-L1 had a modest effect on the survival of ID8-burdened mice (Fig. 6B), consistent with previous studies showing these cells to have modest intrinsic immunogenicity (38, 39). In contrast, the combination of verteporfin and BMN 673 improved outcome compared with either monotherapy that was equivalent to the combination of PD-L1 and BMN 673 (Fig. 6B). Independent of BRCA mutations, PARP inhibition induces immune responses in cancer models including that of lung cancer, which has been exploited for therapeutic intervention (40, 41). Similar to ID8 cells, murine LLC cells exhibited limited sensitivity to verteporfin and no synergism with the BMN 673 PARP inhibitor in vitro (Fig. 6C). Monotherapy with either verteporfin or BMN 673 had a modest effect on tumor growth in a syngeneic setting, whereas the combination of verteporfin and BMN 673 produced a significant reduction in tumor volume (Fig. 6D). In LLC tumors, verteporfin treatment led to marked decreases in PD-L1 expression and increases in CD8⁺ T cells, especially in the vereporfin and BMN 673 treatment group (Fig. 6E; Supplementary Fig. S6A and S6B). In contrast, verteporfin and BMN 673, either administrated alone or in combination, were not effective in immunodeficient nude mice relative to control (P > 0.1 one-way ANOVA; Fig. 6F), emphasizing the role of immune responses in the drug action. The difference in tumor size on day 21 of the verteporfin-treated mice was not statistically significant.

While T-cell infiltration was consistent with the overall outcome of PD-L1 suppression, we observed that PD-L1 expression was highly heterogeneous in the LLC tumor microenvironment. The strongest PD-L1 signal was detected in the tumor vasculature in BMN 673–treated tumors (Supplementary Fig. S6A). Although PARP inhibition did not affect tumor cell expression of PD-L1 (Supplementary Table S1). We then sought to determine whether tumor-associated endothelial cells were the sources of PD-L1 expression in the BMN 673-treated tumors. Although PD-L1 expression appeared to be associated with LYVE1-positive lymphatic vasculature (Supplementary Fig. S6B), closer examination only revealed a relatively weak expression of PD-L1 in LYVE1-positive cells.

Because high endothelial venules (HEV), which are specialized post-capillary structures, are the only known routes of tumor infiltration of lymphocytes (42, 43), we tested the hypothesis that PARP inhibition drives PD-L1 expression in HEVs. PARP inhibition with BMN 673 on lymphatic endothelial cells led to increased protein expression of CIITA and PD-L1 (Fig. 7A). Whereas PARPi-induced PD-L1 expression was suppressed by verteporfin, CIITA expression was not significantly altered even at supraphysiological drug doses (i.e., ≥1 μmol/L; Fig. 7B). Ex vivo analysis of tumors stained for the selective HEV marker MECA-79 formed poorly developed vascular-like structures in control tumors but highly organized structures were detected in BMN-673–treated tumors with strong luminal MECA-79 staining (Fig. 7C). Strikingly, despite a weak coexpression of PD-L1 and MECA-79 in control tumors, strong PD-L1 expression was detected in MECA-79–positive cells in BMN 673–treated tumors (Fig. 7D), suggesting that PARP inhibition induced PD-L1 expression in HEVs. PD-L1 expression on the HEVs thus inhibited effector and proliferating T cells as they exited from the peripheral circulation into the tumor microenvironment (Fig. 7E).

High-throughput screening identified verteporfin as a lead candidate for inhibiting PD-L1, suggesting the potential repurposing of the FDA-approved ophthalmological drug verteporfin for oncologic therapeutic benefit. Verteporfin is a photosensitizer commonly prescribed for retinopathy but also has antitumor properties in various experimental settings (44, 45). Verteporfin possesses intrinsic antitumor activities through targeting YAP1 and other mechanisms (44, 45). Given the patient safety profile and relative low cost ($1700 per injection, wholesale), verteporfin could be considered as an alternative approach to PD-L1 blockade. PD-L1 blockade has therapeutic benefit in non–small cell lung carcinoma in a randomized phase III study (46). Notably, the patients with lung cancer with the highest expression of PD-L1 derived the greatest benefit from PD-L1. PARP inhibitors, such as BMN 673, upregulated PD-L1 on HEVs and the combination of PARP inhibitors and PD-L1 blockade increased the therapeutic effects in vivo. Our preclinical data may provide a strong rationale for the combinatorial use of verteporfin and a PARP inhibitor in patients with lung cancer, especially given the high frequency of PARP1 gain in lung cancer. In many other cancers such as GBM, PD-L1 is low (47) and the lack of immune effector responses in the tumor microenvironment may not necessarily benefit from this strategy.

Our study revealed two unique mechanisms of action of verteporfin on PD-L1 expression; an autophagy-dependent mechanism mediated by the oligomeric Golgi complex of the COG proteins (specifically COG2 and COG3) for the removal of intrinsically expressed PD-L1 and suppression of IFNγ-induced PD-L1 expression. This selective downregulation of PD-L1 with verteporfin maintained the proinflammatory IRF1–CIITA–MHC II signaling cascade. IRF1 is a key regulatory factor for IFNγ, TLR, and TNF (48–50); thus, IRF1 plays an important role in inflammatory responses. Indeed, IRF1 knockout in mice leads to a multifaceted defect in immune surveillance (51). Depending on the specific target gene, IRF1-mediated transactivation can occur with or without its cofactor STAT1 (52). In the case of PD-L1 expression, both IRF1 and STAT1 are essential (4). We showed that verteporfin disrupted the IRF1–STAT1 interaction, thereby suppressing IFNγ-induced PD-L1 transcription, but it increased IRF1 binding to the PD-L1 promoter in response to IFNγ. It remains unclear as to whether or not this promoter-trapping effect is required for suppression of PD-L1 transcription in addition to loss of STAT1 binding. Supporting this concept, IRF1 is involved in target gene silencing (48). Our study also unveiled a previously unrecognized role for TRIM28 in inhibiting the IRF1–CIITA–MHC II axis but not PD-L1 expression. Decommissioning of this role of TRIM28 upon verteporfin treatment provided another layer of selectivity, allowing precision targeting of the IRF1–PD-L1 axis. More studies are needed to determine whether TRIM28 contributes to the development of resistance to adaptive immune surveillance in human cancer and whether TRIM28 represents a therapeutic target for reversing the immunoediting process associated with suppression of the CIITA-MHC genes (53).

We also found that the PARP inhibitors induced PD-L1 expression on specialized tumor-associated endothelial structures known as the HEVs, the “gateway” for tumor-infiltrating lymphocytes (43). The presence of HEVs correlate with more favorable patient outcomes in multiple cancer types (54). As shown in other tumor-associated endothelial cells (55, 56), HEV expression of PD-L1 was probably modulating the effector and proliferative function of the T cells as they move from circulation into the tumor microenvironment. This mechanism would be in addition to the previously documented role of PARP inhibition, the accumulation of double-stranded DNA breaks in cancer cells with increases in mutation loads or the presence of damaged chromosomal DNA in the cytoplasm triggering the cGAS-innate immune response pathway (57). STING pathway activation in this context also likely drives the robust T-cell infiltration we observed in the tumors (40). We found a strong inverse correlation between the mRNA expression of PARP1/2 and the proinflammatory CIITA-MHC cascade genes in multiple human cancers. Thus, PARP inhibition was likely to activate the CIITA-MHC cascade, immunogenicity, and T-cell priming through multiple pathways, which consequently elicited immune responses and adaptive PD-L1 expression.

Despite the strong induction of both HEV neogenesis and PD-L1 in response to PARP inhibition, different mechanisms might be involved in these processes. In addition to DNA repair, PARP1 regulates gene expression through multiple mechanisms (58) and can act as a cofactor of multiple transcriptional regulators including NFκB and TNFα (59, 60). Indeed, the expression of a number of NFκB-dependent genes are impaired in PARP1^−/− cells as well as in response to PARP inhibition (61). Whereas the lymphotoxin–β-LTBR pathway is essential for HEV neogenesis, the signaling of the closely related cytokine TNFα has the opposite role in experimental melanoma (62). Future studies will be directed at ascertaining the applicability of modulating HEV PD-L1 expression across cancer lineages and its relevance as a biomarker for response to PD-L1 strategies.

Enrichment markers for response to PD-L1 strategies have not had high predictive value for therapeutic response, which is likely the scenario for responses to verteporfin as it may also have direct antitumor activities through targeting YAP1 (44, 45). At the dosage range that abrogated PD-L1 expression used in our study verteporfin had little effect on either YAP1 expression or its phosphorylation, and exhibited minimal cell-autonomous cytotoxicity on cancer cells. Verteporfin in combination with PARP inhibition produced a strong synergistic effect in syngeneic models of both ovarian and lung cancers but not in cell culture, suggesting a therapeutic efficacy independent of cell-autonomous anticancer activity. Nevertheless, verteporfin may prove uniquely effective in cancers that rely on oncogenic signals from YAP1. In addition, the striking coexpression pattern of MECA-79 and PD-L1 suggested that HEV staining could be considered as a potential biomarker for PD-L1 blockade, especially with a combination of PARP inhibitors and verteporfin within the context of a clinical trial.

H. Liang is a consultant/advisory board member for and has ownership interest (including patents) in Precision Scientific. G.B. Mills is a scientific advisory board member for AstraZeneca, Chrysalis, ImmunoMET, Ionis, Lilly, PDX Pharma, Signalchem Lifesciences, Symphogen, and Tarveda and has ownership interest (including patents) in ImmunoMET, Signalchem Lifesciences, Tarveda, Catena Pharmaceuticals, and Spindletop Ventures. No potential conflicts of interest were disclosed by the other authors.

The funding agencies had no role in the data analysis, interpretation of the results, or writing of this article.

Conception and design: J. Liang, L. Wang, C. Wang, Y. Lu, J. Chen, G.B. Mills, A.B. Heimberger

Development of methodology: J. Liang, L. Wang, C. Kassab, K.J. Jeong, Y. Lu

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Liang, L. Wang, J. Shen, B. Su, D. Fang, C. Kassab, Y. Lu, Z. Zhou, C. Lu, Z.-X. Xu, Q. Yu, S. Shao, X. Chen, M. Gao

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J. Liang, L. Wang, J. Shen, C. Kassab, K.J. Jeong, W. Zhao, Y. Lu, A.K. Jain, H. Liang, S.-C. Sun, X. Chen, F.X. Claret, Z. Ding, P. Chen

Writing, review, and/or revision of the manuscript: J. Liang, L. Wang, D. Fang, C. Kassab, K.J. Jeong, S.-C. Sun, M.C. Barton, G. Peng, G.B. Mills, A.B. Heimberger

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): L. Wang, Y. Lu, M. Gao, J. Chen, P. Chen, G.B. Mills, A.B. Heimberger

Study supervision: J. Chen, M.C. Barton, G. Peng, G.B. Mills, A.B. Heimberger

Other (dosing the animals for the study): A.L. Marisetty

Other (provided RPPA analysis on this study): Y. Lu, Q. Yu

Other (provided the plasmid DNA materials and expertise in data interpretation): A.K. Jain

The authors thank Kenneth Dunner Jr at the High Resolution Electron Microscopy Facility for assistance in performing electron microscopy; David M. Wildrick, PhD, for editorial assistance; and Audria Patrick for assisting in manuscript preparation. This project was supported by the Gynecologic SPORE (5P50CA098258, NIH/NCI; to G.B. Mills), the Provost Retention Fund and the Brockman Foundation (to A.B. Heimberger), and the Shanghai Pujiang Program (17PJ1401400; to C. Wang). This research was performed in the Flow Cytometry & Cellular Imaging Facility, which is supported in part by the NIH through MD Anderson's Cancer Center Support Grant CA016672.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.