Induction of Vasculogenic Mimicry Overrides VEGF-A Silencing and Enriches Stem-like Cancer Cells in Melanoma

Human malignant melanoma is a highly aggressive cancer due to its ability to metastasize and its resistance to conventional anticancer treatments. Angiogenesis supports melanoma tumor growth and metastasis by supplying the tumor with oxygen and nutrients and providing tumor cells with an entry route into circulation (1). Given that the key proangiogenic factor is considered to be VEGF-A, anti-VEGF therapies for the treatment of melanoma have been studied extensively (2, 3). Unfortunately, despite showing initial antitumor effects in a subset of patients, anti-VEGF strategies achieved limited efficacy in patients due to the rapid development of resistance (3–6). This adaptive resistance is associated with increased invasion and metastasis of cancer cells (6).

It has long been speculated that to escape anti-VEGF therapies, tumors may use alternative vascularization mechanisms, including vasculogenic mimicry (VM; refs. 7–9). VM is a phenomenon in which tumor cells mimic endothelial cells by undergoing transendothelial differentiation, characterized by their increased expression of vascular markers, such as CD144 (VE-cadherin), EphA2, Tie-1, and MMP-2 (10–13). VM channels provide an alternate mechanism for nutrient supply and act as a potential access point for metastases, and thus correlate with tumor aggressiveness and patient mortality (10). Gene-expression analysis has revealed that VM-engaging melanoma subsets express genes associated with undifferentiated embryonic-like cells, suggesting the participation of melanoma stem-like cells (MSLC) with the phenotypic plasticity to serve an endothelial function (8–18). In agreement with these findings, we previously demonstrated that CD133⁺ MSLCs colocalize to CD144⁺ VM-engaging melanoma subsets in a perivascular MSLC niche (19). Thus, in response to VEGF inhibition, MSCLs with the intrinsic ability to differentiate and form VM channels may be selected. Although experimental evidence to validate this proposal is limited, treatment of human breast cancer xenografts with bevacizumab, an anti-VEGF antibody, resulted in an increase in the cancer stem-like cell population (20). Furthermore, Wang and colleagues (21) showed that blocking VEGF-A or its receptor, VEGFR-2, inhibited the maturation of tumor endothelial progenitors into endothelium, but did not prevent the transendothelial differentiation of CD133⁺ glioblastoma cells.

Despite speculation that VM induction and MSLC maintenance might affect the effectiveness of anti-VEGF therapies, little research supports this claim, and the mechanisms of adaptive resistance following VEGF-A inhibition remain largely unknown. In this study, we examined the adaptive response of WM1617, C8161, and A2058 melanoma cell lines to VEGF-A inhibition in the context of VM niche morphogenesis and MSLC subsets. We hypothesized that tumors overcome VEGF-A inhibition by promoting niche morphogenesis through VM induction and MSLC enrichment. Better understanding of the adaptive response to VEGF-A inhibition will provide valuable insights for designing new approaches to treat melanoma through effective combination therapies.

Melanoma cell lines, WM1617, WM983C, 1205Lu (obtained from Dr. M. Herlyn at the Wistar Institute, Philadelphia, PA), and C8161 (a kind gift from Dr. M.J.C. Hendrix at the Stanley Manne Children's Research Institute, Northwestern University, Chicago, IL), and A2058 cells (obtained from ATCC) were cultured as previously described (22). All cell lines were tested within 6 months using the PowerPlex 18D system (Promega BioSciences).

VEGF-A silencing in WM1617, C8161, and A2058 melanoma cells was achieved by SureSilencing shRNA pGeneClip plasmids (SABiosciences). Lipofectamine 2000 (Invitrogen) was used according to the manufacturer's protocol. Stable transfectants were selected with 600 μg/mL G418 (Gibco Laboratories) and subsequently subcloned to select the most efficient knockdown cells. VEGF-A knockdown (VEGF-A KD; sh1) and VEGF-A KD (sh2) denote two independent shRNA sequences targeting VEGF-A.

Cell lysates were extracted in RIPA buffer (Pierce) and quantified by a BCA Protein Assay Kit (Pierce) according to the manufacturer's protocol. Equal amounts (100 μg) of protein were subjected to the Human VEGF-A ELISA Kit (R&D Systems) according to the manufacturer's instructions. In separate experiments for baseline VEFG-A measurement, tissue culture supernatant was collected from melanoma cells in serum-free medium (5 mL/2 × 10⁶ cells over a period of 14 hours) and subjected to the Human VEGF-A ELISA Kit.

Growth assays were conducted in 6-well plates. Cells were plated at the density of 10,000 cells per well. Cells were counted after 4, 7, and 10 days. Each cell line was plated in triplicate and the experiment was repeated at least three times with consistency. Data presented represent results from one experiment, and data were analyzed using the Student t test.

Three-dimensional soft agar clonogenic assays were conducted in 6-well plates as previously described (23). Cells were plated at the density of 10,000 cells per well. Colonies with more than five cells were counted in 20 randomly chosen fields (×100) after 10 days using an inverted microscope and the percentage of colony formation was calculated. Each cell line was plated in triplicate and the experiment was repeated three times with consistency. The represented data are compiled from the independent repeats, and data were analyzed using the Student t test.

Cells were plated at the density of 10,000 cells per well in 96-well plates coated with Matrigel (BD Biosciences). Luminal area per field (×40) was observed after 24 hours using an inverted microscope and analyzed using ImageJ software (National Institutes of Health, Bethesda, MD). Each cell line was plated in triplicate, and the data were analyzed using the Student t test.

RNA from melanoma cells and frozen tumor xenografts was extracted using an RNAeasy kit (Qiagen) and reverse transcribed using the SuperScript III RT cDNA Synthesis Kit (Invitrogen) according to the manufacturer's protocol. Real-time quantitative PCR (qRT-PCR) was performed subsequently on a StepOnePlus Real-Time PCR System (Applied Biosystems) using human-specific primers. The primer sequences were as follows: CD133 forward 5′-TTCTTGACCGACTGAGAC-3′ and reverse 5′-CCAAGCACAGAGGGTCAT-3′; CD144 forward 5′-ATATGTCAGTGATGACTA-3′ and reverse 5′-CTTACCAGGGCGTTCAGG-3′; CD271 forward 5′-ACTCACTGCACAGACTCT-3′ and reverse 5′-GAAGCTTCTCAACGGCTC-3′; MMP-2 forward 5′-TTTCCATTCCGCTTCCAGGGCAC-3′ and reverse 5′-TCGCACACCACATCTTTCCGTCACT-3′ (24); Tie-1 forward 5′-CACGACCATGACGGCGAAT-3′ and reverse 5′-CGGCAGCCTGATATGCCTG-3′ (25); and GAPDH forward 5′-CGACCACTTTGTCAAGCTCA-3′ and reverse 5′-AGGGGAGATTCAGTGTGGTG-3′. All samples were run in triplicate and normalized to the housekeeping gene, GAPDH. Data were analyzed using the 2^−ΔΔC_t method (26).

Cell lysates and xenograft tissue homogenates were extracted in RIPA buffer (Pierce) and quantified by a BCA Protein Assay Kit (Pierce) according to the manufacturer's protocol. Equal amounts (40–100 μg) of protein were subjected to electrophoresis and transferred to nitrocellulose. Membranes were probed overnight at 4°C with mouse anti-CD133 (Miltenyi Biotech Inc.; clone W6B3C1) at 1:200, rabbit anti-CD271 (Alomone) at 1:1,000, rabbit anti-CD144 (Cell Signaling Technology) at 1:1,000, or mouse anti–β-actin (Abcam) at 1:5,000 followed by probing with the appropriate secondary antibody conjugated to horseradish peroxidase (Jackson Immunoresearch). Immunoreactive bands were visualized by SuperSignal West Pico Chemiluminescent Substrate (Pierce). Densitometry measurements were performed using ImageJ software (National Institutes of Health, Bethesda, MD); β-actin was used as a loading control.

Mice are maintained under pathogen-free conditions in an American Association for Accreditation of Laboratory Animal Care-accredited facility at the Boston University Medical Center, under the supervision of the Laboratory Animal Science Center and its staff of veterinarians and support personnel. To determine the effects of VEGF-A downregulation on tumorigenicity, 2 × 10⁵ WM1617 or A2058 melanoma cells transfected with control or VEGF-A shRNA constructs were injected s.c. in the dorsal skin of each severe combined immune-deficiency (SCID) mouse (CB17; Taconic Laboratory; 5 mice/condition). In a separate experiment, to ensure the generation of sizable VEGF-A KD xenografts for various analyses, 2 × 10⁶ WM1617, C8161, or A2058 control or VEGF-A KD melanoma cells were injected per mouse (5 mice/group). Tumor volume was monitored and determined as the volume of ellipsoid: 4/3π(width/2×length/2×height/2). Statistical analyses were performed using ANOVA. Melanoma xenografts were harvested when tumors reached 1 cm³. Xenografts were subjected to various analyses, including immunofluorescence, qRT-PCR, and Western Blot analyses.

Consecutive frozen melanoma xenograft sections were subjected to double indirect immunofluorescence according to standard protocols. The primary antibodies used were rabbit anti-human CD144 (Cell Signaling Technology) at 1:100, biotinylated anti-human CD133 (Miltenyl Biotec Inc.) at 1:20, rabbit anti-human CD271 (Alomone) at 1:100, and rat anti-mouse CD31 (BD Biosciences) at 1:50. The secondary antibodies used were FITC-conjugated donkey anti-rabbit IgG (Accurate Chemical & Scientific Corporation) at 1:100, FITC-conjugated mouse anti-biotin IgG (Jackson ImmunoResearch) at 1:100, and TRITC-goat anti-rat IgG (Jackson ImmunoResearch) at 1:100. For control, isotype-matched immunoglobins (Vector Laboratories) were used in place of the primary antibody. Sections were mounted with Vectashield containing DAPI (Vector Laboratories) and examined under a Nikon Eclipse E400 microscope equipped with FITC and TRITC filters (Nikon) and a Mercury-100W lamp (Chiu Technical Corporation). Host angiogenesis (mCD31⁺) and human melanoma–derived VM channels (hCD144⁺) were quantified as measured by luminal area per field (×100) using ImageJ software, and data were analyzed using the Student t test.

Dewaxed formalin-fixed and paraffin-embedded 5-μm melanoma xenograft tissue sections were subjected to the Periodic Acid-Schiff (PAS) Kit (Sigma) according to the manufacturer's instructions. The tubular length of VM-like channels per field (×200) was quantified using ImageJ software, and data were analyzed using the Student t test.

RNA from melanoma cells and frozen tumors was extracted using an RNAeasy kit (Qiagen). The RT² Profile pathway-focused PCR array specific for human angiogenesis were purchased from SABiosciences. The complete gene lists for the PCR arrays are available on the SABioscience's website (http://www.sabiosciences.com/). Web-based software provided by the supplier was used to analyze the PCR data.

Differences between two groups were analyzed using the Student t test. Differences between more than two groups were analyzed by one-way ANOVA. Two-sided P values of less than 0.05 were considered significant.

Using ELISA, we screened five melanoma cell lines, including WM1617, WM983C, 1205Lu, A2058, and C8161 cells, for VEGF-A expression. On the basis of this screening, the amounts of secreted VEGF-A in the tissue culture supernatant (Fig. 1A left) correlate with those of the cytoplasmic fraction in cell lysates (Fig. 1A, right), and WM1617, C8161 and A2058 cells exhibited the highest level of VEGF-A expression (Fig. 1A). Thus, we chose these cell lines to explore the functional role of VEGF-A with regard to tumor growth and niche maintenance. To test the biologic significance of VEGF-A, we generated stable VEGF-A KD in WM1617, C8161, and A2058 cells using two independent plasmid-based shRNAs (sh1 and sh2). ELISA revealed VEGF-A KD at the protein level in WM1617, C8161, and A2058 cells in vitro compared with control cells expressing nontarget shRNA (Fig. 1B and Supplementary Fig. S1A). Quantitative RT-PCR verified downregulation of VEGF-A message levels in vitro (Fig. 1C).

To explore the effect of VEGF-A KD on cell proliferation in vitro, we performed two-dimensional (2D) growth assays. VEGF-A KD led to a decrease in WM1617 and C8161 cell growth (Fig. 2A and Supplementary Fig. S1B). In contrast, VEGF-A KD in A2058 cells did not modulate cell growth (Fig. 2A and Supplementary Fig. S1B). These results highlight the differential effect of VEGF-A KD on cell growth in different melanoma cell lines. In addition to 2D growth assays, we examined anchorage-independent growth using soft agar colony formation assays. VEGF-A KD did not alter soft agar colony formation in WM1617, C8161, or A2058 cells compared with nontarget control cells (Fig. 2B and Supplementary Fig. S1C). Furthermore, VEGF-A KD did not modulate VM tubule formation in vitro, as measured by VM luminal area, in WM1617 (P = 0.490), C8161 (P = 0.587), or A2058 (P = 0.118) cells (Fig. 2C).

Given that WM1617 VEGF-A KD (sh1), C8161 VEGF-A KD (sh1), and A2058 VEGF-A KD (sh1) cells exhibited the greatest VEGF-A KD efficiency, we chose to use these cells for our in vivo studies. In two separate tumorigenicity assays in which mice were injected with 2 × 10⁶ (Fig. 3A) 2 × 10⁵ (Supplementary Fig. S2A) melanoma cells, VEGF-A KD significantly (P < 0.05) retarded WM1617 tumor growth compared with the nontarget control in vivo. Although growth was delayed, VEGF-A KD xenografts did eventually develop tumors (data not shown). These findings suggest that in the experimental setting WM1617 cells mirror the initial responders in the clinic; that is, some melanomas are initially sensitive to VEGF-A silencing, but subpopulations of cells eventually develop adaptive resistance and survive (4–6). In contrast, VEGF-A KD did not affect C8161 or A2058 tumor growth in vivo, suggesting that C8161 and A2058 cells are intrinsically resistant to VEGF-A silencing and reflect the disease in clinical nonresponders to anti-VEGF therapy (Fig. 3A and Supplementary Fig. S2B). The efficiency of VEGF-A KD was verified in WM1617, C8161, and A2058 xenografts using ELISA (Fig. 3B).

To explore the biologic significance of VEGF-A KD on niche morphogenesis and MSLC homeostasis, we analyzed expression of the VM marker, CD144, and the stem cell markers, CD133 and CD271, in xenografts by real-time qRT-PCR using human-specific primers and by Western blot analysis. We observed an increase in both the message (Fig. 4A) and protein (Fig. 4B) levels of CD144, CD133, and CD271 in VEGF-A KD WM1617 xenografts compared with their nontarget control counterparts, suggesting that VEGF-A KD promotes VM and enriches the MSLC phenotype. Interestingly, although we did not observe an increase in CD144 message levels in VEGF-A KD C8161 xenografts compared with their nontarget control counterparts, we did observe an increase in CD133 and CD271 message levels (Fig. 4A); however, VEGF-A KD C8161 xenografts did not exhibit changes in CD144 or CD271 protein levels (Fig. 4B). VEGF-A KD A2058 xenografts did not exhibit changes in CD144, CD133 or CD271 message levels or CD144 or CD271 protein levels (Fig. 4A and 4B). Attempts to examine CD133 in VEGF-A KD C8161 or A2058 xenografts using Western blot analysis were unsuccessful due to sensitivity issues (Fig. 4B).

Multilabel immunofluorescence staining confirmed that VEGF-A KD WM1617 xenografts exhibit increased CD144⁺ VM-like melanoma channel formation as measured by CD144⁺ luminal area (P = 0.0003; Fig. 5A), whereas no significant change in mouse CD31⁺ blood vessel formation was observed (P = 0.074; Fig. 5A). Thus, WM1617 VEGF-A KD xenografts exhibit a significant increase in the percentage of human tumor–derived VM channels (hCD144⁺) versus host angiogenesis (mCD31⁺; P = 0.0003; Fig 5A). In contrast, VEGF-A KD C8161 xenografts did not display a change in CD144⁺ VM-like melanoma channel formation (P = 0.269), mouse CD31⁺ blood vessels (P = 0.134) or the percentage of human tumor–derived VM vessels versus host angiogenesis (P = 0.417; Supplementary Fig. S3A, left). VEGF-A KD A2058 xenografts also did not display a change in CD144⁺ VM-like melanoma channel formation (P = 0.946), mouse CD31⁺ blood vessels (P = 0.059) or the percentage of human tumor–derived VM vessels versus host angiogenesis (P = 0.287; Supplementary Fig. S3A, right). PAS staining further confirmed that VEGF-A KD increases VM channels as measured by tubular length in WM1617 xenografts (P = 0.0008), but not in C8161 xenografts (P = 0.384) or A2058 xenografts (P = 0.624; Supplementary Fig. S4). Consistent with these findings, VM-associated genes, such as Tie-1 and MMP-2 were significantly upregulated in WM1617 but not C8161 or A2058 VEGF-A KD xenografts compared with the control xenografts in an angiogenesis pathway-specific qRT-PCR array (data not shown). These results were further validated by qRT-PCR (Fig. 5B and Supplementary Fig. S3B).

Immunofluorescence staining also confirmed the increase in CD133⁺ and CD271⁺ MLSCs in the perivascular niche in close association to CD144⁺ melanoma cell–lined VM channels in VEGF-A KD WM1617 xenografts (Fig. 5A). No increase in CD133⁺ or CD271⁺ MLSCs was observed in VEGF-A KD C8161 or A2058 xenografts (Supplementary Fig. S3A). Collectively, these data suggest that WM1617 cells, which are initially sensitive to VEGF-A silencing, eventually acquire adaptive resistance and survive through increased VM formation and MSLC enrichment. In contrast, the C8161 and A2058 cells are intrinsically resistant to VEGF-A silencing and, therefore, fail to show adaptive compensatory mechanisms.

Of note, given that numerous studies have demonstrated that HIF1α regulates VM and the cancer stem cell–like phenotype (27), we observed an increase in HIF1α in WM1617 VEGF-A KD xenografts but not in A2058 VEGF-A KD xenografts in an angiogenesis pathway-specific qRT-PCR array (data not shown). We verified these results using qRT-PCR (Fig. 6A and Supplementary Fig. S3B). Western blot analysis confirmed an increase in HIF1α protein levels in WM1617 VEGF-A KD xenografts (Fig. 6B). Attempts to examine HIF1α expression in C8161 and A2058 VEGF-A KD xenografts were unsuccessful due to sensitivity issues.

As melanomas are heterogeneous in the clinical setting, some melanomas initially respond to anti-VEGF therapy, but eventually develop clinical resistance, the so-called “initial responders,” whereas other melanomas are intrinsically resistant and do not respond to anti-VEGF therapy (4–6). In accord with frequent clinical relapses in “initial responders” following anti-VEGF treatments, we show that VEGF-A silencing initially inhibits WM1617 melanoma xenograft growth in vivo; however, tumors eventually develop adaptive resistance and overcome VEGF-A inhibition. Conversely, mirroring the clinical disease in nonresponders to anti–VEGF-A therapy, we found that C8161 and A2058 cells are intrinsically resistant to VEGF-A silencing. Compensatory survival pathways in the initial responders to anti-VEGF therapies have had a negative impact on their clinical effectiveness; however, the mechanisms responsible for the adaptive resistance to anti–VEGF-A therapies are largely unknown. Our studies reveal that VEGF-A inhibition in WM1617 xenografts leads to an increase in CD144⁺ melanoma cell–lined VM channels and associated VM markers, Tie-1 and MMP-2, providing evidence that VM induction results from an adaptive response to VEGF-A inhibition. C8161 and A2058 xenografts, which were intrinsically resistant to VEGF-A silencing, failed to show adaptive compensatory mechanisms and did not exhibit an increase in VM formation. Taken together, our data suggest that some melanomas develop adaptive resistance to VEGF-A inhibition by increasing VM, an alternate vascularization mechanism to angiogenesis, which is essential in supplying tumors with oxygen and nutrients to support tumor growth (10).

The role of VEGF signaling in regulating VM has been widely studied, however, it remains controversial. Initial reports demonstrated that VEGF-A stimulation did not affect VM (28). Consistent with these reports, VEGF neutralizing antibody bevacizumab failed to modulate VM formation in glioblastoma cells in vitro, and AG28262, a selective inhibitor of VEGFR-1, R-2, and R-3, did not modulate glioblastoma-derived VM channels in vivo (29, 30). Conversely, VEGF-A downregulation abolished VM in osteosarcoma cells in vitro and VEGFR-2 downregulation reduced VM formation in glioblastoma cells in vitro and in vivo (30–33). In melanoma, VEGFR-1 expression on MSLCs has been shown to promote tumor growth via increased VM formation (34). As stated above, our data demonstrate that VM is increased in some melanomas following VEGF-A inhibition. Such discrepancies of the role of VEGF signaling in modulating VM formation may reflect cell type- and/or cell line-specific differences, and emphasize the importance of the cellular context in determining the response to VEGF-A inhibition.

VM and stem cell signaling pathways are intertwined in the tumor microenvironment (35). Gene-expression analysis has revealed that VM-engaging melanoma subsets express genes associated with undifferentiated embryonic-like cells, suggesting the participation of MSLCs with the phenotypic plasticity to serve an endothelial function (10–15, 36). Furthermore, we previously demonstrated that CD133⁺ MSLCs colocalize to CD144⁺ VM-engaging melanoma subsets in a perivascular MSLC niche, and Valyi-Nagy and colleagues (37) showed that VM-forming uveal melanoma cells preferentially express the stem cell marker CD271⁺ (19). Parallel to these reports, we found that VM induction following VEGF-A inhibition leads to a concurrent increase in CD133⁺ and CD271⁺ MSLCs in the perivascular niche of WM1617 xenografts, but not in C8161 or A2058 xenografts, suggesting that some melanomas adapt to VEGF-A inhibition through enrichment of MSLCs, which exhibit the intrinsic ability to differentiate and form VM channels. Highlighting the importance of the microenvironment in supporting VM induction and MSLC maintenance, we did not observe changes in (i) CD144, CD133 or CD271 expression, (ii) VM induction, or (iii) growth in soft agar following VEGF-A KD in WM1617 cells in vitro. These observations support the view that the dynamic niche microenvironment is imperative in providing the essential cues for cell fate determination, including the phenotypic and functional rendering of VM channels and MSLC enrichment (19, 35, 38–40).

As a result of pruning the tumor vasculature, anti-VEGF therapies lead to an increase in hypoxia in the tumor microenvironment (41, 42). Emerging evidence suggests that therapy-induced hypoxia may be a contributing factor to the disappointing clinical results of anti-VEGF strategies (41, 42). Hypoxia, either in the transient or persistent state, is a hallmark of cancer. HIF1α, the main mediator of hypoxia has been shown to regulate pathways involved in the maintenance of the cancer stem cell–like phenotype, angiogenesis and VM (27, 43, 44). Of note, melanomas with high VM capability also display increased HIF1α expression levels (45). Here, we report that VEGF-A inhibition in WM1617 xenografts leads to a concomitant increase in VM and HIF1α, suggesting that hypoxia-induced HIF1α expression may be one important in vivo environmental cue promoting VM induction and MSLC maintenance (Fig. 6C; ref. 27). Parallel to our findings, tumors from glioblastoma multiforme patients treated with bevacizumab, an anti–VEGF-A antibody, exhibit an increase in HIF1α expression (46). Furthermore, daily administration of topotecan, an HIF1α inhibitor, in combination with bevacizumab synergistically attenuates tumor growth and angiogenesis in a glioma xenograft model (47). Future work should further examine whether HIF1α downregulation prevents the increase in VM induction and MSLC maintenance following VEGF-A inhibition in melanoma.

In summary, our work reveals that, although VEGF-A silencing attenuates WM1617 cell growth in vivo, tumors compensate for VEGF-A KD by adopting VM, thereby promoting perivascular MSLC niche morphogenesis and MSLC enrichment potentially via an HIF1α-mediated mechanism. In contrast, C8161 and A2058 cells are intrinsically resistant to VEGF-A silencing and, therefore, fail to show adaptive compensatory survival mechanisms. Our findings are critical in understanding the mechanisms of adaptive resistance in some melanomas following treatment with anti-VEGF agents. Future work should examine the specific signaling pathways involved in the adaptive response to VEGF-A inhibition. Uncovering the specific effectors of VM-induction and MSLC maintenance will be essential in discovering novel therapeutic targets to treat melanoma through combination therapies that simultaneously target angiogenesis and the MSLC niche.

No potential conflicts of interest were disclosed.

Conception and design: C.I. Schnegg, S.K. Ghosh, M.-Y. Hsu

Development of methodology: S.K. Ghosh, M.-Y. Hsu

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C.I. Schnegg, M.-Y. Hsu

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C.I. Schnegg, M.-Y. Hsu

Writing, review, and/or revision of the manuscript: C.I. Schnegg, M.-Y. Hsu

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C.I. Schnegg, M.H. Yang, M.-Y. Hsu

Study supervision: M.-Y. Hsu

This work was supported by the NIH, National Cancer Institute (grant no. R01CA138649 to M.-Y. Hsu).

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.