Enhanced Angiogenesis in Salivary Duct Carcinoma Ex-Pleomorphic Adenoma

Salivary duct carcinoma (SDC) is morphologically similar to breast cancer, with HER2-overexpression reported. With regard to the pattern of disease onset, SDC can arise from de novo or carcinoma ex-pleomorphic adenoma (Ca-ex-PA). Recently, multiple molecular profiles of SDC as well as breast cancer have been reported, with significant differences in HER2 expression between Ca-ex-PA and de novo. We assessed the differences in gene expression between onset classifications. We conducted immunohistochemical analysis and HER2-DISH for 23 patients and classified SDCs into three subtypes as follows: “HER2-positive” (HER2+/any AR), “Luminal-AR” (HER2-/AR+), and “Basal-like” (HER2-/AR-). We assessed the expression levels of 84 functional genes for 19 patients by using a qRT-PCR array. Ten cases were classified as HER2-positive, seven cases as Luminal-AR, and six cases as Basal-like. The gene expression pattern was generally consistent with the corresponding immunostaining classification. The expression levels of VEGFA, ERBB2(HER2), IGF1R, RB1, and XBP1 were higher, while those of SLIT2 and PTEN were lower in Ca-ex-PA than in de novo. The functions of those genes were concentrated in angiogenesis and AKT/PI3K signaling pathway (Fisher’s test: p-value = 0.025 and 0.004, respectively). Multiple machine learning methods, OPLS-DA, LASSO, and RandomForest, also show that VEGFA can be a candidate for the characteristic differences between Ca-ex-PA and de novo. In conclusion, the AKT/PI3K signaling pathway leading to angiogenesis was hyper-activated in all SDCs, particularly in those classified into the Ca-ex-PAs. VEGFA was over-expressed significantly in the Ca-ex-PA, which can be a crucial factor in the malignant conversion to SDC.


INTRODUCTION
Salivary duct carcinoma (SDC) is a highly aggressive type of carcinoma, similar to breast cancer morphologically (1,2). Recently, the histological resemblance of SDC to breast cancers has led to the study of HER2 (human epidermal growth factor receptor 2, also known as ERBB2) expression (3). HER2 overexpression or amplification is seen in 15-20% of patients with invasive breast cancers and is considered to be an adverse prognostic factor (4). Strong immunohistochemical staining for HER2 protein has also been identified in 25-92% of SDCs (5). These findings highlight the similarity between SDC and breast cancer with regard not only to the overall morphology but also the immunophenotype and gene expression profile. Moreover, a few investigators have reported that the treatment regimens, including the use of HER2 antagonists, result in variable clinical benefits to patients with HER2-positive SDCs, although these protocols are not necessarily effective for every patient with SDC (6)(7)(8)(9)(10)(11). These results indicate that the concept of "SDC", as well as that of breast cancer, should not be regarded as a single disease but as a collection of heterogeneous entities with various characteristics. Based on the molecular biological profile, breast cancers can be stratified into multiple subtypes and individualized treatments are indicated for each subtype. In practice, anti-HER2 antibodies have already been applied for HER2-positive breast cancers, with clinical benefits observed in metastatic and adjuvant settings (12,13). In SDCs as well, individualized treatment selection is expected to be based on stratification (14,15).
Although a few clinical trials from Japan, such as combined androgen blockade for AR (Androgen receptor)-positive salivary gland cancer and trastuzumab, an anti-HER2 antibody, plus docetaxel therapy for HER2-positive SDCs, have brought favorable results (11,16), there is no stratification currently available for personalized treatment selection and no consensus regarding a standard treatment or protocol. For the development of therapeutic methods with clinical applications, further understanding of SDC tumorigenesis is necessary.
Moreover, it is well-known that SDC, as well as other salivary gland cancers, can occur de novo or as a malignant component of carcinoma ex-pleomorphic adenoma (Ca-ex-PA) (17). For this reason, we focus on the relationship between the biological profile and onset pattern (de novo or Ca-ex-PA) of SDC based on the hypothesis that SDCs have different biological profiles depending on whether they arise de novo or Ca-ex-PA. This study aims to clarify and explore the etiology and onset mechanism of SDCs. We examined their gene expression profiles and immunohistology, and compared these between Ca-ex-PA and de novo as well as among classifications based on their immunohistological profiles.

Patient Selection and Histological Review
We retrospectively analyzed 23 patients with untreated SDCs, who underwent surgery as a primary treatment in Hokkaido University Hospital, Japan, between 1991 and 2015. All tumors were confirmed to have been diagnosed accurately by two expert pathologists (TS and KH) according to the rigorous histomorphologic criteria for SDC (2). We conducted a histological review of the multi-step sections from the entire tumor in each case to classify SDCs into Ca-ex-PA group and de novo in accordance with the current WHO classification. All patients were treated with surgery as a primary treatment, and most underwent subsequent postoperative irradiation and/ or chemotherapy.

Tissue Microarray
Tissue microarray blocks were constructed using a manual tissue microarrayer (JF-4; Sakura Finetek Japan, Tokyo, Japan) with a 1.5 mm diameter needle. The finalized blocks were sliced into 4 mm-thick sections and mounted on glass slides. To check the histopathological diagnosis and adequacy of tissue sampling, a section from each microarray was stained with hematoxylin and eosin and examined by two expert pathologists (TS and KH).

Immunohistochemistry
For immunohistochemistry (IHC), a polymer-based detection system with heat-mediated antigen retrieval was employed. Diaminobenzidine was applied to detect antigen-antibody reactions. Appropriate positive and negative controls were employed for all conditions. IHC for HER2 (4B5, Ready-To-Use, Ventana) and AR (AR27, 1:50 dilution, Leica) was performed according to the respective manufacturer's recommendations.

HER2/CEN17 Dual Color in situ Hybridization
Dual color in situ hybridization (DISH) analysis using a

Scoring System for Immunostaining Classification
For HER2, the ASCO/CAP scoring system was used as follows: Negative, no membrane staining or <10% of cells stained; 1+, incomplete membrane staining in >10% of cells; 2+, >10% of cells with weak to moderate complete membrane staining; and 3 +, strong and complete membrane staining in >30% of cells (19). Cases were considered HER2-positive if HER2 staining was scored as 3+ or 2+ with HER2 gene amplification as defined by in-situ hybridization. AR expression level was semi-quantitatively counted every 10 percent. Nuclear staining was evaluated as positive. With regard to AR scoring, we considered a nuclear positivity ≥1% as positive according to ASCO/CAP 2013.

PCR Array for Gene Expression
RT2 Profiler PCR Arrays ® are tools for analyzing the expression of a focused panel of genes. Each 96-well plate PCR array includes SYBR ® Green-optimized primer assays for a thoroughly researched panel of relevant, pathway-or diseasefocused genes simultaneously under uniform cycling conditions. Total RNA was isolated using a FFPE Kit (QIAGEN, #217504). cDNA was synthesized using RT2 SYBR Green ROX qPCR MasterMix (QIAGEN, #330522). Four of the 23 samples were excluded due to the low quality of the RNA. The Human Breast Cancer RT 2 Profiler ™ PCR Array (QIAGEN, PAHS-131ZC-12) was used for the reaction in accordance with the manufacturer's instructions. Amplification and real-time analysis were performed with a StepOnePlusTM real-time PCR sy stem (The rm o Fis her Scientifi c, Waltham, MA, USA). Transcript levels were normalized against the GAPDH (Glyceraldehyde-3-Phosphate Dehydrogenase) RNA levels. The relative mRNA expression levels were calculated according to the comparative Ct (DDCt) method. We excluded six genes, including ADAM23 (ADAM Metallopeptidase Domain 23), BIRC5 (Baculoviral IAP Repeat Containing 5), BRCA2 (Breast cancer 2, early onset), CCNA1 (Cyclin A1), RARB (Retinoic Acid Receptor, Beta), and TWIST1 [Twist homolog 1 (Drosophila)], from the analysis as more than half of the values were missing. The function of each gene was referred to the document included with the PCR Array kit ( Table S1). All error bars represent the standard error value across biological replicates divided by the square root of the sample size (SEM). Where technical replicates were conducted, these values were averaged to yield a single value per biological replicate and, subsequently, all data are shown as means ± SEM.

Relationship Between Onset Classification (de novo vs. Ca-ex-PA) and Gene Expression by Using a Two-Group Comparison Test and Multiple Machine Learning Methods
To compare gene expression profiles between Ca-ex-PA and de novo, we performed a two-group comparison test on 78 genes for 19 cases. Subsequently, we searched for characteristic gene expression patterns between the onset classifications by using multiple machine learning methods; Orthogonal projections to latent structures discriminant analysis (OPLS-DA), LASSO, and RandomForest (RF).
OPLS-DA is a supervised machine learning method that employs a linear multivariate discriminant model, following projection of the predicted variables and observable variables to a new space (20)(21)(22). OPLS-DA can clarify the difference between two groups of high-dimensional data. Practically, by generating a principal component that maximizes the distance between the centroid of two groups and quantifying the weight of the principal component as the importance of the factor contributing to classification (VIP: Variable Influence on Projection), we can clarify the characteristic differences between two groups in an interpretable manner (21,(23)(24)(25). Furthermore, this method can avoid multicollinearity issues, which can be often assumed due to the large number of genes relative to that of samples in our study. We performed OPLS-DA with 6-fold cross-validation and 10,000 permutations to explore gene importance as a classifier for onset classification and clarify the mechanism for disease onset. The first principal component of variable importance in the projection (VIP) value above 1.5 is taken as a significant value for classification, suggesting the genes that characterize Ca-ex-PA or de novo.
LASSO is a logistic regression analysis method with L1 regularization in order to enhance the prediction accuracy and interpretability of the statistical model it produces (26). It sets the coefficients of less significant variables to 0 by assigning a L1penalty (lambda) to regression coefficients, and only more significant variables are extracted. A leave-one-out crossvalidation (LOOCV) with the grid-search method is used to determine the parameter lambda for optimization of the area under the ROC Curve (AUC).
RF is an ensemble learning method for classification or regression that functions by constructing a multitude of decision trees during training and outputting the mode of the classes (classification) or the mean prediction (regression) of the individual trees (27). RF is able to deal with high-dimensional data in a nonparametric manner, which allows us to assess interactive and nonlinear (regression) effects. Recently, feature selection based on the random forest classifier has been found to provide multivariate feature importance scores that are relatively effective, and which have been successfully applied to highdimensional data arising from microarrays (28).
We performed RF analysis for onset classification to explore gene importance as a classifier for onset classification and to clarify the mechanism underlying disease onset. In accordance with the default settings of the R package "RandomForest", the number of variables used per decision tree was set at eight. To build a CART model, 13 out of the 19 samples were selected to allow duplication. The number of decision trees was set at 10,000. This final prediction is determined by the principle of majority vote. We adopted MeanDecreaseGini as the variable importance index for classification.

MissForest
Biomedical research based on high-throughput technology often faces the problem of missing data. Algorithms commonly used in the analysis of such large-scale data often depend on a complete set, particularly for some machine learning methods.
Multiple imputation (MI) has been widely used for handling missing data in biomedical research (29). The most prevalent multiple imputation methods, such as k-nearest-neighbors for continuous data, saturated multinomial model for categorical data and multivariate imputation by chained equations for mixed data types, depend on tuning parameters or specification of a parametric model such as a linear data structure; however, realworld data does not necessarily follow these assumptions (30)(31)(32).
MissForest is an iterative imputation method based on a random forest, by averaging over many unpruned classification or regression trees. The characteristics of the non-parametric and randomized method in this algorithm can be applied to realworld data without strict assumptions about the distributional aspects of the data. Practically, it is also known to achieve better performance without tuning parameters.
We impute the missing values for our "gene expression" data including missing values with the R package "missForest" (33). The percentage of missing values was 9.8% (145/1,482), which suggests that imputation by "missForest" is applicable from the viewpoint of imputation quality (34). In accordance with the default settings of the R package "missForest", the number of decision trees was set at 100, and the number of variables used per decision tree was set at eight. We calculated the OOB imputation error rate based on the normalized root mean squared error and evaluated the complement accuracy. The OBB error rate value converged to around 0.483. These imputed data were then applied to OPLS-DA, LASSO, and RandomForest.

Partitioning the Dataset for Machine Learning
Supervised machine learning methods typically require the partitioning of data into training data and test data. According to the "70-30 rule", we randomly split the data into 70% for training and 30% for test with the R package "caret". These training data and test data were then applied to LASSO.

IHC and DISH images are shown in
The proportion of AR-positive cells showed a bimodal distribution and the median percentage of AR-positive cells was 80% (1st-quantile: 20.0, 3rd-quantile: 85.0, IQR: 65.0). Seventeen of 23 cases (73.9%) were considered to be ARpositive ( Table 2).
In the current study, 10 cases were classified as HER2positive, seven cases as Luminal-AR, and six cases as Basal-like ( Table 3).

Relationship Between Di Palma Classification and Gene Expression
We assessed the expression of 78 gene for 19 cases, including 10 HER2-positive, 5 Luminal-AR, and 4 Basal-like cases, by using a quantitative RT-PCR array (Figure 2).
To evaluate whether Di Palma classification reflects the gene expression status, we assessed 14 cancer classification marker genes corresponding to immunostaining status: "HER2-positive , and XBP1 (X-box binding protein 1), and "Basal-Like Cancer": EGFR, KRT5 (Keratin 5), and NOTCH1 (Notch 1). Among these genes, nine genes (ERBB2, GRB7, AR, ESR1, GATA3, KRT18, SLC39A6, and TFF3) were found to have significant differences in expression according to Di Palma classification, with the gene expression patterns showing significant differences being generally consistent with their markers for immunostaining classification ( Table 4). Except for those 14 cancer classification marker genes, there were significant differences in the expression of three genes: GLI1 (GLI family zinc finger 1), KRT19 (Keratin 19), and VEGFA.

Relationship Between Onset Classification (de novo vs. Ca-ex-PA) and Di Palma Classification
Of the 14 cases of Ca-ex-PA, nine were classified as "HER2positive", three as "Luminal-AR", and two as "Basal-like". On the other hand, of the nine cases of de novo SDC, one was classified as "HER2-positive", four as "Luminal-AR", and four as "Basallike". A comparison of de novo and Ca-ex-PA revealed more frequent HER2-positivity in Ca-ex-PA than in de novo (Fisher's exact test: p-value = 0.029) ( Table 5).
Relationship Between Onset Classification (de novo vs. Ca-ex-PA) and Gene Expression by Using a Two-Group Comparison Test and Multiple Machine Learning Methods

Two-Group Comparison Test
Regarding onset classification, seven genes showed significant differences in gene expression between Ca-ex-PA and de novo: VEGFA, ERBB2, IGF1R (Insulin-like growth factor 1 receptor), RB1 (Retinoblastoma 1), XBP1, SLIT2, and PTEN (Phosphatase and tensin homolog). In Ca-ex-PA type SDCs, the gene expression of VEGFA, ERBB2, IGF1R, RB1, and XBP1 was observed to increase, while that of SLIT2 and PTEN decreased. Analysis by genetic function showed that the significant differences in gene expression between Ca-ex-PA and de novo were concentrated in genes associated with angiogenesis and the AKT/PI3K signaling pathway (Fisher's test: p-value = 0.025 and 0.004, respectively) (Table 6, Figure 3).

DISCUSSION
Salivary duct carcinoma (SDC) was first described as morphologically similar to breast cancers in 1968 (1). SDC presents as a rapidly growing mass with the potential for local recurrence and neck and/or distant metastases. The current standard treatment for SDC is complete surgical resection with lymph node dissection and adjuvant radiotherapy. However, the results of standard therapy continue to be disappointing. More than 50% of patients die of disease within 3 to 5 years despite aggressive surgical resection and radiotherapy, and the overall 5year survival rate is 42-55% (35)(36)(37)(38)(39). Moreover, therapeutic options for patients with advanced unresectable primary, recurrent, or metastatic disease are particularly limited (40). In breast cancer, therapeutic strategies based on clinicopathology and molecular biology have already been established, and these have contributed to our understanding of the developmental mechanisms of the disease and have been applied in clinical settings. In SDC, on the other hand, although a new classification based on immune-profiles has been proposed by Di Palma et al. (18), there have been few studies to date on whether the classification reflects the molecular biological status.
Subsequently, we focused on onset classification and analyzed the differences in gene expression between Ca-ex-PA and de novo SDCs. As shown in Table 6, different gene profiles were observed for the two groups, with seven genes showing significant differences in gene expression. Interestingly, their functions were generally associated with angiogenesis and the AKT/PI3K signaling pathway (Figure 3).

Di Palma Classification Represents Gene Expression Profiles, and Subtypes Have Different Gene Expression Profiles
Previous articles have reported that approximately 80 to 90% of patients with SDCs are positive for AR, and 30 to 40% are positive for HER2, which is similar to our results (2,14,(41)(42)(43)(44)(45)(46). Thus, the patients in this study are thought to afford a similar population to those of the previous articles.
We examined whether Di Palma classification, based on immunostaining status, can be applied to gene expression profiles by comparing the immunostaining status with the expression of the corresponding gene. Among 14 cancer classification marker genes corresponding to immunostaining status, nine genes were found to show significant differences in gene expression among Di Palma classifications, suggesting that immunostaining appropriately reflects the gene expression profiles and that each subtype based on Di Palma classification ("HER2-positive", "Luminal-AR", and "Basal-like"), has a different gene expression profile.

Characteristic Differences Between Ca-Ex-PA and de novo
In the current study, HER2-positive cases were observed more frequently in the Ca-ex-PA than in the de novo (Fisher's exact test: p-value = 0.029), suggesting that there is a tendency for characteristic immunostaining differences to exist between the two types of SDC. Subsequently, we searched for characteristic differences in gene expression between the onset classifications. First, we compared the gene expression on 78 genes for 19 cases between Ca-ex-PA and de novo with a two-group comparison test.
Next, we wanted to perform multivariate analysis to remove the influence of confounding factors; however, the number of genes evaluated for the number of cases was so large that it was considered likely to result in a problem of multiple collinearity. Therefore, we applied a number of machine learning methods to search for characteristic differences in gene expression between the onset classifications.
As a result, the two-group comparison test and machine learning methods showed that genes related to angiogenesis could be extracted as candidates displaying characteristic differences between Ca-ex-PA and de novo. In particular, VEGFA, which plays a primary role in angiogenesis, was selected as a candidate to explain the characteristic differences between Ca-ex-PA and de novo in all assays.

Angiogenesis in Salivary Duct Carcinoma
In the current study, VEGFA was over-expressed in SDCs on the whole, and particularly in the SDCs in the Ca-ex-PA group. VEGFA is a member of the VEGF family of genes, which are particularly important to the induction of angiogenesis (47). VEGF has been reported to be expressed at high levels in most cancers (48), and to be associated with increased risk of recurrence, metastasis, and death in NSCLCs and RCCs (49)(50)(51).
Faur et al. investigated whether salivary gland tumors with a different morphology and evolution also differ in terms of neo-vascularization and VEGF expression, and the prognostic value of the results (52). Surgical specimens (8 PAs, 7 Warthin tumors, 5 basal cell adenomas, 6 Ca-ex-PAs, 6 mucoepidermoid carcinomas, 5 acinic cell carcinomas, 4 adenoid cystic carcinomas, and 4 adenocarcinomas not otherwise specified) were immune-stained. Malignant salivary gland tumors showed a significantly higher level of VEGF expression compared to benign tumors (p = 0.001). Fonseca et al. reported VEGF immunostaining for 132 salivary gland tumors (50 PAs, 32 mucoepidermoid carcinomas, 30 adenocarcinomas not otherwise specified, and 20 adenoid cystic carcinomas) and its relationship with their histopathological type (53). VEGF expression was found in the cytoplasm in all cases, proving to be overexpressed in malignant tumors in comparison to PAs, suggesting that VEGF might be associated with salivary gland cancer pathogenesis and aggressiveness. Fernańdez et al. examined the expression of VEGF protein in 66 salivary gland carcinomas and elucidated the relation between VEGF and clinicopathological parameters (54). VEGF expression was seen in 41 tumors (62%) and was correlated with lymph node metastasis (p < 0.005), clinical stage (p < 0.02), cause-specific survival (p < 0.01), and local failure-free survival (p < 0.02). They suggested that VEGF can contribute to the progression of salivary gland carcinomas and seems to be associated with neck node metastasis, worse survival and poor local control of the disease. Soares et al. investigated the angiogenic switch during the malignant transformation of PA into Ca-ex-PA, including 10 PAs, 8 early Ca-ex-PAs, and 8 advanced Ca-ex-Pas, and proved that angiogenesis was gradually but significantly increased from PAs to widely invasive Ca-ex-Pas (55). All of the above reports suggest that VEGF is involved in malignant conversion in salivary gland cancers.  Furthermore, VEGFA is also known to play a key role in the PI3K/AKT pathway (49) (Figure 4). PI3K activation occurs via RAS mutation, loss of PTEN, or by increased expression of growth factor receptors such as ERBB2, EGFR, IGF1R, and VEGFA. On the other hand, the activation of the PI3K/AKT pathway in tumor cells can promote the secretion of VEGFA, both by hypoxia-inducible factor 1 (HIF-1a)-dependent and -independent mechanisms (49). That is, VEGFA causes autonomous proliferation itself via the AKT/PI3K/VEGFA pathway, which has previously been identified as an "autocrine VEGFA signaling loop" in numerous other cancers (56). In the current study, we observed high levels of VEGFA, AKT1, and IGF1R expression, which indicates the possibility of the presence of an "autocrine VEGFA signaling loop" in SDCs. Furthermore, in the Ca-ex-PA SDCs, the expression levels of VEGFA, ERBB2 and IGF1R were increased while that of PTEN was decreased. These results suggested that the AKT/PI3K/VEGFA pathway could be activated more aggressively in Ca-ex-PA than in de novo SDCs, and further promote angiogenesis.
With regard to the mechanism underlying angiogenesis, recent studies have provided tremendous insights into fundamental aspects of angiogenesis that have led to a mechanistic model of vessel branching (57, 58) ( Figure 5). Under quiescent normal conditions, the basal membrane is located between endothelial cells (ECs) and mural cells, preventing resident ECs from leaving their position relative to the coating of mural cells. MMPs, including MMP2 and MMP9, which are expressed by many cell types, including fibroblasts, keratinocytes, and ECs, degrade the basal membrane in cooperation with VEGFA in the extracellular matrix to promote EC migration and generate angiogenesis (59). In this study, we observed high levels of MMP2 and MMP9 expression in SDCs in general, suggesting greater EC migration into the extracellular matrix and the promotion of angiogenesis.
Moreover, ECs transform into tip cells and stalk cells. Tip cells lead to the formation of new sprouts and explore whether the environment is suitable for angiogenesis, while stalk cells adjacent to tip cells form a lumen and support tip cells in the FIGURE 3 | Diagram of the relationship between genes and their genetic functions. Correspondence between each gene and its function is shown by connections between lines. In the "Cancer Classification Markers" section, 17 classification marker genes were categorized in accordance with Di Palma classification. Filled "yellow" highlights represent significant differences in gene expression by immunostaining classification. Filled "violet" highlights represent significant differences in gene expression by onset classification (Ca-ex-PA > de novo). Filled "pink" highlights represent significant differences in gene expression by onset classification (de novo > Ca-ex-PA). Analysis by genetic function revealed that the significant differences in expression between Ca-ex-PA and de novo were concentrated in genes associated with angiogenesis and the AKT/PI3K signaling pathway (Fisher's test: p-value = 0.025 and 0.004, respectively).  (60). As this feedback loop is repeated, the differentiation of ECs into tip cells and stalk cells is promoted, resulting in hyper-vascularization. To regulate this feedback loop, SLIT2-ROBO4 (roundabout guidance receptor 4) works as a "brake", with SLIT2 inhibiting VEGFR and NOTCH1 through the ROBO4 receptor on the cell membrane of tip cells and stalk cells (57,61). In Ca-ex-PA SDCs, we observed a higher level of VEGFA and a lower level of SLIT2  expression, which indicates that the regulation of the VEGFR-Dll4-Notch-VEGFR feedback loop by SLIT2/ROBO4 might have been out of control, resulting in hyper-activation of the feedback loop. These results suggested that uncontrolled angiogenesis could be promoted in SDCs, particularly in Ca-ex-PA type.

LIMITATIONS
There are some limitations to this study. The first limitation is that the sample size is small due to the rarity of SDC. Secondly, we did not provide in vitro result showing the possible relationship between the gene expression and the carcinogenesis of SDC, because there is no established cell line representing SDC. Although the combination of in vivo and in silico methods in this study can reveal the possible involvement of gene expression in the carcinogenesis, further studies including in vitro model will be necessary.

CONCLUSION
Our research indicated that the new classification of SDCs based on immune-phenotype appropriately reflects each gene expression profile. Our analysis of the onset classification of SDCs showed that there were significant differences in the expression of some genes related to angiogenesis and the AKT/PI3K signaling pathway. VEGFA, in particular, appears to be important in the transformation from PA to Ca-ex-PA and the aggressive behavior of SDCs, and affords a new target for approaches to the treatment of SDC.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding author.

ETHICS STATEMENT
The present study (016-0029) was approved by the Institutional Ethics Review Board of the Ethics Committee of our institution, and informed consent was obtained from each patient.