Establishment of PDOs library of SGTs
The workflow for this study was illustrated in Fig. 1A. Totally, 8 categories of SGTs and the adjacent normal tissue, and the adjacent normal tissue from submandibular gland cyst (SGC) were obtained from 61 patients (Table 1). Tumor organoids were cultured in the DMEMF/12 basal medium with an additional supplement, while the organoids medium for NSG was modified from the medium used to derive NSG organoids from induced pluripotent stem cells [19, 20]. Following the protocols, we established a library with 45 PDOs that included 21 benign SGTs (PA: 11/13, BCA: 10/10) and 24 malignant tumors (MEC: 6/8, ACC: 10/12, AciCC: 4/4, SDC: 4/5) (Fig. 1B). In addition, we cultured the matched normal organoids (26 out of 33) from the adjacent normal tissue (Table 1). Overall, the success rate of organoids cultures was 71/85 (84%). Those samples failed to set up the organoid models might due to the variation in tumor stroma. Typically, the unsuccessful cases contained a high percentage of tumor stroma in various morphology, such as the hyaline stroma in ACC, or the chondroid and osteoid stromal components in PA, which had fewer tumor cells than the others (Supplementary Fig. S1A). Unfortunately, under the current culture system, the Warthin tumor, the second most common benign tumor in salivary glands (SGs) with dense lymphoid stroma, and oncocytoma, a rare benign tumor with vascular stroma, could not initiate tumor organoids (Supplementary Fig. S1B and C), and these cases were excluded from our final count. All the clinical characteristics for those tumors applied in this research were shown in Table 1.
The initial stage of SGTs organoids was a process of self-assembly from single cells, and the expansion rates varied according to the tumor types (Fig. 2A). Most types of tumors were in a rapid growth phase in the first week, could be firstly passaged approximately in 14 days, and then reached a platform phase till 21 days (Fig. 2B and Supplementary Fig. S2), which was consistent with the previous report for NSG organoids [25]. These organoids were expanded at least for five passages except for AciCC which has a lower proliferation rate. Notably, we observed that NSG organoids were larger than SGTs, and the secreted mucus often appeared in the middle of NSG organoids (Fig. 2C), which indicates that NSG organoids are functional. Interestingly, tumor organoids did not grow faster than the matched normal organoid counterparts, and in many cases even grew more slowly (Fig. 2B and D). As reported in the other research, this might due to the higher rates of mitotic failures and subsequent cell death of tumor cells [26, 27].
Pdos Recapitulate The Morphologic Features Of Parental Sgts
Current classifications of SGTs are based on histological morphology and the expression of a few biomarkers. Under the brightfield microscope, all types of organoids showed mainly ball-like structure and irregular tumor mass shape in the culture plate (Fig. 2). To compare the morphological and histological features of PDOs with the parental tissues, we performed H&E staining and IF analysis for the matched tumors and tumor organoids. As shown in Fig. 3, we applied the biomarkers used in the routine pathological diagnosis, including the ductal and basal cell markers (CK5, p63), the glandular and luminal cell markers (CK8, AQP5), and the myoepithelial cell markers (CK5, SMA, p63). Interestingly, these organoids exhibited overlapping growth patterns and cell types, as illustrated in Fig. 3 that the CK5+ cells and the CK8+ luminal cells comprised the most mass in almost all the organoids. However, the proportion of these two types of cells in tumor mass was different according to the specific subtype of SGTs, which was consistent with the parental tissue (Fig. 4).
In particular, PA, the benign and most frequently occurring tumor in the SGs, consists of glandular cells and myoepithelial cells that are arranged in ducts, sheets, mucous stroma-like, and chondral stroma-like structures [6]. We observed that the PDOs also formed a similar substructure to the parental tumors, and the expression patterns of biomarkers (CK5+, CK8+, p63+, AQP5+) were consistent with the parental tumors (Fig. 4A). BCA is another type of benign tumor in SGs, characterized by basal cell-like tumor cells that are arranged in the typical palisading structures [6]. Interestingly, the organoids exhibited a palisade-like structure with the outer layer of p63+ basal cells and the inner layer of CK8+ luminal cells together with CK5+ ductal cells, which was similar to the original tumors (Fig. 4B). MEC is the most common malignancy in SGs, featured with squamoid, mucous and intermediate cells with solid and cystic patterns (Fig. 4C). The squamoid cells (CK5+, CK8+) and the mucin-producing cells (CK8+, AQP5+, MUC4+) comprised the tumor mass in both PDOs and the parental tumors, and these cells formed typical luminal structure and nest patterns. Like the parental tumors, the outer layer of the solid pattern in PDOs was CK5+ p63+ basal cells, while the inner layer of tumor mass was CK8+ AQP5+ luminal cells. The mucus-producing marker MUC4 was positive in both tumor tissue and tumor organoids. ACC is the second most common malignancy in SGs, and the tumor cells are composed of ductal epithelial and myoepithelial cells arranged into tubular, cribriform, and solid structures[5]. Notably, our ACC organoids displayed all these three typical structures which are arranged by luminal cells (CK8+, AQP5+), basal cells (CK5+ p63+) and myoepithelial cells (CK5+, SMA+, p63+) (Fig. 4D). AciCC, the third most common malignancy, comprises of acinar and ductal cells. AciCC organoids showed the small acini structures which were positive for acinar cells markers (CK8+, CK5−, AQP5+), and the known biomarkers (DOG1+, SOX10+) (Fig. 4E). These markers in organoids are identical to the parental tumors, indicating that the PDOs maintained the molecular characteristics of AciCC and reconstituted the acinar structure as the original tumors. SDC is another common malignancy in SGs, characterized by comedonecrosis and cribriform in the nest tumor mass. The tumor cells are typically apocrine oncocytoid with abundant cytoplasm and a large pleomorphic nucleus[5]. The cultured organoids showed a solid pattern with the large cytoplasm and nuclei cells, which was consistent with the parental tumors. In both organoids and tumor tissues, CK8+ AQP5+ cells were the major population, which suggested that the tumor cells in SDC were terminally differentiated ductal cells. Expression pattern of HER2 (a marker for SDC) in the organoids were also consistent with their parental tumors (Fig. 4F). Compared with SGTs organoids, NSG organoids showed bigger and more regular luminal structures. Similar to the NSG tissue, the inner layer of organoids locates CK8+ AQP5+ luminal cells, and the outer layer CK5+ p63+ basal cells (Fig. 4G).
Pdos Recapitulate The Transcriptional Features Of Parental Sgts
We performed transcriptome analysis for all these organoids and their corresponding tissue. t-distributed stochastic neighboring embedding (t-SNE) is one of the popular dimensionality reduction techniques to visualize RNA-seq data in two dimensions. By performing t-SNE, the sequencing data of 108 samples were shown in Fig. 5, and all the aberrant samples that did not show the subtype-specific characters were excluded in the following analysis. We found that each type of tumor, organoids, and NSG could form its own clusters, and both tumors and their corresponding organoids could be separated from NSG. This suggested that each subtype of SGTs had its own transcriptional characters. Interestingly, the clusters of AciCC, MEC, and SDC were close to their corresponding organoids, indicating that the cultured organoids could maintain the transcript stability as their parental tumors. However, the t-SNE plot revealed the separation of ACC, PA, and BCA from their corresponding organoids. We think the major change in these organoids might be due to the diminishing of myoepithelial cells that are exclusively present in ACC, PA, and BCA. Of note, in the t-SNE plot, we also included 3 cases of myoepithelioma (ME) and one case of myoepithelial carcinoma (MC), whose main cell types are myoepithelial cells to visualize the benign and malignant tumors with myoepithelial cells. Myoepithelial cells have characteristics of both epithelium and smooth muscle cells. It is known that the current method of organoids culture is selective for growing epithelial cells, therefore the myoepithelial cells might gradually diminish under this condition. Indeed, we did observe that SMA+ myoepithelial cells were almost undetectable in the long-term cultured organoids by IF staining (Fig. 4A, B and D).
In order to characterize the myoepithelial cells in SGTs, we firstly evaluated our samples with biomarkers for myoepithelial cells (CK5+ SMA+), and verified the myoepithelial cells presented only in ACC, PA, and BCA, but not in AciCC, MEC, and SDC (Fig. 5B). Then we further compared these two groups of SGTs, one with myoepithelial cells including ACC, PA, and BCA, and another without myoepithelial cells including ACC, PA, and BCA. With RNA-seq data, 1799 differentially expressed genes (DEGs) (Log FC > = 2, P < 0.05) were selected from these two groups, and we found transcription factors SOX1 and SOX14 were on the top list in the upregulated genes in SGTs with myoepithelial cells (Fig. 5C). Interestingly, two genes EN1 and FAM178B were highly expressed in the SGTs with myoepithelial cells (Fig. 5C), which is consistent with our following analysis that EN1 and FAM178B were related to myoepithelial cells.
Organoids Clustering Reveals The Epithelial Characteristics And Visualizes The Intrinsic Relationship Among The Subtypes
To understand the heterogeneity and overlap of SGTs, we need to comprehensively evaluate the molecular characteristics among the subtypes of SGTs. Firstly, we analyze the 6 most common SGTs by performing a K-means clustering with our transcriptomic database. Although these tumors were classified morphologically into 6 subtypes by pathological diagnosis, both the transcriptomic profiling of PDOs and tumor tissue were clearly clustered into 4 subgroups (Fig. 6A and B). The clusters of organoids model showed a merge of BCA and ACC, but a separation from all other subtypes of SGTs (Fig. 6A). This suggested a clear overlap of epithelial characteristics between BCA and ACC. Interestingly, with the tumor stroma, the entire tumor tissue model reveals a separation of BCA and ACC (Fig. 6B), which indicated the stromal characteristics of BCA and ACC are distinct since they could be clustered into different groups. However, MEC and SDC were clustered into the same group in both the organoid and the entire tissue models, but could be slightly separated in the organoid model, which indicated their epithelial characteristics were distinct, but not the stromal characteristics.
To further understand the correlation among these subtypes of SGTs, we applied hierarchical clustering analysis based on Euclidean distance and complete-linkage with the transcriptional profiles, which could visualize the associated data in a tree-like structure [23]. The tree structure in the organoid model showed that PA was associated with all the other subtypes of SGTs, BCA and ACC were closely associated, and MEC and SDC were closely related (Fig. 6C), which is indeed consistent with the clinical observation. As we know PA, BCA, and ACC are commonly composed of CK5+ basal cells, CK8+ luminal cells, and SMA+ myoepithelial cells, while MEC and SDC are composed of CK5+ basal cells and a large number of CK8+ luminal cells. Notably, PA is the most plastic subtype of SGTs in the clinic, that frequently undergo malignant transformation, such as MEC, ACC, SDC, or NOS [28]. Of interest, with the input from the tumor stroma, the tree-like structure of the entire tumor tissue showed that BCA, ACC, and PA were related to each other, but BCA and ACC were further separated (Fig. 6D), which indicates it is the tumor stroma that makes them more different. In addition, MEC and SDC were associated, while AciCC was a distinct branch in the tissue model as well as in the organoid model. This is also consistent with our observation and the other study that AciCC does not directly develop from CK5+ stem cells in the organoid model, but might develop through the self-duplication [29].
Selection Of The Potential Biomarkers In The 6 Most Common Types Of Sgts
After understanding the similarities and differences among the subtypes, we aim to select the subtype-specific biomarkers. Since the organoid culture selectively maintains the epithelial cells from the parental tumors, tumor organoids are also an ideal model to identify tumor biomarkers without interference from the tumor stroma. For this purpose, we selected the biomarkers that were exclusively highly expressed in these 6 types of SGTs respectively. Firstly, we picked up the DEGs in each type of tumor tissue and the corresponding organoids by comparing them with the normal tissue. By intersecting analysis, the DEGs presented both in tumor tissue and its organoids were selected, which represented the DEGs mainly in tumor epitheliums (Fig. 7A). Upset plot (Fig. 7B) showed that the selected DEGs were exclusively present in the specific type of SGTs. To avoid possible overfitting and multicollinearity in screening the variables in high-dimensional data, we applied the logistic regression model with lasso regularization to define the lowest cross-validation error rates to pick up the most remarkable DEGs (Fig. 7C). Finally, we selected the potential biomarkers exclusively highly expressed in the 6 most common types of SGTs respectively. These candidates identified in tumor organoids showed similar expression patterns to the parental tumor tissues (Fig. 7D), which indicated that tumor organoids kept consistent transcriptomic stability in the culture system in vitro. Notably, the level of EN1 and FAM178B were exclusively high in ACC, but not in the corresponding cultured organoids. We speculate these genes might be related to myoepithelial cells, since myoepithelial cells are the main cell components in ACC and diminish in the long-term organoids culture.
Validation Of The Potential Biomarkers In Mec
MEC is the most common malignant tumor in SGTs, characterized by variable cells and structures. It has overlapping morphologic features with cystadenocarcinoma, adenosquamous carcinoma, squamous cell carcinoma, and so on. Many efforts have been made to identify biomarkers to increase diagnostic accuracy [30, 31]. We found here that NEFL and PTP4A1 were the distinct genes highly expressed in MEC compared with other types of SGTs, therefore, both RNA and protein levels of NEFL and PTP4A1 in SGTs were evaluated. By RT-PCR, NEFL RNA was detected and showed a prominently high level in MEC compared with other types of SGTs (Fig. 8A), which was concordant with the RNA-seq data. The protein level of NEFL was evaluated by IHC staining on TMA sections, with a panel of 367 cases of SGTs, including PA (n = 82), BCA (n = 51), ACC (n = 54), AciCC (n = 39), SDC (n = 61), MEC(n = 80) respectively. As shown in Fig. 8B and C, NEFL was highly expressed in both cell membrane and cytoplasm in MEC, and the percentage of immunoreactivity (IR) was strong 26% (21/80), medium positive 44% (35/80), weak staining 18% (14/80), negative 12% (10/80) respectively. However, it could also be detected in PA 11% (9/82), BCA 49% (25/51), ACC 46% (25/54), and DC 35% (21/61) with strong and medium positive IR (Supplementary Fig. S3A). For PTP4A1, both RNA level and protein level were the highest in MEC (Fig. 8D, E, and F). IHC staining showed that PTP4A1 was positive in MEC 81% (58/72), including medium positive 39% (28/72), weak staining 42% (30/72), and negative 19% (14/72). Interestingly, there was a lower expression of PTP4A1 in AciCC and DC (Supplementary Fig. S3B), and all these tumors that expressed PTP4A1 were the non-myoepithelial types of SGTs, which suggested that PTP4A1 might function in tumor epitheliums, but not myoepithelial cells.