The molecular landscape and microenvironment of salivary duct carcinoma reveal new therapeutic opportunities

Purpose: Salivary duct carcinoma (SDC) is a rare and aggressive salivary gland cancer subtype with poor prognosis. The mutational landscape of SDC has already been the object of several studies, however little is known regarding the functional genomics and the tumor microenvironment despite their importance in oncology. Our investigation aimed at describing both the functional genomics of SDC and the SDC microenvironment, along with their clinical relevance. Methods: RNA-sequencing (24 tumors), proteomics (17 tumors), immunohistochemistry (22 tumors), and multiplexed immunofluorescence (3 tumors) data were obtained from three different patient cohorts and analyzed by digital imaging and bioinformatics. Adjacent non-tumoral tissue from patients in two cohorts were used in transcriptomic and proteomic analyses. Results: Transcriptomic and proteomic data revealed the importance of Notch, TGF-β, and interferon-γ signaling for all SDCs. We confirmed an overall strong desmoplastic reaction by measuring α-SMA abundance, the level of which was associated with recurrence-free survival (RFS). Two distinct immune phenotypes were observed: immune-poor SDCs (36%) and immune-infiltrated SDCs (64%). Advanced bioinformatics analysis of the transcriptomic data suggested 72 ligand-receptor interactions occurred in the microenvironment and correlated with the immune phenotype. Among these interactions, three immune checkpoints were validated by immunofluorescence, including CTLA-4/DC86 and TIM-3/galectin-9 interactions, previously unidentified in SDC. Immunofluorescence analysis also confirmed an important immunosuppressive role of macrophages and NK cells, also supported by the transcriptomic data. Conclusions: Together our data significantly increase the understanding of SDC biology and open new perspectives for SDC tumor treatment. Before applying immunotherapy, patient stratification according to the immune infiltrate should be taken into account. Immune-infiltrated SDC could benefit from immune checkpoint-targeting therapy, with novel options such as anti-CTLA-4. Macrophages or NK cells could also be targeted. The dense stroma, i.e., fibroblasts or hyaluronic acid, may also be the focus for immune-poor SDC therapies, e.g. in combination with Notch or TGF-β inhibitors, or molecules targeting SDC mutations.


30
Patient consents 31 The French patients with SDC diagnosis were consented for tissue collection and research analysis 32 under institutional reviewing board approval at the University Hospital of Montpellier (France). 33 For the Belgian patients, the ethical committee of the University Hospital Liege has approved the 34 use of human material in the current study. All samples were obtained from the institutional 35 biobank of the University Hospital Liege, Belgium. According to Belgian law, patients obtained 36 the information that the residual material could be used for research purpose and the consent is 37 presumed as long as the patient does not oppose (opting-out), which was not the case for those 38 patients.

51
Searching for 50-fold or higher variation between cohort 1 and cohort MSKCC after upper quartile 52 read count normalization yielded a list of 41 genes that was comprised of histones mostly: 53 54 We decided to remove these 41 genes from the study. The final read count matrix combining 57 cohorts 1 and MSKCC was filtered by only keeping the genes expressed with > 5 reads in > 5 58 samples (16,680 genes). Subsequently, the matrix was normalized by total read counts. We 59 observed no significant batch effect between cohorts ( Figure S1). It was also the case in all our 60 subsequent analyses. 61 62 63 64 65 Figure S1. Absence of batch effect. We considered genes with minimal expression at least, imposing an 66 average of 10 reads over all the samples (after data normalization), which left us with 16,129 genes. Then, 67 we selected either the 10,000 or the 5,000 most variable genes based on their coefficient of variation and 68 computed a dendrogram. Both dendrograms showed perfect separation of the normal samples from SDCs. 69 They also mixed Dalin et al. samples (marked with an asterisk) and ours in the different parts of the 70 dendrograms, showing the absence of any notable batch effect. The two dendrograms also clustered samples 71 in a comparable fashion, which was obviously expected. 72 73 74

75
Twenty FFPE tissue sections of five-micrometer thickness were deparaffinized with 1ml of xylen 76 at 60°C for 10 min. Following this, the samples were centrifuged at 20.000g, room temperature 77 (RT) for 5 min and the supernatant was removed. The xylen treatment was reapplied for a total of 78 3 times. Next, 1ml of ethanol was added and the samples were vortexed and centrifuged at 20.000g,

79
RT for 5 min. The supernatant was discarded and the ethanol wash was re-applied to the pellet for 80 a total of 4 times. The samples were then dried using Speed Vacuum and suspended in 500μl of 81 citrate buffer (pH 6) with 1% SDS. Following a sonication step, the samples were incubated for 82 30min at 95°C under vigorous shaking. Next, the samples were allowed to cool down at RT for 83 20-30min and the pH was re-adjusted to 8.5 with 100mM NaOH solution. The samples were 84 centrifuged (20.000g, RT for 5 min) and the supernatant was transferred to a new tube. The peptide samples were analyzed using a 1D-nano-HPLC system (Sciex, Framingham, MA, 103 USA), which was connected on-line with an electrospray Q-TOF mass spectrometer 6600 (Sciex).

104
A total of 1 µg of sample was injected on the C18 analytical column (Acclaim® 75 µm x 150 mm, 105 p/n: 162224; Dionex, California, USA) with a gradient of 0-40% phase B (90 % acetonitrile, 9.9 % 106 water and 0.1 % formic acid) for 100 min at the flow rate of 0.3 µl/min. Two acquisition modes 107 were used, data-dependent (DDA) for the measurement of the library and data-independent (DIA 108 or SWATH) for the samples. In the DDA mode, mass spectral data were acquired over a mass 109 range from 400 to 1600 m/z. One full MS scan was automatically followed by up to 30 MS/MS 110 scans of the most intensive peptides found in this mass range (bearing +2 or +3 charges). The   The list of antibodies used in the present work is outlined in Table S2 below. Following this, the

135
For immunofluorescence, tissue sections were prepared as described above with exception of 136 primary antibody incubation that was conducted at 4°C and over night and the staining that has 137 been performed using Opal system (Perkin Elmer, cat. no.: NEL810001KT). Following the 138 primary antibody incubation, the slides were incubated with the corresponding secondary antibody 139 as described above. The slides were then incubated with 100μL staining solution prepared from 140 2μL Opal dye and 98μL Amplifying Buffer. Following 10 min incubation, the slides were washed 141 three times for 5 min in PBS and then subjected to microwave-assisted antibody removal. Slides 142 were immersed in AR6 buffer and were treated in the microwave for 15 min, maintaining the heat 143 close to the boiling point. After cooling and a wash in PBS buffer for 5 min, the tissues were re-144 blocked with for 30 min in protein block serum-free solution at RT. Tissues were then incubated 145 with the next primary antibody and the staining procedure was repeated as described above using 146 the following Opal dyes: 520, 570, 620 and 690. Finally, slides were mounted using 147 VECTASHIELD® Antifade Mounting Medium with DAPI (Vector, Burlingame, USA). (ifLR-score) and the definition of a threshold above which the interaction is considered positive.

153
Taking the PD-1/PD-L1 interaction as an example, we first determined the average diameters of 154 PD-1+ and PD-L1+ cells independently ( Figure S2A). This allowed us to define a crown-shaped 155 area around each PD-1+ cell. The receptor abundance is estimated by the average PD-1 156 fluorescence inside the inner disc that is centered on the PD-1+ cell and has the corresponding 157 diameter. The ligand abundance is estimated in the crown that has a width equal to half the 158 diameter of a PD-L1+ cell. represents the number of ligands close enough to the PD-1+ cell to 159 engage inhibition. Empirically, we define 160 ifLR-score = 1/3 1/2 /( + 1/3 1/2 ), 161 where and are as above, is the average of the average intensity over the whole ligand image 162 and the average intensity over the whole receptor image (each label results in a separate gray-scale 163 image). The fractional powers account for the ligand and the receptor to reside in a 3-, respectively 164 2-, dimensional space. represents the background signal intensity and its role is to regularize the 165 ifLR-score to obtain values between 0 and 1. Analysis of the PD-1/PD-L1 interaction in 3 SDC 166 allowed us to plot ifLR-score value distribution ( Figure S2B). It is bimodal (or even trimodal for 167 SDC22), with a fist mode corresponding to random signals (low values) followed by a rightmost 168 mode corresponding to overlapping ligand and receptor fluorescence. We empirically set a 169 conservative threshold at 0.4. That is, each PD-1+ cell with ifLR-score > 0.4 is considered in 170 positive interaction with adjacent PD-L1+ cell(s), otherwise the interaction is deemed negative.  181 We also defined a ligand-receptor score meant to assess co-expression in transcriptomics as a 182 proxy for potential true interaction in the sample. The empirical formula is similar to the above: 183 LR-score = 1/3 1/2 /( + 1/3 1/2 ), 184

LR-score (transcriptomics)
where is the ligand read count in log10 (ligand transcript expression), the receptor read count in 185 log10, and the average log10 read count over all the genes and all the SDC transcriptomes. See