Epithelium

Representative tumor regions were annotated across all 312 H&E-stained slides by an experienced pathologist (MBL) using QuPath’s manual annotation tools. A script was then applied in batch to automatically identify and set the average background intensity for the red, green and blue channels of each image, which varied markedly according to the scanner used. A second script was then run over all images to apply QuPath’s SLIC superpixel segmentation command to subdivide each annotated region into ‘superpixels’ based upon simple linear iterative clustering^{33 (link)}. This script additionally calculated both the average hue for each superpixel along with Haralick texture features³⁴ from optical density values using QuPath’s Add intensity features command. QuPath’s Add smoothed features command was also applied to calculate a Gaussian-weighted sum of the features of neighboring superpixels, and append these to the existing features for each superpixel. This provide additional contextual information extending beyond the superpixel itself.
A subset of 40 ‘training’ images was then identified for the pathologist to interactively train a random trees classifier to distinguish between tissue areas comprising tumor epithelium, stroma and ‘other’ (e.g. whitespace, mucin, normal muscle or necrosis). This required drawing around regions containing tissue of each class and annotating these accordingly. During this process, QuPath used all available features to train the classifier in a background process and thereby provide immediate feedback on classification performance. Once the classification was considered adequate across the training images, the classifier was applied to all images within the set and the total area of superpixels for each class was exported. The tumor stromal percentage (TSP) was then calculated as $$\mathrm{TSP}=\mathrm{AS}/(\mathrm{AE}+\mathrm{AS})\times 100\%$$
where AS represents the total area classified as stroma, and AE represents the total area classified as epithelium.

Below we give a brief summary of the datasets used in the construction of the reference databases (see also Table 1).Table 1

Main Illumina 450k DNAm datasets used. We list the main datasets used in this study, the cell-types/tissue profiled, whether the data was used for reference database construction (if yes, we specify which cell-types were used), whether the data was used for validation/evaluation purposes (if yes, we specify which cell-types were used) and the reference/citation

Dataset Name	Tissue/cell-types	Use in Reference DNAm Database	Testing/Evaluation	Reference
Reinius et al.	WB, PBMC, NK, B, CD4T, CD8T, Monoc, Neutro, Eosino. (n = 6 of each)	NK, B, CD4T, CD8T, Monoc., Neutro., Eosino.	WB & PBMC	[24 (link)]
Liu et al.	WB (n = 335 controls, n = 354 rheumathoid arthritis cases)	No	Average Flow Cytometry estimates for cases and controls	[2 (link)]
Koestler et al.	WB (n = 18)	No	12 Reconstructed WB mixtures + 6 WB samples with Flow Cytometry estimates	[20 (link)]
Zilbauer et al.	PBMC, CD4T, CD8T, NK, B, Monoc, Neutro. (n = 6 of each)	No	In-silico mixtures of purified blood cell subtypes	[27 (link)]
ENCODE	Various	HMEC, HRCE, IMR90, Liver	No	[22 (link)]
Slieker et al.	Various	Pancreas	Liver	[26 (link)]
SCM2	Various	No	HRCE, Pancreas, IMR90	[29 (link)]
Lowe et al.	Various	No	HMEC	[28 (link), 35 (link)]
Teschendorff et al.	WB (n = 152)	No	Smoking associated DMCs	[31 (link)]

Abbreviations: DNAm = DNA methylation, WB = whole blood, PBMC = peripheral blood mononuclear cells, HMEC = human mammary epithelial cells, HRCE = human renal cortical epithelia, IMR90 (fetal lung fibroblast), SCM2 = Stem-Cell-Matrix Compendium-2, DMCs = differentially methylated CpGs, NK = natural killer cells, B = B-cell, Monoc = Monocytes, Neutro. = Neutrophils, Eosino = Eosinophils, CD4T = CD4+ T-cells, CD8T = CD8+ T-cells

Deconvolution was performed using the CibersortX algorithm at cibersortx.stanford.edu [74 (link)]. Single-cell transcriptomic profiling dataset of cells in the embryonic pancreas [39 (link)] was used as a reference, including count matrix and metadata labels. Particularly, only cells with pancreatic epithelial or endocrine cell fate were used, corresponding to the annotation of five broader cell types—α cells, β cells, endocrine progenitors, trunk epithelium and tip epithelium [39 (link)]. The reference matrix was built out of the 2589 cells and gene list of 18,565 gene features, as deposited by [39 (link)]. Each cell population counted > 250 cells. The units of the reference matrix were UMI counts. Calculation of the scRNA-seq signature matrix was done in default mode (quantile normalization disabled, minimal expression of 0.75, replicates of 5, sampling of 0.5). Imputation of cell fractions and group-mode expression were used in default settings, with S-mode batch correction enabled, quantile normalization disabled and n = 100 permutations for significance analysis. Sample mixture file was submitted with unfiltered gene list 27,124 features for Isl1CKO and in UMI counts.

Pancreatic cancer cell lines (AsPC-1 and BxPC-3) were cultured in RPMI-1640 (Corning, NY, USA) with 10% fetal bovine serum (FBS) and 1% penicillin–streptomycin. Two additional pancreatic cancer cell lines (PANC-1, MIA Paca-2) were cultured in DMEM (Dulbecco’ modified eagle medium) (Gibco, Grand Island, NY, USA) supplemented with 10% FBS and 1% penicillin–streptomycin. Human pancreatic ductal epithelium (hTERT-HPNE) cells were cultured in Medium D with mixtures of M3 and DMEM medium containing one volume of medium M3TM Base F culture media (InCell Corp., San Antonio, TX, USA), three volumes of glucose-free DMEM, 5% FBS, 5.5 mM glucose, 10 ng/ml EGF, and 50 µg/ml gentamycin [26 (link)]. All these cells were cultured at 37 °C in a humidified atmosphere containing 5% CO2. RNA was extracted from tissues using the TRIzol reagent (Invitrogen, Carlsbad, CA, USA) and was reverse-transcribed into cDNA using the PrimeScript RT Master Mix (Takara, Otsu, Shiga, Japan). RT-qPCR analyses were quantified with PowerUp™ SYBR® Green Master Mix (Applied Biosystems, Austin, TX, USA), and expression levels were normalized to GAPDH levels. Proteins were extracted in RIPA buffer supplemented with a complete, EDTA-free protease and phosphatase inhibitor single-use cocktail (Thermo Scientific). Proteins were separated by SDS-PAGE and blotted onto a PVDF membrane. Anti-TSC22D2 (1:1000 dilution, #25,418–1-AP, Proteintech) was used as primary antibodies for immunoblotting. Reacted antibodies were detected using an enhanced chemiluminescence detection system.