Pancreas single-cell RNA-seq analysis input for DUET: ductal, endothelial and stellate cells from PDAC and a non-tumor reference (GEO, HPAP)

Abstract

This dataset is the pancreas analysis input used in the manuscript describing DUET (Detection–Expression Unified Test), a Python implementation of two-part hurdle testing for single-cell differential expression. It contains the 47,187 ductal, endothelial and stellate cells used for the agreement with R MAST, the computational benchmark, the p-value underflow analyses, the donor-partition null experiments, the pseudobulk validation and the effect-size comparison. Contents. One AnnData file (h5ad): 47,187 cells × 17,750 genes. X: log1p(normalize_total(counts, target_sum = 10,000)), natural logarithm, float32 sparse matrix. layers["counts"]: integer UMI counts (int32 sparse matrix). obs: Sample (Reference = non-tumor reference, pancreas = PDAC), group, celltype, donor, patient, library, series, tissue, source_file, runs_merged, total_counts, n_genes_by_counts. var: gene symbols. A CSV table lists every input file of the integrated dataset with its series, sample, tissue, donor and library. Cell type Non-tumor reference PDAC Ductal cell 8,797 11,476 Endothelial cell 5,252 9,195 Stellate cell 2,074 10,393 Donors: 25 non-tumor, 121 PDAC. Sources. PDAC samples are from GEO series GSE154778, GSE155698, GSE156405, GSE197177, GSE205013, GSE212966, GSE214295, GSE217845, GSE229413, GSE231535, GSE242230, GSE263733, GSE278688 and GSE335452. The non-tumor reference comprises organ-donor pancreas (GSE229413, including samples with pancreatic intraepithelial neoplasia lesions), tumor-adjacent pancreas (GSE155698, GSE197177, GSE229413, GSE263733, GSE278688) and islet preparations of Human Pancreas Analysis Program (HPAP) donors HPAP-077, HPAP-092, HPAP-099, HPAP-101 and HPAP-104, whose count matrices were generated from the FASTQ files available through PANC-DB (https://hpap.pmacs.upenn.edu). Repeated sequencing of one library. Several input files re-sequenced one library (the 14 HPAP count files of five donor libraries, 16 GEO libraries deposited in both GSE155698 and GSE229413, and one library deposited as two files in GSE263733). Each file had been UMI-collapsed separately, so summing them would count a molecule sequenced in more than one run more than once. Files of one library were identified by shared cell barcodes with correlated UMI totals, and each cell is kept once, with the counts of the file with the most UMIs. Processing. Count matrices were read per sample, files of one library combined as above, and genes measured in every sample kept. Cells with fewer than 600 detected genes, genes detected in fewer than three cells, cells outside the 2nd–98th percentile of detected genes or with at least 20% mitochondrial counts, doublets (Scrublet) and cells with an estimated ambient-RNA fraction above 0.5 (DecontX) were removed. Expression was normalized to 10,000 counts per cell and log1p-transformed; samples were integrated with scVI using the sequencing library as batch; Leiden clusters (resolution 1.0) were annotated with decoupler and CellMarker marker sets, and cancer-labeled cells of non-tumor samples were reassigned after infercnvpy analysis. The file contains the three cell types analyzed; the integration embedding is not included. Changes from version 1.0. New integration with more sources and a broader non-tumor reference. In version 1.0 the repeated HPAP sequencing runs were summed by donor and cell barcode, which counted molecules sequenced in more than one run more than once; version 2.0 counts each cell once. Limitation. Tumor and non-tumor samples differ in data source and tissue preparation (islet preparations and organ-donor pancreas are non-tumor only). The data are suitable for methodological comparisons, not for biological inference about tumor versus non-tumor expression. Verification. Every row of X and counts is identical to the object analyzed in the manuscript, X equals log1p(CP10k) of counts, and DUET run on the ductal cells reproduces the manuscript results for all 8,179 genes (maximum absolute difference in −log10 p 5.7 × 10−14). Software. DUET 1.1.0: https://github.com/Pahi95/duet, https://doi.org/10.5281/zenodo.22877336 How to cite and acknowledge. Please cite this dataset, the DUET manuscript, the original GEO studies and the HPAP publications (Kaestner et al., Diabetes 2019; Shapira et al., Cell Metab 2022; Fasolino et al., Nat Metab 2022; Patil et al., Nat Metab 2023), and include the HPAP acknowledgement: "This study used data from the Human Pancreas Analysis Program (HPAP; RRID:SCR_016202; PMID: 31127054; PMID: 36206763). HPAP is part of the Human Islet Research Network (RRID:SCR_014393) consortium (UC4-DK112217, U01-DK123594, UC4-DK112232, and U01-DK123716)."