ProstaTrend Meta-Analysis

The development of the (revised) ProstaTrend signature is described in the main publication. For an overview of the cohorts used for the meta-analysis, see “Datasets.” All pre-processing stept are described in the main publication. We included all ProstaTrend genes annotated with an Ensembl ID and present in at least 2 cohorts in the meta-analysis. For each gene, a univariate Cox-regression model was applied to standardized expression values in log-space for each cohort individually. For the ProstaTrend training cohorts [1], it should be noted that due to a matched study design, patients from FF_seq_RP were included in FF_array_RP. Therefore, we used a weighted Cox-regression with weights 1/N, where N is the number of samples for the specific patient included across the two cohorts. Similarly, 40 patients are included in both the CancerMap_2017_Luca and CamCap_2016_Ross_Adams cohorts. However, since these patients are not identifiable, no adjustment could be made for these duplicates. Given the small number of duplicate patients and the large number of cases (n= 1821), the effects of higher weights for matched samples are negligible. Using the log hazard ratio values and the respective standard errors calculated from the coxph function, we estimated a combined effect size for each gene. We used a random effects model because we did not assume that there is one true effect size, which is shared by all the included cohorts, but rather a range of true effect sizes with additional sources of variation, such as different platforms (RNA-Seq and microarray), clinical or demographic variables, etc. The model was fitted with the restricted maximum-likelihood estimator using the R package meta. We declared that a prognostic gene from the (revised) ProstaTrend signature had a significant combined effect size when the FDR adjusted p-value (Benjamini-Hochberg method) was <0.05.

TCGA/GTEx

RNA-Seq data from the the Cancer Genome Atlas Project (TCGA) were downloaded using the recount2 database [2]. For this, we used the download_study('TCGA', type = 'rse-gene') function implemented in the R/Bioconductor package recount [3] to obtain gene feature-level count matrices. RNA-Seq data from human tissues of the Genotype-Tissue Expression (GTEx) project [4] were downloaded using the command download_study('SRP012682', type = 'rse-gene'). Transcripts per million (TPM) normalized gene expression values were extracted using the getTPM function. For TCGA cox proportional hazards models were performed for continuous standardized gene expression (in log-space) values of the ProstaTrend signature in a univariate regression model using the coxph function from the survival package. Analysis was performed using samples from primary tumors. Significance was adjusted for multiple testing using the Benamini-Hochberg method. The primary endpoint, with the exception of PRAD, was death from disease. For PRAD, biochemical relapse was used as the primary endpoint.

PCa cell atlas

Raw read counts from scRNA-Seq data of 5 studies were downloaded from NCBI GEO. This included the studies Chen et al. (GSE141445, [5]), Dong et al. (GSE137829, [6]), Ma et al. (GSE157703, [7]) and Song et al. (GSE176031, [8]). Read counts from the study Tuong et al. [9] were obtained from https://www.prostatecellatlas.org. We excluded one sample from the study by Chen et al. because it was a biopsy of a lymph node metastasis that cannot be directly compared with primary tumor. Pre-processing, integration, clustering, quality control, cell cluster annotation, estimation of copy number variations (CNVs) and differential gene expression analysis (DGEA) are described in Additional File 1 of the main publication. Using the study-wise standardized, normalized expression values of the scRNA-Seq datasets, we applied for each cell a simplified ProstaTrend Transcriptomc risk score, which was the mean of all genes at increased risk minus the mean of genes at reduced risk. The simplification of the score was necessary due to the low expression levels and high drop-out rate of individual genes in the scRNA-Seq data, which makes the transferability of the weights from the bulk analyses not robust.

[1] Kreuz M, Otto DJ, Fuessel S, et al. ProstaTrend-A Multivariable Prognostic RNA Expression Score for Aggressive Prostate Cancer. Eur Urol. 2020;78(3):452-459. doi:10.1016/j.eururo.2020.06.001

[2] Collado-Torres L, Nellore A, Kammers K, et al. Reproducible RNA-seq analysis using recount2. Nat Biotechnol. 2017;35(4):319-321. doi:10.1038/nbt.3838

[3] Collado-Torres L, Nellore A, Jaffe AE. recount workflow: Accessing over 70,000 human RNA-seq samples with Bioconductor. F1000Res. 2017;6:1558. Published 2017 Aug 24. doi:10.12688/f1000research.12223.1

[4] Carithers LJ, Moore HM. The Genotype-Tissue Expression (GTEx) Project. Biopreserv Biobank. 2015;13(5):307-308. doi:10.1089/bio.2015.29031.hmm

[5] Chen, S. et al. Single-cell analysis reveals transcriptomic remodellings in distinct cell types that contribute to human prostate cancer progression. Nat Cell Biol 23, 87–98 (2021)

[6] Dong, B. et al. Single-cell analysis supports a luminal-neuroendocrine transdifferentiation in human prostate cancer. Communications Biology 3, 778 (2020)

[7] Ma, X. et al. Identification of a distinct luminal subgroup diagnosing and stratifying early stage prostate cancer by tissue-based single-cell RNA sequencing. Molecular Cancer 19, 147 (2020)

[8] Song, H. et al. Single-cell analysis of human primary prostate cancer reveals the heterogeneity of tumor-associated epithelial cell states. Nat Commun 13, 141 (2022)

[9] Tuong, Z. K. et al. Resolving the immune landscape of human prostate at a single-cell level in health and cancer. Cell Reports 37, 110132 (2021)