A reliable transcriptomic risk-score applicable to formalin-fixed paraffin-embedded biopsies improves outcome prediction in localized prostate cancer

Abstract

Background: Clinical manifestation of prostate cancer (PCa) is highly variable. Aggressive tumors require radical treatment while clinically non-significant ones may be suitable for active surveillance. We previously developed the prognostic ProstaTrend RNA signature based on transcriptome-wide microarray and RNA-sequencing (RNA-Seq) analyses, primarily of prostatectomy specimens. An RNA-Seq study of formalin-fixed paraffin-embedded (FFPE) tumor biopsies has now allowed us to use this test as a basis for the development of a novel test that is applicable to FFPE biopsies as a tool for early routine PCa diagnostics.

Methods: All patients of the FFPE biopsy cohort were treated by radical prostatectomy and median follow-up for biochemical recurrence (BCR) was 9 years. Based on the transcriptome data of 176 FFPE biopsies, we filtered ProstaTrend for genes susceptible to FFPE-associated degradation via regression analysis. ProstaTrend was additionally restricted to genes with concordant prognostic effects in the RNA-Seq TCGA prostate adenocarcinoma (PRAD) cohort to ensure robust and broad applicability. The prognostic relevance of the refined Transcriptomic Risk Score (TRS) was analyzed by Kaplan-Meier curves and Cox-regression models in our FFPE-biopsy cohort and 9 other public datasets from PCa patients with BCR as primary endpoint. In addition, we developed a prostate single-cell atlas of 41 PCa patients from 5 publicly available studies to analyze gene expression of ProstaTrend genes in different cell compartments.

Results: Validation of the TRS using the original ProstaTrend signature in the cohort of FFPE biopsies revealed a relevant impact of FFPE-associated degradation on gene expression and consequently no significant association with prognosis (Cox-regression, p-value > 0.05) in FFPE tissue. However, the TRS based on the new version of the ProstaTrend-ffpe signature, which included 204 genes (of originally 1396 genes), was significantly associated with BCR in the FFPE biopsy cohort (Cox-regression p-value < 0.001) and retained prognostic relevance when adjusted for Gleason Grade Groups. We confirmed a significant association with BCR in 9 independent cohorts including 1109 patients. Comparison of the prognostic performance of the TRS with 17 other prognostically relevant PCa panels revealed that ProstaTrend-ffpe was among the best-ranked panels. We generated a PCa cell atlas to associate ProstaTrend genes with cell lineages or cell types. Tumor-specific luminal cells have a significantly higher TRS than normal luminal cells in all analyzed datasets. In addition, TRS of epithelial and luminal cells was correlated with increased Gleason score in 3 studies.

Conclusions: We developed a prognostic gene-expression signature for PCa that can be applied to FFPE biopsies and may be suitable to support clinical decision-making.

Keywords: Molecular diagnostic testing; Molecular pathology; Personalized medicine; Prognostic biomarker; Prostate cancer; Transcriptome.

Fig. 1

Analysis workflow. A Overview of the included tissue samples we used to develop the ProstaTrend-ffpe signature. Shown are the number of patients included in the study and reasons for exclusion. For 185 patients, we performed a strand-specific transcriptome-wide sequencing from FFPE biopsy tissue. A total of six samples did not meet the quality criteria for RNA-Seq data. In addition, we excluded three samples due to missing clinical follow-up data. The final cohort included 176 patients, for 75 of whom BCR was observed within the follow-up time. FFPE formalin-fixed paraffin-embedded, PCa prostate cancer, DoD death of disease, BCR biochemical recurrence. B For the development of the PCa single-cell atlas, we used scRNA-Seq data of PCa patients from 5 publicly available studies (Chen et al. ; Dong et al. ; Ma et al. ; Song et al. ; Tuong et al. 2021). Spatial transcriptomics data of a human PCa biopsy (GS = 3 + 4) were downloaded from the 10× Genomics database. C The prognostic value of the Transcriptomic Risk Scores (TRS) using ProstaTrend(-ffpe) was evaluated by survival analyses in 9 publicly available cohorts (Li et al. ; Fraser et al. ; Luca et al. ; Long et al. ; Gerhauser et al. ; Jain et al. ; Ross-Adams et al. ; Taylor et al. 2010) and a meta-analysis with a total of 13 cohorts

Fig. 2

Prognostic value of ProstaTrend(-ffpe) in the FFPE biopsy cohort FFPE_Bx. We assessed the prognostic value of the ProstaTrend (A, C) and ProstaTrend-ffpe TRS (B, D) signature by Kaplan–Meier analysis and Cox proportional hazard regression (BCR as the primary endpoint). A, B Kaplan–Meier curves for patients with TRS > 0 (increased risk) compared to patients with TRS ≤ 0 (reduced risk). Color shades depict the 95% confidence intervals for Kaplan–Meier curves. The curves were truncated if the number of patients at risk dropped below 10 in both groups. The colored numbers above the x-axis indicate the number of patients at risk. Log-rank tests were performed to evaluate probabilities of BCR-free survival between these two groups. The numbers under the log-rank p-values indicate the number of patients and cases with BCR. C, D Univariate Cox-regression analysis for TRS on a continuous scale (top) and multivariable Cox-regression for TRS adjusted for Gleason grading group (GGG > 2) of the biopsies (bottom). logHR log hazard ratio, HR hazard ratio, CI confidence interval

Fig. 3

Analysis of ProstaTrend genes in the developed PCa single-cell atlas. A We embedded about 90,000 cells of 41 PCa patients from 5 publicly available datasets (Chen et al. ; Dong et al. ; Ma et al. ; Song et al. ; Tuong et al. 2021) into a two-dimensional space by the t-distributed stochastic neighbor embedding (tSNE) method. Each dot represents a single cell. Cells are colored according to cell identity. B We applied a simplified TRS to each cell from the tumor samples using the ProstaTrend and ProstaTrend-ffpe signatures. Each area containing cells on the tSNE was divided into hexagonal bins, and cells within each bin were averaged. The bins are colored according to TRS. C For tumor samples, TRS were grouped based on the ProstaTrend(-ffpe) signatures and colored according to luminal and tumor-specific luminal (T-luminal) cells (****p-value < 0.0001, Wilcoxon rank-sum test). PT ProstaTrend, PT-ffpe ProstaTrend-ffpe. The y-axis depicts the TRS. D UpSet plot of ProstaTrend(-ffpe) DE genes (FDR < 0.05) of four cell lineages. The squares in the matrix represent unique or overlapping DE genes for the cell lineages. The stacked bar graph above the matrix summarizes the number of ProstaTrend(-ffpe) DE genes for each unique lineage. The top stacked bar plot shows the fraction of DE genes with log hazard ratio > 0 or < 0. E Heat map of the highest ranked DE genes for the ProstaTrend(-ffpe) genes. Genes are ranked by log2 fold change. 15 genes (if present) are depicted for each cell lineage. Each column depicts the average expression value for one patient, grouped by cell lineage and tissue source. Average gene expression values are standardized. F UpSet plot of ProstaTrend(-ffpe) DE genes for cell types. G Pathological annotations of a human prostate stage III adenocarcinoma biopsy, which was FFPE preserved and processed using the Visium spatial gene expression for FFPE workflow. H Standardized overall enrichment (OE) scores of cell type markers for each spot were estimated using the AddModuleScore function implemented in Seurat. I OE of DE genes (T-luminal vs. luminal cells from the PCa Cell Atlas) for each spot. J TRS was applied to gene expression spots using the ProstaTrend(-ffpe) signatures

Fig. 4

Validation of the ProstaTrend and ProstaTrend-ffpe signature in 9 publicly available PCa cohorts. A For the ProstaTrend-ffpe signature, Kaplan–Meier curves for patients with TRS > 0 (increased risk) compared to patients with TRS ≤ 0 (reduced risk) are shown. The numbers under the cohort IDs indicate the number of patients and cases with BCR. Color shades depict the 95% CI for Kaplan–Meier curves. The curves were truncated if the number of patients at risk dropped below 10 in both groups. The colored numbers above the x-axis indicate the number of patients at risk. Log-rank tests were performed to evaluate probabilities of BCR-free survival between the two groups. The numbers in the plot (above the log-rank p-values) indicate how many ProstaTrend-ffpe genes are available in the datasets. B Forest plot of the overall logHRs and corresponding 95% confidence intervals (95% CI) estimated by Cox-regression on a standardized continuous scale. Significant logHRs with a p-value < 0.05 are highlighted in green. C We generated 1000 random gene sets each for ProstaTrend and the ProstaTrend-ffpe signature, followed by Cox-regression analysis for each cohort (see “Methods”). The estimated −log10(p-value) for the random gene sets are shown as boxplots. The p-values for ProstaTrend(-ffpe) gene sets are shown as colored dots (same p-values as in B). D The x-axis depicts the combined effect size of logHRs for genes from a univariate random-effects meta-analysis approach of the ProstaTrend training cohorts. The same meta-analysis approach was performed for the 9 validation cohorts (y-axis). Genes whose logHRs from the meta-analysis of the training cohort showed no significance (FDR ≥ 0.05) are colored in gray. The logHRs of the ProstaTrend(-ffpe) genes are colored accordingly. E The dots depict the -log10(p-values) estimated from a Cox-regression model with a two-sided Wald test. Prognostic signatures that were among the top 5 (sorted by p-value) in more than 3 cohorts were highlighted by a label. The white numbers represent the rank by p-value. The dashed vertical line indicates a p-value of 0.05. F The average rank of each prognostic gene set across all cohorts from the log-rank and Cox-regression analysis

Fig. 5

Results of the univariate random-effect meta-analysis using Cox proportional hazard models. A We estimated a combined effect size of logHRs for each gene across 13 cohorts. Shown are the combined effect sizes of all prognostic ProstaTrend genes (n = 1376). The x-axis represents the combined effect sizes and the y-axis the adjusted p-values (Benjamini–Hochberg correction for multiple testing) in log-space. The dashed horizontal line indicates an FDR of 0.05. For the highest ranked genes (by adjusted p-value) from the meta-analysis, 5 gene labels each with a combined effect size > 0 and < 0 are shown. The number in parentheses indicates the number of cohorts in which the respective gene was available. The numbers at the top of the plot indicate the number of significant ProstaTrend(-ffpe) genes with a combined effect size > 0 and < 0. B The y-axis shows the number of cohorts in which the ProstaTrend(-ffpe) genes had a significant combined effect size and were available. The x-axis indicates in how many cohorts the significant genes have a consistent direction of the logHRs. The labels indicate the number of genes in each case. C Scatter plot of all prognostic ProstaTrend genes (n = 1376). As in A, the x-axis represents the combined effect size of the logHRs. The y-axis represents the combined effect size of the log odds ratios derived from the logistic regression model predicting GS > 7 vs. ≤ 7. The meta-analysis for the log odds ratio was performed in the same way as for the logHRs. The colored dots indicate whether the combined effect sizes were significant (FDR < 0.05) in both meta-analyses approaches, in one meta-analysis or in none. The correlation coefficient is displayed in the upper right corner of the plot. D Forest plots for the 5 highest ranked genes by adjusted p-value. The logHR for each cohort is represented by a square. A horizontal line depicts the confidence interval (CI). The size of the square corresponds to the weight of the cohort in the meta-analysis. The combined effect and the CI are represented by a diamond. ProstaTrend-ffpe genes are highlighted with green colored horizontal bars; the blue ones are part of the original ProstaTrend signature