Molecular sampling of prostate cancer: a dilemma for predicting disease progression

Abstract

Background: Current prostate cancer prognostic models are based on pre-treatment prostate specific antigen (PSA) levels, biopsy Gleason score, and clinical staging but in practice are inadequate to accurately predict disease progression. Hence, we sought to develop a molecular panel for prostate cancer progression by reasoning that molecular profiles might further improve current clinical models.

Methods: We analyzed a Swedish Watchful Waiting cohort with up to 30 years of clinical follow up using a novel method for gene expression profiling. This cDNA-mediated annealing, selection, ligation, and extension (DASL) method enabled the use of formalin-fixed paraffin-embedded transurethral resection of prostate (TURP) samples taken at the time of the initial diagnosis. We determined the expression profiles of 6100 genes for 281 men divided in two extreme groups: men who died of prostate cancer and men who survived more than 10 years without metastases (lethals and indolents, respectively). Several statistical and machine learning models using clinical and molecular features were evaluated for their ability to distinguish lethal from indolent cases.

Results: Surprisingly, none of the predictive models using molecular profiles significantly improved over models using clinical variables only. Additional computational analysis confirmed that molecular heterogeneity within both the lethal and indolent classes is widespread in prostate cancer as compared to other types of tumors.

Conclusions: The determination of the molecularly dominant tumor nodule may be limited by sampling at time of initial diagnosis, may not be present at time of initial diagnosis, or may occur as the disease progresses making the development of molecular biomarkers for prostate cancer progression challenging.

Figure 1

Study design. From 1256 men of a Watchful Waiting Cohort, we selected the "Extreme" cases: those who died of prostate cancer or men who lived more than 10 years without signs of progression. We also filtered out some patients based on tumor tissue availability, sample quality or because they were treated. Finally, we randomly divided the patients in a Learning and Validation sets, ensuring that similar proportions of lethals and indolents are present in the two groups.

Figure 2

Schematic of silhouette widths, i.e. homogeneity scores, and silhouette plots. A. (left) Given an element in a group (the orange cross surrounded by a diamond) the distances from elements in the same group (magenta lines) and from those in the other group (green lines) are computed. The homogeneity score can be viewed as the difference between the averages of the inter-group distance (green) and the intra-group distance (magenta). (right) The homogeneity score of each sample is plotted on a horizontal bar, after sorting the samples within each group. The average of the homogeneity scores is computed for each group yielding an estimation of the homogeneity of the cluster. B. Four different categories of homogeneity (left) and the corresponding silhouette plots (right) are depicted. Specifically: Scenario 1. two homogeneous and well-separated groups; Scenario 2. one homogeneous and one heterogeneous group, well-separated; Scenario 3. one homogeneous and one heterogeneous group, overlapping; Scenario 4. two heterogeneous overlapping groups. The empirical interpretation of the average homogeneity score for a group is shown at the bottom.

Figure 3

Supervised analysis. A. Results of logistic regression on the Validation dataset. On top are reported the AUCs of the models, whereas on the bottom the parameters that are used in the corresponding model are shown. A colored square means that the parameter was used in the model, whereas a white square means that the parameter was not used. The last row reports the number of genes that were used by the model, if any. Models including clinical and molecular parameters are reported only if they improved on the corresponding models using clinical parameters only. Models are sorted from left to right according to their AUC. We estimated the Confidence intervals (CIs) for models including genes using the sampling distribution of AUCs generated by the iterative cross-validation procedure on the Learning set. For the other models, a bootstrap estimation of CIs was computed on the Validation set. The genes that are involved in the models are reported in Additional file 1, Table S4. B. Contingency table showing ERG rearrangement status association with clinical outcome. In parenthesis the expected numbers of cases if no association is assumed.

Figure 4

Homogeneity analysis. A. Silhouette plot for Burkitt's lymphoma (left) and prostate cancer (right). The numbers report the average homogeneity score for each group. B. Average homogeneity score for different cancer data sets.