Harnessing sample preparation for RNA-sequencing toward a reliable bioinformatics analysis

Dear Editor,

We aim to clarify the issues raised by Tan et al. [1] regarding our recent publication “Cancer-associated fibroblasts (CAFs) gene signatures predict outcomes in breast and prostate tumor patients” [2], which provided changes in the transcriptomic milieu of these malignancies largely occurring in women and men, respectively [3]. We appreciated that the letter of Tan and colleagues dealt with our experimental model and data, giving us the opportunity to further detail the evidence shown.

The first issue was related to the patient’s characteristics and sample processing before RNA-sequencing (RNA-seq) analysis. Regarding the patient data, we selected 20 female patients with luminal invasive breast cancer characterized by estrogen receptor (ER)-positivity, human epidermal growth factor receptor 2 (HER2)-negativity and Ki67 ≥ 30% as well as 20 male patients with prostate cancer characterized by Gleason score at least of 8 or PSA > 20 mg/l. Concerning the sample processing method, we specified the experimental procedure in the “cell cultures” paragraph of the Methods section of our manuscript [2]. In particular, we stated that breast and prostate CAFs were isolated respectively from 20 mammary ductal carcinomas and 20 prostate adenocarcinomas. Then, a unique population of breast and prostate CAFs was obtained by pooling the 20 isolated cell cultures of each tumor type. From the single population of breast and prostate CAFs, we obtained three biological replicates, as previously recommended for the RNA-seq experiments [4]. Next, the RNA extraction from each biological replicate was performed by employing the same sample processing method.

As it concerns the grouping settings of our analysis, we aimed to identify the differences in the gene expression profiles occurring in CAFs of the two malignancies considered, similarly to previous studies that evaluated different types of tumors by RNA-seq analyses [5]. Therefore, comparing the differentially expressed genes in CAFs from breast and prostate cancer, we uncovered the unique molecular signatures and pathways that may trigger the action of CAFs in these malignancies. Our data may also help in pinpointing potential specific biomarkers and therapeutic targets of the breast and prostate tumor microenvironment, thereby supporting the improvement of personalized medicine and targeted strategies.

We esteemed the suggestion of Tan and colleagues [1] regarding the use of specific tools that may allow the identification of molecular and biological functions of the differentially expressed genes. However, we preferred the use of the enrichment analyses rather than to explore gene correlation patterns. Additionally, we clarify that our study did not compare tumor versus normal TCGA samples, as stated in the letter of Tan and colleagues. Instead, we specifically compared breast versus prostate samples of the TCGA dataset patients, thereafter we intersected the identified genes with those obtained in CAFs.

Next, we remark that the cumulative impact of the genes belonging to the identified signatures was robustly held by k-means analysis. Importantly, we demonstrated that a common prognosis characterizes patients clustered according to comparable gene expression patterns. The reliability of our analysis was further validated through a classification task, which provided attribute usage for each gene along with metrics for accuracy, recall and precision. The attribute usage percentage served as a quantitative measure of each gene’s influence in patient classification, thereby assigning varying degrees of importance to the different genes. This metric reflects the frequency by which each gene is employed as a decision criterion in our predictive models, thus showing its significance toward patient outcomes. We also appreciated the suggestion of Tan and colleagues regarding the time-dependent ROC analysis. Considering that our clustering approach inherently stratifies patients on the basis of gene expression patterns leading to distinct groups with built-in prognostic differences, we believe that a time-dependent ROC analysis is not suitable for our current workflow. Of note, we carried out our tests by (i) exploiting the k-cross validation, (ii) computing multiple evaluation metrics (accuracy, precision and recall) and (iii) showing the confusion matrices for the different cases. This methodology guarantees a statistically robust estimation of classification model performances as well as a complete explainability in terms of true positive, true negative, false positive and false negative rates. Overall, we recognize the value of the analyses proposed by Tan et al., therefore it would be useful to take into consideration their suggestions for subsequent studies.