Pharmacogenomic Modeling in Pancreatic Cancer—Letter

Yu and colleagues (1) assessed the pharmacogenomics profiling (PGx) of circulating tumorigenic and invasive cells to stratify/predict treatment response in pancreatic ductal adenocarcinoma (PDAC). They suggested the PGx assay as an appropriate tool for choosing a potentially effective targeted agent and first-line therapy for PDAC patients. Nevertheless, in our opinion, some aspects of the study need to be discussed in more detail.

Clinical research in the field of PDAC is of paramount importance because it remains a major unsolved health problem. The treatment option is still controversial, and upfront determinants for effective therapeutic approaches are greatly needed. In the current study, a total of 50 participants were initially enrolled (1), but merely 35 patients were ultimately deemed eligible for final evaluations. For their treatments, the patients were divided into five groups with various chemotherapeutic regimens, i.e., 5-FU–irinotecan–oxaliplatin; 5-fluorouracil–oxaliplatin; gemcitabine–oxaliplatin; gemcitabine–capecitabine, and gemcitabine–docetaxel–capecitabine. The first point is treatment heterogeneity, low number of patients in treatment groups, nonrandomized selection of patients for a specific regimen as to the physicians discretion, and lack of the validation of emerging data in an independent cohort, preferably multicenter settings, which justifies the need for the confirmation of their findings and demonstrates the reproducibility and sensitivity of the model system.

Second, the statistical methods used for the data analysis seem not clearly described, whether Yu and colleagues (1) performed multivariate analyses, and whether they adjusted for different variables with already known prognostic value. This is particularly the case for “age” that was revealed to be significantly different in resistant groups. Moreover, the associations of 11 variables, e.g., treatment, performance status, location/number of disease sites, stage, and CA19.9, with overall survival/progression-free survival were compared between groups; however, according to the “rule of tens,” of logistic regression and Cox models, there should be a minimum of 10 events per variable/condition, to prevent overfitting and gain enough confidence in reported findings (2, 3). Thus, for 35 patients analyzed, only three variables could have reliably been fitted. Furthermore, it is beneficial to know whether those who performed data analysis/interpretation were also blinded to PGx prediction.

Finally, the authors (1) tested the central hypothesis using a similar PGx model for paclitaxel, oxaliplatin, irinotecan, 5-fluorouracil, erlotinib in the NCI60 cell line. However, it would be worthwhile to perform this experiment in primary cell cultures or freshly generated xenograft from tumor tissues, because these models are better candidates for mimicking the genetic characteristics of disease and might be better predictors of drug activity (4, 5).

In spite of the fact that the findings by Yu and colleagues (1) are of utmost importance, additional analysis with incorporation of proper methodology, larger/uniformly treated populations according to powered statistical analysis, and integration with functional data with appropriate confirmation may help to obtain solid evidence to introduce an authentic predictor for treatment response in clinical practice.

See the Response, p. 1498