The research presented in this thesis describes studies into the individual biological tumor properties of head and neck cancer, using messenger- and microRNA data to predict which tumors will be more radioresistant and to gain insight into the mechanisms behind this. Eventually this should lead to a better understanding of the causes for radiotherapy failure allowing an up-front adaptation of therapy to give each individual patient the best chance of survival with the least amount of toxicity (1, 2, 3).

Since 2002, after the publications of Van de Vijver en Van ‘t Veer et al. showing that pre-treatment gene expression can be used for the successful prediction of survival in breast cancer patients (4, 5), there has been a huge influx of papers trying to replicate these results for different tumor sites. Various authors reported the discovery of a gene expression profile to predict outcome in head and neck cancer (6, 7, 8, 9, 10, 11, 12). Most series were small and very heterogeneous in terms of patient characteristics and treatment regimens used. Often gene expression profiles were not validated on independent series, which is particularly important when prognostic genes are selected from a set of almost 20,000 genes, even when the correct statistical methods are applied. Additionally, the reported prognostic gene expression profiles were not tested in a model with clinical factors that were already known to be prognostic. In the worst case scenario, one of these gene expression profiles would be a very complicated method to tell the gender of a patient (as mentioned previously being male is prognostically unfavorable) and not at all useful.

Keeping this in mind, we first questioned whether gene expression would be able to add prognostic power to known clinical factors in head and neck cancer. In chapter 2, we show that gene expression (HPV-status and a profile published by Chung et al.) can improve the prediction model and adds valuable information to known clinical factors. However, this series was heterogeneous (different subsites, HPV positive and negative tumors) and chemotherapy was administered concomitantly with radiotherapy.

In order to find a true predictor of response to radiotherapy, the next step was to study a more homogeneous series of patients, preferably all treated with only radiotherapy. Since gene expression could at the time of sample collection only be done on fresh frozen material, these scarce samples were recruited from various Dutch hospitals to collect a matched series of small larynx cancers, described in chapter 3. With the analysis of this small, but homogeneous series, we preferred a hypothesis-driven approach (test gene sets for known biological processes), as opposed to a data-driven approach (test all ~20,000 genes) for two reasons. Firstly, this reduces the number of tests: 10 gene sets versus ~20,000 separate genes, making the statistics more robust. To illustrate this: using a p-value of 0.05 (which is of course not advised for the analysis of 20,000 genes) the chance of finding a false positive is 5%, meaning less than 1 out of 10 gene sets, but 1,000 false positives out of 20,000 tested genes would be found. Secondly, the hypothesis-driven approach will give results that are directly correlated to biological processes that could possibly be targeted to improve therapy. In this series we found cancer stem cell marker CD44 to be the only predictor of response to radiotherapy, which was validated on an independent series using immunohistochemistry (protein level). Since then many other authors have published this same finding, also in larger and non-laryngeal head and neck cancers (13, 14, 15, 16, 17, 18, 19).

A problem with the use of the hypothesis-driven approach is the acquisition of useful gene sets that correctly portray important biological processes. In neither of our patient series intrinsic radiosensitivity came up as a significant factor, while we know from clinical data that radiosensitivity measured by colony assays correlates with outcome after radiotherapy (20). We therefore concluded that we were not using an accurate messenger RNA set as a representative of this process and resolved to generate such a set. Another possibility was that messenger RNA levels alone were giving an incomplete picture of the active processes in the cell, since more factors can influence translation to protein. Among these are microRNAs, small pieces of RNA that can single handedly inhibit the translation of many messenger RNAs. The fact that it was reported that microRNA profiles were more accurate than messenger RNA profiles in the classification of poorly differentiated tumors (21), led us to hypothesize that they might also be more accurate in the prediction of intrinsic radiosensitivity.

Chapter 4 describes the discovery of a microRNA (miR-203), which downregulation strongly correlates with intrinsic radiosensitivity in cell lines and response to radiotherapy in a series of laryngeal cancer patients. The loss of miR-203 correlates with a biological process called epithelial to mesenchymal transition (EMT). The induction of EMT in cell lines is shown to decrease radiosensitivity.

Although a link between EMT and cancer stem cell marker CD44 has been described (22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34), we observed no correlation between CD44 expression and intrinsic radiosensitivity (chapter 3), nor a correlation between CD44 and miR-203 in 34 laryngeal cancer patients (unpublished results chapter 4, after acquisition of messenger RNA data for the same patients). This suggests that although there might be a link between EMT and cancer stem cells, not all cancer stem cells possess the same radiosensitivity and therefore both factors are independently important in the prediction of response to radiotherapy.