A well thought-out research plan and the accrual of reliable data is of the greatest importance for the generation of relevant, replicable results. Many factors should be taken into consideration when studying response to radiotherapy on pre-treatment tumor material.
6.3.1 The pre-treatment sample
Heterogeneity and tumor percentage
Different parts of a tumor could consist of cells with different genetic characteristics and radiosensitivity that are not being detected when only sampling a small part of the tumor (89, 90). In our studies we have used conclusions from biopsies of several millimeters as a surrogate for a tumor of several centimeters. Had tumor heterogeneity been an enormous problem, we would not have been able to use pre-treatment biopsies for outcome prediction at all. However, it seems reasonable to assume that part of the information on the whole tumor is lost with this approach. Toustrup et al. tested how much information gets lost due to head and neck cancer heterogeneity by studying hypoxia gene expression in multiple (2-4) samples from 20 tumors (91). They showed that in 70% of the tumors all replicate samples were awarded the same hypoxia score. However, when only samples with the highest percentage of tumor cells were selected, only 10% of patients would have wrongfully been classified as having less hypoxia. This is another difficulty with tumor biopsies: it will mostly consist of both tumor cells and stroma, different percentages of these two in a studied biopsy might lead to different results. Roepman et al. conclude that there was a poor signature performance for a head-neck expression signature that predicts the presence of lymph node metastasis on samples that contain less than 50% tumor cells (92).
Monitor during treatment?
It is plausible that biology changes during treatment. Still, it appears that we are fairly capable of predicting the response to radiotherapy on a pre-treatment sample, for example chapters 2 and 3, ref. (20, 93) and many others. As shown in chapter 4, not the changes in gene expression after irradiation, but the baseline microRNA levels in unirradiated cells correlated with radiosensitivity. Similarly, we know that fast repopulation of tumors only starts around the fifth week of radiotherapy (94), but benefit from accelerated radiotherapy can be predicted on a pre-treatment sample (83, 84). However, we might miss some biological changes during treatment that would be useful to improve treatment by adaptation during therapy. A study taking multiple biopsies during treatment is hard to conduct and not very patient-friendly. Imaging modalities like MRI or PET are more convenient to study biology during treatment and possibly adapt treatment for non-responders (95, 96), although the monitoring of multiple biological processes will be far more challenging. There have been some reports suggesting that a change in certain PET tracers early during a course of radiotherapy better predicts treatment outcome than only pre-treatment uptake values (97, 98, 99). However, the opposite has been reported as well (100). Another possibility would be to monitor biomarkers in saliva or blood (101, 102, 103, 104).
6.3.2 Just (messenger) RNA?
The studies in chapter 2 and 3 have used just messenger RNA to study the active biological processes in a tumor. As mentioned in the introduction, just messenger RNA might not entirely depict what happens in a cell. Therefore, microRNAs were integrated in the analysis in chapter 4, and one of them was shown to be the most useful predictor of radiosensitivity. Perhaps this is a result of the absence of a correct messenger RNA set for the same process, or the fact that there is less degradation of microRNAs during sample-handling, but could just as well result from the fact that messenger RNA alone is not enough, as has been shown by Jung et al. By combining data on methylation, DNA copy number, messenger RNA and microRNA they were able to better select patients at risk for metastases that with any of those methods alone (105). Another possibility is that we lose information because of the complicated statistics involved in the analysis of gene expression data. For example, before the final analysis, all samples in chapter 4 were normalized using the assumption that the total amount of microRNAs is the same in every sample, while there is evidence that levels of microRNA differ between samples (62, 106).
Ideally, all possible pre-treatment information for a large group of patients would be collected (DNA methylation, DNA and RNA sequencing, protein levels and their phosphorylation status, different CT/MRI/PET scans, blood and saliva parameters) to filter out the most useful biomarkers for different therapeutic approaches (107). But even with all this information, it remains crucial to know which markers reliably represent certain processes and how we can target these processes to improve radiotherapy.
6.3.3 Need for adequate biomarkers of processes
Critics of gene expression profiles argue that many gene sets are not ready for clinical use because of the large differences between reported sets in literature. Results are not reproducible and therefore not deemed useful (1). According to our data, this is partly based on the misconception that different sets of genes per definition classify patients differently. In chapter 5 we show that for hypoxia different sets of genes have been reported, with almost no overlapping genes. However, almost entirely different sets of genes can come to the same conclusion. While this is true for hypoxia, there are other processes that are still lacking reliable methods to assess the absence or presence of a factor causing radioresistance. Another problem illustrated in chapter 5, is that while it was assumed by most authors that they were studying both acute and chronic hypoxia, the gene sets only corresponded with an in vitro chronic hypoxia profile, which has a different supposed origin (lack of perfusion and not diffusion) and could have consequences for the appropriate therapeutic intervention.