Response to neoadjuvant therapy is heterogeneous in LARC patients (2, 4). Currently, neoadjuvant therapy is indicated for all LARC patients, even though a significant subset of patients is therapy resistant. Adequate prediction of the response to neoadjuvant treatment might improve personalized treatment, and avoid unnecessary waiting periods and therapy related toxicity in non-responding patients. In addition, it might open up new opportunities to apply neoadjuvant therapy in a selected subset of well responding early stage tumors (without an indication for neoadjuvant therapy according to current guidelines). To ensure clinical translation of this response prediction, the use of readily available clinical, radiological and/or pathological parameters is desired. Genomic mutational status can be assessed on pre-operative biopsies which are routinely obtained in the diagnostic work up.
As this and previous research indicates, the use of a single predictive biomarker often underestimates the complex mechanisms involved in response prediction. In this study, genomic mutations in pre-operative biopsies were compared to 4 other locations within the same tumor using next generations sequencing. In 36% of the patients, evaluation of genomic mutational status on a single pre-operative biopsy has shown to be inadequate. This illustrates the genomic variability in rectal cancer and explains the difficulties in obtaining reliable biomarkers for response prediction. These results are in line with previous evidence supporting the presence of intra-tumoral genomic heterogeneity in a considerable proportion of rectal cancers (30). Three previous studies have compared genomic mutations in up to 3 intra tumoral locations. Hardiman et al. reported up to 10 coding variants uniquely corresponding to one of 3 of the tumor locations in their study of 6 patients (30). In the study of Bettoni et al., only 27% of the observed mutations corresponded to all three samples of a single rectal adenocarcinoma in one patient (31). On the other hand, Dijkstra et al. reported no differences in mutational status between deep and superficial colorectal cancer tissue in 30 patients (32).
This study has several limitations. First of all, the small sample size and limited targeted next generation sequencing panel might influence the interpretation of results. The number of disconcordant cases might actually be higher, as this targeted gene panel only provides information on a selected number of mutations. Furthermore, there is no 100% certainty the found mutations were not germ-line mutations, however considering the observed allelic frequency this is very unlikely. Furthermore, since the main aim of this study was to determine the reliability of response prediction based on a single preoperative biopsy rather than define the degree of intra-tumoral genomic heterogeneity, the influence on our results would be limited. To increase the reliability of the biopsy analysis, the use of multiple and possibly even deeper/larger preoperative biopsies might provide a better representation of intra-tumoral heterogeneity, but might also increase the risk of procedure related complications. A second possibility might be the application of whole exome sequencing or larger targeted gene panels (such as the TSO500, Illumina, San Diego, USA), as this possibly provides a more elaborate analysis of genomic mutations, as compared to next generation sequencing using a limited targeted gene panel. Using these techniques, the mutant-allele heterogeneity (MATH) score was developed to quantitatively assess the spread of allele frequencies, and has been correlated to response (19, 33). However, as sampling errors are innate to the biopsy technique, parameters derived from full tumor imaging might be preferable to incorporate characteristics of all genetic sub clones present in these cancers.
Although the use of a single predictive parameter may be the most clinically usable method, it seems to underestimate the complexity of cancer biology. Previous research on the application of other predictive parameters (e.g. MRI, PET, radiomics feature analysis, metabolomics (34–37)) has up until now not resulted in the routine use of such predictive parameters, as these do not provide sufficient accuracy yet (13).
To overcome this challenge and address these complexities, data should probably be combined into a response predictive algorithm. The first steps towards identifying reliable parameters to be included in such models have already been made by Santos et al. in 2017, using multivariable analysis of both blood and tissue biomarkers (38). In such models, features derived from clinical, radiological, pathological and more advanced analysis such as radiomics and metabolomics can be incorporated. The downside of such studies is the need of large datasets acquired under a variety of conditions, as to be applicable in a myriad of clinical scenarios.