AbstractPATIENTS AND METHODS
A total of 16 patients with invasive squamous cell carcinoma of the uterine cervix were included in this study. All patients were treated with standardized radiotherapy alone. Ten of the tumors were clinically radiosensitive and six were radioresistant. Total RNA, extracted from tumor specimens obtained prior to treatment, was hybridized onto an oligonucleotide microarray with probe sets complementary to over 20,000 transcripts. The genes were first subjected to a statistical filter to identify genes with statistically significant differential expression levels between those that were radiosensitive and those that were radioresistant. A back-propagation neural network was then constructed to model the differences so that patterns could be easily identified.RESULTS
Although a number of genes were found to express differentially between radiosensitive and radioresistant tumors; the 10 most discriminating genes were used to construct the model. Using the expressions from these 10 genes, we found that neural networks constructed from random subsets of the whole data were capable of predicting radiotherapy responses in the remaining subset, which appears stable within the dataset.DISCUSSION
This study shows that such an approach has the potential to differentiate tumor radiosensitivity, although confirmation of such a pattern using other larger independent datasets is necessary before firm conclusions can be drawn.