Development of Gene Expression Signatures for Radiation Exposure Prediction using Cross-Platform Microarray Data

• Main Authors: Yi-Yao Hsu, 許逸堯
• Other Authors: Eric Y. Chuang
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/70722013808234904592

碩士 === 國立臺灣大學 === 生醫電子與資訊學研究所 === 101 === Exposure of ionizing radiation (IR) can cause DNA damages in cells due to DNA double-strand break. It is well known that exposure to IR could lead to the increased risks of cancer and other acute radiation sickness, which may eventually cause rapid death in high exposure dose. In a radiological emergency, such as Fukushima Daiichi nuclear disaster in 2011, exactly evaluating the exposure dose would be highly required for better health care in the early treatment. In recent years, several studies had suggested some gene expression signatures associated with radiation doses, which can serve as effective predictors for the exposed radiation doses by using high-throughput microarray technique. However, most of them only tested the performances in their original studies and no external datasets were further used for validation. These observations clearly revealed that results based on one microarray dataset was prone to suffer from systematic biases, and thus was limited in biological interpretation. On the other hand, most of the studies that only focused on a single cell line typically had high false positive rates since the transcriptional responses to IR were cell line-dependent. To overcome these issues, we proposed a comprehensive approach by performing a meta-analysis of publicly available microarray datasets. First, differentially expressed genes in responding to different radiation doses were identified in the respective dataset by using statistical models. Next, the identified genes that selected from differential expression analyses were further filtering by the frequency in which they are present among different datasets. A machine learning algorithm, support vector machine, was utilized to develop prediction models distinguishing high-dose and low-dose exposure. Lastly, the performances of the meta-signature was evaluated by using cross-validation in internal datasets and independently validated in external studies. In this study, a 29-gene meta-signature was identified that can distinguish cells with either high-dose exposure (> 8 Gy) or low-dose exposure (< 2 Gy). The gene functions related to the apoptosis regulating pathway and the p53 signaling pathway were significantly enriched in the identified meta-signature. Compared to the previous studies, our findings had superior performance in predicting IR exposure levels with a total accuracy of 85% (6–14% higher than previous studies) in the internal cross-validation. Even in an external validation, this meta-signature correctly predicted the samples with an overall accuracy of 84% (81/94). In conclusion, we provide a comprehensive approach to dissect the radiation responses among different doses in the independent studies. The findings may facilitate exploration in biological functions regulated by IR. Above all, the results shown improvement in the robustness of prediction and may be applied into practical usage.