Deep learning-based algorithm for analysis of next-generation genomic data in cancer

• Main Authors: MOI SIN HUA, 魏芯樺
• Other Authors: YANG, CHENG-HONG
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/3w3ke6

博士 === 國立高雄科技大學 === 電子工程系 === 107 === Cancer genomic are extremely complex, and the process is usually regulated by different types of related genes; thus, studying the cancer mutations is challenging. Recently, a major breakthrough in genetic detection technology revealed that genetic mutations are significantly associated with various diseases and cancers. This breakthrough facilitates precision medicine as a new treatment method in clinical settings. Precision medicine aims to guide the most appropriate treatment for patients according to the results of genetic detection, thereby achieving early effective treatment and ensuring efficient medical resource use. In precision medicine, identifying the association rule of genetic mutations, detecting the high-risk oncogene set, and predicting the prognosis risk of genetic mutations are crucial. Clinical evidence on cancer gene mutations is insufficient and lead the application of gene analysis might be limited due to the high proportion of Type I errors (false positive rate). This thesis proposed a fuzzy improved deep learning (DL)-based algorithm to identify high-probability missense mutation variants and candidate genes for cancer mortality or cancer immunologic signature by using next-generation genomic data in cancer. Abstracted weights were obtained from a network of deep-learning, odds ratios and hazard ratios were calculated by regression model. The model-based risk scores were yield by employing the matrix operation to combine the abstracted weights with the odds ratios and hazard ratio. The fuzzy improved deep learning-based model were generate based on the model-based risk score by using the fuzzy logic system. The cancer genome atlas (TCGA) open assess genomic datasets (BRCA: breast invasive carcinoma and HNSC: head and neck squamous cell carcinoma) were used to evaluate the cancer mortality-associated missense mutation variants and candidate genes, and identified the significant canonical pathways using gene ontology (GO) and gene set enrichment analysis (GSEA). The results indicate all of the cell lines have a common significant pathway, caveolar-mediated endocytosis, which is playing role in the innate immune mechanism. The fuzzy improved deep learning-based algorithm can improve stratification for binary outcome through consideration of comprehensive or classification effects among multiple features, which could exhibit more comprehensive classified weight estimation for binary outcome. We found cancer mortality is associated with three pathways in TCGA-BRCA (Epithelial–mesenchymal transition, hepatic stellate cell activation, and caveolar-mediated endocytosis) and three pathways in TCGA-HNSC (IL-17A signaling, IL-8 signaling and the ERK5 signaling pathway). Furthermore, we used the oral squamous cell carcinoma (OSCC) cell line genomic datasets (OC4, OC5 and OECM-1) to identify the highly immunologic associated missense mutation variants and candidate gene. The fuzzy improved DL-based fuzzy model could achieved the more precise distinguishability for missense mutation variants and candidate genes indicating a high risk of mortality and immunologic associated signature using genomic data in cancer, which might offer new targets for anticancer therapy via immunologic or anti-angiogenic mechanisms, in order to accomplish precision medicine goals.