Discovering genetic factors associated with human diseases through applying various computational approaches on genome-wide association study

Ph.D.Through high-throughput microarray technologies, biologists can perform data-driven researches like genome-wide association study (GWAS) and examine millions of genomic variations over a large population in a reasonable time and cost. Although many disease-associated single nucleotide polymorphisms (SNPs) have been discovered through traditional statistical tests, it has been shown that non-linear SNP-SNP interactions may also be associated with genetic diseases. In this thesis, we aim at discovering high-risk genetic factors for genetic diseases through exhaustively searching for statistically significant SNP-SNP interactions from different GWAS datasets.First, we performed two case studies on psoriasis and schizophrenia. Driven by eight dominance-dominance and recessive-recessive SNP-SNP interaction patterns [1, 2], exhaustive searches for statistically significant second-order SNP-SNP interactions were performed on two psoriasis and three schizophrenia GWAS datasets requested from the database of Genotypes and Phenotypes. Our analysis of high-ranking statistically significant SNP-SNP interactions revealed 27 potential psoriasis-associated SNPs and 11 potential schizophrenia-associated SNPs where most of them were associated with known psoriasis-related genes and schizophrenia-related genes, respectively. Furthermore, we inferred a gene network for psoriasis and discovered a novel long-distance interaction between XXbac- BPG154L12.4 and RNU6-283P, which has not been previously reported. Finally, we discovered a replicable SNP-SNP interaction among three schizophrenia datasets, which could be interpreted as a specific interaction between FHIT (protein-coding) and LINC00969 (long non-coding RNA), and both of these two genes were reported to have expression in brain.After that, we developed a genetic programming algorithm called GPPi 2.0 for learning classifiers with high human interpretability. GP-Pi 2.0 has a penalizer to increase diversity and an SNP initializer to prioritize high potential disease-associated SNPs. Moreover, it has a novel terminal set driven by the eight SNP-SNP dominance-dominance and recessive-recessive interaction patterns. New hypotheses of high order disease-associated SNP-SNP interactions can be constructed after the well-trained classifiers are interpreted by experts. Furthermore, we explored the potential of utilizing convolutional neural network (CNN) to learn classifiers mainly for predicting the health status of any given sample. Experiments with synthetic datasets showed that GP-Pi 2.0 outperformed GPAS, J48, and its predecessor GP-Pi in most cases, especially under a large number of SNPs or samples. Furthermore, we tested GP-Pi 2.0 and CNN on a rheumatoid arthritis dataset downloaded from the North American Rheumatoid Arthritis Consortium. The best classification tree built by GP-Pi 2.0 was consistent with existing literature and potentially discovered some new disease-associated SNPs. Meanwhile, we showed that CNN could accurately predict the health status of samples in both testing and unseen data. A novel wrapper approach for feature selection may be developed based on our CNN in the future to prioritize a group of interacting SNPs.Finally, we accelerated our exhaustive search algorithm with the following two acceleration techniques. First, we accelerated the computation of contingency tables of SNP pairs through GPU parallelization. Second, we explored the potential of pruning similar SNP pairs through approximating the similarity between SNPs in each SNP combination with locality sensitive hashing (LSH) algorithm. Proof-of-concept experiments on these two techniques were performed with synthetic datasets, and these two techniques were shown to be potentially effective in accelerating the exhaustive search algorithm.通過應用像DNA 微陣列這樣的高通量定序技術，生物學家可以收集一大群樣本，並在合理的成本和時間下測量每一個樣本百萬個基因體序列變異後進行像全基因組關聯分析（GWAS）等數據導向型研究。儘管歷年來通過傳統統計學已經發現了有很多和遺傳疾病有關聯性的單核苷酸多態性（SNP)，現在卻有硏究表明SNP 之間的非線性相互作用也能和遺傳疾病也可能有關聯性。在這一份畢業論文中，我的研究方向是通過把SNP 和SNP 之間的相互作用找出來進行測試，並把那些和遺傳疾病有統計學上明顯有關聯性的相互作用找出來，藉此找出能引起遺傳疾病的高危遺傳因子。我對銀屑病和思覺失調症這兩個病進行了研究. 通過應用基於「顯性-顯性」和「隱性-隱性」基因組的相互作用所提出來的SNP 和SNP之間的八個二元相互作用模式，我對從基因型和表型數據庫(dbGAP)下載回來的兩個銀屑病和三個思覺失調症的GWAS 資料集進行了SNP和SNP 之間的所有二元相互作用的統計學測試，藉此找出那些和遺傳疾病有統計學上明顯有關聯性的相互作用。在我的分析下，我發現了27 個和銀屑病可能有關的SNP。同時我也發現了11 個和思覺失調症可能有關的SNP。然後, 我通過建立一個銀屑病有關的基因網絡，發現了XXbac-BPG154L12.4 和RNU6-283P 這兩個基因的相互作用。最後，我找到一個SNP 之間二元相互作用，不但能被三個思覺失調症的GWAS資料集證實，同時能被解讀成FHIT 和LINC00969 這兩個在人腦中有表達的基因之間的相互作用。我也開發了一個名叫GP-Pi 2.0 的遺傳編程算法，並希望透過建立一個能有高度人類可解釋性的模型，能讓専家發現並對SNP 之間潛在的更高元相互作用建立假說。GP-Pi 2.0 內建一個能提升題解多樣性的懲罰機制，也有一個能挑選有潛力和遺傳疾病有關的SNP 的題解生成器。GP-Pi 2.0 的題解樹末端節點集合中的部分節點也有考慮到SNP 和SNP 之間的八個二元相互作用模式。我也嘗試用了一個簡單的卷積神經網（CNN）來去建立一個能預估任何一個樣本的健康狀態的模型。通過人工模擬資料集，我證明了GP-Pi 2.0 比GPAS, J48 和前身GP-Pi 的表現更好. 當人工模擬資料集的SNP 數量比較大時, GP-Pi 2.0 的優勢則更大。我也使用了一個從北美類風濕性關節炎協會下載的類風濕性關節炎資料集來測試GP-Pi 2.0 和CNN, GP-Pi 2.0 所找出來的分類樹不但和現有的文獻吻合, 而且也發現和類風濕性關節炎可能有關的SNP. 而且CNN 也成功建立了一個能預估任何一個測試數據（Testing Data）和未見數據（Unseen Data）樣本的健康狀態的模型。這個CNN 也可以被改良為一個包裹式的特徵選擇算法，用於選擇有相互作用的SNP。我也嘗試了利用兩種不同的加速技術來加速SNP 和SNP 之間的所有二元相互作用的統計學測試的過程。首先, 我應用了圖形處理器（GPU）的平行運算能力去加速每一對SNP 的相互作用的列聯表運算。然後, 我也嘗試設計了一個基於LSH 算法的剪枝算法，能去快速判斷相類似的SNP 並跳過這些SNP 之間的相互作用的統計學測試來節省時間。我使用了人工模擬資料集來對這兩種不同的加速技術進行了概念驗證。在人工模擬數據的測試下，我展示了這兩種不同的加速技術有可能是有效的。Lee, Kwan Yeung.Thesis Ph.D. Chinese University of Hong Kong 2020.Includes bibliographical references (leaves 191-207).Abstracts also in Chinese.Title from PDF title page (viewed on March 31, 2022)