以演化式最佳化方法建立量化基因調控網路模型與整合型平台

Abstract

推論量化之基因調控網路(quantitative gene regulatory networks, qGRNs)來發現一些重要的調控關係，在後基因體時代扮演很重要的角色，尤其在生物化學、生化工程學以及製藥學上更是舉足輕重。有效率的qGRNs推論涉及到兩個重要步驟：1)選擇合適的數學模型來描述調控網路；2)發展能準確估算數學模型參數的演算法。網路分析(NCA)是已知以模型分析來有效率建立調控網路的方法，但其在模型參數估演算法上卻有著無法充分有效利用已知調控關係與產生不穩定參數解的問題。 本論文首先提出基於直交退火演算法的全域最佳化演算法OptNCA，在充分使用已知調控關係的前提下，限制解空間的搜尋並充分利用NCA模型的優點。與目前NCA模型衍生方法的比較中，OptNCA從已發表的大腸桿菌實驗與由其衍生的合成資料集所重建的qGRNs，在已知調控關係與估算基因表現量資料的正確性上都有優異的結果。OptNCA在沒有初始調控關係下，相較與其他非NCA衍生方法在推論定性調控網路上，亦有不錯的結果。此外，在OptNCA的可用性分析中，對於所需要的觀測表現值數量也被詳細討論。 本論文建立軟體框架GeNOSA，提供使用OptNCA所需的前處理與後期分析工具集來建立qGRNs。GeNOSA成功展示：1)應用在推論cAMP劑量控制實驗中，受CRP調控且在RegulonDB中記載為雙向調控的調控關係，2)使用以劑量反應與時間序列的表現量資料，推論出的qGRNs，其子網路表現量估算值與實驗觀測值高度相關，以及3)預測未知的調控關係並證實得以由實驗方法驗證。 最後，本論文實作以GPU加速OptNCA計算評估函數值的cuOptNCA，得以改善大型調控網路中參數增加後的執行效能。同時考量到GeNOSA需要藉助於大量運算，並要能逐漸增加已知調控關係資料庫的豐富性，最後建立一個整合性並方便使用的qGRNs平台(GRNP)，以期能幫助相關研究領域的學者能快速地找出隱藏於調控網路中的知識。

The establishment of quantitative gene regulatory networks (qGRNs) plays an important role in biochemistry, bioengineering, and pharmaceutics in the post genomic research. An efficient approach involves two steps: 1) choosing mathematical models to present GRNs, and 2) developing algorithms accurately estimating parameter values. The network component analysis (NCA) approaches suffer from shortcomings such as usage limitations of connectivity information and the instability of inferred qGRNs. In this dissertation, a global optimization algorithm (OptNCA) bases on orthogonal simulated annealing was proposed to utilize advantages of the NCA model without matrix reduction to confine the solution space of the decomposition problem. OptNCA performs well against existing NCA-derived algorithms in terms of utilization of connectivity information and accuracy of estimated expression profiles using an experimental dataset of Escherichia coli and its derived datasets. For comparisons with non-NCA-derived algorithms, OptNCA is evaluated in terms of qualitative assessments that are considerable. Furthermore, the usability analysis for the required number of observed expression profiles are discussed in details. The OptNCA-based framework (GeNOSA) was proposed in this dissertation to provide pre- and post-processing scripts to establish qualitative GRNs and analyze inferred qGRNs, respectively. GeNOSA was successfully demonstrated in several applications: 1) deducing condition-dependent regulations of dual regulation from dose-response data, 2) inferring high-consensus qGRNs resulting in highly correlated estimations of expression levels of a sub-network for both dose-response and time-course data, and 3) predicting a novel regulation that was confirmed through experimental validations. A GPU-accelerated OptNCA (cuOptNCA) was implemented in this dissertation to improve the performance of calculating fitness values with increasing numbers of parameter values in large-scale qGRNs. Eventually, an integrated GRN platform (GRNP) with cuOptNCA, incrementally refined databases of regulations, and the user-friendly interface was developed to assist researchers revealing hidden information under interactions of regulations for various species.