Feature Selection for Genomic Data Using Known and Novel Biological Information

描述（由申请人提供）：我们的长期目标是基于临床和基因组数据等多种数据，通过建立准确的癌症风险和预后预测模型，制定个性化的预防和治疗策略，降低癌症风险。我们当前研究的直接目标是开发创新的统计方法，结合生物信息，包括先验知识基因通路和新型microRNA （miRNA）调控网络，识别与癌症风险和预后相关的基因组特征。本研究的基本原理是：1)高维数据，如基因组生物标志物，已经在许多研究中获得，并可能在可预见的未来在实践中随时可用；2)为了利用高维基因组数据建立良好的预测模型，特征选择是必不可少的；3)结合已知和新的生物信息，可以在特征选择中借鉴信息，从而获得更大的能力；4)半参数方法比参数方法在特征选择方面具有更强的鲁棒性。这些考虑导致了四个具体目标：1)结合已知和新的生物信息，在癌症结局（例如，癌症复发或死亡的时间）的半参数加速失效时间（AFT）模型中开发高维生物标志物的分层特征选择；2)通过对miRNA调控网络的综合分析，结合已知和新的生物信息，在AFT模型中开发针对癌症结局（如癌症复发或死亡时间）的高维生物标志物的贝叶斯特征选择；3)开发高效的算法和用户友好的软件，目标是将它们传播给癌症研究人员；4)通过广泛的数值研究，包括模拟和实际数据分析，对所提出的方法进行系统评估。我们提出的方法与现有方法的区别在于，我们使用已知和新的生物信息来指导特征选择，我们研究了两种替代方法，半参数和全贝叶斯联合建模，每种方法都有自己的优缺点。所有目标的进展都将由激励前列腺癌和脑肿瘤数据以及广泛的模拟研究来指导和评估。提出的方法将使研究人员能够识别关键的基因组特征以及预测癌症风险和预后的生物学途径，从而导致潜在的药物靶点和随后有效的个性化治疗。它们承诺在经常遇到类似数据和生物信息的广泛生物医学科学环境中也有类似的好处。

DESCRIPTION (provided by applicant): Our long-term goal is to reduce cancer risk by building accurate prediction models for cancer risk and prognosis and developing individualized prevention and treatment strategies based on diverse data including clinical and genomic data. Our immediate goal in the current study is to develop innovative statistical methods to identify genomic features in relation to cancer risk and prognosis with incorporation of biological information including prior-knowledge gene pathways and novel microRNA (miRNA) regulatory network. The underlying rationale for this research is that: 1) high-dimensional data such as genomic biomarkers have been obtained in many research studies and will likely be readily available in practice in the foreseeable future; 2) feature selection is imperative in order to buid good prediction models using high-dimensional genomic data; 3) incorporating known and novel biological information allows information-borrowing in feature selection, resulting in greater power; and 4) semiparametric methods are more robust to model misspecification than parametric methods that have dominated the literature in feature selection. These considerations lead to four specific aims: 1) develop hierarchical feature selection of high-dimensional biomarkers in semiparametric accelerated failure time (AFT) models for cancer outcomes (e.g., time to cancer recurrence or death) with incorporation of known and novel biological information; 2) develop Bayesian feature selection of high- dimensional biomarkers in AFT models for cancer outcomes (e.g., time to cancer recurrence or death) with integrative analysis of the miRNA regulatory network and incorporation of known and novel biological information; 3) develop efficient algorithms and user-friendly software with the goal of disseminating them to cancer researchers; and 4) perform systematic evaluation of the proposed methods through extensive numerical studies including simulations and real data analyses. Our proposed methods distinguish themselves from existing approaches in that we use both known and novel biological information to guide feature selection, and we investigate two alternative approaches, semiparametric and fully Bayesian joint-modeling, each of which has its own strengths and weaknesses. Progress on all aims will be guided by and evaluated on motivating prostate cancer and brain tumor data, and by extensive simulation studies. The proposed methods will allow investigators to identify key genomic signatures as well as biological pathways that are predictive of cancer risk and prognosis, leading to potential drug targets and subsequently effective personalized treatments. They promise similar benefits to a wide range of biomedical science settings where similar data and biological information are often encountered.