CAREER: Theory and Application of Small-Sample Error Estimation in Genomic Signal Processing

NSF提案#0845407题目：基因组信号处理中小样本误差估计的理论和应用P.I.：基因组信号处理（GSP）是一门研究生物信号建模和统计问题的工程学科，这些生物信号是通过高通量技术测量的，如基因表达微阵列或蛋白质丰度质谱。GSP的研究通常涉及使用模式识别或机器学习方法发现用于疾病诊断和预后的可靠分子标记。这种方法依赖于用于分类和预测的误差估计的准确性。由于GSP应用中常见的小样本量，这一点尤为重要。GSP中需要新的稳健的小样本误差估计方法，以实现可重复的科学发现，从而导致真正的医学进步。本研究的目标是解决小样本误差估计中存在的重要计算和统计问题。为了获得线性连续分类器估计误差和真实误差的联合抽样分布的精确和近似表示，（2）研究离散分类器的误差估计，包括二元决定系数（CoD）的估计，采用分析和完全枚举两种方法;（3）发展支持误差估计的方法学，解决高维空间中的应用问题，并采用自适应核函数，重点是特征选择;（4）与翻译基因组学（TGen）的医学合作者合作，将这些误差估计技术应用于癌症和传染病诊断和预后的生物标志物发现问题，约翰霍普金斯医学院和巴西奥斯瓦尔多克鲁兹基金会（FIOCRUZ）。

NSF Proposal #0845407Title: Theory and Application of Small-Sample Error Estimation in Genomic Signal ProcessingP.I.: Ulisses Braga-NetoProject AbstractGenomic Signal Processing (GSP) is the engineering discipline thatstudies modeling and statistical issues related to biological signalsmeasured by high-throughput technology, such as gene-expressionmicroarrays or protein-abundance mass spectrometry. Research in GSPtypically involve the discovery of reliable molecular markers fordisease diagnosis and prognosis, using pattern recognition or machinelearning approaches. Such approaches rely on the accuracy of errorestimation for classification and prediction. This is particularlycritical due to the small sample sizes that are common in GSPapplications. Novel robust small-sample error estimation methodologiesin GSP are needed in order to enable reproducible scientific discoverythat leads to genuine medical advancement.This research has as its goal solving significant computational andstatistical problems that exist in small-sample error estimation.Among the open problems that will be addressed are (1) to obtain exactand approximate representations of the joint sampling distribution ofthe estimated and true errors for linear continuous classifiers, whichwill lead to better-performing error estimators and practical tools toassess significance of results; (2) to study error estimation fordiscrete classifiers, including the binary coefficient ofdetermination (CoD), using both analytical and complete enumerationapproaches; (3) to develop the methodology of bolstered errorestimation, addressing the application in high-dimensional spaces andwith adaptive kernels, with an emphasis on feature selection; (4) toapply these error estimation techniques to the problem of biomarkerdiscovery for diagnosis and prognosis in cancer and infectiousdiseases, in partnership with medical collaborators at TranslationalGenomics (TGen), the Johns Hopkins Medical School, and the OswaldoCruz Foundation, Brazil (FIOCRUZ).