CLASSIFIERS FOR HIGH RESOLUTION CELL SORTING

尽管能够收集复杂的多参数列表模式 数据，直线或位图单元格排序边界仍然是 通常以相当随意的方式手动选择。通常情况下 试验者在绘制边界之前执行视觉聚类， 没有成功分类的统计预测。 复杂的多参数数据的可视化也很困难。单程 要处理可视化问题，请查看前三个 数据的主成分，并使用此信息来估计 用于“引导式”聚类分析的质心数目和近似质心。聚类 然后，将使用成员概率来进行排序决策。 处理将排序边界放置在 任意“视觉分类”的基础是应用统计学 细胞分类方法，如贝叶斯判别分析 决策界限。将计算判别函数和贝叶斯 决策边界将用于根据以下条件对单元格进行排序 将通过以下方式实时计算的判别函数分数 硬件和/或软件查找表。错误分类的代价将是 也包括在细胞分类决定中。 对于所有开发的分类器系统，分类器性能将是 通过ROC(“接收器工作特性”)分析测量 真阳性和假阳性。为了实现这一点，我们将使用 定义良好的数据系统和模型单元系统，所有 可以对照“标记的”参数检查分类器的正确性。 所有分选的模型细胞都可以通过PCR(聚合酶)进行明确的鉴定 链式反应)或FISH(荧光原位杂交)。 虽然该提案的主要重点是开发实时小区 分类器对细胞分类很有用，许多技术也可以 被其他研究人员用于传统列表模式的离线分析 流式细胞仪数据。因此，这些技术中的许多应该证明 对其他研究人员很重要，即使他们无法执行 这项提议中描述的复杂的细胞分类。 为了证明这些新技术对许多问题的重要性 在生物学和医学中，我们将尝试将这些新技术应用于 几个重要的应用包括：(1)高分辨率分选 从人母血中提取单个胎儿细胞用于产前诊断； 癌基因、抑癌基因、转移基因的分子特征 少见转移性乳腺癌中多药耐药基因的表达 从外周血和骨髓中高速分离细胞 浓缩或高分辨细胞分选；以及(3)骨髓净化 允许乳房自体移植的转移细胞的数量 接受大剂量化疗的癌症患者。

Despite the capability to collect sophisticated multiparameter listmode data, both rectilinear or bit-map cell sorting boundaries still are usually chosen manually and in a rather arbitrary fashion. Usually the experimenter performs visual clustering prior to drawing boundaries which have no statistical prediction of successful classification. Visualization of complex multiparameter data is also difficult. One way to deal with the visualization problem is to view the first three principal components of the data and use this information to estimate the number and approximate centroids for "guided" cluster analysis. Cluster membership probabilities will then be used to make sort decisions. Another way to deal with the problem of placing sort boundaries on the basis of arbitrary "visual classifications" is to apply statistical methods of classifying cells, e.g. discriminant analysis with Bayes decision boundaries. Discriminant functions will be calculated and Bayes decision boundaries will be used to sort cells on the basis of discriminant function scores which will be calculated in real-time by hardware and/or software lookup tables. A cost of misclassification will also be included in the cell sorting decision. For all classifier systems developed, classifier performance will be measured through ROC ("receiver operating characteristics") analyses of true-positives and false-positives. To accomplish this we will use a well-defined system of data and model cell systems whereby all classifiers can be checked for correctness against "tagged" parameters. All sorted model cells can be unequivocally identified by PCR (polymerase chain reaction) or by FISH (fluorescence in-situ hybridization). While the main focus of the proposal is to develop real-time cell classifiers useful for cell sorting, many of the techniques can also be used by other researchers for off-line analysis of conventional listmode flow cytometry data. Hence many of these techniques should prove important to other researchers even if they are unable to perform the sophisticated cell sorting described in this proposal. To demonstrate the importance of these new techniques to many problems in biology and medicine we will attempt to apply these new techniques to several important applications including: (1) high-resolution sorting of single fetal cells from human maternal blood for prenatal diagnosis; (2) molecular characterizations of oncogene, tumor suppresser, metastatic, and multi-drug resistance genes in rare human metastatic breast cancer cells isolated from peripheral blood and bone marrow by high-speed enrichment or high-resolution cell sorting; and (3) bone marrow purging of metastatic cells to allow for autologous transplantations in breast cancer patients undergoing high-dose chemotherapy.