Machine learning for bio- and medical-informatics

许多领域都有大量的数据-从Web浏览器的日志文件到设备的遥测数据，再到医院记录。这些数据中的模式可能非常有价值;例如，这些模式可以识别哪些机器需要预防性维护，用户可能想要访问哪些站点以及他/她可能购买哪些产品，甚至可以识别哪些患者对给定的治疗反应良好。不幸的是，由于数据的数量和复杂性，这些模式可能难以提取。机器学习（ML）领域为这项任务提供了许多技术和工具。这项工作的大部分涉及理解、开发和使用将实例的“标记数据集”作为输入的工具（例如，一组患者，加上他们各自的临床结果），并返回一个分类器，该分类器可以将新的实例映射到标签--例如，以确定具有给定微阵列的患者，我将探索将ML应用于生物和医学信息学主题的方法：设计和构建自动化工具，[1]学习与脑肿瘤相对应的大脑磁共振图像中的模式;[2]将患者基因组的模式（包括单核苷酸多态性）与乳腺癌和其他患者状态相关联; [3]了解蛋白质的特性（例如，一般功能，亚细胞位置和特定代谢途径中的外观）;[4]将疾病与代谢组学谱和微阵列数据相关联，分别基于尿液/血液样本和活检。我还将探索相关的基础问题，提供分析，[1]使我们能够有效地学习和使用概率模型，如信念网; [2]帮助研究人员设计更经济有效的实验，通过确定哪些测试运行在哪些训练科目，以提供相关信息;[3]帮助研究人员设计实验，以提供相关信息。和[3]使用双聚类技术来有效地学习高维数据中的模式，例如微阵列和代谢组学谱。

Many fields have massive amounts of data -- ranging from log files from web browsers, to telemetry data from equipment, to hospital records. The patterns within this data can be very valuable; eg these patterns may identify which machine needs preventative maintanence, which sites a user may want to visit and which products s/he is likely to purchase, or even identify which patients will respond well to a given treatment. Unfortunately, these patterns may be difficult to extract, due to both the quantity, and complexity, of the data. The field of machine learning (ML) provides many technologies, and tools, for this task. Much of this work involves understanding, developing and using tools that take as input a "labeled dataset" of instances (eg, a set of patients, coupled with their respective clinical outcomes), and returns a classifier, which can map novel instances to labels -- eg, to determine whether a patient, with a given microarray, will respond well to some treatment.I will explore ways to apply ML to topics in bio- and medical-informatics: designing and building automated tools that [1] learn patterns in Magnetic Resonance images of brains that correspond to brain tumors; [2] correlate patterns of a patients genome (including Single Nucleotide Polyomorphisms) with breast cancer and other patient states; [3] learn properties of proteins (eg, general function, subcellular location and appearance within specific metabolic pathways); and [4] relate diseases to metabolomic profiles and microarray data, based respectively on urine/blood samples and biopsies. I will also explore relevant foundational issues, providing analyses that [1] enable us to effectively and efficiently learn and use probabilistic models, like belief nets; [2] help researchers design experiments more cost-effectively, by determining which tests to run on which training subjects, to provide the relevant information; and [3] use bi-clustering techniques to effectively learn patterns in high-dimensional data, such as microarrays and metabolomic profiles.