Unsupervised machine learning methods that discover the molecular programs underlying cellular biology

十年前，人类基因组被测序。也就是说，我们了解了组成我们遗传密码的30亿个字母A/C/T/G。从那时起，我们收集了数千个关于基因组活性的全基因组测量数据。这些测量跨越了基因组的数百个生化特性，例如测量给定蛋白质与哪些碱基对结合，并跨越了数百个人体组织。我们总共进行了10^13次基因组测量。然而，尽管有这些庞大的数据集，基因组的功能仍然未知。我的研究目标是发现新的基因组功能类型。为了做到这一点，我开发了自动发现高维数据集模式的计算方法；这些算法被称为无监督机器学习算法。在过去，我开发了一类无监督的基因组注释算法，并使用它们来发现人类基因组中几个新的功能活动类别。受到这一成功的启发，我将运用同样的基于发现的研究理念，通过(1)利用新可用的数据类型，(2)开发新的无监督机器学习算法，可以发现过去可能出现的更复杂的模式，继续提高我们对基因组生物学的理解。这项工作将以两种独立的方式产生影响。首先，我将发现基因组生物学的新机制，这将反过来提高我们对人类遗传特征、疾病和进化的理解。其次，我将为无监督机器学习开发新的通用方法。这些方法将为自动化发现铺平道路，使其成为其他可获得高维数据的领域的标准实践。

A decade ago, the human genome was sequenced. That is, we learned the string of three billion letters A/C/T/G that make up our genetic code. Since then, we have collected thousands of genome-wide measurements about the activity of the genome. These measurements span hundreds of biochemical properties of the genome, such as the measurement of which base pairs a given protein binds to, and span hundreds of human tissues. In total, we have on the order of 10^13 genomic measurements. Yet despite these massive data sets, much of how the genome functions remains unknown. The goal of my research is to discover new types of genomic functions. To do this, I develop computational methods that automatically discover patterns in high-dimensional data sets; these algorithms are known as unsupervised machine learning algorithms. In the past, I developed a class of unsupervised genome annotation algorithms and used them to discover several new categories of functional activity in the human genome. Inspired by this success, I will apply the same discovery-based research philosophy to continue to improve our understanding of genome biology, through (1) leveraging newly available types of data, and (2) developing new unsupervised machine learning algorithms that can discover more complex patterns that was possible in the past. This work will have impact in two independent ways. First, I will discover new mechanisms of genome biology, which will in turn improve our understanding of human heritable traits, disease and evolution. Second, I will develop new general-purpose methods for unsupervised machine learning. These methods will pave the way for automated discovery to become standard practice in other fields where high-dimensional data is available.