III: Small: Integrating and Interpreting Heterogeneous Genomic Data Through Deep Learning

对基因组中编码的所有功能元件进行全面鉴定是基础和应用生物学研究的基本需要。虽然基因组的编码区已经被很好地了解了，但非编码区，占哺乳动物基因组的98%以上，研究得很少，但它是理解基因调控、进化、复杂表型的遗传基础等的关键。本项目的目标是开发计算方法，通过利用来自公开可获得的基因组数据的大量数据和来自机器学习的最新算法来推断非编码序列的功能。这些算法可以极大地扩展现有基因组数据的实用性，提高非编码变异体致病能力的标注精度，并为研究非编码序列编码的基因调控文法提供了一种新的途径。该项目还将创造机会，促进生物学家和计算机科学家之间的互动，并为本科生和研究生提供跨学科培训，特别是那些来自传统上代表性较低的群体的学生。该项目的目标是开发一个基于深度学习的新计算框架，以理解非编码序列。在过去的几年里，研究人员已经生成了数千个基因组规模的数据集，涉及染色质可获得性、组蛋白修饰、DNA甲基化、蛋白质结合等，涵盖了广泛的组织和细胞类型。该项目将整合这些异质数据集，通过基于卷积和递归神经网络的创新机器学习算法以及深度生成模型来推导出非编码序列的全面表征。PI将开发深度学习算法来映射非编码序列和不同基因组测量之间的关系，学习染色质状态并从这些测量中发现新的功能元件，并预测非编码遗传变体的影响。用大量数据训练灵活和可扩展的学习模型提供了一种以无偏和稳健的方式表征非编码序列的方法，并提供了比传统方法更好的机会来提取编码在非编码序列中的复杂调控规则。该项目将为基因组学社区提供一个通用的、模块化的、开放源码的软件包工具箱，目的是大大提高当前基因组分析的准确性。

Comprehensive identification of all functional elements encoded in genomes is a fundamental need in both basic and applied biological research. Although the coding regions of genomes are well understood, the noncoding regions, representing over 98% of mammalian genomes, are far less studied, but hold the key to understanding gene regulation, evolution, genetic basis of complex phenotypes, etc. The goal of this project is to develop computational methods to infer the function of noncoding sequences by leveraging the plethora of data from publicly available genomic data and state-of-the-art algorithms from machine learning. These algorithms can greatly expand the utility of existing genomic data, improving the accuracy of annotating pathogenicity of noncoding variants, and offering a new way of studying grammars of gene regulation encoded by noncoding sequence. The project will additionally create opportunities to facilitate interactions between biologists and computer scientists, and offer interdisciplinary training for both undergraduate and graduate students, especially those from traditionally underrepresented groups.The goal of this project is to develop a new computational framework based on deep learning to understand noncoding sequences. Over the past few years, researchers have generated thousands of genome-scale datasets on chromatin accessibility, histone modifications, DNA methylation, protein-binding, and others, spanning a broad range of tissue and cell types. This project will integrate these heterogeneous datasets to derive a comprehensive characterization of noncoding sequence through innovative machine learning algorithms based on convolutional and recurrent neural nets, and deep generative models. The PI will develop deep learning algorithms to map the relationship between noncoding sequences and the diverse genomic measurements, learn chromatin states and discover novel functional elements from these measurements, and predict effects of noncoding genetic variants. Training a flexible and scalable learning model with large amounts of data provides a way of characterizing noncoding sequences in an unbiased and robust fashion, and offers a better chance of extracting complex regulatory rules encoded within noncoding sequences than conventional methods. This project will provide the genomics community with a versatile, modular, open-source toolbox of software packages, with the goal of greatly improving the accuracy of current genome analyses.