Collaborative Research: GOSTRUCT: modeling the structure of the Gene Ontology for accurate protein function prediction

科罗拉多州立大学获得了一项资助，用于开发预测蛋白质功能的机器学习方法。蛋白质功能注释的可用性支持生物学家在多个领域的日常工作-从生物医学发现到植物抗旱性研究，以及生物燃料生产中有用的细菌设计。为已测序基因组中的蛋白质分配功能是一项重大任务，随着每天都有新的生物体被测序，通过实验确定这些生物体中所有蛋白质的功能是不现实的，需要对尚未在实验室中研究过的蛋白质进行功能计算分配。二十多年来，计算科学家一直在研究函数预测问题。然而，在此期间，蛋白质功能预测的基本方法并没有太大变化，仍然是使用BLAST等序列比较方法从具有已知功能的蛋白质进行“注释转移”。蛋白质功能预测有几个特性，使得最先进的机器学习方法难以应用于该问题，例如大量的潜在函数（数千个可能的术语），蛋白质可以具有多种功能的事实，以及基因本体（GO）中术语之间的层次关系，这是用于描述蛋白质功能的关键字的标准系统。在这项工作中，用氧化石墨烯术语注释蛋白质的问题将使用“结构化输出的核方法”的方法明确地建模为分层分类问题，该方法允许对复杂的预测问题进行建模。这种方法将允许pi整合各种基因组信息-序列数据，基因表达，蛋白质-蛋白质相互作用，以及从生物学文献中挖掘的信息。该奖项将带来一种具有最先进精度的功能预测方法。该项目将以可下载软件和在线功能预测服务器的形式向生物信息学和生物学社区提供GOstruct方法，从而产生广泛的影响。通过将该工具纳入生物学家编程和核心方法的新课程，教育将受到影响。

Colorado State University is awarded a grant to develop machine learning methods for predicting protein function. The availability of protein function annotations supports the everyday work of biologists in multiple areas---from biomedical discovery to the study of plant drought resistance, and the design of bacteria useful in biofuel production. Assigning function to proteins in sequenced genomes is a major undertaking, and with new organisms being sequenced daily, experimentally determining the function of all the proteins in those organisms is not practical, requiring computational assignment of function to proteins that have not been studied in the lab. Computational scientists have been considering the problem of function prediction for over two decades. Yet, the basic methodology for protein function prediction has not changed much during this time and remains that of "annotation transfer" from proteins with a known function using a method for sequence comparison such as BLAST. Protein function prediction has several properties that make it difficult to apply state-of-the-art machine learning methods to this problem, such as the large number of potential functions (thousands of possible terms), the fact that proteins can have multiple functions, and the hierarchical relationship between terms in the Gene Ontology (GO), which is the standard system of keywords used to describe protein function. In this work the problem of annotating proteins with GO terms will be explicitly modeled as a hierarchical classification problem using the methodology of "kernel methods for structured outputs", which allows the modeling of complex prediction problems. This methodology will allow the PIs to integrate a variety of genomic information - sequence data, gene expression, protein-protein interactions, and information mined from the biological literature. The award will lead to a function prediction method with state-of-the-art accuracy. The project will have broad impact by providing the GOstruct method to the bioinformatics and biology communities in the form of downloadable software and an online-accessible function prediction server. Education will be impacted through the incorporation of the tool into new courses in programming for biologists and on kernel methods.