データ融合によるタンパク質切断解析および疾患との関連性発見

蛋白酶-蛋白质靶标预测问题是生物信息学中的一个重要研究课题。然而，现有的算法要么是非常特异的（即它们仅与特定的蛋白质或蛋白质家族，如半胱天冬酶一起工作），要么仅仅基于一级结构，因此非常容易提供假阳性（即非切割对错误地标记为切割）。我们的工作包括设计一个蛋白酶-蛋白质靶向算法，通过数据融合的应用，将一般的蛋白质切割机制包裹起来。我们从MEROPS数据库中提取了多达9000对切割的、湿实验室测试的蛋白酶-蛋白质靶向对。除了使用蛋白质相似性（BLAST）之外，通过融合相关但直接相关的切割信息来设计模型。通过收集公开可用的数据库，如KEGG，BioGRID，STRING，Domine和Interpro，我们将结构域，途径，基因和蛋白质知识纳入我们的模型。通过在测试数据上评估我们的模型，不参与训练和调整阶段，我们展示了它如何优于最先进的蛋白酶切割靶点预测软件。与最先进的方法不同，该算法是通用的，并不专用于特定的蛋白酶，因此它可以用于探索研究不足的蛋白酶，例如二级和三级结构完全未知的蛋白酶。

The problem of protease-protein target prediction has been extensively studied in Bioinformatics. However, existing algorithms are either very specific (i.e. they work only with specific proteins or protein families, such as Caspases) or solely based on the primary structure, therefore very prone to provide false positives (i.e. non-cleaving pairs wrongly labeled as cleaving). Our work consisted in the design of a protease-protein target algorithm, wrapping up the general protein cleavage machinery, through the application of data fusion.We extracted up to 9000 pairs of cleaving, wet lab tested protease-protein target pairs from the MEROPS database. Beside the use of protein similarity (BLAST), the model was designed by fuse relevant, but directly related cleavage information. By harvesting publicly available data bases such as KEGG, BioGRID, STRING, Domine and Interpro, we included domain, pathway, gene and protein knowledge to our model. By assessing our model on test data, not involved in the training and tuning phase, we showed how it outperforms state-of-the-art software for the protease cleavage target prediction. Unlike state-of-the-art approaches, this algorithm is general and not dedicated to specific proteases, therefore it can be used to explore poorly-studied proteases, where for example secondary and tertiary structure are completely unknown.