Developing novel machine learning algorithms for network biology

在过去的几十年里，基于网络的方法来探索生物过程得到了广泛的研究。其中许多方法只在规模相对较小的基准数据集中进行了研究。目前，随着高通量技术的进步，更多的基因组规模的基因组数据已经变得容易获取。一个有趣的研究领域是如何挖掘这些高维和海量的数据。基于网络的计算方法被认为是一种很有前途的技术，可以建立尽可能好的学习系统来理解复杂生物网络中相互作用的不同生物和病理过程。然而，在许多生物学应用中，它们的学习过程的计算复杂性远远高于其他替代学习方案。在这个建议中，我主要集中在三个方面：(1)联合预测生物网络中的蛋白质功能和疾病基因类型。虽然基于网络的基因/蛋白质功能和疾病基因预测已经被广泛研究，但这两项任务是独立解决的。我将实现一个计算机学习系统来联合预测生物网络中的两种标签类型；(2)识别网络生物标记物。已经开发了不同的方法来选择基因和蛋白质生物标记物。然而，许多研究表明，当这些类型的生物标志物应用于其他独立研究时，它们是不可重现的。我将开发新的计算机系统来识别在生物和功能上相关的子网络生物标记物，这被认为是更稳健的；以及(3)从生物数据推断不同生物条件下子网络的样本特异性激活模式。为了了解使用已识别的子网络生物标记物的生物学机制，我将开发计算方法来推断子网络激活和共激活的模式。这项研究计划将为加拿大和国际IT行业推进基于网络的大规模自动知识发现技术。

In the past few decades, network-based approaches to explore biological processes have been extensively studied. Many of these methods have been investigated only in  relatively small-scale benchmark data sets. Currently, with the advances in high-throughput technologies, more genome-scale genomic data have become readily accessible. An interesting research field is how to mine these high-dimensional and voluminous data. Network-based computational approaches have been regarded as a promising technique to build the best possible learning systems to understand different biological and pathological processes that interact in a complex biological network. However computational complexity of their learning process is much higher than other alternative learning schemes in many biological applications. In this proposal I focus on three major areas: (1) Jointly predict protein functions and disease gene types in biological networks. Although network-based gene/protein function and disease gene predictions have been widely studied, these two tasks are solved independently. I will implement a computer learning system to joint prediction of the two label types in a biological network; (2) Identify network biomarker. Different approaches have been developed for selecting gene and protein biomarkers. However, many studies have shown these types of biomarkers are not reproducible when they are applied in other independent studies. I will develop novel computer systems to identify biologically and functionally relevant subnetwork biomarkers, which is thought to be more robust; and (3) Infer sample-specific activation patterns of subnetworks under different biological conditions from biological data. To understand biological mechanisms using the identified subnetwork biomarkers, I will develop computational methods to infer patterns of subnetwork activation and co-activation. This research program will advance large-scale network-based automatic knowledge discovery technologies for Canadian and international IT sectors.