Network-based machine learning framework for for data integration in medical applications

点击翻译按钮获取中文摘要

Recent technological advances have made it possible to assemble very large data collections – Big Data. Aside from the sheer scale, we now have access to multiple data types each describing the same phenomena in their own way. For example, in computer vision, it is now common to combine unlabeled images with text to label images more accurately. In biology and medicine, it has recently become cost-effective to collect genomic, transcriptomic, epigenetic, microbiome and other measurements that describe the state of cells and human bodies in health and disease. It has thus become essential to develop robust and efficient machine learning methods that can integrate multiple data types to gain deeper understanding of the phenomena measured by these data. Majority of existing integrative methods have limitations, e.g. they often require significantly more samples than features (not readily available in biological and medical applications); do not scale to the large number of available features requiring ad hoc feature pre-selection (Shen et al, 2009); do not deal with missing data and noise requiring substantial data pre-processing. In human studies, especially in childhood diseases, where the goal is to combine multiple types of measurements to understand disease mechanisms and reasons for phenotypic heterogeneity, the number of patients is very limited, whereas the number of measurements available for each patient is very large. Majority of existing methods are not applicable in this scenario. It is thus essential to ensure that new methods for biological and clinical data integration are scalable and robust to small sample sizes and many features. Together with my advisees we have developed a robust unsupervised approach to integrate multiple types of biological data, called Patient Network Fusion (PNF) (Wang et al, 2013) that addresses the issues above. Our results on five cancers show that we obtain more clinically relevant subtypes than those previously reported. We believe that networks in general and our approach for patient network fusion in particular set a perfect foundation for the comprehensive integrative framework that we are proposing to build in the course of the next five years. Our research program addresses the problem of data integration from several angles. Building on our recent successes we will develop methods to integrate more data types, specifically, low-signal-to-noise ratio data, such as single nucleotide polymorphisms, into our network-based framework. We will also extend non-negative matrix factorization to perform data integration of biological data using network regularization. These developments will serve as a broad base for the integration framework that will be tested on novel data obtained through our established clinical collaborations. Further, we will investigate several methods for integrative feature selection, i.e. identifying a small set of features stemming from multiple data types that explain majority of the variation in the data. Such methods are essential in shedding light onto the inner workings of biological processes and disease mechanisms. I work in close collaboration with clinicians at the Hospital for Sick Children and internationally to make sure that the methods are applied to real data and can be used and evaluated by clinicians in the process of their development. By working with a selected set of end-user collaborators, we will evaluate and refine our machine learning methods and user-interfaces, and ultimately develop a system that will impact the larger community of researchers interested in biological and medical data analysis.