Biological Data Mining and Its Applications in Healthcare

The following sections are included:

• Preface
• Contents

With the amount of data deluge as a result of high-throughput sequencing techniques and structural genomics initiatives, there comes a need to leverage the large-scale data. Consequently, the role of computational methods to characterize genes and proteins solely from their sequence information becomes increasingly important. Over the past decade, development of sensitive profile-based sequence database search algorithms has improved the quality of structural and functional inferences from protein sequence. This chapter highlights the use of such sensitive approaches in recognition of evolutionary related proteins when the amino acid sequence similarity is very low. We further demonstrate the use of sequence database mining based remote homology detection methods in exploring the repertoire of functions and three dimensional structures of parasitic proteins in Trypanosoma brucei brucei, causative agent of African sleeping sickness. With an emphasis on various metabolic pathways, sequence-function and structure-function relationships are investigated. Integrating the information of parasitic proteins in metabolic pathways along with their homology to targets of FDA-approved drugs, attractive drug targets have been proposed.

Genomic sequences contain many functional regions, such as exons and promoters. Identification of these regions is a very important task in molecular biology. It is costly and time consuming to carry out such identification based on biological experiments. Computational methods to identify the functional regions provide a good alternative. In this chapter, we briefly review the current methods of exon and promoter prediction. Then we present our work on the identification of short human exons based on autoregressive (AR) models and multi-feature spectral analysis, and eukaryotic promoter predictions based on both sequence and structural properties using Isomap and support vector machines. Methods for the recognition of disease-related genes and regulatory regions are also introduced. Our computational methods for functional region identification have been tested on many genomic datasets with good results.

Next generation sequencing has led to a deluge of biological sequence data in the public domain. Almost everything in bioinformatics depends on the inter-relationship between sequence, structure and function (all encapsulated in the term relatedness), which is far from being well understood. In particular, pairwise statistical significance has been recognized to be able to accurately identify related sequences (homologs), which is a very important cornerstone procedure in numerous bioinformatics applications. However, it is both computationally and data intensive. To prevent it from becoming a performance bottleneck, we resort to high performance computation techniques for accelerating the computation. In this chapter, we first present the algorithm of pairwise statistical significance, then describe several high performance algorithms, which enable significant acceleration of pairwise statistical significance estimation.

Biological networks demonstrate how different biochemical entities interact with one another to perform vital functions in an organism. Due to recent advances in high throughput computing and experimental techniques, several network databases, each containing large number of biological networks, have emerged. We need efficient methods for accessing and querying these databases. As these networks are generally represented as graphs in theory, several graph indexing methods are developed for answering queries on them. We group these methods in three categories, namely feature, tree and reference based indexing. In this chapter, we describe these network indexing categories on large biological network databases using example algorithms. Moreover, we provide a review of the literature and research directions on this topic.

In this chapter, we consider the knowledge completion problem, in which given existing knowledge is modified to be consistent with observed data. In particular, we focus on the knowledge completion problem on Boolean networks, and mathematically formalize the problem. Although basic versions of this problem are known to be NP-complete, they are solvable in polynomial time if the amount of observed data is not very large and the topology of given networks is a tree or a tree-like structure. We also consider the problem of detecting topological changes in signaling pathways after cell state alteration, and present an integer programming-based method. Our method is applied to a data set of gene expression profiles of colorectal cancer downloaded from the Gene Expression Omnibus and the signaling pathway data of colorectal cancer downloaded from the KEGG database.

Graphs have become widely adopted for representing biological data. Examples of biological data represented as graphs include chemical compounds, protein tertiary structure, protein-protein interaction networks, gene coexpression networks, etc. Graph mining techniques have proven to be powerful in discovering useful patterns in the data. The area of graph mining addresses the problem of discovering interesting subgraph patterns in a database of graphs or a single graph. The set of interesting subgraphs have immediate applications in graph clustering, graph classification, and graph indexing…

Scientific literature in biology and medicine contains a wide amount of knowledge which is unfortunately difficult to exploit and integrate with data generated with high-throughput technologies. This is particularly relevant when we address system-level problems characterized by large complexity and dimensionality. This chapter aims to focus on recent results on the integration of prior knowledge and genomic data for the inference of regulatory networks. First we present existing tools to retrieve prior knowledge from different sources such as PubMed and structured biological databases. Then we focus on the current state-of-the-art of network inference methods combining genomic data and prior knowledge. In the second half of the chapter we present a case study based on two publicly available cancer data sets using the tool predictionet showing the usefulness of prior knowledge for network inference.

Classification is one of the most popular and widely used supervised learning tasks, which categorizes objects into predefined classes based on known knowledge. Classification has been an important research topic in machine learning and data mining. Different classification methods have been proposed and applied to deal with various real-world problems. Unlike unsupervised learning such as clustering, a classifier is typically trained with labeled data before being used to make prediction, and usually achieves higher accuracy than unsupervised one…

In this chapter, we introduce classification techniques that are used to predict group membership to which a data object belongs. In the first half of the chapter, we summarize several start-of-the-art classification algorithms. In the latter section, we show examples of applications of the classification techniques for prediction of protein-protein interactions in human proteome.

Trend analysis often refers to the techniques for extracting an mode of behavior in a time series which would be partly or nearly completely hidden by noise. Now, the techniques of trend analysis have been widely used in detecting outbreaks and unexpected increases or decreases in disease occurrence, monitoring the trends of diseases, evaluating the effectiveness of disease control programs and policies, and assessing the success of health care programs and policies et al. common methods of trend analysis have been introduced in this which include Age-period-cohort (APC) model, Joinpoint regression, Time series analysis, Cox-Stuart test, Runs test and Functional data analysis (FDA) et al. The theoretical foundation of each method has simply described for convenience of readers. Specific details of each method include five parts, such as introduction, mathematical model, empirical study, problems and solutions, and a conclusion. Last, the of methods-related software and the procedure codes are described as supplements.

Three dimensional (3D) medical imaging technique is a new field that is rapidly evolved over the last years, leading to a major improvement in patient care. Starting from 3D models of the patient, anatomical structures can be identified and extracted, and diagnosis and surgical simulation shall be supported. Given that the computational reconstruction of 3D models is critical for medical diagnosis and treatment, and the complex imaging processing requires considerable resources and advanced training, we introduce 3D image acquisition, segmentation and registration in this chapter, for all of which, the main purpose is to effectively transform the unstructured image data to structured numerical data for further data mining tasks.

The massive flow of scholarly publications from traditional paper journals to online outlets has benefited scientists by providing ease to access to published work. Many publishers are now making the full texts of articles publicly available online. However, due to the sheer volume of available literature, researchers are finding it increasingly difficult to locate information of interest. Recent advances in textmining technology have aimed to accelerate literature curation, maintain the integrity of information, and ensure proper linkage of data to other resources. This chapter will give an overview of essential textmining technologies that can be applied in biomedicine and healthcare including: entity recognition, entity linking, relation extraction, and co-reference resolution. Different perspectives on the application of these techniques for researchers, clinicians and patients will be presented, and we will see how these technologies can be integrated in healthcare administration. Finally, a case study of a biomedical text mining database system that automatically recognizes and collects cardiovascular disease related genes will be demonstrated to illustrate the application of some of the principles discussed in this chapter.

In the text mining field, obtaining training data requires human experts' labeling efforts, which is often time consuming and expensive. Supervised learning with only a small number of positive examples and a large amount of unlabeled data, which is easy to get, has attracted booming interests in the field. A recently proposed relabeling method, which assumes unlabeled data as negative data for text classification, has been shown successful in identifying relevant biomedical documents on a specific topic. However, it's not known whether and how feature selection affects the performance of the method. In addition, no extensive research has been conducted to evaluate how the performance of the method changes when the proportion of positive examples in the unlabeled dataset varies. Following the relabeling method, we train Support Vector Machines using positive and unlabeled data to rank incoming documents based on their probability values of being predicted as positive. Using an RNA-protein binding (RNAPB) dataset collected from PubMed, we conduct a series of experiments to evaluate the performance of the proposed text ranking algorithm. Our experimental results show that 1) feature selection is helpful in improving the performance of the algorithm; 2) the increase of the proportion of the positive examples in the unlabeled dataset decreases the performance of the algorithm, but when no reliable negative training data is available, the classifiers, built based on the unlabeled data that contains a small proportion of positive examples, may have comparable performance with the classifiers built based on pure negative data.

Elucidation of protein interaction networks (PINs) is essential for understanding disease-related biological processes and mechanisms. While significant information regarding PINs is available in the literature as free-text, such information is difficult to retrieve and synthesize. The following chapter describes machine learning techniques and web-based tools for literature-based extraction of protein interactions and their networks. Two case studies are provided to illustrate the efficient use of automated text-mining based applications in biomedical research using the tools PIMiner (http://www.biotextminer.com/PPI/index.html) and CPNM (http://www.biotextminer.com/CPNM/).

The following section is included:

• Index