Biological Pattern Discovery with R

The following sections are included:

• Dedication
• Preface
• Contents

Classical approaches have been used for solving biological systems since the Newton time. Most of the earlier effort has reduced a system to the simplest levels for the study. Such an effort is referred to as the reductionist [Mazzochhi, 2008]. Having recognised that a biological system is one of the most complicated systems on the earth, whether a biological system can be easily simplified for a meaningful research has been debated for a few decades [Mazzochhi, 2008]. Moreover, decomposing a biological system, such as a single organism or a single cell, to unrelated subparts to reduce the component number for the study has been greatly challenged nowadays [Glaeser, 1994; Wood, et al., 2004; Schenkel and Leggewie, 2015; Nussinov, 2015]. The challenges not only result from the aspects before an analysis such as experiment condition, data creation and storage, but also the methods used in analysis including the computational approaches, the computing facility as well as the interpretation and explanation after an analysis…

A phenomenon change will usually occur in a species after an appropriate stress has been applied. What researchers want to know is how an observed phenomenon change in a species is determined by or associated with some genetic function. It has been well-researched that it is the genetic function of a species that responds to a stress directly and leads to the observed phenomenon change. Rather than employing all of genetic elements, such as genes, a biological system only activates a small proportion of genes to respond to a stress. Moreover, different genes may have different strengths to respond to a stress. Filtering out non-responsive genes to discover responsive genes thus plays a key role in researching and investigating the mechanism by which a biology system responds to a stress. From this, interpreting, explaining and extrapolating the discovered mechanism can follow. No doubt, computational models are required to help discover these responsive genes through a research of the pattern by which a phenomenon change is associated with genetic functions. And the unique function of the computational models constructed by machine learning algorithms can never be ignored and neglected for this purpose. A task of responsive gene discovery normally has no a priori knowledge about how such a pattern is formed, and hence discovered. This therefore fits to one typical machine learning mechanism, i.e., unsupervised learning. This chapter will introduce several unsupervised machine learning algorithms such as density estimation and cluster analysis and demonstrate how these algorithms can be used for responsive gene discovery. A typical responsive gene discovery problem, i.e., the essential gene discovery problem, will be introduced and discussed in this chapter. How essential genes can be discovered using unsupervised machine learning algorithms will be demonstrated.

A protein functions when it interacts with molecules or chemicals. Among many formats of interactions, protease cleavage is one of the widely researched subjects for several decades. This type of research aims to build up a predictive model based on collected laboratory data to discover novel interactions. Such a model is commonly established based on laboratory-verified protease cleavage data, in which the association knowledge between protease cleavage structure and protease cleavage function can be examined. A protease cleavage structure commonly means a primary sequence (or sub-sequence or peptide) which is believed to contain a specific amino acid composition pattern or trend in relationship with the protease cleavage function. In other words, the protease cleavage pattern must not show a random amino acid composition. Instead, the composition of the amino acids in a data set of protease cleaved peptides should demonstrate a trend for a specific protease to recognise for the interaction. To make a protease cleavage pattern discovery model to work efficiently, two types of peptides are collected and pooled together for constructing a model. They are the cleaved peptides and the non-cleaved peptides. Non-cleaved peptides definitely must have no trend of the amino acid composition at all. Instead, they must show random distribution of the amino acids. By the contrast comparison between a data set with a trend and a data set without any trend, a pattern by which two types of data can be discriminated can thus be discovered and can thus be formulated as the knowledge or the intelligence. The purpose of discovering such intelligence is, no doubt, for the prediction. The protease cleavage pattern discovery problem is a classification problem in machine learning. This chapter will introduce several classification analysis or discriminant analysis algorithms. This chapter will discuss how these algorithms can be used for the protease cleavage pattern discovery. Importantly, how to encode peptides so as to make encoded data biologically sound is the key to make a protease cleavage pattern discovery task successful. This chapter will therefore introduce the bio-basis function, a new and cutting-edge approach for encoding peptides. Based on the bio-basis function, several advanced algorithms will be introduced for the protease cleavage pattern discovery.

A biological system is very complex, in which different components interplay all the time. The interactions include gene-gene interaction, gene-metabolite interaction, pathway interaction and genetic-epigenetic interaction. It has been found in several studies that the genetic-epigenetic interplay may contribute to the development of several diseases. How the genetic-epigenetic interplay contributes to the disease development thus requires descent research about how genetic signatures and epigenetic signatures interplay. To determine how thousands of epigenetic signatures are associated with a genetic signature and to determine how thousands of genetic signatures are associated with an epigenetic signature is a regression problem. Importantly, the genetic-epigenetic interplay may happen between local or remote contributing components. Therefore, it is desirable to discover and focus on the most important signatures among thousands of signatures instead of scanning every contributing signature. In this way, whether the interplay happens locally or remotely can be discovered. Therefore, a regression analysis algorithm with well-equipped variable ranking mechanism is desirable. Only in such models, the quantitative evidence of local interplay or remote interplay between genetic signatures and epigenetic signatures can be well-researched. This chapter will introduce several regression algorithms which can rank variables for genetic-epigenetic interplay pattern discovery.

Analysing a spectrum to discover chemicals and molecules has been one of the important subjects in biochemistry research. However, spectrometric data is complex because it is often a mixture between a number of signals and a complicated background. The latter is also called a baseline. The signals mixed with the background of a spectrum can be well discovered only after the background of a spectrum has been accurately identified. Background estimation or baseline removal is thus the very first step to go in the area of spectra pattern discovery. The difficulty, however, is that a baseline of a spectrum is hardly an easily estimated linear function (a straight curve) or a simple function. Instead, it is commonly a complex, unknown and a non-analytic function. Many algorithms have therefore been developed for estimating the baseline of a spectrum in the hope to extract peaks and thus discover the signals as accurately and as correctly as possible. Only when a baseline has been accurately estimated and removed, the number of the falsely discovered signals can be minimised and the number of the true signals can be maximised. Among many algorithms, the Whittaker-Henderson smoother is one of the best for spectra pattern discovery. This chapter will introduce this algorithm and its variants as well as other algorithms which are used for spectra pattern discovery. How these algorithms can be applied to some real spectrometric data for spectra pattern discovery will be introduced in this chapter. How to accurately extract peaks, discover signals, separate signals from artifacts and align discovered signals across replicates will also be introduced in this chapter.

Differentially expressed genes are treated as the major signal to understand how a stress results in a phenomenon change in a species. Scientists hypothesised that it is the genetic signature that ultimately causes the observed phenomenon change. The major objective of gene expression pattern discovery is thus to discover differentially expressed genes. In this chapter, the approaches for gene expression pattern discovery will be introduced for two major types of data, i.e., the microarray gene expression data and the RNA-seq sequencing count data. Gene expression pattern discovery may encounter the challenges of the outlier problem or the subpopulation problem. Therefore this chapter will also introduce the alternatives for gene expression pattern discovery. Moreover, insufficient replicate number is the common issue in gene expression pattern discovery. The dual-scale Gaussian mixture algorithm for discovering differentially expressed genes in a data set with insufficient replicates will be introduced in this chapter.

Either a DNA sequence or a protein sequence is the basic component for most genomics research. It is well understood that a sequence is the main carrier of genetic information for any species. It is also no doubt that genomics research can hardly be successful without looking into sequence constituents. Importantly, most novel and significant discoveries in biology or medicine are based on sequencing data analysis nowadays. This chapter will introduce the commonly used sequence analysis approaches from the basic ones to the advanced ones and mainly focus on the sequence comparison approaches for whole genome pattern discovery. Moreover, this chapter will show how the sequence comparison approaches can be used to analyse the SARS-CoV-2 pandemic data.

When analysing the patterns of protease cleavage sites or posttranslational modification sites based on peptide data, a key question is whether it is possible to discover interpretable and explainable as well visible rules by which how peptides are classified can be well-understood. A linear model benefits a better interpretation between experimental data sets such as peptides and peptide labels. However, the relationship between peptides used in either protease cleavage pattern discovery or posttranslational modification pattern discovery and peptide labels may not always be simple. Moreover, peptides are non-numerical data. On the other hand, most nonlinear models such as neural network models do not offer sufficient insight into data. The decision-tree algorithms or the random forest algorithms are capable of providing a better interpretation to a model. However, in order to discover the optimal models, an expensive exhaustive enumeration has to be considered. This is why the evolutionary computation approaches have provided a better way and have been well-employed in many areas for generating optimal or near optimal models with a better interpretation capability. This chapter will introduce a different type of machine learning approaches for this kind of biological pattern discovery. It is the genetic programming algorithm, which is one type of the evolutionary computation approaches. This chapter will introduce how the genetic programming algorithm can be used for discovering the interpretable rules for a peptide data set. This chapter will also show how the rules developed by the genetic programming models can interpret the residue interplay within peptides.

Biological/medical systems are highly complex. It is still a challenge to improve the accuracy, robustness and interpretation capability of approaches or models for biological pattern discovery. The previous chapters in this book have only covered a very small proportion of biological pattern discovery subjects. Even for these subjects, there have been ongoing advanced studies. This last chapter is dedicated to a brief introduction of several promising new trends in this area.

The following sections are included:

• References
• Index