INTRODUCTION

The design of experiments is one of the most important areas of microarray bioinformatics and is a long-standing topic in classical statistics. The reason for good experimental design is that it allows you to obtain maximum information from an experiment for minimum effort – which translates into time and money. The alternative to good experimental design is to performmicroarray experiments which produce data that cannot be analysed.

You might ask why it is that this topic appears at this point in the book, after data analysis rather than earlier in the book, alongside the material on the design of microarrays themselves. There are two reasons for this. The first is that the topics in this section use concepts from some of the earlier chapters, most importantly the ideas of hypothesis tests and p-values introduced in Chapter 7. But there is also amore philosophical reason why I have chosen to place the material on experimental design after the material on data analysis. In my view, it is absolutely critical to understand the scientific questions you are trying to answer, or even the scientific hypotheses you are seeking to generate, before you design your experiment. To this end, you should have a clear idea of the structure of the data you are seeking to produce and the types of data analysis you intend to employ before you design an experiment.

This chapter considers three areas of experimental design:

1. Section 10.2: Blocking, Randomisation and Blinding, looks at the statistical problems of confounding and bias, and the methods that are used to resolve these issues.

2. […]

INTRODUCTION

Microarrays are a genomic technology. Genomics is different from genetics in that it looks not at genes in isolation, but at how many genes work together to produce phenotypic effects. In Chapter 7 we saw how microarrays can be used to study genes in isolation. But much of the real power of microarrays is their ability to be used to study the relationships between genes and to identify genes or samples that behave in a similar or coordinated manner. This chapter looks at a number of analysis techniques to find and verify such relationships.

We will use two example data sets to examine the ideas of this chapter.

EXAMPLE 8.1 YEAST SPORULATION DATA (DATA SET 8A)

Budding yeast can reproduce sexually by producing haploid cells through a process called sporulation. Yeast was placed in a sporulating medium, and samples were taken at six time points following the start of sporulation and hybridised to microarrays. We want to identify groups of genes that behave in a coordinated manner in this time series.

EXAMPLE 8.2 DIFFUSE B-CELL LYMPHOMA SUBTYPES (DATA SET 8B)

Samples were taken from 39 patients suffering from diffuse large B-cell lymphomas and hybridized to microarrays. We want to identify genes that are co-regulated in this disease. We are also interested in whether there are groups of patients with similar gene expression profiles.

This chapter discusses methods that can be used to answer such questions; it is organised into the following five sections:

1. Section 8.2: Similarity of Gene or Sample Profiles, looks at different methods for quantifying the similarity or dissimilarity of gene expression profiles. We show how the different methods can give different results and, hence, the need to think carefully about choosing the method you use.

2. […]

INTRODUCTION

In this book we have described many different types of microarray experiment. All these experiments generate large volumes of complex data. As scientists, we need to be able to communicate the results of our experiments with other scientists. There are many reasons why scientists seek to share data, including the following:

1. ▪ To verify the results of a published microarray experiment. It is accepted that for scientific results to be published in a refereed journal, it is necessary to provide sufficient information so that others can reproduce the experiment.

2. ▪ To perform further experimental work based on the results. Microarray data frequently generate hypotheses that require further experimental investigation; often, microarray experiments are performed precisely to generate such hypotheses.

3. ▪ To undertake further data analysis of the results. Sometimes it is possible to perform further data analysis beyond the analyses carried out by the researchers in their original paper, which requires full access to the data.

4. ▪ To compare the results with other functional genomics data. It is valuable to make comparisons either between different microarray experiments, or between microarray data and data from other sources (e.g., proteomics).

5. ▪ To developnovel data analysis methods. Bioinformatics researchers developing novel data analysis methods need data sets for testing their methods.

Data sharing is not simply a function of scientists. A microarray laboratory will typically run a number of different computer applications to capture, store, publish and analyse microarray data (Figure 11.1). In order for the laboratory to operate successfully, each of these computer applications needs to be able to exchange data with the other.

INTRODUCTION

Normalisation is a general term for a collection of methods that are directed at resolving the systematic errors and bias introduced by the microarray experimental platform. Normalisation methods stand in contrast with the data analysis methods described in Chapters 7, 8 and 9 that are used to answer the scientific questions for which the microarray experiment has been performed. The aim of this chapter is to give an understanding of why we need to normalise microarray data, and the methods for normalisation that are most commonly used. The chapter is arranged into three further sections:

1. Section 5.2: Data Cleaning and Transformation, looks at the first steps in cleaning and transforming the data generated by the feature extraction software before any further analysis can take place.

2. Section 5.3: Within-Array Normalisation, describes methods that allow for the comparison of the Cy3 and Cy5 channels of a two-colour microarray. This section is only relevant for two-colour arrays.

3. Section 5.4: Between-Array Normalisation, describes methods that allow for the comparison of measurements on different arrays. This section is applicable both to two-colour and single channel arrays, including Affymetrix arrays.

SECTION 5.2 DATA CLEANING AND TRANSFORMATION

The microarray data generated by the feature extraction software is typically in the form of one or more text files (Table 4.2). Before you use the data to answer scientific questions, there are a number of steps that are commonly taken to ensure that the data is of high quality and suitable for analysis. This section describes three stages of data cleaning and transformation:

1. ▪ Removing flagged features

2. ▪ Background subtraction

3. ▪ Taking logarithms

DNA microarrays are devices that measure the expression of many thousands of genes in parallel. They have revolutionised molecular biology, and in the past five years their use has grownrapidly in academia, medicine, and the pharmaceutical, biotechnology, agrochemical and food industries.

One of the principal features of microarrays is the volume of quantitative data that they generate. As a result, the major challenge in the field is how to handle, interpret and make use of this data. The field of bioinformatics has come to mean the applications of mathematics, statistics and information technology in the biological sciences, and the bioinformatics of microarrays is the answer to that challenge.

This book is a comprehensive guide to all of the bioinformatics you will need to successfully operate DNA microarray experiments. It is written for researchers, clinicians, laboratory heads and managers, from both biology and bioinformatics backgrounds, who work with or who intend to work with microarrays. The book covers all aspects of microarray bioinformatics, giving you the tools to design arrays and experiments, to analyze your data, and to share you results with your organisation or with the international community. It has been inspired by the Microarray Bioinformatics professional course at Oxford University, and thus would also be suitable for teaching the subject at postgraduate or professional level.

The book assumes a minimum knowledge of molecular biology, computer use and statistics. On the biology front, readers will find it helpful if they have an understanding of the basic principles of molecular biology, i.e., DNA, RNA, transcription and translation, as well as the notions of genome sequencing and the existence of sequence databases.

This appendix has been reproduced in full with permission from MGED. It is also available from:

http://www.mged.org/Workgroups/MIAME/miame_glossary.html

Age The time period elapsed since an identifiable point in the life cycle of an organism. (If a developmental stage is specified, the identifiable point would be the beginning of that stage. Otherwise the identifiable point must be specified such as planting) [MGED Ontology Definition]

Amount of nucleic acid labeled The amount of nucleic acid labeled

Amplification method The method used to amplify the nucleic acid extracted

Array design The layout or conceptual description of array that can be implemented as one or more physical arrays. The array design specification consists of the description of the common features of the array as the whole, and the description of each array design elements (e.g., each spot). MIAME distinguishes between three levels of array design elements: feature (the location on the array), reporter (the nucleotide sequence present in a particular location on the array), and composite sequence (a set of reporters used collectively to measure an expression of a particular gene)

Array design name Given name for the array design, that helps to identify a design between others (e.g., EMBL yeast 12K ver1.1)

Array dimensions The physical dimension of the array support (e.g., of slide)

Array related information Description of the array as the whole Attachment How the element (reporter) sequences are physically attached to the array (e.g., covalent, ionic) Author, laboratory, and contact Person(s) and organization(s) names and details (address, phone, FAX, email, URL)

Biomaterial manipulation Information on the treatment applied to the biomaterial

INTRODUCTION

Data analysis is seen as the largest and possibly the most important area of microarray bioinformatics. Reflecting this, there are three chapters in this book describing data analysis methods, which themselves answer three sets of scientific questions that are asked of microarray data:

1. Which genes are differentially expressed in one set of samples relative to another?

2. What are the relationships between the genes or samples being measured?

3. Is it possible to classify samples based on gene expression measurements?

In this chapter, we describe the methods for the first of these questions: the search for up- or down-regulated genes; Chapters 8 and 9 answer the other two questions. This chapter covers a variety of techniques, drawn from both classical statistics and more modern theory, to give a detailed account of how to analyze DNA microarray data for differentially expressed genes. We start the chapter with three examples to illustrate what we mean by the identification of differentially expressed genes.

EXAMPLE 7.1 DATA SET 7A

Samples are taken from 20 breast cancer patients, before and after a 16-week course of doxorubicin chemotherapy, and analyzed using microarrays. We wish to identify genes that are up- or down-regulated in breast cancer following that treatment.

EXAMPLE 7.2 DATA SET 7B

Bone marrow samples are taken from 27 patients suffering from acute lymphoblastic leukemia (ALL) and 11 patients suffering from acute myeloid leukemia (AML) and analyzed using Affy metrix arrays. We wish to identify the genes that are up- or downregulated in ALL relative to AML.

INTRODUCTION

A DNA microarray consists of a solid surface, usually a microscope slide, 1 onto which DNA molecules have been chemically bonded. The purpose of a microarray is to detect the presence and abundance of labelled nucleic acids in a biological sample, which will hybridise to the DNA on the array via Watson-Crick duplex formation, and which can be detected via the label. In the majority of microarray experiments, the labelled nucleic acids are derived from the mRNA of a sample or tissue, and so the microarray measures gene expression. The power of a microarray is that there may be many thousands of different DNA molecules bonded to an array, and so it is possible to measure the expression of many thousands of genes simultaneously.

This book is about the bioinformatics of DNA microarrays: the mathematics, statistics and computing you will need to design microarray experiments; to acquire, analyse and store your data; and to share your results with other scientists. One of the features of microarray technology is the level of bioinformatics required: it is not possible to performa meaningful microarray experiment without bioinformatics involvement at every stage.

However, this chapter is different from the remainder of the book. While the other chapters discuss bioinformatics, the aim of this chapter is to set out the basics of the chemistry and biology of microarray technology. It is hoped that someone new to the technology will be able to read this chapter and gain an understanding of the laboratory process and howit impacts the quality of the data.

INTRODUCTION

Chapter 5 described a number of methods to correct for unwanted systematic variability either within an array or between different arrays. In this chapter, we describe methods to measure and quantify the random variabilities introduced by the microarray experiment. The common sources of variability (Figure 6.1) are

1. ▪ The variability between replicate features on the same array

2. ▪ The variability between two separately labelled samples hybridised to the same array

3. ▪ The variability between samples hybridised to different arrays

4. ▪ The variability between different individuals in a population hybridised to different arrays

Estimates of these variabilities are essential to gaining an understanding of how well the microarray platform you are using is performing. They are also important parameters for determining the number of replicates required for a microarray experiment – a topic that is discussed in full in Chapter 10.

The first two levels of variability – between replicate features or samples hybridised to the same array – are meaningful only for two-colour arrays.However, the second two levels of variability – between hybridisations to different arrays and between individuals in a population – are meaningful both for two-colour arrays and Affymetrix arrays.

MEASURING AND QUANTIFYING MICROARRAY VARIABILITY

The variabilities between different features on an array, between two samples hybridized to the same array or between samples hybridized to different arrays are all introducedby the microarray experimental process. In contrast, the variation between individuals in the population is independent of the microarray process itself. Experimental variability is measured with calibration experiments; population variability is measured with pilot studies.

Calibration Experiments

The aim of a calibration experimentis to identify and quantify the sources of variability in your microarray experimental platform.

In the previous chapter, we concentrated on the flow of information from DNA to RNA to protein sequence and emphasized RNA's role as a “messenger” carrying copies of DNA's protein recipes to the ribosome for production. In this function, RNA's role is similar to the paper in a photocopy machine: though perhaps not always flat, it must be flattened out and dealt with in a linear fashion when used. RNA also plays catalytic roles in which it more resembles the paper used in origami: the shape into which it is folded is what matters most.

Messenger and Catalytic RNA

RNA is sufficiently versatile as a catalyst that some scientists postulate that the origins of life lie in an “RNA world” preceding both DNA and proteins. Other scientists have invented new RNA catalysts in the laboratory by experimentation with random RNA sequences. Important categories of catalytic RNAs in higher organisms include the following.

In Chapters 3 and 5 we saw how to develop an alignment scoring matrix and, given such a matrix, how to find the alignment of two strings with the highest score. In Chapter 6, we learned about some of the large genomic databases available for reference. Perhaps the most commonly performed bioinformatic task is to search a large protein sequence database for entries whose similarity scores may indicate homology with some query sequence, often a newly sequenced protein or putative protein.

The Needleman–Wunsch algorithm described in Chapter 3 constructs global alignments – alignments of the entireties of its two input sequences. In practice, homologous proteins are not similar over their entire lengths. This is because of differences in the importance of different segments of the sequence for the function of the protein. A typical protein has one or more active sites that play crucial roles in the chemical reactions it catalyzes. Acceptance of mutations in the midst of an active site is infrequent, since such mutations are likely to disrupt the protein's function. The segments intervening between the active sites help give the protein its peculiar shape but do not form strong bonds with other molecules as the protein performs its function. Mutations in these regions are more easily tolerated and thus are more common. Hemoglobin provides a good example; it easily tolerates mutations on its outer surface, but mutations affecting the active sites in its interior can destroy its ability to hold the iron-binding heme group essential to its role as oxygen carrier.

The urge to record or to reconstruct “family trees” seems to be a strong one in many different areas of human activity. Animal breeders have an obvious interest in pedigrees, linguists have grouped human languages into families descended from a common (and in many cases unattested) ancestor, and, when several manuscripts of the same text have been recovered, biblical scholars have tried to piece together which ones served as sources for which others. Even before the advent of the theory of evolution, naturalists attempted to discern the “Divine Plan” by assigning each known organism to its correct place in a system of nesting categories known as a taxonomy.

More modern biologists have tried to reconstruct the course of evolution by building trees reflecting similarities and differences in relevant features or characters of various species. Whereas early work relied upon morphological features such as shape of leaf or fruit for classification, most recent efforts focus on less subjective molecular features. In this chapter, we develop a program for constructing the evolutionary tree – or phylogeny – that best accounts for the differences observed in a multiple sequence alignment.

Parsimonious Phylogenies

Broadly speaking, there are two approaches to reconstructing the phylogeny of a group of species. One approach first reduces the similarities and differences among the n species to n(n – 1)/2 numerical scores; it then finds the phylogeny that optimizes a certain mathematical function of those scores.

In Chapter 3 we considered the problem of aligning DNA sequences that had been read from the same source but with errors introduced by laboratory procedures. Rather arbitrarily, we assigned a reward of +1 for a match and penalties of -1 and -2 for mismatches and gaps. In this chapter, we will examine how to align protein sequences that differ as a result of evolution itself rather than owing to experimental error. The outcome will be a method for constructing substitution matrices for scoring alignments. Potentially, these matrices can assign a different reward or penalty for each of the 210 possible unordered pairs of amino acids that may appear in a column of an alignment.

The function of a protein is determined by its shape and charge distribution, not by the exact sequence of amino acids. During DNA replication, various mutations can alter the protein produced by a gene. Some types of mutations are:

• point mutations, in which the machinery of replication randomly substitutes an incorrect nucleotide for the correct one;

• indels, or insertions and deletions, in which extra bases are randomly inserted or bases are omitted;

• translocations, in which longer pieces of DNA – possibly including one or more entire genes – are moved from one part of the chromosome to another part, or to another chromosome; and

• duplications, in which long pieces of DNA are copied and integrated into a chromosome.

Once a family of homologous proteins has been identified, it is often useful to arrange their sequences in a multiple alignment such as the one in Figure 9.1.

A multiple alignment is useful for constructing a so-called consensus sequence, which – while probably differing from every individual sequence in the family – is nonetheless a better representative of the family than any of its actual members. Multiple alignments can also form the basis of more abstract statistical models of the protein family called profiles.

By examining which elements of the consensus are present in most or all family members and which exhibit a greater degree of variability, we can also find clues to the protein's function. Highly conserved regions are likely to have been conserved because they form active sites crucial to function, while more variable regions are more likely to have merely structural roles.

We have already seen in Chapter 3 that the number of ways in which a mere two sequences of only moderate length can be aligned is comparable to current estimates of the number of atoms in the observable universe. The addition of more sequences only increases the number of possibilities. We need both a criterion for evaluating multiple alignments and a computational strategy that will allow us to eliminate large sets of alignments at one stroke.

To describe our evaluation criterion, we will rely on the notion of projection of a multiple alignment.