Datasets:

SouthernCrossAI
/

Arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 132k rows

Filename stringlengths 22 64	Paragraph stringlengths 8 5.57k
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	an initial list of proteins targeted by the autoimmune reaction in APS1 patients, we implemented a protein microarray-speciﬁc pre-processing pipeline as well as performed differential analysis.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	We aimed at a more profound understanding of the mechanisms behind APS1 condition and autoimmunity in general. Therefore, protein targets identiﬁed in the previous publication were studied further [2]. We analysed multiple open databases and public protein datasets to determine common biological factors be- hind selected protein targets. We used a web server for the functional enrichment analysis to validate our results.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	A focal point of this PhD work is the protein microarray web-explorer (PAWER) – an R-based web-tool, developed to enable semi-automatic protein microarray analysis [3]. PAWER incorporates all the relevant computational methods im- Its intuitive user interface and step-by-step plemented in the previous papers. workﬂow are designed to help perform protein microarray analysis with ease.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Finally, in the fourth publication included in this thesis, we have explored the value of applying machine learning models along with classical statistical meth- ods discussed in the previous publications. Here we analysed a case-control study of endometriosis [4]. Prior statistical analysis had shown that no individual pro- teins are capable of distinguishing endometriosis patients from controls based on proteins found in the blood. We used several powerful machine learning metods to evaluate the predictive performance of combinations of proteins. In line with sta- tistical test results, neither model achieved performance signiﬁcantly better than the random chance. Therefore, machine learning results supported the hypothe- sis that neither measuring individual proteins nor in combination with others can predict endometriosis and thus help to diagnose the disease in the given samples.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Life is beautiful in its complexity, and proteins are some of the most fundamental biological elements that enable this complexity. Often referred to as “workhorses” of the cell, proteins are responsible for almost every imaginable item on organ- isms’ to-do list [5]. Proteins carry out tasks ranging from building tissue and replicating deoxyribonucleic acid (DNA) to enabling timely immune response and facilitating oxygen delivery, albeit different types of proteins are at work. The number of proteins present at any given moment is strictly regulated by cells [6], as any signiﬁcant deviation from the norm may cause malfunction and even dis- ease. Therefore, information about protein abundance can offer valuable insight into the mechanisms of various diseases. In this work, we focused on quantify- ing protein abundance in a human body via protein microarray technology. The current chapter will present the biological context relevant to this thesis.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	One of the key principles behind the scientiﬁc approach is being open to new evidence that contradicts established doctrines. However, there is one scientiﬁc dogma that remained present in the discourse over the years – the central dogma of molecular biology. Coined by Francis Crick in 1957 and published in 1958 [7], it states that DNA in the cell nucleus gets transcribed into ribonucleic acid (RNA), which in its turn is used to produce proteins. Although being proven wrong on a number of occasions, e.g., transcription factor proteins that regulate RNA production, the central dogma remains a useful approximation for the most important biological interactions in a cell.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	DNA is a long double-stranded molecule that consists of four nucleotides: ade- nine (A), thymine (T), guanine (G), and cytosine (C). Nucleotides form pair-wise bonds (A with T and C with G), helping to hold two strands of DNA together. Hu- man DNA is made up of approximately 3.6 billion nucleotide pairs forming our complete genetic blueprint. Albeit an impressive number, only a fraction of DNA has been associated with relevant biological functions in the organism. These functional regions are called genes. The role of the majority of DNA remains largely unestablished. Because DNA is a static molecule that never leaves the cell nucleus, to execute biological functions, genes need to send instructions to the rest of the organism. They do it via the process of transcription, which transfers information stored in genes into a molecule called RNA. RNA travels to structures called ribosomes, where both take part in synthesising proteins.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	modest 20,000 proteins, if assumed that one gene is responsible for one protein (i.e., canonical proteins) to hundreds of thousands if the combinatorial nature of gene expression and alternative splicing is taken into account [10]. While some proteins leave the native cell to operate elsewhere (e.g., pancreas produces in- sulin to help regulate blood sugar level), others remain to help facilitate domestic processes, including regulating the gene expression. Transcription factor proteins bind to DNA and either suppress or enhance RNA production of nearby genes, di- rectly violating the basic premise of the central dogma of molecular biology. This creates a cycle of regulation, where genes create RNA that initiates the production of proteins, which in their turn regulates the gene expression.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Proteins do most of the work in cells and tissues and are required for many critical processes in the body [8]. For example, the immune system employs special types of proteins – antibodies to recognise foreign substances and ﬁght infections. Most of the work presented in this thesis is focused on antibodies and their role in immune response, therefore the following section will dive into immunity.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Despite continuous advances in the understanding of the immune system, the sub- ject of the matter remains so vast and complex that we consider an in-depth dis- cussion of immunity to be well beyond the scope of this thesis. The section below is meant to introduce readers to the most central concepts that are essential for the autoimmunity discussion that will follow.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	The term immunity comes from Latin imm¯unit¯as, which referred to legal pro- tection offered to Roman senators during their time in the ofﬁce [11]. In biology, immunity is deﬁned as the defense system of an organism from infections and other intruders. The immune system is a large network of cells, tissues, and or- gans that tirelessly work together to protect the organism from anything that is recognised as an ‘invader’ or ‘foreign’, for example, bacteria, virus, parasite, can- cer cell, or toxin [12]. While not all invaders are necessarily harmful, those foreign substances that have disease-causing potential are called pathogens.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 1. Immune system can be broadly divided into two subsystems: innate and adap- tive immunity. Two systems supporting each other in their own ways to ﬁght intruders. With the latter being slower and more speciﬁc, while the former is faster and more gen- eral.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	The two main types of cells involved in the adaptive immune system are B and T cells. While the T cells target intracellular pathogens by directly trigger- ing cell death mechanisms in the pathogen-infected cells, the B cells are mainly involved in targeting extracellular pathogens. After recognising foreign antigens on the surface of the pathogen, B cells differentiate into plasma B cells that pro- duce antibodies. An antibody is a large Y-shaped protein that binds to a speciﬁc antigen. The structure of a typical antibody is presented in Figure 2. Produced antibodies bind to the cognate antigen which results in neutralization of this par- ticular pathogen. Remaining antibodies are then circulating in the bloodstream where they can initiate the immune response.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 2. Structure of a typical antibody. Antibodies use their variable regions (coloured in red) as arms to bind to potential intruders. By binding to a molecule, antibodies send a signal to other immune cells to take action against the intruder.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	or are mild [15]. However, more systematic failures accumulate and may lead to serious damage, causing various pathologies or even, in some circumstances – premature death, and more so among women [16].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	partly T cell-dependent, with defective T cells contributing to the production of autoreactive B cells [20, 21]. In healthy situations, B cells produce antibodies for our protection against pathogens but in case of bypassing defensive mechanisms of the immune system, self-reactive B cells produce antibodies that tag the organ- ism’s own cells and may trigger autoimmunity. Such self-reactive antibodies are called autoantibodies. Thus, failure to recognise and mitigate self-reactive T and B cells may potentially lead to the accumulation of autoantibodies and result in autoimmune disorders [22].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Autoimmune diseases affect about 5% to 7% of the world population, with the ma- jority of patients being women [23]. The most famous examples of autoimmune diseases are type 1 diabetes (T1D), celiac disease and multiple sclerosis [24]. For instance, the onset of type 1 diabetes is caused by an autoimmune reaction against insulin-producing β cells in the pancreas [25], exogenous gluten proteins are linked to the autoimmune reaction in celiac disease [26, 27] while patients with multiple sclerosis harbor autoantibodies against myelin i.e. fatty tissue in the brain and spinal cord that facilitates neurotransmission [28].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	the autoimmune disorders – autoimmune polyendocrinopathy- candidiasis-ectodermal dystrophy (APECED) or autoimmune polyendocrine syn- drome type 1, is of particular importance for the researchers and doctors who study autoimmunity. APECED is a rare disorder caused by a small modiﬁcation in genetic code. Mutations in the AIRE gene alter negative selection mechanisms that normally prevent the body from producing harmful autoantibodies. As a re- sult, a wide range of autoantibodies is released into the bloodstream, causing var- ious types of damage to the organism’s own cells and tissues. Many of these au- toantibodies that are produced in APECED are shared with other diseases such as autoantibodies against β cells in T1D [2]. Precisely due to such high diversity of self-reactive antibodies, APECED is considered an important disease model that helps to understand the processes that drive autoimmunity in general [29]. The study of APECED is central for two publications included in this thesis, which we will examine in later chapters.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	One of the earliest attempts to experimentally detect the presence of an anti- body engaged in a reaction against a patient’s own tissue was successfully carried out in 1955 by Dr. Henry Kunkel [36]. Dr. Kunkel used antibodies tagged with a ﬂuorescent marker (also known as secondary antibodies) to detect autoantibodies in lupus erythematosus cells extracted from the serum of patients with systemic lupus erythematosus disease [37]. Later, a number of simpler and more accurate methods have been developed: radiobinding assay, western blot, and enzyme- linked immunosorbent assays (ELISA). These methods allowed for the detection of antibodies associated with a pre-deﬁned antigen [38]. A need for an a pri- ori hypothesis about the antigen was a signiﬁcant limiting factor, preventing the discovery of autoantibodies against new previously deemed unrelated antigens. Therefore, in order to discover novel autoantibodies, researchers needed a way to screen a much wider range of potential candidate molecules in a high-throughput manner [24].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	DNA microarray technology was developed in the 1990s by the American re- searcher Patric O. Brown and has revolutionised the analysis of biological sys- tems [38, 39]. DNA microarray is a collection of microscopic DNA fragments, short sections of genes printed on a solid surface [38]. A biological sample con- taining ﬂuorescently labeled complementary DNA or RNA molecules can be then applied to each array enabling researchers to quantify the expression of each gene i.e. gene productivity. DNA microarrays were the ﬁrst high-throughput technol- ogy enabling quantiﬁcation of gene expression in a parallel fashion using minimal sample input requirements [38]. Soon after the ﬁrst DNA microarrays appeared it was demonstrated that similar technology involving protein binding molecules can be used to estimate the number of proteins, including autoantibodies from pa- tients’ blood [38]. Therefore DNA microarrays have played a pivotal role in the emergence and the development of protein microarrays.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	number of clinical applications steadily growing [42]. In contrast to functional microarrays, analytical or capture arrays utilize panels of antibodies attached to the slide to detect and measure proteins from the sample [41]. Instead of print- ing an arbitrary set of antibodies or proteins on the slide, in reverse-phase protein microarrays (RPPA or also known as lysate arrays), all proteins from a speciﬁc cell interior (lysate) are printed and the antibody binding from the sample is de- tected [41, 43]. In this thesis, we are going to focus solely on functional protein microarrays that are used to detect and measure the abundance of autoantibodies in the blood [38].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	In order to use functional protein microarrays to accurately quantify the amounts of autoantibodies, at ﬁrst, a sample in the form of plasma, serum or other solu- tion from a human subject is applied to protein microarray slides. Autoantibodies from human sample bind to proteins immobilized on microarray surface in spots or tiny cavities in the glass. Array slides are then washed and dried to get rid of all molecules that failed to bind and remained on the surface. Later, autoan- tibodies that have genuinely bound to spotted proteins are detected by applying ﬂuorescently labeled secondary antibodies. Microarrays are then again washed and dried. Fluorescent signals coming from each well are acquired with a mi- croarray scanner and later analysed using computer software [34]. The amount of light coming from each well is associated with the levels of autoantibody binding to the speciﬁc protein (from this well). Schema of a typical protein microarray is presented in Figure 3.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Fabrication and further handling of functional protein microarrays is a complex process. It involves multiple consecutive steps including printing and immobiliza- tion of proteins on the slide surface, incubation with a sample, repeated washing, and drying and scanning of arrays. Hence, a substantial number of technical fac- tors inﬂuence the quality of the resulting autoantibody binding signal [44]. Uncal- ibrated printing machinery may result in uneven distribution of proteins in spots or in proteins being carried over to neighbouring sites by the printing pin. Irregular washing and drying of the slides can also cause the variation of the actual pro- tein content [45]. Several studies have reported a cross-talk between neighbour- ing spots that resulted in unlikely highly correlated signals between neighbouring spots [2, 46]. Technical variability introduced by mechanical liabilities can mask the true underlying biological signal [47]. Therefore, sufﬁcient care must be taken in design, fabrication, and subsequent analysis to account for possible technical biases.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 3. Schema of the exemplary protein microarray. In functional protein microarrays, proteins and protein fragments are printed inside spots (blue circles) arranged in rows, columns and blocks.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Scientiﬁc in 2014. ProtoArray® includes more than 9,000 full-length human pro- teins as well as several thousand control proteins. All proteins are spotted twice on the array to enable quality control. More than 6,100 proteins that are included on the chip are potential drug targets and thus, relevant to disease processes. Data from ProtoArray® experiments can be analysed by a number of publicly available tools including the manufacturer’s own – ProtoArray Prospector. ProtoArray® has a broad spectrum of potential applications, including discovering novel dis- ease biomarkers via analysing autoimmune reactions, discovery, validation, and development of novel drug targets [44]. Nevertheless, the majority of studies that used the technology have focused on discovering and characterizing autoimmune targets from the blood in a speciﬁc disease [1, 2, 24, 33, 34, 48].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	available data remains valuable for many researchers who plan to either reproduce old results or make their own analysis of protein microarray data.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	HuProt human protein microarray. HuProt is an actively maintained alterna- tive platform to ProtoArray®. Initially developed by CDI laboratory at John Hop- kins University in 2012 [44, 51], HuProt contained 16,368 full-length functional proteins, representing 12,586 protein-coding genes [44, 52]. At the time of writ- ing, the most recent Human Protein Microarray v4.0 in total contains more than 21,000 unique human proteins and protein variants, covering more than 81% of the canonical human proteome. Similar to ProtoArray®, HuProt has mostly been used for detecting and evaluating autoimmune reaction in patients across vari- ous disorders, including: primary biliary cirrhosis [51], ovarian [53] and gas- tric cancers [54], neuropsychiatric lupus [55] and Behcet’s syndrome [56]. As a data generation platform, HuProt was shown to be susceptible to similar biases as ProtoArray®, including non-speciﬁc binding and printing contamination [44].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Custom protein microarrays. Both equipment and constituents necessary for creating custom protein microarrays are commercially available from private and public vendors. Also, fully assembled protein microarrays are available on the market. Coverage of commercial microarray products ranges from arrays with few dozens of carefully selected proteins to vast collections that include almost the entire known proteome [38].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Protein microarray chips are high-throughput platforms capable of measuring thousands of protein interactions in parallel across multiple samples [57]. Compu- tational analysis of a typical protein microarray study starts with acquiring high- resolution images of stained protein arrays, using the special microarray scanner (e.g. GenePix Microarray scanner). Signal information about each spot on the array is then extracted into a ﬁle by segmentation and registration software (e.g. GenePix Pro 7). Each array usually results in one ﬁle. Data from such ﬁles is then used to assemble a data matrix, which serves as an input to the analysis pipeline. Each column in such a matrix represents an individual sample while a row is as- sociated with a protein. This initial matrix is called raw data, as it has not been “puriﬁed” by pre-processing methods. Techniques such as background correc- tion, signal transformation, outlier detection, and normalisation are essential for further statistical analysis as they help to remove or at least minimise the effects of technical noise and thus, enable a fair comparison between samples. Normalised and pre-processed data can be visualised and explored further. If a study follows a case-control design [58] (including Publications I, II, and IV in this thesis), it is possible to compare protein signals in patients with those in healthy individu- als. This enables researchers to pinpoint proteins in which concentration levels can reliably differentiate patients from controls. Such differential proteins can both be used as clinical biomarkers as well as reveal important insights about the mechanisms of the disease.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Protein microarray analysis has notably beneﬁted from the set of analytical methods developed for DNA microarrays, as both technologies enable measure- ment of numerous molecules immobilised on the slide surface, e.g. the array scanning approach [47]. But as it soon turned out, not all statistical methods designed for the DNA microarrays may directly be applied to the protein microar- rays, as the latter is grounded on different biological assumptions. Namely, DNA microarrays assume comparable levels of gene expression across individuals re- gardless of their clinical condition. While this may be true for genes, the number of autoantibodies present in the blood of a healthy person and a patient is vastly different [44, 47, 57]. Such discrepancy between assumptions has motivated the use of a different normalisation strategy from the one used in DNA microarrays, which will be discussed in the sections below.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 4. A rough outline of data analysis methods used for protein microarray experi- ments. Raw data after the acquisition is assembled into a large matrix. This matrix passes through a number of steps, including background correction, signal transformation, out- lier detection, and normalisation before being used as an input for statistical analysis and machine learning. Next, results of statistical analysis and machine learning modelling are interpreted in the context of the existing body of knowledge.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	machine learning methods alongside classical statistical algorithms described in previous publications with a goal to increase the choice of methods available for analysis. Below, we describe each essential part of our pipeline in detail.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	individual array are saved into GenePix Result (GPR) ﬁle – a de facto standard format for storing protein microarray data [24]. Each collected sample normally corresponds to one GPR ﬁle. Typical protein microarray studies collect dozens or even hundreds of samples, resulting in a corresponding number of GPR ﬁles.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	GPR ﬁles are text ﬁles in disguise, hence, they are tab-delimited text ﬁles that can be read by most popular spreadsheet programs such as Microsoft Excel. GPR ﬁles contain a header with relevant meta-information about the experiment and the data matrix, which contains raw ﬂuorescent intensity values of each spot on the chip. If several ﬂuorescent molecules with different wavelengths were used in the experiment, foreground and background signals are measured and reported for each. Different types of arrays may have vastly different contents of both meta- information and intensity matrix. In this thesis, mostly GPR ﬁles from ProtoArray and HuProt platforms were analysed, thus we will focus on them.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	At the beginning of the protein microarray analysis, researchers extract individual raw signal values from GPR ﬁles and combine them into a large matrix of raw data. Multiple studies have shown that raw protein microarray data should not be used directly in the computational analysis [24, 44, 47]. Various technical issues discussed previously can introduce a signiﬁcant amount of noise into the output signal [24, 44, 47]. This noise can hinder the analysis by masking the true signal, rendering the ﬁnal results indecisive. Therefore, raw data must be carefully pre- processed prior to any further analysis. Pre-processing helps to identify and get rid of technical noise at the same time preserving valuable biological signal. Below we discuss a set of common pre-processing strategies, which can be applied in various orders depending on the experimental setup.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	In functional protein microarrays, proteins of interest are immobilized in rows and columns on the glass surface forming a grid of spots [59]. After a slide is incu- bated with a sample solution (usually blood), autoantibodies bind to immobilized proteins. One of the technical challenges related to correctly quantifying the ﬂuo- rescent signal emitted by the labeling antibody is to discriminate signal produced by the genuine biological reaction and local residuary background light [44, 59]. The true signal is usually derived by subtracting the median background intensity of the spot, i.e. background signal, from the amount of ﬂuorescent signal reg- istered within the spot, i.e. foreground signal (Figure 5) [24]. However, several other more elaborate background corrections methods have been proposed in the past [45,60,61]. For example, it has been suggested that instead of using an imme- diate local background intensity it is beneﬁcial to consider a wider neighbourhood of the spot when calculating the median background signal [45].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 5. Foreground signal, background signal and wider neighbourhood of the spot. The ﬂuorescent signal of the protein is determined by subtracting the median of the im- mediate background of the spot from the median of its foreground signal. Alternatively, background value can be calculated taking a wider neighbourhood of the spot into con- sideration.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Another commonplace practice in biomedical research is to apply one of the data transform techniques. For example, log-transformation is known to make fold changes symmetric around zero, reduce the skew in the data, and provide a good approximation for the normal distribution – desirable property for many methods [24,62], including protein microarray speciﬁc normalisation strategy that we are going to discuss in a later section. In spite of some researchers revealing negative effects of logarithm-based data transformation [62], it remains popular and was used in a number of recently published works related to protein microar- ray analysis [46,48], including Publications I and II included into this thesis [1,2].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	In order to satisfy assumptions imposed by most of the statistical methods, protein expression levels registered by a panel of protein microarrays should follow a normal distribution: most of the values being close to the average signal with few very low and high values in the tails. However, in practice, due to a multitude of technical factors e.g. inattentive handling of microarray slides, sample quality, or manufacturing errors, proteins can exhibit expression levels vastly inconsistent with the rest of the data. Such proteins are known as outliers or “anomalous data points”. In protein microarray experiments, both individual proteins and entire microarray slides may exhibit abnormal expression levels and thus considered anomalous. The presence of outliers can be unfavourable for the downstream analysis and resulting conclusions [63]. Hence, once detected, such values usually are either removed completely or substituted with a reasonable approximation e.g. average or median of the corresponding protein.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Several methods exist to automatically detect outliers. One of the most popular and suitable for data that follows normal distribution is to label as outliers all data points that fall outside three standard deviations from the corresponding mean (i.e. three-sigma rule or empirical rule). The common assumption is that such extreme values are unlikely to be generated by the same biological process as the rest of the data. This reasoning is based on the deﬁnition of the normal distribution, for which 95.45% of its data lie within two standard deviations from its mean, while 99.73% within three standard deviations. If the value is either larger or smaller than the aforementioned threshold of three standard deviations, it has only 0.27% of the chance to come from the same distribution as other values. This line of thought is valid only in case data follows the normal distribution. In other cir- cumstances, the above calculations may not apply. In practice, it has been shown that protein expression proﬁles vary a lot between individuals rarely resulting in signal values that follow the normal distribution [1,2,64,65]. Therefore data from protein microarray experiments must be transformed (e.g. using log-transform) if the three-sigma rule is to be applied.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	plots have at least two relevant merits. Firstly, they offer a natural way to visualise the data, and secondly, they can be used to detect outliers in a way that is indiffer- ent to the underlying data distribution. Boxplot summarises data using ﬁve quan- titative measures: minimum, three quartiles, and maximum. While the second quartile (Q2) corresponds to a median (middle value), the ﬁrst (Q1), and the third (Q3) quartiles enclose the ﬁrst 25% and 75% of data distribution respectively. The distance between the ﬁrst and third quartiles is called the interquartile range (IQR) and given by Q3 −Q1. Genuine data points must be larger than Q1 −1, 5∗IQR and smaller than Q3 + 1, 5 ∗ IQR. Data points outside this range are considered to be outliers. Boxplots are usually rendered as rectangles (hence the name “boxplot”) with a ﬁxed width, and length equal to IQR, with outliers, visualised as circles outside of the box either at the top or bottom of the ﬁgure. Although boxplots can be plotted side by side to compare distributions of multiple features, they are not suitable for identifying outliers from multivariate data (i.e. data points character- ized by more than one feature).
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Clustering techniques in combination with various visualisation strategies can be used to recognise outliers in multivariate data, such as protein microarray read- outs. Hierarchical clustering is an algorithm that recursively groups data into clusters based on a predeﬁned distance metric e.g. Euclidean distance [67]. The result of hierarchical clustering is a dendrogram. Dendrogram visually shows the arrangement of clusters produced by the algorithm. Anomalous samples or pro- teins will stand out far from the rest of the clusters on the dendrogram, making them easy to spot and remove. Dendrograms are often coupled with another visu- alisation approach named heatmaps (Figure 6). Heatmaps use colour to represent the magnitude of individual signals. Heatmaps supplement dendrograms with an additional context about single expression values. Despite enabling rich visual- isations, hierarchical clustering does not label samples as outliers automatically. One of several metrics can be used on top of hierarchical clustering results to detect clusters and therefore identify outliers (e.g. elbow method and silhouette score). Density-based spatial clustering of applications with noise (DBSCAN) is another clustering method that uses the spatial density of points as a factor for cre- ating clusters [68]. Data points from the low-density regions, far from established clusters are considered to be outliers or noise. Therefore, unlike hierarchical clus- tering, DBSCAN can detect outliers without human intervention. However, DB- SCAN requires several key parameters to be ﬁxed to work. Although DBSCAN is the only technique mentioned in this section that was not explicitly applied in publications included in this thesis, we consider it to be an important addition, potentially valuable for the readers that may decide to use it in their work.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	One of the primary goals of protein microarrays is to compare the amount of bind- ing between individual samples or groups of samples (e.g. healthy and controls).
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 6. An example of a heatmap obtained from a protein microarray experiment using ClustVis tool. Each cell of the heatmap is coloured based on the signal level of the corre- sponding protein (in rows) across all samples (in columns). Additional meta-information available about samples and proteins is visualised using extra colour legends. Hierarchi- cal clustering results are presented in the form of dendrograms at the top and on the left of the heatmap.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	For this process to yield realistic results, the protein binding signal measured from multiple arrays must be comparable. This can be problematic due to the poten- tial difference in the number of proteins printed on the slides and other techni- cal factors that can introduce systematic biases [45]. Such biases may signiﬁ- cantly distort or shift signal distribution for some or all proteins on one or several protein microarrays. Therefore, unlike the above-mentioned data pre-processing approaches, signal normalisation acts globally combining information about sig- nal variation from all arrays and proteins to successfully eliminate non-biological differences. The most commonly used approaches for protein signal normalisa- tion were adopted from the DNA microarray context. These techniques, make strong assumptions about underlying signal distribution, which are not always in line with biological mechanisms at work in protein microarrays [24, 47, 60]. Pro- tein microarray-speciﬁc normalisation strategy based on robust linear model [47] makes use of control proteins printed on each array and block. Control proteins can be positive or negative, but in either case, they are assumed to exhibit constant signal levels across all samples. Any differences in signal values of these proteins are considered to be technogenic and thus, corrected for. Below we present some of the most popular approaches for signal normalisation implemented in the com- putational tools used for protein microarray analysis [3, 44, 69].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Robust linear model. Robust linear model (RLM) [47] makes use of special sets of proteins – controls that are often built into protein microarrays to enable normalisation of the signal. Controls can be either positive, that are guaranteed to have a high signal regardless of the blood content, or negative that should not react with serum under any circumstances, for example on microarray slides it is common to use empty spots as negative controls. Controls are present on every array as well as in every block of proteins on each array. Any signiﬁcant deviation in the control signal may indicate the presence of unwanted noise or bias. Hence, RLM is employing controls to quantify and control for potential biases associated with arrays and blocks.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	for all possible i and j values. Figure 7 describes the normalisation process using RLM. In publications presented in this thesis, RLM normalisation was imple- mented in R, using functions from MASS [74] and limma [75] packages.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 7. Normalising protein microarray signal using the robust linear model and control proteins. First, intensity values of the control proteins are modelled as a linear combina- tion of the array (α), block (β ), type of control protein, and noise using a robust linear model (purple line). The resulting coefﬁcients for arrays and blocks are considered to be associated with unwanted technical biases. Then the normalised signal is obtained by subtracting the corresponding coefﬁcients associated with a speciﬁc array and block.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	where y0 is modelled signal of the corresponding control protein. Random noise ε is sampled from the normal distribution for each control protein independently. All in all, in a generic case we get yi jkr = αi + β j + τk + εr, where i, j, k and r are indexes of corresponding array, block, protein type and protein signal. This ﬁnal equation is equivalent to 3.1.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	RLM is considered a preferred normalization strategy for protein microarrays as it exploits protein microarray speciﬁc control proteins and does not assume the near equal signal distribution across arrays [47]. On the other hand, RLM assumes a normal distribution of the underlying signal to work well. Logarithm transfor- mation that was discussed earlier, can be applied to the raw protein microarray data to approximate the normal distribution.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Unfortunately, a source code of the RLM method was not available at a time when work that laid the foundation of this thesis was performed. RLM was imple- mented as part of Prospector software – a standard analysis tool provided by the manufacturer of ProtoArray. At the same time, Prospector had a prohibitive limit on the number of input samples, rendering it futile for larger studies as ours [76].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Later, RLM was introduced in a protein array analyser (PAA) – an R package developed by Michael Turewicz [69]. However, for a signiﬁcant stretch of time RLM was nowhere to be found and thus, we had to create our own implementa- tion of the RLM normalisation module for publications I, II, and later released it to the research community as part of the web-tool (publication III of this thesis). The absence of a source code, as well as a pseudo-code of the RLM method in the original publication [47], made re-implementation of the RLM one of the most challenging part of the protein microarray analysis pipeline built in this thesis.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	After protein data was properly pre-processed, relevant statistical methods can be applied. Statistical analysis is a vast ﬁeld with a large number of techniques avail- able at researchers’ disposal. Characteristics of data and the research question determine the choice of the statistical method. For example, often researchers are looking for proteins that are capable of reliably distinguishing between two (or more) groups of samples, e.g. disease versus controls. Such protein features, in which intensity levels are sufﬁciently different between studied conditions are considered signiﬁcantly differential and can be used as important biological mark- ers suggesting the presence or absence of the condition in question often called outcome variable. To assess the relationship between the intensity of a single protein and the outcome variable, univariate analysis tools are used. If simple uni- variate analysis yields no results or there is a good reason to believe that multiple proteins in combination can be predictive of the sample’s outcome, multivariate analysis can be performed. Finally, once inﬂuential proteins are identiﬁed with either multivariate or univariate techniques, the enrichment analysis can be used to discover their common properties. The following sections focus on statistical methods used in this thesis, while adjacent methods are described only brieﬂy.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Although statistical tests that we are going to talk about further, work slightly differently, some basic notions remain universal. The common starting point in hypothesis testing is deﬁning a baseline or a null hypothesis (H0) – a general state- ment about the absence of the assumed phenomenon. For example, H0 may be formulated as observing no difference between means of protein intensity signals of two groups of samples. Namely H0 : µ1 = µ2, where µ1 and µ2 are means of the group 1 and group 2 respectively. An opposite to H0, alternative hypothesis (H1) thus can be formulated as H1 : µ1 (cid:54)= µ2. The statistical test usually results in either rejecting the H0 and thus favouring the alternative hypothesis or failing to reject the null hypothesis. In either case, employing statistical tests inevitably entails a number of background assumptions, which are made either about the data or about the ways how the data has been gathered. An example of a data-based assump- tion could be that protein intensity signals follow a normal distribution. Invalid assumptions may lead to invalid test results, therefore it is of utmost importance to establish correct assumptions and choose an appropriate test.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Statistical tests use a numeric quantity derived from data to perform the hy- pothesis test. This quantity is commonly referred to as test statistic. The observed value of the test statistic can be calculated from the data at hand and compared with a known theoretical distribution of the test statistic under the null hypothesis (null distribution). If the observed value of the test statistic is at the far ends of the distribution i.e. either much larger or smaller than most of the values in the distri- bution, it is considered to be sufﬁcient evidence for rejecting the null hypothesis. However, instead of relying on vague notions such as “far ends” or “much larger”, researchers compute a probability of the observed value of the test statistic to be sampled from the null distribution - a p-value. If the p-value is less than some pre-deﬁned threshold value, the corresponding H0 should be rejected. Common threshold values are 5% and 1%, more about this in the following sections.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Finally, most of the statistical tests can generally be divided into two large categories: one-sample and two-sample tests. A one-sample test explores the pos- sibility of the mean of the sample being statistically different from the known population mean. Two-sample tests are used to assess the signiﬁcance of the dif- ference between means of two groups (e.g. patients and controls). Two-sample tests can be paired or unpaired (or independent). The unpaired two-sample test assumes no overlap between tested groups (e.g. two independent groups of mice). As it follows from the name, a two-sample test can compare only two groups. If there is a need to estimate the signiﬁcance of the difference between three or more groups, an analysis of variance can be performed (also known as ANOVA). The sizes of groups to be compared and corresponding variances also inﬂuence the choice of the test. In this thesis, we make use of both one-sample and independent two-sample tests, which are discussed further in more detail.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Many studies involving protein microarrays follow case-control study design (in- cluding Publications I and II in this thesis), where protein concentrations in pa- tients can be compared to those in healthy individuals [58]. Proteins that can reliably differentiate between patients and controls are referred to as differential and process of identifying such proteins – differential analysis [24, 57]. More for- mally differential proteins are the proteins for which the probability to observe the corresponding value of the test statistic to be sampled from the null distribu- tion is below the acceptable signiﬁcance threshold leading to rejection of the null hypothesis.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Differential proteins can be used to explore the mechanisms of the disease (such as in Publications I and II) or as a screening tool – measuring autoantibody reaction to these proteins in the general population can ideally reveal individu- als at risk (Publication IV). These individuals could be treated early and less ag- gressively, increasing their chances for long-term well-being. Below we discuss different statistical methods used in this thesis to detect differential proteins.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Z-score analysis. A substantial number of protein microarray-based studies (including Publications I and II in this thesis) have used an approach referred to as Z-score analysis or Z-test to deﬁne differential proteins [1, 2, 46, 57, 59].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	x − µ0 σ This approach is similar to the three-sigma rule described in the outlier detec- tion section, as protein concentrations with a Z-score of more than 3 or less than -3 in one or more samples are considered to be signiﬁcantly differential [1, 2, 57]. However, due to the potentially high number of tests, such a strategy may lead to a substantial number of proteins deemed differential by mistake (see a section on multiple testing correction). Such risk can still be justiﬁed in the case the goal is not to identify proteins that are consistently differential across all patients, but proteins that have abnormally higher concentration values only in a handful of patients. This is the case for APECED patients, who develop heterogeneous sets of autoantibodies that may differ from patient to patient and thus target vastly different sets of proteins spotted on protein microarray slides.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Two main assumptions should be met in order for Z-test to be applicable: Z- scores should follow a normal distribution and population parameters i.e. mean and standard deviation should be known in advance, which is not always possible. Often population mean and standard deviation can be estimated using the sample mean and standard deviation, which transforms Z-test into a t-test.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Student’s t-test. The Student’s t-test (or a t-test) is one of the most popular statistical approaches for hypothesis testing. The test was named after William Sealy Gosset who published the method under the alias “Student”.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Above, sp is a pooled standard deviation (3.6), n1 and n2 represent the number of samples in group 1 and group 2 respectively, while s1 and s2 are the standard deviations of these two groups. The formula for t (3.5) is valid as long as there is a good reason to believe that groups have similar sizes and are sampled from the populations with equal variances. In other circumstances, slightly different formulas for t and sp must be applied. From the deﬁnition, it follows that t is the distance between group means in units of pooled standard deviation. If the null hypothesis is true the value of t should be close to 0, suggesting no difference between the two means. However, the larger the value of t, the less likely is H0. Thus, using computed t-value it is possible to quantify the p-value by comparing t to a null-distribution. If the p-value is less than a predeﬁned threshold, the null hypothesis is considered to be false and can be rejected. It is common to use 0.05 as a threshold imposed on p-values. Rejection of H0 under 0.05 threshold can be interpreted as that there is less than 5% chance that the observed difference between groups is due to random chance.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	powerful alternatives to classical t-test can be used, such as moderated t-test [77] or Mann–Whitney U test [78] discussed in the next section.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	The t-test can be expressed in terms of linear models discussed in the previ- ous sections and can be formulated as y = Xw + ε (3.2). Here we will pay no attention to ε term as it is independent for each sample and cannot be accounted for. In our case, X is an indicator of whether a sample was drawn from the ﬁrst or second group and thus can be written simply as xi. The above equation (3.2) can be reformulated as follows: yi = xiw, where yi is predicted signal value of i-th sample. This equation can be further simpliﬁed yi = w0 + w1 ∗ xi. If an i-th sample is drawn from the ﬁrst group, xi becomes 0 and the whole equation transforms into yi = w0 + w1 ∗ 0 or simply yi = w0. Hence, w0 is a predicted signal for the samples in the ﬁrst group. Since the best way to summarise a set of points is via their mean, w0 represents a mean signal of the ﬁrst group. When sample is from the second group, the xi equals to 1 and thus the core equation changes to yi = w0 + w1. With this, we model the second group by adding w1 to the mean of the ﬁrst group w0, and therefore w1 is the difference between the means of the two groups. The null hypothesis can be formulated accordingly as H0 : w1 = 0. Not only the linear model formulation of the t-test can help to understand the procedure better, but it also facilitates the implementation using programming languages. Various statis- tical software packages e.g. limma in R, uses linear model formulation as a basis for t-test implementation. T-test-based expression analysis performed and pre- sented in Publication II of this thesis was implemented in limma and formulated in terms of the linear model.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Mann–Whitney U test. Mann-Whitney U test (also known as the Wilcoxon rank-sum test) is an alternative to the two-sample equal variance t-test discussed before [78]. Contrary to the t-test, the Mann-Whitney U test belongs to the family of non-parametric tests that do not rely on any particular parameterized distri- bution for hypothesis testing. Therefore, it is applicable to data that does not necessarily follow the normal distribution as is often the case in biology. Under null hypothesis H0 the two compared distributions should be considered equal. More formally, the probability of an observation from the ﬁrst group to be larger (or smaller) than an observation from the second group is not consistently differ- ent from the probability of the opposite, namely, that observation from the second group being larger (or smaller) than an observation from the ﬁrst group. Thus, the Mann-Whitney U test assumes that observations are comparable, i.e. it is possible to say if one is bigger than the other. In linear model formulation, Mann–Whitney U test is very similar to the standard t-test, except the model is built on ranks of x and y instead of actual values: rank(yi) = w0 + w1 ∗ rank(xi). In this thesis, the Mann-Whitney U test has been used in an attempt to identify differential cytokines (Publication IV).
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	data. In the case of two independent groups, we may decide to calculate the dif- ference between two means µ1 and µ2 of two groups with n1 and n2 samples respectively. Hence, d fo = µ1 − µ2 is considered an observed value of the test statistic for the original data. In order to obtain a distribution of the test statistic under a null hypothesis H0 : µ1 = µ2, the permutation test performs the following steps. First, it randomly shufﬂes all the data and assigns n1 ﬁrst observations to the new ﬁrst group. The remaining n2 samples are assigned to the second group. Next, the difference between means of randomly created groups d fr is calculated. If obtained value d fr is larger than d fo, the pre-initialized counter i is increased by 1. Later, the data is reshufﬂed again and all the same, steps are repeated a large number of times (e.g. 10,000), each time a new d fr is computed. To estimate the corresponding p-value, the observed value of test statistic d fo should be compared to the distribution of test statistic under the null hypothesis, thus the distribution of d fr. This can be done by dividing the resulting value of the counter i by the number of repetitions that were performed. For example, if after 10,000 repeti- tions only on seven occasions d fr was larger than d fo, the probability to observe d fr as extreme as d fo under the null hypothesis is 0.0007, which is less than a classical signiﬁcance threshold of 0.05, and therefore small enough to reject the null hypothesis.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	There is no need to calculate all possible permutations of the original data, as this number can be extremely large (e.g. two groups with 30 observations in each will result in 1.18 ∗ 1017 possible permutations) [80]. Instead, a large enough random sample of all possible combinations would be sufﬁcient. The larger the sample, the more precise estimate it will generate. Such a sampling procedure is usually referred to as the Monte Carlo approach. Modern software tools, as well as processing hardware, enable researchers to shufﬂe their data enough times to obtain sufﬁciently precise estimates in almost no time, making randomization tests a practical solution to hypothesis testing.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	In the research presented in this thesis (in Publication II) we used a permutation test to test a hypothesis that proteins targeted by the autoantibodies in the blood of APECED-positive patients originate from genetically more conservative (i.e. those that accumulate fewer mutations over time) regions of the DNA.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	alyze a pool of proteins that are targeted by released autoantibodies, trying to identify properties and functions that are common among the targets. Pinpointing these properties and functions might shed some light on autoantibodies and the autoimmune process, for example, it may provide clues as to autoantibodies’ ori- gin. In general, the process of determining properties that are over-represented in a group of proteins or genes is usually referred to as enrichment analysis. Enrich- ment analysis is often performed by quantifying the size of the overlap between a group of proteins with a known biological property, e.g. proteins expressed in lymphoid cells, and a group in question. A statistical test is then used, e.g. hyper- geometric test, to estimate the probability that this overlap or larger was observed by random chance. If such probability is deemed sufﬁciently low (< 5%), the overlap between groups is considered signiﬁcant and therefore, genuine. In this case, the group in question is said to share the same biological property as a group with which it was compared. A large number of public databases, such as Gene Ontology [49], KEGG [81], Reactome [82], Human Phenotype Ontology [83] and Human Protein Atlas [84] are available with protein and gene groups char- acterized with various biological properties and functions. Hence, in practice, enrichment analysis means comparing the obtained group of target proteins to hundreds or even thousands of groups stored in public databases. A number of po- tential databases and datasets that can be searched to ﬁnd relevant terms has long become prohibitively large for humans to manually work through. Thus many software tools, e.g. g:Proﬁler [85] were developed to automate the enrichment analysis, saving dozens of researchers’ work hours. In this thesis, speciﬁcally in publication II, we have used g:Proﬁler as well as a hypergeometric test to identify enriched terms in the group of targeted proteins.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	where K is the number of such proteins in total on the platform. The number of proteins in our list is n and N is the number of proteins overall on the platform. By inserting all possible values of k (from 0 to n) into 3.7 it is possible to obtain corresponding hypergeometric distribution. By accumulating parts of the distri- bution that correspond to actual kactual observed from the data, it is possible to estimate the probability to observe k as extreme as kactual or larger by chance. If this probability is low enough (less than 0.05) we reject the null hypothesis and assume that there are more proteins associated with a biological property among extracted proteins than what we can expect at random. Although we have used g:Proﬁler to determine biological processes common among targeted proteins, we still applied the hypergeometric test in Publication II to be able to perform en- richment analysis on datasets that were not part of the g:Proﬁler tool and also to cross-check our ﬁndings.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	each test, a p-value will be produced, and the null hypothesis is rejected if the cor- responding p-value is less than 0.05. If we assume the absence of true differential proteins on the platform, 5%, namely 1000 (0.05*20,000) proteins will be consid- ered signiﬁcant simply by chance (because the number of tests is so large). For example, for the ProtoArray study discussed in Publication II, this would mean that half of the proteins in the positive group are false. Therefore, the number of tests needs to be taken into account when performing statistical analysis. A num- ber of methods have been proposed to adjust the classical statistical signiﬁcance threshold of 5%.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Bonferroni correction. The simplest multiple testing correction approach is called the Bonferroni correction [87]. This method adjusts the p-value threshold by dividing it by the number of experiments. In the above example of HuProt array with ∼ 20000 proteins and threshold of 0.05, the new signiﬁcance threshold becomes 0.05/20000 = 2.5 ∗ 10−6. Applied to n tests with a p-value threshold of α Bonferroni correction ensures that the probability of observing at least one false signiﬁcant result is α [88], while it is usually sufﬁcient to optimize for some acceptable proportion of false predictions. Thus this procedure was commonly regarded as overly strict for most of the practical applications [86, 88].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Benjamini and Hochberg correction. The method by Benjamini and Hochberg (also known as FDR correction) attempts to keep the number of falsely signiﬁcant results at a certain predeﬁned level (e.g. 5%) [89]. The proportion of falsely admitted associations (false positives) among all signiﬁcant results is called a false discovery rate. The method starts with ordering all m unadjusted p-values in a descending order [86]. Then for the i-th p-value pi, the algorithm checks if this value is less or equal to (i/m) ∗ α, where α is acceptable level of FDR. As soon as such pi was found consider it to be a signiﬁcant threshold. We used the method by Benjamini and Hochberg to correct unadjusted p-values for multiple testing in publications II and IV.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 8. Decision tree algorithm uses information about individual protein intensities (PA and PB) in order to explain the outcome variable. For example, the decision tree algo- rithm may assign a class ”control“ to a sample for which protein A (PA) has an intensity value less than 4.5.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	learning algorithms have been developed for each of these categories [91]. Some of the most popular machine learning methods capable of working with both con- tinuous and discrete outcome variables are decision trees [92], and random for- est [93].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	The decision tree algorithm. Decision tree algorithm [94] uses values of input features (e.g. protein intensities) to infer the value of the outcome variable. A decision tree has a recursive structure with a root at its origin and leaves at the bottom. Each node can potentially have two children nodes. All nodes except leaves encode conditions in a form of questions. For example one of the nodes may inquire if the normalised intensity of protein A has a value of more than 4.5 in order to decide on a value of the outcome variable (Figure 8). The tree starts with checking if the input data satisﬁes the initial condition at the root and descends down the tree depending on the outcome. Leaf nodes of the tree determine the outcome of the algorithm: class in the case of classiﬁcation or continuous value for regression. To build the tree the decision tree algorithm uses values of all available features.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Figure 9. Random forest algorithm uses predictions produced by individual decision trees in order to predict an outcome variable. Decision trees that are part of a random forest ensemble are built by randomly discarding a pre-deﬁned number of features and observa- tions. For example, decision tree 1 has been built after discarding protein A intensity and the ﬁrst control sample.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	come variable in parallel. Predictions from the individual trees are then combined together to obtain the ﬁnal joint prediction. Such approach is often called ensem- ble learning or ensembling. The random forest algorithm has another important difference from the decision tree algorithm. The individual trees are constructed using a random subset of input features and input data points (e.g. proteins). For example, it is common to use random 80% of samples for each tree in the forest and when building a tree, nodes are optimized using only random 80% of the fea- tures (Figure 9). This has been shown to make random forest models extremely robust to noise and highly generalizable to unseen data.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	mance when training machine learning algorithms and tweaking their parameters is a very important step in creating a viable data analysis pipeline. The perfor- mance of a model is usually evaluated using independent parts of data (referred to as validation or test set). This approach implies detaching a substantial part of data from the training set, which may not be used for model training. However, if the number of samples is limited, which is usually the case in biological research, creating a separate validation set can be prohibitively data-expensive. Another way to evaluate the performance of the model is called the cross-validation (CV) algorithm. It starts by randomly splitting the initial data set into a ﬁnite number of chunks (folds). The total number of folds depends on the size of the data set, of- ten ranging between 3 and 6. At every iteration of the CV algorithm, the machine learning model is trained on all folds except one, the remaining fold is used as a validation set, allowing to take a snapshot of the model’s performance. Part of the data that serves as a validation set is changed at every iteration, allowing to obtain several performance estimates using one data set. The cross-validation algorithm provides a safe way to estimate the unbiased performance of the model under dif- ferent sets of parameters without letting the model memorize training data. Once the perfect combination of parameters was found and its performance estimated, the model can be trained on the entire dataset. The cross-validation approach was used in the publication IV of this thesis as well as in our previous work [95] to efﬁciently evaluate machine learning models.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	At present, it is impossible to imagine a section on machine learning that would ignore a connection between machine learning and deep learning tech- niques. Deep learning algorithms are machine learning methods that use a popular type of machine learning models – neural networks to tackle the most challeng- ing problems, often previously unsolvable by humans. Since the success of the AlexNet neural network in 2010 [96] these methods have been considered to be the most advanced form of machine learning. Although deep learning approaches are considered to be state-of-the-art in many areas, including biology [97,98], due to several substantial limitations of neural networks (e.g. low transparency and data-intensive nature of training), here we have focused on less complex, yet still powerful machine learning methods that have been used in publication IV of this thesis.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	The AIRE gene has an important role in central tolerance as it is responsible for assembling self-antigens used in T cell maturation [18, 99, 100]. These antigens are presented to T cells during the so-called negative selection phase. Binding be- tween prospective T cells and self-antigens indicates the potentially self-reactive tendency of T cells. As a result, such T cells are deemed dangerous for the organ- ism and normally are removed from the pool of potential immune cells.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Occasional genetic mutations can alter or even cease the function of AIRE, jeopardising central T cell tolerance [100] and thus resulting in the accumulation of self-reactive immune cells in the bloodstream. T cells have been linked with activation of B cells [101] that produce autoantibodies. The theory emerged that genetically deﬁcient AIRE gene may distort not only the selection of T cells but also may indirectly create autoimmune B cells and as a result – disease-causing autoantibodies. Despite such an important role, precise molecular mechanisms of AIRE have remained poorly understood [29]. Genetic alternations of AIRE cause the APECED/APS1 autoimmune condition mentioned previously. This disease is characterised by large numbers of autoantibodies against self-antigens expressed in the peripheral organs present in the patients’ blood [102]. Although this condi- tion is rather rare in the general population (depending on the country it can range from 1 in 25,000 to 1 in 1,000,000), it is often considered as a model disease for human autoimmunity. We, therefore, reasoned that studying autoantibodies from a sufﬁciently large set of APS1/APECED patients may help to gain a better under- standing of the interplay between AIRE and autoimmunity, for example, collect evidence for the hypothesis that speciﬁc protein features may be an indirect cause of B cell autoimmunity [1, 2, 29].
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Overall, eighty-one APS1/APECED patients of ﬁve distinct geographical ori- gins were recruited into the study that was published in 2016 [1]. Some of the patients were sampled several times over the course of the study, resulting in a total of ninety-seven patient samples. Control samples were extracted from nine ﬁrst-degree relatives and twelve healthy volunteers. In total, data from a hundred eighteen samples were analysed. Protein microarray chips from Fisher Scientiﬁc (ProtoArray) were used to quantitatively measure the presence of autoantibodies in these samples.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	and dividing by the standard deviation of the combined healthy and ﬁrst-degree relatives group. We considered a protein to be a positive hit if the corresponding Z-score had a value of 3 or larger in at least one sample. This resulted in a high number of positive targets (3,731) jointly recognised by the group of patients and 406 proteins recognized by controls. This observation was in concordance with a widely held view that APS1/APECED patients are very heterogeneous in the nature of their autoimmune response with only a handful of proteins (such as type I interferon family) being recognised by the majority of patients [64, 65], while other constituting a private set.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	To conclude, this paper had at least two major ﬁndings. Firstly, our statistical analysis showed that APS1/APECED patients as a group develop a unique set of autoantibodies that recognise approximately a hundred body’s own proteins. Al- most all samples contained high concentrations of autoantibodies against a small set of proteins (about 10), such as type I interferon or interleukin-22 (IL22). The remaining proteins were collectively recognised by autoantibodies present only in a handful of individuals. Thus, it was observed that blood from all 81 pa- tients collectively contained autoantibodies against more than 3,700 human pro- teins (which is about 14% of the canonical proteome). Secondly and perhaps even more importantly, the presence of autoantibodies against type I interferons had a surprising negative correlation with type I diabetes. In this publication, our main contribution is of two folds: implementation of a protein microarray speciﬁc data pre-processing pipeline, including re-implementation of the robust linear model for normalisation and identiﬁcation of positive hits using Z-scores. A more com- prehensive analysis of autoimmune targets was performed in Publication II, which will be discussed in the following chapter.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	The comment published in eLife [103] and our subsequent response [104] enabled us to appreciate the complexity involved in devising a precise rule for identifying true-positive autoantibody targets in protein microarray experiments. Where the classical statistical methods are deemed unsuitable for the task due to the stochastic nature of the autoimmunogenesis, the ad hoc solutions may lack the desirable precision. Arguably the most common way to respond to such a chal- lenge is to employ several orthogonal assays to conﬁrm the ﬁndings. Hence, there seems to be a need for the non-parametric technique that would enable robust yet non-restrictive analysis of inherently variable signal intensities often exhibited by the protein microarrays.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Autoantibodies have been shown to play an important role in the onset and pro- gression of various autoimmune diseases [109,110]. Yet, while recently a lot more has been discovered about cellular and genetic factors that contribute to the emer- gence of autoimmunity, our understanding of properties of involved autoantibod- ies remains limited [2]. We made an initial attempt to characterise autoantibodies and their targets in Publication I [1], where we showed that patients’ blood con- tains high concentrations of autoantibodies against a set of well-known proteins. Some of these proteins such as type I interferons (especially IFN-α), became di- agnostic markers for APECED [18]. Autoantibodies against other proteins, like IL17A, IL17F and IL22 were shown to contribute to the onset of chronic muco- cutaneous candidiasis – another distinctive feature of APECED [2, 111]. Previous studies suggested that there are autoantibodies that are shared between APECED and other complex autoimmune diseases, such as Addison’s disease [109] and T1D [110]. But the number of such commonly targeted and widely known pro- teins, is small, comparing to the total number of proteins collectively targeted by autoantibodies in all patients (3,731 in total). Hence, in Publication II we focused on an in-depth analysis of autoimmune targets identiﬁed in Publication I.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	Protein data from the same samples as in Publication I was pre-processed us- ing an earlier developed pipeline with few minor modiﬁcations. As before, the background-subtracted signal (we used basic median local background subtrac- tion) was log-transformed before being normalised using RLM. The resulting values were standardized using the mean and standard deviation of control sam- ples. After several studies that have employed ProtoArrays highlighted a danger of cross-contamination between neighbouring protein spots [46] we decided to add another quality control step into our pipeline. Thus, unlike the workﬂow of the ﬁrst publication, here, we removed 31 proteins with unexpectedly highly cor- related signals (with Pearson’s coefﬁcient of 0.6 or higher) with the expression of nearby proteins or well-known protein targets (e.g. IFN-α). The remaining proteins with a standardized score of 3 or higher in three or more patients were considered to be true positive targets [2]. Additional ﬁltering criteria of three patients were introduced in this work to reduce the number of possible false pos- itives that could distort the analysis. This requirement narrowed down the list of positive proteins from an initial 3,731 to 963, which we later referred to simply as the “positive group”. Although the positive group was signiﬁcantly decreased in size, it remained big enough to include proteins recognised “privately” i.e. by fewer patients, and therefore capable of providing details on the mechanisms of the autoimmunity.
Processed_Developing_a_data_analysis_pipeline_for_automated_.txt	available meta-information about samples. In the course of this research, we looked for biological processes that are over-represented in the positive group. We quantiﬁed the number of single nucleotide polymorphisms (SNPs) and APECED related mutations as well as the level of evolutionary conservation of relevant gene regions. We also performed differential and clustering analysis with respect to as- sociated clinical conditions, which however revealed few notable results. Finally, we ran a longitudinal analysis of protein expression patterns in the positive group. Performing all these experiments involved a number of technical challenges, so- lutions to which, are our main contributions to this publication.
Processed_Mapping_the_Conformal_Window:_SU(2)_with_4,_6_and_.txt	expected to lie on both sides of the edge of the conformal window, where the theory has an infrared ﬁxed point. We observe that the coupling grows with the length scale at four ﬂavors, implying QCD-like behavior. At ten ﬂavors the results are compatible with a Bank-Zaks type ﬁxed point. The results at six ﬂavors remain inconclusive: the running is slow towards the infrared but the range and accuracy of the study are insufﬁcient for determining the existence of a ﬁxed point.
Processed_Mapping_the_Conformal_Window:_SU(2)_with_4,_6_and_.txt	c(cid:13) Copyright owned by the author(s) under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike Licence.
Processed_Mapping_the_Conformal_Window:_SU(2)_with_4,_6_and_.txt	Figure 1: The phase diagram of SU(N) gauge theory as a function of the number of colors, ﬂa- vors and fermion representations (F = Fundamental, 2A = 2-index antisymmetric, 2S = 2-index symmetric, Adj = Adjoint). The shaded bands indicate the estimated conformal windows.
Processed_Mapping_the_Conformal_Window:_SU(2)_with_4,_6_and_.txt	Within the possible phase space of gauge theories there is a group of models with a non- trivial infrared ﬁxed point. In these conformal models, under renormalization group evolution, the coupling runs to smaller values at small distances, exhibiting asymptotic freedom, but runs to a constant at large distances. They have applications in phenomenological model building, such as for technicolor theories [1, 2, 3], where the Higgs sector is replaced with a strongly interacting sector with chiral symmetry breaking. From purely theoretical point of view, mapping the phase diagram of gauge theories in the number of colors N and fermion ﬂavors N f is interesting for understanding their nonperturbative dynamics from ﬁrst principles. Many lattice studies of the conformal window have already appeared in the literature: for example SU(2) with fundamental representation fermions [4], SU(2) with adjoint fermions [5, 6, ?, 8, 9, 10] and SU(3) with fermions in the fundamental [11, 12, 13] or in the two-index symmetric [14], i.e. the sextet, representation. In ﬁgure 1 we sketch a phase diagram for SU(N) gauge theories as a function of N and N f for model with fermions interacting with the fundamental, two-index (anti)symmetric and ad- joint representations of the gauge ﬁeld. The upper boundary corresponds to the loss of asymp- totic freedom, when the ﬁrst coefﬁcient of the perturbative expansion of the beta function is zero: b 0 = 11/3N −4/3N f T (R) = 0, where T (R) is the group theory factor for the fermion representation R. Just below the upper bound the value of the ﬁxed point is expected to be small and perturbation theory applicable. When the number of ﬂavors is lowered, however, the ﬁxed point is expected to move to higher coupling. Finally, as one comes to the lower limit, the critical coupling for the chiral symmetry breaking becomes smaller than the expected ﬁxed point and the model becomes chirally broken. The lower bound is therefore and inexact approximation in a region where pertur- bation theory may not be applicable, and needs to be checked using nonperturbative methods. This provides and interesting challenge for the lattice community.
Processed_Modeling_Multiday_Extreme_Precipitation_Across_Eas.txt	Abstract. The purpose of this paper is to illustrate new techniques for computing mul- tiday extreme precipitation taken from recent theoretical advancements in extreme value theory in the framework of dynamical systems, using historical precipitation data along the eastern coast of Australia as a case study. We explore the numerical pitfalls of apply- ing standard extreme value techniques to model multiday extremes. Then, we illustrate that our data conforms to the appropriate setting for the application of recently derived extreme value distributions for runs of extremes in the dynamical framework and adapt these to the non-stationary setting. Finally, we use these distributions to make more informed predictions on the return times and magnitudes of consecutive daily extreme precipitation and find changes in the dependence of increasing consecutive daily rain- fall extremes on the Southern Oscillation Index. Although our case study is focused on extreme precipitation across eastern Australia, we emphasize that these techniques can be used to model expected returns and magnitudes of consecutive extreme precipitation events across many locations.
Processed_Modeling_Multiday_Extreme_Precipitation_Across_Eas.txt	Summary of extremes in rainfall across Australia. In recent decades, eastern Aus- tralia has experienced flooding attributed to multiday extreme rainfall, for example, the historic 1974 floods in Brisbane, the Queensland and Victoria floods occurring through 2010 and 2011, and the 2022 flooding across New South Wales and Queensland. These catastrophic flooding events are considered some of the most substantial economic risks to the country, with the Insurance Council of Australia reporting that the 2022 floods were the costliest in Australia’s history, resulting in an estimated $3.35 billion in insured losses [29]. Accurately estimating returns and magnitudes of multiday extreme rainfall can inform infrastructure, city planning, and agriculture to help combat expected future loss.
Processed_Modeling_Multiday_Extreme_Precipitation_Across_Eas.txt	five-day precipitation extremes in Australia, in contrast to the global trend of increasing precipitation extremes. However, Alexander and Arblaster (2017) [2] discovered that global climate models show increased precipitation extremes across Australia. In another study, Hajani et al. (2017) [17] examined the annual maximum of sub-daily to daily precipitation in New South Wales, Australia, and found that the number of stations showing significant trends decreased when the impact of climate variability modes was considered. Jakob and Walland (2016) [19] revealed that there has been no significant change in daily precipita- tion extremes in Australia over the past century, but the ENSO significantly impacts the occurrence of these extremes. Laz et al. (2014) [25] analyzed extreme rainfall patterns from 1950-2010 in southeast Australia on both sub-daily and daily scales and discovered that annual maximum rainfall events lasting less than 12 hours increased, while rainfall events lasting 12-72 hours decreased. Westra and Sisson (2011) [35] reported a similar result. In particular, they analyzed the annual maxima precipitation in eastern Australia from 1965 to 2005 and reported a statistically significant increase in short-duration rainfall over this period but no trend for daily and a decreasing trend in multiday precipitation. Many of these studies have explored trends in extreme precipitation, demonstrating the nonstationarity in precipitation time-series in relation to ENSO.
Processed_Modeling_Multiday_Extreme_Precipitation_Across_Eas.txt	A common approach for modeling extremes in a sequence is to use results from the field of Extreme Value Theory which provide distributional assumptions for the extremes of time- series under certain weak dependence conditions, provided the sequence is stationary or does not change in time or space. Consequently, the spatial and temporal nonstationarity observed in weather data introduces complexities in modeling using standard extreme value statistics; however, many studies have successfully addressed this issue by including non- stationary dependencies directly into the model to create a unified spatiotemporal model of rainfall extremes. Recent studies developed a spatiotemporal nonstationary model for extreme precipitation across regions in Australia, including large-scale modes of climate variability and topology variables [20, 31, 35, 36]. However, there are limited studies con- cerning consecutive days of extreme precipitation in the model. Despite [20, 35] applying a nonstationary generalized extreme value (GEV) distribution to model extreme precipi- tation across several durations from 6 minutes through to 72 hours annual maxima, the studies did not provide the goodness of fit of the model. In addition, it is not always clear how concurrent daily extreme precipitation is influenced by the commonly used indices for monitoring extreme precipitation, such as annual maximum consecutive 5-day precip- itation (Rx5day), annual total precipitation from days greater than the 99th percentile (R99p), or maximum number of consecutive days with daily precipitation greater than 1 mm (CWD) (see e.g., [2, 17, 25, 33, 35, 36]). For example, it is known that increasing the number of consecutive days one considers for the model will reduce the data available for fitting the model by the same factor. Consequently, the reliability of the statistical model decreases significantly as the investigated number of consecutive days of extreme rainfall increases.
Processed_Modeling_Multiday_Extreme_Precipitation_Across_Eas.txt	Dynamical systems and recent advancements in modeling extreme weather. Recent advancements in the field of dynamical systems have provided insights into key factors for more accurately modeling the consecutive nature of extreme weather events. To explain the impact of these results, we begin by introducing some background on the relationship between extreme value theory for dynamical systems and their role in deriving extreme value statistics for climate applications.
Processed_Modeling_Multiday_Extreme_Precipitation_Across_Eas.txt	Beginning in the late 1970s with the work of Leadbetter and Chernick, [10, 26] an extreme value law guaranteeing the existence of an extreme value distribution was proven in the limit for weakly dependent sequences satisfying a mixing and recurrence condition called D(un) and Dk(un), respectively [10]. Moreover, Leadbetter [26] showed that under these conditions the resulting extreme value law was similar to the one derived for independent and identically distributed random variables apart from the introduction of a new pa- rameter, θ ≤ 1 called the extremal index, which characterizes the short-term dependence observed in the weakly-dependent sequence. The authors in [18] showed under general conditions that the returns of a the sequence to an extreme or rare event can be under- stood as some (compound) Poisson process and 1/θ represents the expected cluster size of extremes. The formal definition of the Extreme Value Law is provided below.

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.