3.1 Preprocessing

In order to compare mass spectra for the identification of proteomic biomarkers, multiple steps need to be performed to standardize all given spectra to a common format. Therefore, a preprocessing pipeline is used for the transformation of data into a proper format for analysis (shown in Figure 28.1) (Handler et al., 2011).

Some of the modules in this workflow use the PROcess library (Li, 2005) of the Bioconductor R package (R Development Core Team, 2009). For the integration of processing steps into KD³ submodules (functional objects), a parallel execution of the algorithms was made possible, which can efficiently compute numerous MALDI spectra on systems with multiple cores, for example, on a computational cluster or workstation. The figures demonstrating the preprocessing results were created using the JFreeChart library (JFreeChart, 2012).

Binning of m/z values Due to unequally distributed m/z values in the compared spectra, a direct comparison between intensities of different spectra is difficult. In our approach, therefore, we included a binning step to map the intensities of the given spectra onto equally distributed m/z bins. In particular, the dimension (number of m/z values) of high-resolution spectra can be reduced using this step (Ressom et al., 2005).

Using the Binning module, the width w of the resulting bins and the method for calculation of the bin representative can be selected. As intensity value of the bin representative I(x) at m/z value x the mean/maximum intensity of all intensities in the range $\left[x-\frac{w}{2};x+\frac{w}{2}\right)$ is calculated. Furthermore, the user can choose whether the intensities of empty bins should be linearly interpolated by neighboring intensities or set to 0.

After setting these parameters, spectra can be assigned to this module. As a result, the Binning module delivers spectra with equally distributed m/z bins (see Figure 28.2).

Adjustment of m/z ranges Spectra with different m/z ranges cannot be matched with each other over the whole m/z domain. To use most of the spectra for further preprocessing, the adjustment step is applied, which removes all values from the spectra outside the boundary of the highest common minimal m/z value x_min and the lowest common maximum m/z value x_max of all spectra.

The Adjustment module allows the user to set the values x_min and x_max manually. The user can also decide individually whether spectra are selected for further preprocessing that do not lie within the defined boundaries. If, for example, a spectrum starts at a higher m/z value as defined by x_min, this spectrum can be filtered out by this step. The m/z lower and upper bounds of residual spectra are equalized by removing all m/z bins outside the defined boundaries. Therefore, after this step, the m/z range of all spectra is now equal and consistent (see Figure 28.3).

Baseline subtraction Chemical noise in the energy absorbing molecule solution and ion overload can lead to an elevated baseline within the spectra (Li et al., 2005). Different algorithms are available for effective removal of baselines from spectra. For example, Ressom et al. (2005) used a spline approximation on local minima of a spectrum to estimate its baseline. In this work, the baseline subtraction algorithm described by Li et al. (2005) was implemented, which estimates the baseline by local regression to the points below a certain quantile or to local minima of a moving window. After the baseline is estimated, it is subtracted from the original spectrum. The Baseline subtraction module allows the configuration of parameters, which are directly passed to the bslnoff operation of the PROcess package of R (Li, 2005). The user can set parameters such as the number of breaks on the log m/z scale for finding the local minima or intensity values below a defined quantile necessary for local regression calculation. Basically, it can be selected between local regression or linear interpolation for smoothing the estimated baseline. The user can also specify the bandwidth for the local regression method. Figure 28.4 depicts a baseline corrected spectrum returned by this module.

Normalization Effects of experimental noise can cause variations in the amplitude of spectra. To remove this artifact from spectra the total ion normalization procedure was used (Li et al., 2005) to rescale intensity values of the spectra. In this step, the area under the curve (AUC) is calculated for each spectrum above a user-defined cutoff. This cutoff is relevant because of high noise signals at low m/z values. The AUC of a spectrum is calculated as the sum of all intensities of a spectrum (considering the given cutoff value) if the intensities are represented by equally distributed m/z bins (Li et al., 2005). A normalized spectrum N_i based on the spectrum U_i is calculated as

$N_i\left(x\right)=U_i\left(x\right)\cdot \frac{\mathit{AU}{C}_M}{AUC\left(U_i\right)},$

where

$AUC\left(U_i\right)={\sum }_{j=1}^n{U}_i\left(j\right)$

defines the AUC of the spectrum U_i, n is the number of m/z values in the spectrum U_i, and AUC_M defines the median AUC of all spectra (see Figure 28.5). Alternatively, the user has the possibility to define the AUC as a parameter directly.

Peak detection Peaks in the spectra represent specific, abundant polypeptides in the sample (Li et al., 2005). These peaks define reasonable intensities compared to intensities of other m/z values, which appear as noise (Ressom et al., 2005). To identify peaks in the spectra, the function isPeak of the PROcess package is used (Li, 2005). Three different parameters are considered for the selection of the peaks:

• The signal-to-noise ratio (SNR), defined as the local smooth divided by the local estimate of variation

• A detection threshold below which the intensities of the spectrum are considered as zero

• The shape ratio, which is defined as the ratio between the area under the curve within a small distance of a peak candidate, as already identified by the first two criteria, and the maximum of all such peak areas of a spectrum

In addition to these parameters, the user can define window widths for the estimation of local variance, for smoothing the spectrum before peak detection, and for the calculation of the area under the peak, which are passed to the isPeak operation of the PROcess library.

Quality assessment Spectra of poor quality may cause reduction of statistical significance of selected masses. Therefore, a quality assessment step on preprocessed spectra is recommended. In our workflow, we used the quality operation of the PROcess package (Li, 2005).

The quality of a spectrum is evaluated using the following three measures: Quality, Retain, and Peak. For computation of the Quality and Retain measures, a noise envelope is used, which is computed as follows: First, the noise is estimated as the difference between the spectrum and its moving average with a window size of 5 points starting from a user-defined cutoff value. Afterward, the noise envelope is computed as three times the standard derivation of the previously calculated noise in a 250-point window.

For the computation of the Peak measure, the peak information retrieved by the previous step is required.

The three measures for the quality estimation are defined as follows (Li et al., 2005):

• Quality: The measure of the separation of signal from noise, defined as the ratio of the AUC before and after subtraction of the noise envelope.

• Retain: The number of high peaks in a spectrum is quantified by comparing the number of intensities more than five times the noise envelope over the total number of points in a spectrum.

• Peak: The number of peaks of the current spectrum is compared to the average number of peaks in all spectra.

The boundary parameters for the three measurements can be defined by the user. If a spectrum does not exceed any of the defined boundaries, the spectrum is removed.

Alignment to internal standards Because of measurement variations, peaks in different spectra that correspond to the same protein may be located at different m/z values (Li et al., 2005). In the peak alignment of spectra, peaks are identified across spectra, which are likely to represent the same protein. For the alignment, different algorithms are available (e.g., Li et al., 2005; Ressom et al., 2005). In our workflow, a method for peak alignment was integrated that used internal standards as reference points in the spectrum. By definition, an internal standard is a compound added to a sample in known concentration to facilitate the qualitative identification and/or quantitative determination of the sample components (Ettre, 1993). In this approach, two internal standards are used. To align a spectrum by using this method, corresponding peaks have to be found close to the defined m/z values of the internal standards. Either the closest peak or the peak with maximal intensity in a user-defined window is chosen. The m/z values of the aligned spectrum are subsequently calculated by linear interpolation:

$\hat{\mathrm{x}}=\left(x-I{S}_l\right)\cdot \frac{{\widehat{IS}}_u-{\widehat{IS}}_l}{I{S}_u-I{S}_l}+{\widehat{IS}}_l$

where x is the original m/z value, $\hat{x}$ is the aligned m/z value, IS_l and IS_u denote the m/z values of the corresponding lower and upper internal standard peaks of the processed spectra, and ${\widehat{IS}}_l$ and ${\widehat{IS}}_u$ denotes the m/z values of the lower and upper internal standard (see Figure 28.6).

The peak information and the m/z values of two internal standards, including a defined window size around the standards, are required for running the alignment. If no peaks are found with a smaller or equal distance than half of the window, the spectrum is removed and is not considered for further analysis. Note that the maximum peak or the closest peak within a window can be chosen as an internal standard for the alignment.

Recalibration of preprocessed spectra Due to the alignment to internal standards, parts of the initial preprocessing modality (i.e., binning, adjustment, and normalization) need to be repeated to ensure that the m/z values of the remaining spectra are (i) equally binned, (ii) within the same m/z bounds, and (iii) have the same AUC.

Generation of mean spectra If multiple spectra per sample are available, these spectra can be averaged to a mean spectrum by calculating the mean intensities over the entire m/z value range. By this action, the noise can be reduced and the SNR significantly improved. Note that spectra of the same sample are identified by the file name of the spectra that included the sample IDs.

Formatting of spectra for feature selection approach For the biomarker identification, some algorithms from the Weka software package are used (Hall et al., 2009). To assign the spectra to different classes and transform the spectra information into ARFF (Attribute-Relation File Format), two additional modules were implemented (class assignment by file name and conversion to ARFF).

3.2 Identification of biomarker candidates

After preprocessing, the spectra have comparable m/z and intensity values and can be defined as a set of tuples, $T=\left\{\left(c_j,m\right)|c_j\in C,m\in M\right\}$ with $C=\left\{\mathit{case},\mathit{control}\right\}$, where C is the set of class labels and M is the set of features (m/z values in the spectrum). In order to identify those masses in M that show highest discriminatory ability according to class C, we adjusted an algorithm previously published by our group (Plant et al., 2006).

In the first step, a filter approach is used to select relevant features (m/z values). The resulting features from step 1, however, contain regions of adjacent features that are highly correlated, as most of them are redundant, representing the same information of the spectra. Consequently, in step 2, we identify a representative for every region in the spectra. Finally, in step 3, a wrapper approach further reduces the dimensionality of the result set of step 2 by optimizing the discriminatory ability using a classifier.

Note that for evaluating these analysis steps, synthetic spectra were created using the spectrum generator of mMass 5.0 (Strohalm et al., 2010). Overall, we generated 50 case and 50 control spectra and inserted randomly two artificial, well-discriminating m/z regions.

Step 1: Selecting relevant features In this work, we used a Student t-test as a filter approach. In addition to the resulting P-value, we calculated a second parameter Δ representing the ratio of the mean intensities in each class ${\overline{x}}_{c_i}$ relative to the maximum intensity I_max in all spectra. The parameter Δ was defined by

$\mathrm{\Delta }=\frac{\left|{\overline{x}}_{c_1}-{\overline{x}}_{c_2}\right|}{I_{\mathit{max}}}$

This parameter is important to ensure that differences in the intensity can also be technically detected. A feature is defined as relevant if the following two conditions are fulfilled:

$\mathrm{P}-\mathrm{value}<\alpha$

$\Delta >\mathrm{\Gamma }.$

The parameters α and Γ are set by the user.

Step 2: Selecting region representatives We used a forward selection strategy to identify a representative for every region in the spectra. The representative feature is the feature with the highest quality (i.e., the discriminatory ability according to the filter method applied in step 1) within the region. The size s of the region depends on the index of the feature representing the m/z value, which is due to technical reasons, as different fragments of peptides with low molecular weight cause many narrow peaks in the spectral region of low m/z values.

Step 3: Selecting the best features To further reduce the number of features, a wrapper-based approach is used, including a classifier and a search strategy to find a smaller feature subset while keeping the discriminatory ability at least constant. We apply logistic regression (Le Cessie and Van Houwelingen, 1992) as the classifier and a modified binary search (MBS) (Plant et al., 2006) as the search strategy. The area under the receiver operating characteristic (ROC) curve is selected as the measure for assessing the discriminatory ability. As introduced, selected features using this approach represent highly discriminating m/z values in the spectra. Finally, the local maximum in the neighborhood of a selected m/z value is determined to ensure that the selected mass represents a real peak in the spectra. Figure 28.7 demonstrates a snapshot of two matched spectra when comparing two different groups of mass spectrometry data (e.g., cases versus controls). This three-step strategy is able to identify highly discriminating mass peaks between the spectra (in this example, with superior sensitivity and specificity). In particular, in this example, we were able to identify our two predefined artificial biomarker candidates.

After applying this computational strategy, identified well-discriminating mass peaks need to be verified and validated as biomarkers by subsequent database verification, lab experiments, and clinical trials before selected biomarker candidates can go into clinical application.

4.1 KD³ composition

The implemented KD³ application consists of four main parts. A screenshot of the application is depicted in Figure 28.8. The application can be divided into four parts. The first part shows the available functional objects, which are loaded using the Java reflection application programming interface (API) and are grouped using a hierarchical structure. In a workspace window, the user can drop functional objects and parameterize them by setting up annotated constructors and methods. In the center of the window, the workflow is visualized.

4.2 KD³ functional object

In general, a functional object is composed as follows: From the user’s perspective, a task is a functional object, which consists of in- and out-ports that have to be assembled and parameterized to fulfill their purpose. In order to extend the functionality of KD³, the software engineer has to create a subclass of the superclass FunctionalObject containing the required algorithms. The abovementioned in- and out-ports get the data from another object and send the processed data to another object. For instance, the BinningStep functional object in Figure 28.8 receives the data from ReadSpectra and finally sends the preprocessed data to the AdjustmentStep. For performing the computation, an abstract method named execute() is available, which must be overridden in derived FunctionalObject classes in order to solve a specified problem.

4.3 KD³ workflow

A workflow is defined as a repeatable pattern activity, which is used to solve a defined process using different parameters and data sets. The user, therefore, can drop the functional objects into the workspace window and can specify the parameters. After the workflow is designed, it can be executed directly from KD³. Finally, if the workflow is properly designed, it can be saved as a GraphML (http://graphml.graphdrawing.org/) file and deployed (e.g., to biomedical researchers to analyze newly available data). In this work, we extended KD³ to allow the preprocessing and biomarker identification of MS data (see also Figure 28.8). Our proposed workflow-oriented architecture results in high usability. KD³ allows the user to exchange different methods for all the aforementioned steps, or the straightforward implementation and integration of newly developed algorithms.

As the integration of new algorithms should be as simple as possible, the developer has to develop the new algorithm in the programming language Java. Therefore, the only need for integrating new algorithms into KD³ is an elementary understanding of Java. KD³ automatically generates the graphical user interface (GUI) for each of the implemented methods, but programmers can also develop a specific GUI for each of those methods. The extended version of KD³ for preprocessing and biomarker identification of MS data is available at http://www.umit.at/kd3ms.

KD³ can be used in the analyses of other data types as well. Therefore, more than 100 different methods and algorithms have been integrated into this application. As our group mainly deals with research questions in computational biomarker discovery, we applied KD³ successfully in several research projects, such as addressing the search for metabolic biomarkers in cardiovascular disease or the identification of human breath gas markers in liver disease using ion-molecule reaction-mass spectrometry (Pfeifer et al., 2007; Netzer et al., 2009, 2011; Baumgartner et al., 2010; Millonig et al., 2010).