Fishing for biomarkers: analyzing mass spectrometry data with the new ClinProTools™ software

Introduction

Mass spectrometry-based search for biomarker patterns is widely recognized as a valuable research tool for predictive medicine and pharmacological monitoring (1). Of particular interest is the identification of tumor markers for early detection and the diagnosis of cancer to improve the clinical prognosis of patients. Various disciplines may take advantage of this new technique in their clinical studies (i.e., oncology, urology, psychiatry, neurology, toxicology, pharmacology, and others). Additionally, this application may even be adopted beyond clinical applications (i.e., from food and seed production to pathogen and mycotoxin profiling).

The pursuit to recognize specific polypeptide biomarker patterns for certain diseases adds a new level to proteomics that demands sophisticated evaluation (2) and valuable statistical tools. Advances in sample preparation and instrumentation enhance the requirements for comprehensive computational methodologies to screen and evaluate data sets. Universal, intuitive, and flexible bioinformatics solutions are needed to satisfy current and upcoming analytical needs.

Materials and methods

Sample Collection and Preparation

Serum samples were collected from 13 patients who were admitted to our hospital because of pneumonia. Meanwhile, two groups of other patients with approximate age and gender matching were enrolled as controls, including 16 patients presented with fever, but without clinical or radiological evidence of pneumonia, and 13 obviously healthy individuals as disease controls and healthy controls, respectively. Serum samples were collected in acute phase of disease for the pneumonia patients, and during the episodes of fever for the disease control group. All serum samples were collected after at least 8 h fasting. Sera were snap-frozen and stored at -80°C in aliquots until use. All experiments were conducted in duplicate.

Prefractionation

Prefractionation has been performed using different surface functionalities of ClinProt™ microparticle beads (Bruker Daltonik, Leipzig, Germany) to generate a diversity of biomarker patterns for the profiling. We have used hydrophobic interaction (MB-HIC8), weak cation ion exchange (MB-WCX), and immobilized metal-affinity chromatography containing copper ions (IMAC-Cu). All preparations have been performed according to the manufacturer's instructions.

Robotics

The complete magnetic bead fractionation has been performed on an automated robotic platform (ClinProt Robot; Bruker Daltonik). This included all pipeting steps and matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) sample plate loading.

Mass Spectrometry

Mass spectrometry has been performed using the ClinProt system, equipped with an Ultraflex™ TOF/TOF instrument (Bruker Daltonik). Spectra have been collected automatically using the AutoXecute™ software (Bruker Daltonik; see Reference (3) for fuzzy-controlled adjustment of critical instrument settings to generate raw data of optimized quality.

Data Interpretation

The ClinProTools™ software (Bruker Daltonik) has been used for all data interpretation steps, which start with a raw data pretreatment, including normalization of a set of spectra derived from a patient cohort, internal signal alignment using prominent internal signal peaks, and a peak picking procedure. The whole data pretreatment has been completed using default settings and was performed automatically, without any user interaction. The pretreated data have been used for visualization and statistical analysis in ClinProTools. Peak statistics has been performed by means of a Welch's t-test.

Results

The ClinProTools software was used for the data interpretation of MALDI-TOF spectra derived from serum samples of different patient groups. A subset of the patients has been affected by the target disease, pneumonia. We have used two different control groups, namely disease controls and healthy controls, which enabled us to distinguish general disease marker candidates from those that are specific for the target disease. The samples have been prefractionated using magnetic beads with three different surface functionalities, namely MB-HIC8, MB-WCX, and MB-IMAC-Cu.

Visualization of Mass Spectrometry Data

ClinProTools offers a variety of viewer options for the analysis of clinical profiling data (Figure 1). A virtual gel view gives an overview representation of data sets derived from large sample cohorts. This viewer is the master navigation tool for a large data set. The virtual gel view can be displayed in different color scales. A representation of the signal intensities in a rainbow color scale allows for detection of faint differences between the classes due to the changing colors with increasing peak height (Figure 1).

An alternative viewer shows a conventional spectrum display, which can be customized to use individual spectra or prototype spectra of the different sample classes. After peak selection, a “box & whiskers” plot graphically displays the intensity distribution of the relevant signals in the different sample classes.

Further options are offered by a stack plot viewer to get an impression of signal intensity scattering in a 3-dimensional view and in a contour plot viewer that allows defining multiple intensity threshold in the display.

Classical and Modern Approaches to Detect Pattern Differences: Peak Statistics and the Genetic Algorithm

Two independent statistical approaches assist the user in the detection of peak pattern differences of assigned spectra to the different classes, namely classical statistics of peak signal intensities and a modified genetic algorithm. Moreover, data export allows applying other statistical approaches like decision tree analysis (Salford Systems, San Diego, CA, USA) or in-house bioinformatics.

ClinProTools provides a list of peaks sorted according to the statistical significance to differentiate between both classes (Figure 2). On the basis of a Welch's t-test, a P value is calculated, which indicates the probability that the observed intensity differences of the individual peaks are not based on coincidence. These calculations are done independently for peak heights and peak areas. A list of all peaks sorted according to the statistical separation strength will be created as an output file. Figure 2 shows the results for two classes of the test serum profiles. For examination, the respective masses can easily be inspected in the spectrum view and in the virtual gel view (Figure 2).

User Interactivity, Validation, and Class Prediction

Peaks with high separation power in the Welch's t-test may be used to generate a biomarker pattern model. Peaks may also be added manually by the user. Classical peak statistics and the genetic algorithm may be used in combination. Once saved, pattern models may be validated in a cluster analysis with an independent test sample set.

If the user did not already establish an independent test data set of sufficient cohort size for a validation, cross validation (4) of the cluster analysis may be used to investigate the quality of specific signal patterns. For the validation of a model, independent test data sets representing the classes can be selected by the user. Validation determines the predictive capability of a model as a percentage of the correctly classified test data. Finally, new unknown samples may be classified in a class prediction, which can be carried out with a previously established pattern model.

ClinProTools accompanies the user through the whole process—from acquiring reasonable biomarker patterns, to the final step of classification of unknown samples. Comprehensive visualization tools, supplemented with sophisticated mathematical algorithms, deliver reproducible results that can instantly be validated and used in class prediction.

Interpretation of Data Derived from Pneumonia Patients

We have performed an initial study concerning visualization and statistical analysis of a data set containing 52 spectra derived from duplicate preparations of 13 disease patients and 13 healthy controls. The aim of this study was to test the visualization capabilities of the software for the identification of some biomarker candidates, which would be promising for a more detailed analysis in subsequent studies with a real life data set. It was explicitly not the aim of this study to establish diagnostic biomarker patterns; a larger data set is needed for such a confirmation. Moreover, identification via TOF/TOF fragment analysis should be performed to provide biological relevance to the statistical analysis.

We have found several mass regions with differences between the two sample classes upon visual inspection (Figures 1 and 2). When using WCX beads for the prefractionation, one obvious region of multiple signals showed differential intensities between the two classes in the range of 11,500 Da. Zooming into this region revealed that it is composed of multiple differential signals (Figure 1). Upon detailed inspection of the complete data set, further differential signals have been found in the mass range from 1500 to 10,000 Da (data not shown). Subsequently, we have used the statistical features of ClinProTools to evaluate those peaks with seemingly high power to differentiate between the two classes by means of a Welch's t-test. The result of this test is a P value, which gives an estimation of the probability that the measured peak signal distributions can be observed by chance. Accordingly, the lower the P value, the better a respective peak signal is suited to be used to separate the two classes. The generally accepted limit of the P values to consider a result significant is defined as 0.05 (respectively, 0.01 for highly significant results). The output table of the statistical analysis in Figure 2 shows that several peaks in the data set resulted in P values indicating high importance. The combined availability of visualization and statistical analysis allows a direct feedback to the original data. Accordingly, the peaks labeled as significant to distinguish between the two classes have been directly controlled using the visualization features. This is shown in Figure 2 for the two peaks at 4662 and 11,525 Da. In both cases, the peaks of interest show different signal intensities between the two classes, while adjacent peaks have nearly identical intensities, thus serving as internal controls (see asterisks in Figure 2).

It is of special interest to distinguish between biomarker candidates for a general disease condition and those that are specific for the target disease. To demonstrate the respective software functionality, we have used three independent groups, namely healthy control, disease control, and target disease. Figure 3 shows the combined results after fractionation using the hydrophobic C8 magnetic beads. Statistical analysis gives a number of peaks, which can be used to separate healthy controls from both disease groups (labeled red in Figure 3). Some further signals are specific for the target disease, as they can be used to separate the respective samples of both control groups (labeled green in Figure 3). These different kinds of potential biomarkers can be especially well visualized in the mass region between 3100 and 3400 Da. Here we found biomarker candidates of both types.

Discussion

The presented initial data interpretations of samples derived from pneumonia patients demonstrate that the combined visualization and statistical features of the ClinProTools software allows for an easy and efficient identification of biomarker candidates. Such candidates can pattern recognition models to analyze independent data to evaluate the quality of these biomarkers to be used in a sample classification for clinical research. Moreover, the statistical approach can be combined with use of the modified genetic algorithm. Here again, the results of both approaches can be directly compared by means of the powerful visualization tools. Peaks used in models generated by the genetic algorithm are highlighted with red or grey bars in the viewer (Figure 2).

Currently, the opinion of the scientific community is divided concerning the appropriate approach for biomarker analysis. One major doctrine says that each potential biomarker has to be identified to be of any further clinical use (5). A second doctrine says that a diagnostic pattern, which will give a better classification than existing tests, could be finally applied in the clinic (6). In both cases, an effective visualization of large patient data sets is crucial to get a confident indication of which potential biomarker signals are promising for peptide identification or for future routine application in diagnostic research. Potential biomarker peaks can be selected for in-depth analysis after further enrichment to reduce the sample complexity and analyzed by tandem mass spectrometry (TOF/TOF) analysis. Instruments equipped with a TOF/TOF analyzer, like the UltraFlex used in this study, can be directly used for profiling and for in-depth TOF/TOF analysis. Here, the signal of a potential biomarker is selected in an according mass filter, whereas other signals are discarded. A postacceleration step allows analysis of all metastable fragments of the biomarker candidates (7), which gives a chance for identification in an according database.

Conclusively, ClinProTools is a very powerful software tool for the data interpretation in the field of clinical proteomics. It combines efficient visualization with automated data pretreatment and intuitive statistical analysis. We have used it to determine a number of biomarker candidates in a preliminary data set within a few hours. Last but not least, ClinProTools does not require a deep knowledge of statistics and mathematics to be efficiently used.