Exploring liquid chromatography-mass spectrometry fingerprints of urine samples from patients with prostate or urinary bladder cancer

Data processing and analysis have become true rate and success limiting factors for molecular research where a large number of samples of high complexity are included in the data set. In general rather complicated methodologies are needed for the combination and comparison of information as obtained from selected analytical platforms. Although commercial as well as freely accessible software for high-throughput data processing are available for most platforms, tailored in-house solutions for data management and analysis can provide the versatility and transparency eligible for e.g. method development and pilot studies. This paper describes a procedure for exploring metabolic fingerprints in urine samples from prostate and bladder cancer patients with a set of in-house developed Matlab tools. In spite of the immense amount of data produced by the LC-MS platform, in this study more than 10(10) data points, it is shown that the data processing tasks can be handled with reasonable computer resources. The preprocessing steps include baseline subtraction and noise reduction, followed by an initial time alignment. In the data analysis the fingerprints are treated as 2-D images, i.e. pixel by pixel, in contrast to the more common list-based approach after peak or feature detection. Although the latter approach greatly reduces the data complexity, it also involves a critical step that may obscure essential information due to undetected or misaligned peaks. The effects of remaining time shifts after the initial alignment are reduced by a binning and 'blurring' procedure prior to the comparative multivariate and univariate data analyses. Other factors than cancer assignment were taken into account by ANOVA applied to the PCA scores as well as to the individual variables (pixels). It was found that the analytical day-to-day variations in our study had a large confounding effect on the cancer related differences, which emphasizes the role of proper normalization and/or experimental design. While PCA could not establish significant cancer related patterns, the pixel-wise univariate analysis could provide a list of about a hundred 'hotspots' indicating possible biomarkers. This was also the limited goal for this study, with focus on the exploration of a really huge and complex data set. True biomarker identification, however, needs thorough validation and verification in separate patient sets. (c) 2011 Elsevier B.V. All rights reserved.