Statistical methods for mass spectrometry data analysis and identification of prostaste cancer biomarkers

BACKGROUND Prostate Cancer (PCa) is the most common cancer among males in Europe. Patients developing early PCa sometimes refer non-specific symptoms, namely lower urinary tract symptoms (LUTS), and they usually undergo medical investigations based on Prostate Specific Antigen (PSA) and Digital Rectal Examination (DRE). Suspicious results of one or both testings are prerequisite to Prostate Biopsy. However, due to PSA low sensitivity/specificity in predicting positive prostate biopsy, the identification of new PCa biomarkers is actually a real need. MALDI-TOF/MS protein profiling could be a valuable technology for biomarkers identification. However, up to now its use is laden with lack of reproducibility that confounds scientific inferences and limits its broader use. AIMS Goal of this study is to analyze urine collected after prostatic massage in patients referring LUTS, to identify candidate biomarker for PCa, by using MALDI-TOF/MS. We considered important aspects of MALDI-TOF/MS label-free proteomic profiling, in order to assess features reproducibility and to propose appropriate strategy to handle both measurement error and limit of detection (LOD) problems. The study results should aid in reducing the number of worthless first-biopsied and assist Urologists on differential diagnosis of PCa. METHODS In a cross-sectional study, we collected urine obtained after DRE from 205 patients that referred LUTS to consultants at the Urological Unit at University of Padova. All patients undergone to prostate biopsy for suspicious PCa. Urines were dialyzed and analyzed by MALDI-TOF/MS in reflectron mode. For the MALDI-TOF/MS reproducibility evaluation, we analyzed a urine pooled from 10 reference samples, spiked with 12.58 pmol of a 1589.9 m/z internal standard (IS) peptide. For the inter-run variability assessment, 14 aliquots were dialyzed by MALDI-TOF/MS. For the intra-run study, an aliquot was divided into 26 separate sub-aliquots and analyzed by MALDI- TOF/MS. To estimate the signal detection limit (sLOD), serial dilution up to 1/256 of a urine pool were analyzed in triplicate. We evaluated the sLOD and adjusted the data appropriately to reduce its variability. We investigated six data normalization approaches - the mean, median, internal standard, relative intensity, total ion current and linear rescaling normalization. Between-spectrum and the overall spectra variability were evaluated by the coefficient of variation (CV). An optimized signal detection strategy was also evaluated to overcome peak detection algorithms errors. Measurement errors and with-in subject variances were evaluated by an external dataset, made of urine repeatedly collected from 20 reference subjects. Intra class correlation coefficient (ICC), Regression Calibration (RCAL) and SIMEX analyses were used to estimate unbiased logistic regression coefficients relating MALDI-TOF/MS features with Patients biopsy outcome. Monte Carlo simulations were used to estimate influence of different LOD adjustment methods on ICC and RCAL. RESULTS Initially, we evaluated the intra- and inter-run on data obtained from automatic peak detection. Normalization methods performed almost similarly in both studies, except IS, which resulted in an increased CV. Calculated sLOD varied with spectra m/z. After sLOD adjustment, raw and normalized data showed a reduction in CVs, while median and mean normalizations performed better, especially in the intra-assay study. However, by optimizing the peak signal detection, the overall features variability drastically decreased. Median normalization with sLOD correction remained the preferable choice for further analyses. Evaluating the external dataset, we found that most of the MALDI-TOF/MS variability is intrinsic to the biological matrix. By using substitution of below LOD values by LOD/2, simulation studies showed that ICC estimations were poorly affected by LOD, when measurement error σ is less that 0.36 and values below LOD are less that 50%. Comparing results from naïve logistic regression, RCAL and SIMEX, measurement error appeared to cause a "bias toward the null". However, SIMEX estimations seemed to correct for a smaller amount of bias than RCAL. Overall, we found eight MALDI-TOF/MS features associated with positive biopsy results. CONCLUSION Findings from the reproducibility study showed that the major contributing factor for MALDI-TOF/MS profiling variability is the peak detection process. So, a new algorithm suited for MALDI-TOF reflectron mode is desirable for its applications in profiling studies. However, normalization strategies aid in increasing MALDI-TOF/MS label-free data reproducibility, especially with sLOD correction. Despite urine does not seem to be a promising biological fluid for proteomic biomarker discovers, RCAL and SIMEX appeared valuable approaches to obtain regression coefficients adjusted for biological and instrumental errors on MALDI-TOF/MS features.