Plasma peptide biomarker discovery for amyotrophic lateral sclerosis by MALDI-TOF mass spectrometry profiling.
Bottom Line: In our study, we looked for peptide biomarkers in plasma samples using reverse phase magnetic beads (C18 and C8) and MALDI-TOF mass spectrometry analysis.These two SVM-based models end up in excellent separations between the 2 groups of patients (recognition capability overall classes > 97%) and classify blinded samples (10 ALS and 10 healthy age-matched controls) with very high sensitivities and specificities (>90%).Some of these discriminant peaks have been identified by Mass Spectrometry (MS) analyses and correspond to (or are fragments of) major plasma proteins, partly linked to the blood coagulation.
Affiliation: Exploratory Unit, Sanofi, Toulouse, France.
The diagnostic of Amyotrophic lateral sclerosis (ALS) remains based on clinical and neurophysiological observations. The actual delay between the onset of the symptoms and diagnosis is about 1 year, preventing early inclusion of patients into clinical trials and early care of the disease. Therefore, finding biomarkers with high sensitivity and specificity remains urgent. In our study, we looked for peptide biomarkers in plasma samples using reverse phase magnetic beads (C18 and C8) and MALDI-TOF mass spectrometry analysis. From a set of ALS patients (n=30) and healthy age-matched controls (n=30), C18- or C8-SVM-based models for ALS diagnostic were constructed on the base of the minimum of the most discriminant peaks. These two SVM-based models end up in excellent separations between the 2 groups of patients (recognition capability overall classes > 97%) and classify blinded samples (10 ALS and 10 healthy age-matched controls) with very high sensitivities and specificities (>90%). Some of these discriminant peaks have been identified by Mass Spectrometry (MS) analyses and correspond to (or are fragments of) major plasma proteins, partly linked to the blood coagulation.
Related in: MedlinePlus
Mentions: Developing a clinical diagnostic involving more than a dozen signals is not realistic. Therefore, we sought to select the most discriminating peaks to develop predictive models based on a minimum number of signals. To overcome the limited number of patients included in our study, data were randomly separated into two data sets, one for the discovery step (training set, with 30 ALS patients and 30 healthy controls) and the second for the validation (classification of blinded samples, 10 ALS patients and 10 healthy controls); and this process was repeated 10 times (Figure 3). From each of the 10 generated training sets, predictive models were calculated using a supervised method based on support vector machine (SVM). In this process, automatic detection was selected for determining the best number of peaks to be integrated in a model. To access the ability of each individual SVM-based model to correctly classify blinded samples, the corresponding validation sets were then used (The SVM model n°1 against the validation set n°1 and so on). Results are summarized in tables 2 and 3. From C18 data, the number of selected signal changes from 2 to 21 according to the model while for C8 data the number of selected signal changes from 7 to 23 (Table 2). Whichever the SVM-based model, the cross validation values and the recognition capacity are very high; at least greater than 94% (Table 3). What is more, the values of sensitivity and specificity obtained from the classification of the validation data sets are also high. C18 SVM-based models correctly classify all the spectra of the control group and at least 89% of the spectra from ALS group. Results with C8 SVM-based models are similar even although the values are a little less good with 68.8% and 90.5% of well classified spectra for ALS and control groups respectively. Under these results, the minimum of peaks for well classified data is at least 2 peaks for C18-samples (1101 and 1426) and 7 peaks for C8-samples (1101, 1426, 1769, 3883, 4964, 7765 and 8141). The high discriminant power of some C18- or C8-peaks is also highlighted by the high p-values obtained from the analysis of the peak variance (ANOVA) using log-2 transformed data (Table S3). 69.5% of the C8-peaks and 86% of the C18-peaks show p-value <0.05; with 1/3 of the peaks that are ALS up-regulated signals. Only one peak has a 2-fold change among C8-signals (peak 7765) while 14 signals pass this fold change value among C18-data (10 ALS up- and 4 down-regulated: 1426, 2230, 2483, 2511, 3469, 4426, 4920, 4979, 4964, 5004 and 1079, 1101, 1127, 2755). A comparison between this ANOVA test and the selected peaks over all the SVM-based models shows that, as expected, SVM models are based on the most discriminating signals. The expression levels of some of the most discriminating peaks are presented in Figure 4 with the corresponding receiver operating characteristic curves.