MALDI-TOF analysis of blood serum proteome can predict the presence of monoclonal gammopathy of undetermined significance

Abstract

Monoclonal gammopathy of undetermined significance (MGUS) is a plasma cell dyscrasia that can progress to malignant multiple myeloma (MM). Specific molecular biomarkers to classify the MGUS status and discriminate the initial asymptomatic phase of MM have not been identified. We examined the serum peptidome profile of MGUS patients and healthy volunteers using MALDI-TOF mass spectrometry and developed a predictive model for classifying serum samples. The predictive model was built using a support vector machine (SVM) supervised learning method tuned by applying a 20-fold cross-validation scheme. Predicting class labels in a blinded test set containing randomly selected MGUS and healthy control serum samples validated the model. The generalization performance of the predictive model was evaluated by a double cross-validation method that showed 88% average model accuracy, 89% average sensitivity and 86% average specificity. Our model, which classifies unknown serum samples as belonging to either MGUS patients or healthy individuals, can be applied to clinical diagnosis.

Fig 1. 20-fold cross-validation scheme.

Fig 2. Double cross-validation scheme.

It highlights the two nested loops. The outer cross-validation loop provides 10 performance estimates from predicting the corresponding test set by the optimized model built in the inner 20 fold cross-validation loop. The data set used to build and tune the model in the inner cross-validation loop is completely independent of the test set used in the outer iteration.

Fig 3. Classification process of an unknown serum sample.

n technical replicates from serum sample raw mass spectra are pre-processed and features selected. The correlation threshold (r_th) sets the k technical replicates passing the intra-experimental quality control (QC). The SVM predictive model classifies the k technical replicates. The majority voter assigns the serum sample predicted class. The parameters determined from the processing of the training set and the building of the predictive model are shaded.