Feature Selection Methods for Early Predictive Biomarker Discovery Using Untargeted Metabolomic Data

Abstract

Untargeted metabolomics is a powerful phenotyping tool for better understanding biological mechanisms involved in human pathology development and identifying early predictive biomarkers. This approach, based on multiple analytical platforms, such as mass spectrometry (MS), chemometrics and bioinformatics, generates massive and complex data that need appropriate analyses to extract the biologically meaningful information. Despite various tools available, it is still a challenge to handle such large and noisy datasets with limited number of individuals without risking overfitting. Moreover, when the objective is focused on the identification of early predictive markers of clinical outcome, few years before occurrence, it becomes essential to use the appropriate algorithms and workflow to be able to discover subtle effects among this large amount of data. In this context, this work consists in studying a workflow describing the general feature selection process, using knowledge discovery and data mining methodologies to propose advanced solutions for predictive biomarker discovery. The strategy was focused on evaluating a combination of numeric-symbolic approaches for feature selection with the objective of obtaining the best combination of metabolites producing an effective and accurate predictive model. Relying first on numerical approaches, and especially on machine learning methods (SVM-RFE, RF, RF-RFE) and on univariate statistical analyses (ANOVA), a comparative study was performed on an original metabolomic dataset and reduced subsets. As resampling method, LOOCV was applied to minimize the risk of overfitting. The best k-features obtained with different scores of importance from the combination of these different approaches were compared and allowed determining the variable stabilities using Formal Concept Analysis. The results revealed the interest of RF-Gini combined with ANOVA for feature selection as these two complementary methods allowed selecting the 48 best candidates for prediction. Using linear logistic regression on this reduced dataset enabled us to obtain the best performances in terms of prediction accuracy and number of false positive with a model including 5 top variables. Therefore, these results highlighted the interest of feature selection methods and the importance of working on reduced datasets for the identification of predictive biomarkers issued from untargeted metabolomics data.

Keywords: biomarker discovery; feature selection; formal concept analysis; machine learning; metabolomics; prediction; univariate statistics; visualization.

Figure 1

General feature selection process.

Figure 2

General framework. Main phases of metabolomic data treatment.

Figure 3

Detailed approach. Representation of the different steps of the proposed approach, from the untargeted metabolomics original dataset to the identification of predictive biomarkers. It includes (i) data transformation, i.e., noise filtering, scaling, to generate suitable datasets for feature selection methods; (ii) data reduction for feature selection, which consists in identifying relevant features for further use in predictive models; (iii) a prediction and validation step for discovering the best predictive markers. The numbers in the circles refer to the different sections of the manuscript for a more detailed description.

Figure 4

Experimental design. Experimental design for comparisons of feature selection methods (RF, RF-RFE, SVM-RFE, ANOVA), applied either on original dataset or after filters, based either on correlation coefficient (Cor) or on mutual information (MI), and used with different classifiers (MdGini, MdAcc, Kappa, W, p-value). It resulted in 10 different subsets, with different feature rankings.

Figure 5

The concept hierarchy derived from 48 × 10 binary table of Supplementary Table 1. It highlights the relationships existing between top-ranked features and selection methods. It allows visualizing the common features (ions) selected by a set of methods.

Figure 6

Correlation network between the 11 top-predictive features. Network built using Pearson correlation coefficient (indicated on the edges) between the best predictive features. Red edges: positive correlations, Blue edges: negative correlations. It highlighted two highly correlated sub-networks (yellow and green).