Esquemas de seleção dinâmica para a identificação de bactérias a partir de dados de m/z virtuais de proteínas ribossomais

Abstract

A gram of soil can contain up to 8.3 million different bacterial species, among which are those that favor agricultural productivity, promoting plant growth and protecting against pests and diseases. One approach that has been used in identifying these microorganisms is through the fingerprint ribosomal proteins in bacterial mass spectra extracted by a chemical analytical technique called MALDI-TOF. This technique extracts charge mass information (m/z) from the molecules of a sample, which can be subjected to digital classifiers that label the sample data according to their taxonomy level. However, in most cases, these data sets have multiple classes and high imbalance ratio, making it difficult to develop digital classifiers. Thus, the multiple classifier systems provide ways to treat these problems. In this context, dynamic selection schemes that use meta-learning explore a set of meta-features, extracted from the training set, to estimate the level of competence of the base classifiers and then select the most suitable set of classifiers to predict a sample. Considering the above, this work evaluates the performance of different dynamic selection schemes for the identification of bacterial genera from their fingerprint in terms of m/z of ribosomal proteins. The work also presents the scheme called METADES-i that extends the METADES scheme, through the use of meta-features sensitive to imbalanced data. The performance of the tested dynamic selection schemes was measured in terms of average accuracy, balanced accuracy, geometric mean and overfitting in a PUKYU synthetic data set. In the experiments different scenarios were used, defined as subsets of the data set PUKYU. In addition, the influence of the composition of the base classifiers: homogeneous or heterogeneous was analyzed. The results of the experiments show that the proposed scheme was significantly superior to the others in terms of balanced accuracy and geometric mean. In terms of average accuracy, the METADES-i scheme was superior only when using the homogeneous composition. Regarding overfitting, the schemes with the best performance were METADES-i, KNOP and KNORA-U. The multiobjective analysis between balanced accuracy and overfitting indicated that the METADES-i scheme participated in the dominance frontier in all scenarios. The result related to the meta-featus selection procedure, indicated that when applying the Relief method to METADES-i, there was an improvement in assertiveness metrics, but there was a decrease in performance in terms of overfitting. Regarding the composition of the classifiers, the heterogeneous composition proved to be superior in most cases. Finally, the results suggest that the adequacy of the meta-features subset of dynamic selection schemes based on meta-learning, can increase the performance of multiple classifier systems in terms of assertiveness in imbalanced data sets.