Análise empírica da utilização de técnicas de aprendizagem de máquina para classificação de sequências de proteínas de Metarhizium anisopliae

• Autor(a) principal: Dias, Maria Fernanda Ribeiro
• Data de Publicação: 2014
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/198
• topic: Bioinformática; Ubiquitinas; Bioinformatics; Ciência da Computação; Ciências Biológicas

Metarhizium anisopliae} is an entomopathogenic fungus used as biological insecticide possessing proteins linked to infection processes and unknown control mechanisms, many of which may be controlled by the ubiquitination system. In this work we used machine-learning techniques to predict {\it M. anisopliae~ isolate E6} ubiquitination-prone proteins. One hundred fifty-one ubiquitinable peptides and one hundred fifty-one non-ubiquitinable peptides from {\it S. cereviseae} and {\it H. sapiens} were used as the training set.\!\footnote{(http://iclab.life.nctu.edu.tw/ubipred/)} These peptides were composed by 21 amino acids flanking a central lysine residue. Each of these peptides was represented as a numerical vector corresponding to the average value for their constituent amino-acids of each of the 31 physicochemical properties previously used in Ubipred. Hierarchical clustering of ubiquitinable proteins from this dataset showed evidence of correlation between several physicochemical properties, indicating redundancy in these features. Redundant features may cause model overfitting and increase computational cost. We used the classification algorithm Weighted-Voting (W-V), with cross validation, to disclose the minimal set of features best correlated with the probability of a given peptide being ubiquitinable. WeightedVotingXvalidation performed similarly for vectors of 10 and 31 dimensions. Indeed, each of the 10 features in the minimal set correlates with most of the remaining 21 features, as confirmed by pairwise Pearson correlation test (coefficients ranging from -0.95 to -0.40 and 0.40 to 0.98). We then compared the performance of the algorithms W-V and Support Vector Machine (SVM) with radial kernel for vectors with 10 or 31 dimensions. In both cases, SVM outperformed W-V. The resulting of recall, precision and accuracy with 10 features were respectively, 67\%, 65\% and 66\% for SVM against 65\%, 55\% and 47\% for W-V and with 31 features were 71\%, 71\% and 71\% for SVM against 60\%, 55\% and 52\% for W-V. Processing time in ASUS K43U, Process with AMD Dual Core C60 1.0 GHz, 2 x 512 KB of cache memory, 4 GB of RAM. Processing time was 8h and 22h, for SVM with 10 or 31 physicochemical features, respectively. Considering the risk of overfitting the model due to features redundancy, we applied the SVM trained with 10 features to search for ubiquitination-prone proteins in the predicted proteome of {\it M. anisopliae~ isolate E6} (10,775 proteins). The 160,694, 21-amino acids long peptides flanking a central lysine residue extracted from these proteins were represented as a 10 dimensional vector for the training dataset. Forty-four of these proteins with no lysine were automatically excluded from this analysis. The classifier predicted 9,314 proteins as being ubiquitination-prone. The small loss of performance of the SVM after dimensionality reduction is compensated by the significant reduction in processing time. The feature correlations suggest a lower risk of overfitting for the 10-dimensions model.