Identificação de snoRNAs usando aprendizagem de máquina

Machine learning methods have been widely used to identify and classify different families of non-coding RNAs. Many of these methods are based on supervised learning, where some previous known attributes, called features, are extracted from a sequence, and then used in a classifier. In this work, we present two methods to identify the two main classes of snoRNAs, C/D box and H/ACA box: snoReport 2.0, a significant improvement of the original snoReport version; and snoRNA-EDeN, a new method based on EDeN, a decompositional graph kernel. On one hand, snoReport 2.0 is a method that, using features extracted from candidate sequences in genomes, combines secondary structure prediction with Support Vector Machine (SVM) to identify C/D box and H/ACA box snoRNAs. H/ACA box snoRNA classifier showed a F-score of 93% (an improvement of 10% regarding to the previous version), while C/D box snoRNA classifier a F-Score of 94% (improvement of 14%). Besides, both classifiers exhibited performance measures above 90%. In the validation phase, snoReport 2.0 predicted 67.43% of vertebrate organisms for both classes. SnoReport 2.0 predicted: for Nematodes, 29.6% of C/D box and 69% of H/ACA box snoRNAs; and for Drosophilids, 3.2% of C/D box and 76.7% of H/ACA box snoRNAs. These results show that snoReport 2.0 is efficient to identify snoRNAs in vertebrates, and also H/ACA box snoRNAs in invertebrates organisms. On the other hand, instead of using known features from a sequence (difficult to find in general), a recent approach in machine learning is described as follows. Given a region of interest of a sequence, the objective is to generate a sparse vector that can be used as micro-features in a specific machine learning algorithm, or it can be used to create powerful features. This approach is used in EDeN (Explicit Decomposition with Neighbourhoods), a decompositional graph kernel based on Neighborhood Subgraph Pairwise Distance Kernel (NSPDK). EDeN transforms one graph in a sparse vector, decomposing it in all pairs of neighborhood subgraphs of small radius at increasing distances. Based on EDeN, we developed a method called snoRNA-EDeN. On the test phase, for C/D box snoRNAs, snoRNA-EDeN showed a F-score of 93.4%, while for H/ACA box snoRNAs, the F-score was 72%. On the validation phase, for C/D box snoRNAs, snoRNA-EDeN showed a better capacity of generalization, predicting 94.61% of vertebrate C/D box snoRNAs and