Deep learning on genomic and phylogenetic data

Metode globokega učenja v praksi dosegajo izjemne rezultate pri reševanju problemov na različnih področjih, med drugim tudi v genomiki. V diplomski nalogi smo se ukvarjali z razvrščanjem genskih zaporedij bakterij v taksonomske razrede. Cilj je bil zgraditi model, ki bo znal bakterijo na podlagi zaporedja njenega gena 16S rRNA razvrstiti v pravo deblo, razred, red, družino in rod. Z uporabo metod globokega učenja smo zgradili več klasifikacijskih modelov in vrednotili njihovo uspešnost na podlagi klasifikacijske točnosti in mere F1. Med seboj smo primerjali konvolucijske nevronske mreže, preproste rekurenčne nevronske mreže, dvosmerne rekurenčne nevronske mreže, kombinirane modele z rekurenčnimi in konvolucijskimi nevronskimi mrežami ter metodo naključnih gozdov. Eksperimente smo izvedli na dveh različno velikih množicah podatkov, preverili pa smo tudi, kako se modeli obnesejo pri klasifikaciji, če imajo na voljo le krajši del genskega zaporedja. Rezultati kažejo, da so za reševanje tovrstnih problemov najbolj primerne konvolucijske nevronske mreže.Deep learning methods have been achieving amazing results in solving a variety of problems in many different fields, a very important one of them being genomics. In the thesis, deep learning methods have been used to classify bacterial DNA sequences into taxonomic ranks. The goal was to build a classification model based on the bacteria\u27s 16S rRNA sequence and classify a bacteria by phylum, class, order, family and genus. The performance of five different models has been compared in terms of accuracy and F1 score. A model with convolutional neural networks, simple recurrent neural network, bidirectional neural network, a hybrid model that combines convolutional and neural network and a model using random forests have been built. Two experiments have been conducted. In the first one classification was based on the whole sequence. In the second one only a small sequence fragment was used. We evaluated the performance of the models based on two datasets of different sizes. Results show that convolutional neural networks outperformed other models in all the cases