Klasifikacija virusnih in bakterijskih genomov s konvolucijskimi nevronskimi mrežami

Alič, Anže (author), Curk, Tomaž (mentor)

V diplomski nalogi smo razvili nevronsko mrežo za napovedovanje uvrstitve v taksonomijo organizma na podlagi kratkih izsekov genoma, ki jih pridobimo s sekvenciranjem. Ovrednotili smo uporabo dodatnih opisov genoma in uporabo dodatnih relacij med ciljnimi razredi. Za reševanje problema in gradnjo napovednih modelov smo uporabili globoke konvolucijske nevronske mreže. Predlagano metodo smo preizkusili na referenčnih genomih virusov in bakterij. Eksperimentalni rezultati pokažejo, da smo zastavljen problem rešili uspešno, saj smo, v najboljšem primeru, dosegli ROC AUC 0.92. Dodatni opisi genoma ne izboljšajo uspešnosti napovednega modela. Uporaba dodatnih relacij v obliki taksonomije zelo izboljša napovedni model, saj nevronska mreža, kljub manjšemu številu učnih podatkov, uspešno generalizira zastavljen problem.

• Language: Slovenian
• Keywords: CNN, bioinformatika, klasifikacija virusov in bakterij, genomski pripisi
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science ; FMF - Faculty of Mathematics and Physics
• Year: 2020
• PID: 20.500.12556/RUL-114438
• COBISS.SI-ID: 1538544067
• Publication date in RUL: 28.02.2020

Secondary language

• Language: English
• Title: Viral and bacterial genome classification using convolutional neural networks

Abstract

We have developed a neural network for assigning taxonomic labels based on short genome sequences. Sequences are collected with next-generation sequencing. We evaluated the contribution of extra attributes of genome and the influence of using extra dependencies between target classes, which are encoded as a taxonomy. We designed a deep convolutional neural network and evaluated it on reference genomes of viruses and bacteria. Experiments show that at best we can achieved a ROC AUC score of 0.92. We observed, that the use of extra attributes did not improve accuracy. However, using the information on extra dependencies among target classes decreased the number of training data needed for training the convolutional neural network.

• Keywords: CNN, bioinformatics, virus and bacteria classification, genome annotation