Deep learning-based detection of viral subtypes

Rješavanje problema detekcije podtipova virusa omogućilo bi lakšu analizu genomskih sljedova virusa, kao i pronalazak epidemioloških i geografskih poveznica između pacijenata te bi potencijalno olakšalo osmišljavanje terapije prilagođene za infekciju pojedinim podtipom. Trenutno se detekcija podtipova virusa provodi algoritmima strojnog učenja poput skrivenih Markovljevih lanaca ili raznih algoritama grupiranja. U ovom radu iskorištena je arhitektura dubokog učenja transformer, koja se pokazala iznimno uspješnom u području obrade prirodnog jezika. Arhitektura transformer je proširena na način da neovisno trenira pojedine podnizove ulaznoga slijeda, što dodatno omogućuje i detekciju rekombinanata virusa. Dobiveni rezultati se nisu pokazali boljim od već postojećih rješenja, ali demonstriraju velik potencijal primjene transformera za rješavanje ovog problema.Solving the problem of virus subtype detection would enable easier analysis of viral genomic sequences, as well as the discovery of epidemiological and geographical links between patients. It also has the potential to develop therapies for specific subtype infections easier. Currently, virus subtype detection is performed using machine learning algorithms such as hidden Markov models or various clustering algorithms. In this study, the transformer architecture, which has shown exceptional success in the field of natural language processing, was used. The transformer architecture was extended to independently train individual subsequences of the input sequence, which enables the detection of viral recombinants. The obtained results were not better than the existing solutions, but they demonstrate the great potential for solving this problem