DNA Nanopore Sequencing Basecaller

Sekvenciranje nanoporama je jedna od vodećih tehnologija sekvenciranja danas. Prolaskom DNA kroz poru mijenja se ionska struja koja teče porom. Uslijed veličine pore, obilno se u pori nalazi 5 nukleotida (5-torka) koji utječu na mjereni signal. Svaka od 1024 moguće 5-torke uzrokuje različit signal i ta informacija se koristi za određivanje baza (pretvaranje sirovog signala u slijed nukleotida). Oblik signala je pravokutan jer se 5-torke mijenjaju po jedan nukleotid, ali je prisutan šum. Cilj rada je razvoj nove metode za pretvaranje signala dobivenog sekvenciranjem nanoporama korištenjem suvremenih arhitektura dubokog učenja, pritom imajući na umu mogućnost korištenja samo-nadziranog učenja. Arhitektura se temelji na transformerima. Za evaluaciju su korišteni javno dostupni skupovi podataka. Rješenje je implementirano u programskom jeziku Python uz korištenje PyTorch biblioteke. Izvorni kod je dostupan na Github repozitoriju na github.com/StanislavPavlic/attentioncall.Nanopore sequencing is one of the state-of-the-art sequencing technologies. It passes a DNA sample through a pore which changes the ionic current in the pore. Due to the size of the pore, there are usually five nucleotides (5-mer) present in the pore influencing the measured signal. Each of the 1024 possible 5-mers produces a different signal, and this information is used for basecalling (converting the raw signal to a sequence of nucleotides). The signal is approximately rectangular because the 5-mer changes one nucleotide at a time, but there is a lot of noise present. The goal of this thesis was to develop a DNA nanopore sequencing basecaller using modern deep learning architectures with self-supervised learning in mind. The architecture is mainly based on transformers. The basecaller was evaluated on publicly available datasets. The solution called AttentionCall was implemented in Python and the PyTorch library. The source code is available on GitHub at github.com/StanislavPavlic/attentioncall