Detecting False Positive and False Negative Overlaps in Assembly Graphs

Među središnjim problemima bioinformatike je rekonstrukcija genoma. Manji dijelovi genoma, zvani očitanjima, trebaju biti sastavljena u potpuni genom. Očitanja se međusobno preklapaju u svrhu određivanja onih koja su zaista susjedna u originalnom genomu. Alati za rekonstrukciju ponekad uzimaju u obzir preklapanja koja u stvarnom genome ne postoje, ili propuste točna preklapanja. Ovaj rad nastoji klasificirati takva očitanja u točno pozitivnu, lažno pozitivnu ili lažno negativnu klasu. Koristi brojčane i tekstualne značajke preklapanja, poput sličnosti, dužine preklapanja i CIGAR niza. U sklopu rada predstavljeno je nekoliko arhitektura dubokih modela, poput neuronske mreže i transformatora za obradu jezika.One of the biggest problems in bioinformatics is genome assembly. Smaller parts of genome, called reads, must be combined to reconstruct the whole genome. Reads are mutually overlapped to understand which ones are adjacent. Assembly algorithms sometimes use overlaps not actually of adjacent reads in the genome or miss some overlaps. This thesis tries to classify those overlaps into true positive, false positive, or false negative classes. For classification, it uses numerical and text features of overlaps, like overlap length, similarity, and CIGAR string. The thesis introduces multiple deep learning architectures, like neural networks and transformers for language processing