Deep learning-based pairwise alignment of protein sequences

Poravnanje sljedova važan je postupak u bioinformatici koji se može opisati kao određivanje korespondencija (podudaranje, nepodudaranje, umetanje ili brisanje znakova) između pojedinih elemenata dvaju ili više bioloških sljedova. Svrha poravnanja je procjena sličnosti sljedova na temelju čega se može odrediti jesu li sljedovi homologni, odnosno jesu li evolucijski srodni. Ovaj rad istražuje mogućnost poravnanja parova proteinskih sljedova korištenjem transformer modela. Transformer modeli originalno su korišteni za zadatke u području obrade prirodnog jezika, međutim pokazali su se vrlo uspješnima i u drugim područjima, poput računalnog vida, obrade zvuka i bioinformatike. U sklopu rada napravljena je vlastita implementacija transformer modela korištenjem okvira TensorFlow i Keras. Za potrebe treniranja i testiranja modela pripremljeni su skupovi neporavnatih i poravnatih proteinskih sljedova hemoglobina, a referentna poravnanja dobivena su algoritmom Needleman-Wunsch. Uspoređeni su različiti načini reprezentacije poravnatih i neporavnatih sljedova kao i različite arhitekture transformer modela. Ocjena podudaranja stupaca koristila se kao mjera za usporedbu s referentnim poravnanjem, a za najbolji model na skupu za testiranje iznosila je 71%.Sequence alignment is an important process in bioinformatics that can be described as the assignment of correspondences (match, mismatch, insertion or deletion) between nucleotides or amino acids. The goal of sequence alignment is to assess the similarity between aligned sequences, which can later be used to determine whether the sequences are homologous. This thesis explores the possibility of using transformer models for pairwise protein sequence alignment. Transformer models were originally used for natural language processing tasks, but they proved to be successful in many different fields, such as computer vision, sound analysis and bioinformatics. In this thesis TensorFlow and Keras libraries were used for transformer model implementation. Datasets consisting of unaligned and aligned hemoglobin protein sequences were created for model training and testing. Reference alignments were obtained using the Needleman-Wunsch algorithm. Different ways of representing unaligned and aligned sequences were compared as well as different model architectures. Column score was used as a measure of similarity with reference alignments. Best model achieved a column score of 71% on test data