Ablation study of deep neural networks for predicting epigenetic modifications

Glavni cilj bioinformatike je pronalazak novih odgovora u području analize genoma, koje ima važnu ulogu u očuvanju i poboljšanju kvalitete života. Neka pitanja na koja se traži odgovor tiču se DNA-metilacije, koja je svojevrsna modifikacija DNA niti. Danas se istraživanja vezanu uz tu modifikaciju sve više provode korištenjem algoritama strojnog učenja. Rockfish je jedan takav model koji se temelji na transformatoru koji koristi neuronske mreže za pronalaženje metiliranih elemenata u DNA niti. U ovom radu, korištenjem programskog jezika Python, neki dijelovi modela su uklonjeni, i time se provodi ablacija, a ponekad i dodatno trenirani kako bi se primijetio utjecaj različitih komponenti i utvrdile njihove važnosti. Rezultati ablacije proučavaju se kroz četiri različite usporedbe, gdje se procjenjuju točnost, preciznost, odziv i specifičnost svakog modela. Modeli se zatim uspoređuju, a komponente modela rangiraju se prema njihovim rezultatima. Zaključci se donose na temelju dobivenih rezultata.The main goal of bioinformatics is to find new answers regarding genome analysis and the field is an important factor in preserving and improving the quality of life. Some questions that need to be answered are regarding DNA methylation, which is a modification of the DNA strand. Nowadays, research is increasingly conducted by utilizing machine learning algorithms. Rockfish is a transformer based-model that uses neural networks to find which elements of the DNA strand are methylated. In this thesis, by using Python programming language, some components of the model are removed, thus performing ablation, and sometimes trained additionally so that the impact of the given components and their importance could be noticed. The results of the ablation study are inspected through four different comparisons, where the accuracy, precision, recall, and specificity of each model are evaluated. The models are then compared and the components of the model are ranked based on their results. Conclusions are made based on the results that were obtained