Search

Interakcije med proteini in RNA imajo ključno vlogo pri velikem številu celičnih procesov. Eksperimentalna analiza 3D struktur molekul je počasna in zahtevna, zato obstaja velika potreba po računskih metodah, ki uspešno napovedujejo mesta ter strukturo molekul v interakciji. V magistrskem delu smo definirali vrsto značilk, ki opisujejo lokalne lastnosti interakcij protein-RNA, na podlagi podatkov o 3D strukturah molekul protein-RNA. Razvili smo metodo, ki združuje strojno učenje in optimizacijski postopek za napovedovanje mesta interakcij med proteinom in RNA. Napovedi strojnega učenja se uporabijo za določanje začetnega stanja optimizacije. Optimizacijski postopek nato uporabi ocenjevalne funkcije osnovane na porazdelitvi 3D strukturnih značilk in tako predlaga najverjetnejšo pozicijo molekule RNA. Predlagani napovedni model dosega natančnost, ki je primerljiva z uspešnostjo najboljših obstoječih metod. Protein-RNA interactions have an essential role in many cellular processes. Experimental analysis of 3D molecular structure is slow and difficult process. Consequently, computational methods, which successfully predict interaction sites and molecular conformations are needed. In this thesis we have defined a number of attributes to describe local properties of protein-RNA interactions using data on 3D structure of protein-RNA molecules. We have implemented a method that uses machine learning and optimization algorithm for prediction of protein-RNA interaction sites. Machine learning predictions are used to generate initial positions for optimization. Optimization algorithm uses scoring functions based on the distribution of 3D structural attributes to identify most likely positions of the RNA molecule interacting with a given protein. The accuracy of the proposed prediction model is comparable to results obtained with best existing methods.

Proteini so pomembni akterji v mnogih celičnih procesih. Interakcije proteinov in RNA imajo pomembno vlogo pri uravnavanju izražanja genov in posledično njihove funkcije. V interakcijo z RNA navadno vstopa več tako posameznih proteinov kakor tudi skupin proteinov. Metoda iCLIP omogoča do nukleotida natančno detekcijo mest interakcij na RNA z izbranim proteinom. V diplomskem delu smo razvili metodo, ki sprejme nabor mest na RNA, ki so v interakciji z izbranim proteinskim kompleksom. V okolici teh mest nato poišče vzorce interakcij iz nabora interakcij proteinov z RNA, za katere imamo podatke. Najdeni vzorci vključujejo tako posamezne komponente (proteine) danega kompleksa kakor tudi proteine, ki danega kompleksa ne tvorijo, ampak vseeno vplivajo na interakcije kompleksa z RNA. Razvita metoda temelji na postopeku faktorizacije nenegativnih matrik. Metodo smo uspešno preizkusili na podatkih o kompleksu SmB, kjer smo uspešno določiti nekaj proteinov, ki so soudeleženi v kompleksu SmB. Proteins play an important role in many processes in a cell. Protein-RNA interactions greatly affect the balancing of gene expressions and consequently their functions. Interaction with RNA may occur with a single protein or a protein complex. iCLIP method is able to detect the protein-RNA crosslink spots with nucleotide resolution. We developed a method that takes a set of crosslink spots that intersect with the choosen protein complex as an input. Method is searching the crosslink spots neighbourhood for the protein-RNA interaction patterns. Found patterns include parts (proteins) of the choosen complex and other proteins that are not a part of the choosen complex, but still affect the interactions of the choosen complex with the RNA. Method that we developed is based on the non-negative matrix factorization. We successfully tested the method on a SmB complex where we found a few proteins that cooporate with SmB complex.