Izboljšanje analize dinamike RNA z uporabo strojnega učenja na multiomskih podatkih

The RNA lifecycle greatly influences and directs the functionality of the cell. Its progression is orchestrated by multiple RNA-binding proteins (RBPs), forming complex and interconnected networks of transcript regulation. To disentangle this network and extract the RNA features influencing its fate, we are often presented with a multitude of single or multiomics information for the unique biological groups we are attempting to differentiate. Commonly, multiple statistical analyses and workflows are employed to manually extract and compare the differentiating features. However, this can introduce confirmation bias, and techniques able to compare the raw data need to be further explored. In this masters thesis, we have developed a machine learning (ML) workflow capable of classifying transcripts based on positional features such as sequence, structure, RNA-binding protein binding and methylation. Employing this workflow we were able to efficiently classify the stabilized and unstabilised transcripts by LIN28A in mouse embryonic stem cells (mESC), finding multivalent AU-rich regions toward the ends of 3’UTRs to be predictive of LIN28A mediated destabilization of transcripts. Expanding this methodology to include multiomics datasets in the second case study, we were able to extract the common features of the condensation prone RNA in mESC to be mainly structured C-rich coding regions with highly protein bound ends of 3’ untranslated regions. This demonstrated an effective use of ML to offer unique biological insights into the features governing different networks of RBP-RNA interactions. The use of such models is therefore anticipated to further expand the bioinformatics toolset and enable further unbiased view into diverse roles of RNA regulation with highly predictive and explainable ML models.