Transfer learning for prediction of transcription start sites across different plant speciesMiškić, David (Avtor) Curk, Tomaž (Mentor) Zrimec, Jan (Komentor) transcription start sitepolymerasebioinformaticstransformertransfer learningTranscription start site (TSS) prediction is a classification problem at the intersection of machine learning and laboratory gene expression measurement methods. The site is significant as it represents the location where the first nucleotide is transcribed by RNA polymerase and can help characterize the genome of an organism. We have developed two variants of prediction models in the plant model organism \textit{Arabidopsis thaliana} based on an existing expression model Enformer, using upscaling and a custom loss function that proved crucial for training success. The GFF model type uses genome annotation information to supplement the context, and this has proven to facilitate the transfer between plant organisms, demonstrated by transfer learning on corn. The MultiTSS model type uses DNA sequence alone with no substantial performance degradation compared to the GFF model, demonstrating that it is able to capture and learn important motifs that characterize a TSS. We show that the developed methods are comparably better than existing approaches and can be applied without retraining as well. We also describe the procedure and pitfalls of the problem area with potential solutions.20232023-09-29 14:30:00Magistrsko delo/naloga151112sl