engleski

End-to-End Deep Learning Model for Base Calling of MinION Nanopore Reads

The MinION device by Oxford Nanopore Technologies is the first portable DNA sequencing device. Main advantages include producing longer reads than competing technologies and real-time data analysis making it suitable for a wide array of possible applications. Although long reads of up to 882 000 bp can be achieved, this comes at a cost - an error rate of 10% or higher. The goal of this thesis is to explore novel basecaller training technique using multi- task training and autoencoder loss as a secondary task to improve performance of basecalling. The model has been trained on R9.4 E.Coli dataset and has been compared with contemporary solutions on Klebsiella pneumoniae dataset. The complete source code is available on https://github.com/nmiculinic/minion-basecaller/.

base calling ; Oxford Nanopore Technologies ; MinION ; deep learning ; seq2seq ; convolutional neural network ; residual network ; CTC loss

nije evidentirano