Unsupervised Barcode Demultiplexing
)
The current approaches to demultiplexing of barcoded reads typically use base called sequences and tend to render up to 20% of the reads unusable due to base calling errors. In contrast, Deepbinner (Wick et al., 2018) works with the raw signal by employing a convolutional neural network and loses only ≈ 5% of reads, while retaining the precision of ≈ 98%. We present a novel approach that also operates in the signal space, but is based on unsupervised learning.