Melchior: A Hybrid Mamba-Transformer RNA Basecaller

AbstractSequence transduction from raw nanopore signals is notoriously difficult because the signal level does not naturally correspond to a single base, but rather many adjacent nucleotides. Thus, we introduce Melchior, an RNA basecaller that uses a hybrid Mamba-Transformer backbone to achieve global visual context at a lower computational complexity. This is in contrast to temporal convolutions, which collate features by fusing both spatial and channel features in the local receptive field. Melchior is also able to exploit the full complementarity between local and global features, unlike Vision Transformers, which have empirically been observed to ignore local features. Augmenting a selective structured state-space sequence model with self-attention unlocks unprecedented performance gains, particularly in homopolymer regions, by modeling fine-grained details in both short and long-range spatial dependencies.