BaseNet: A transformer-based toolkit for nanopore sequencing signal decoding.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Nanopore sequencing offers rapid, high-throughput nucleic acid analysis.
• Accurate basecalling is critical for reliable downstream genomic data interpretation.
• Existing methods like HMM, RNN, and CNN have limitations in accuracy and reliability.

Purpose of the Study:

• Introduce BaseNet, an open-source toolkit for advanced nanopore sequencing signal decoding.
• Utilize transformer models to improve basecalling accuracy and reliability.
• Provide state-of-the-art, accessible algorithms for the research community.

Main Methods:

• Developed BaseNet, incorporating autoregressive and non-autoregressive transformer decoders.
• Employed cross-attention mechanisms to map signal-to-sequence relationships.
• Implemented joint loss training with forward and reverse decoders and utilized large-scale pre-trained models.

Main Results:

• Demonstrated effective mapping of signal-to-sequence relationships using cross-attention weights.
• Showcased improved model convergence through joint loss training.
• Achieved superior decoding accuracy with large-scale pre-trained transformer models.

Conclusions:

• BaseNet advances nanopore sequencing signal decoding with transformer-based approaches.
• The toolkit provides novel concepts and tools for enhanced basecalling accuracy.
• This work contributes to the technological progress in high-throughput sequencing analysis.