DeepConsensus

DeepConsensus uses gap-aware sequence transformers to correct errors in Pacific Biosciences (PacBio) Circular Consensus Sequencing (CCS) data.

Key Features:

• Gap-aware transformer-encoder: Uses an alignment-based loss to train a gap-aware transformer-encoder for sequence correction.
• Error reduction: Reduces read errors by 42% compared to pbccs, the standard hidden Markov model approach.
• HiFi read yield improvement: Increases the yield of PacBio HiFi reads at Q20 by 9%, at Q30 by 27%, and at Q40 by 90%.
• Generalizable models: Models can be trained to analyze alignments of other sequence types, such as unique molecular identifiers or genome assemblies.

Scientific Applications:

• Genome assembly: Improves hifiasm assembly contiguity (NG50 from 4.9 megabases to 17.2 megabases) and assembly base accuracy (Q43 to Q45).
• Gene annotation: Increases gene completeness (94% to 97%) and reduces false gene duplication rate (1.1% to 0.5%).
• Variant calling: Reduces variant-calling errors by 24%.

Methodology:

DeepConsensus uses an alignment-based loss to train a gap-aware transformer-encoder for sequence correction.