DeepConsensus improves the accuracy of sequences with a gap-aware sequence transformer.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Pacific Biosciences (PacBio) circular consensus sequencing generates long, accurate 'HiFi' reads using a hidden Markov model (pbccs) for consensus generation.
• Accurate DNA sequencing is crucial for high-quality genome assembly and variant analysis.

Purpose of the Study:

• To introduce DeepConsensus, a novel deep learning approach for improving DNA sequence accuracy in PacBio HiFi reads.
• To evaluate the performance of DeepConsensus compared to the standard pbccs method.

Main Methods:

• DeepConsensus employs an alignment-based loss function to train a gap-aware transformer-encoder model for sequence correction.
• The model was evaluated on HG003 genome data using PacBio HiFi reads.

Main Results:

• DeepConsensus reduced read errors by 42% compared to pbccs.
• This led to significant increases in the yield of high-quality PacBio HiFi reads (Q20, Q30, Q40).
• DeepConsensus-generated reads improved genome assembly contiguity (NG50 from 4.9 Mb to 17.2 Mb), gene completeness (94% to 97%), and base accuracy (Q43 to Q45), while reducing variant-calling errors by 24%.

Conclusions:

• DeepConsensus offers a substantial improvement over standard methods for PacBio HiFi read consensus generation.
• The developed deep learning models show potential for broader applications in sequence alignment analysis.