Back

DeepConsensus: Gap-Aware Sequence Transformers for Sequence Correction

Baid, G.; Cook, D. E.; Shafin, K.; Yun, T.; Llinares-Lopez, F.; Berthet, Q.; Wenger, A. M.; Rowell, W. J.; Nattestad, M.; Yang, H.; Kolesnikov, A.; Topfer, A.; Ammar, W.; Vert, J.-P.; Vaswani, A.; McLean, C. Y.; Chang, P.-C.; Carroll, A.

2021-08-31 bioinformatics
10.1101/2021.08.31.458403 bioRxiv
Show abstract

Pacific BioScience (PacBio) circular consensus sequencing (CCS) generates long (10-25 kb), accurate "HiFi" reads by combining serial observations of a DNA molecule into a consensus sequence. The standard approach to consensus generation uses a hidden Markov model (pbccs). Here, we introduce DeepConsensus, which uses a unique alignment-based loss to train a gap-aware transformer-encoder (GATE) for sequence correction. Compared to pbccs, DeepConsensus reduces read errors in the same dataset by 42%. This increases the yield of PacBio HiFi reads at Q20 by 9%, at Q30 by 27%, and at Q40 by 90%. With two SMRT Cells of HG003, reads from DeepConsensus improve hifiasm assembly contiguity (NG50 4.9Mb to 17.2Mb), increase gene completeness (94% to 97%), reduce false gene duplication rate (1.1% to 0.5%), improve assembly base accuracy (Q43 to Q45), and also reduce variant calling errors by 24%.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

1
Genome Biology
○ 637 papers in training set
Top 0.4%
17.9%
Similar papers in this journal
2
Nature Communications
○ 5641 papers in training set
Top 13%
14.6%
Similar papers in this journal
3
Nature Biotechnology
○ 172 papers in training set
Top 0.1%
11.5%
Similar papers in this journal
4
Bioinformatics
◐ 1204 papers in training set
Top 3%
9.3%
Similar papers in this journal
50% of probability mass above
5
Nature Methods
○ 385 papers in training set
Top 1%
9.3%
Similar papers in this journal
6
Genome Research
● 468 papers in training set
Top 1.0%
6.0%
7
NAR Genomics and Bioinformatics
◐ 242 papers in training set
Top 1%
3.3%
8
BMC Bioinformatics
○ 457 papers in training set
Top 3%
2.6%
9
Nucleic Acids Research
◐ 1281 papers in training set
Top 7%
2.3%
10
Communications Biology
○ 993 papers in training set
Top 11%
2.0%
11
Briefings in Bioinformatics
◐ 354 papers in training set
Top 5%
1.7%
12
Cell Systems
○ 201 papers in training set
Top 3%
1.7%
13
Cell Genomics
○ 172 papers in training set
Top 3%
1.4%
14
Scientific Reports
○ 3612 papers in training set
Top 58%
1.4%
15
GigaScience
◐ 212 papers in training set
Top 3%
1.4%
16
BMC Genomics
○ 406 papers in training set
Top 5%
1.3%
17
iScience
○ 1154 papers in training set
Top 29%
1.1%
18
Bioinformatics Advances
◐ 203 papers in training set
Top 4%
1.1%
19
Nature Computational Science
○ 55 papers in training set
Top 2%
0.8%
20
Nature Machine Intelligence
○ 70 papers in training set
Top 3%
0.6%
21
PLOS Computational Biology
● 1863 papers in training set
Top 22%
0.6%
22
The American Journal of Human Genetics
○ 234 papers in training set
Top 3%
0.6%

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.