Using PacBio Circular Consensus Sequencing (CCS) for Highly Accurate Assemblies Gene Myers Chair of Systems Biology MPI for Cell Biology and Genetics Dresden, DE MPI CBG CSBD 

HQ Genomics Today & Tomorrow 

PB only or PB + 1 would be a significant savings 2019/2020 1.2K EU 3.5K EU 50X Illumina in 10X read clouds Bionano restriction maps 50X Illumina in Hi-C read pairs Scaffolding Technologies 60X Pacbio long reads 10K EU 5K EU 2017 Assume 1Gbp Genome PB + Bionano: 6 Bats Project: 10Mbp Contig N50 100Mbp Scaffold N50 HQ Genomics But favor PB + Hi-C 

At least 2 ways to improve: HQ Genomes Tomorrow: ✦ Scrubbing to remove artifacts ✦ Repeat/Haplotype separation based on heterogeneity ✦ Repeat detection and modeling HIFI CCS protocol: ~ 3x loss in throughput and cost over raw But each insert wrapped ~ 8x 㱺 ~ 0.2% error rate Which is better? - 15Kbp reads at 99.8% - 50Kbp reads at 90% - some combination of both? • Longer or more accurate reads (CCS) • Better Algorithms 

Conceptual Effect of Read Accuracy on Assembly 

String Graph: The “Reality” “Hairball” … … ? How do you get through ? 

String Graph: The “Reality” 10% error 30% alignment threshold 㱺 10%-repeats entangle .5% error 2% alignment threshold 㱺 only <1%-repeats entangle 

… … small large Solving Repeats spanning reads suffice … … microhet’s could get you through (should be easy(er)) All the power of long reads has thus far been due to this ⟹ longer is better This has not been done Requires ability to id. microhet’s ⟹ more accurate is better 

Preliminary Work 

Reads: Chimers Adaptamers Low Q dropouts 90% average Reads: No Chimers No Adaptamers No Low Q dropouts 99.999% uniformly Haplotype Phased Perfect PB reads with Scrubbing Solves: Artifacts Haplotypes Low Copy Repeats (≤ 5) Scrubber Long Read Assembler Task is easier, but still necessary: 99.8% average .5% of reads are chimers .02% of reads have no adaptamer 15% of reads have a low Q “panel” 100bp with 5% or more error 

Daligner: Switch from 14-mers to 40-mers Take 1 out of ever 10 mers (at random) Compute Time Reductions • 99.999% Sensitivity (Alignment between ≧1000bp with .5% error in each read) (R. Durbin) (and uses ≦ 8Gbp memory) • Can use 1Gbp blocks vs. 1/4Gbp (vs 2000+ for raw reads) • 90 CPU hrs for 30X HG002 (on this laptop) 

Concluding Remarks 

Accelerates assembly compute time HiFi reads likely to improve diploid assembly Likely to be quite effective at haplotype phasing / repeat separation Better CCS algorithms are needed 

Acknowledgements PacBio Paul Paluso James Drake Kevin Corcoran Jonas Korlach Mike Hunkapillar Dresden-Concept Genome Center CRTD TU-D Andreas Dahl MPI-CBG Sylke Winkler Martin Pippel German Tischler