- What resources does Canu require for a bacterial genome assembly? A mammalian assembly?
- How do I run Canu on my SLURM / SGE / PBS / LSF / Torque system?
- My run stopped with the error
'Failed to submit batch jobs'
- My run of Canu was killed by the sysadmin; the power going out; my cat stepping on the power button; et cetera. Is it safe to restart? How do I restart?
- What parameters should I use for my reads?
- Can I assemble RNA sequence data?
- My assembly continuity is not good, how can I improve it?
- What parameters can I tweak?
- My asm.contigs.fasta is empty, why?
- Why is my assembly is missing my favorite short plasmid?
- Why do I get less corrected read data than I asked for?
- What is the minimum coverage required to run Canu?
- My circular element is duplicated/has overlap?
- My genome is AT (or GC) rich, do I need to adjust parameters? What about highly repetitive genomes?
- How can I send data to you?
Canu will detect available resources and configure itself to run efficiently using those resources. It will request resources, for example, the number of compute threads to use, Based on the genome size being assembled. It will fail to even start if it feels there are insufficient resources available.
A typical bacterial genome can be assembled with 8GB memory in a few CPU hours - around an hour on 8 cores. It is possible, but not allowed by default, to run with only 4GB memory.
A well-behaved large genome, such as human or other mammals, can be assembled in 10,000 to 25,000 CPU hours, depending on coverage. A grid environment is strongly recommended, with at least 16GB available on each compute node, and one node with at least 64GB memory. You should plan on having 3TB free disk space, much more for highly repetitive genomes.
Our compute nodes have 48 compute threads and 128GB memory, with a few larger nodes with up to 1TB memory. We develop and test (mostly bacteria, yeast and drosophila) on laptops and desktops with 4 to 12 compute threads and 16GB to 64GB memory.
Canu will detect and configure itself to use on most grids. You can supply your own grid options, such as a partition on SLURM or an account code on SGE, with
gridOptions="<your options list>"which will passed to every job submitted by Canu. Similar options exist for every stage of Canu, which could be used to, for example, restrict overlapping to a specific partition or queue.
To disable grid support and run only on the local machine, specify
It is possible to limit the number of grid jobs running at the same time, but this isn’t directly supported by Canu. The various gridOptions parameters can pass grid-specific parameters to the submit commands used; see Issue #756 for Slurm and SGE examples.
The grid you run on must allow compute nodes to submit jobs. This means that if you are on a compute host,
qsub/bsub/sbatch/etcmust be available and working. You can test this by starting an interactive compute session and running the submit command manually (e.g.
If this is not the case, Canu WILL NOT work on your grid. You must then set
useGrid=falseand run on a single machine. Alternatively, you can run Canu with
useGrid=remotewhich will stop at every submit command, list what should be submitted. You then submit these jobs manually, wait for them to complete, and run the Canu command again. This is a manual process but currently the only workaround for grids without submit support on the compute nodes.
Canu is designed to be universal on a large range of PacBio (C2, P4-C2, P5-C3, P6-C4) and Oxford Nanopore (R6 through R9) data. Assembly quality and/or efficiency can be enhanced for specific datatypes:
- Nanopore R7 1D and Low Identity Reads
With R7 1D sequencing data, and generally for any raw reads lower than 80% identity, five to ten rounds of error correction are helpful:canu -p r1 -d r1 -correct corOutCoverage=500 corMinCoverage=0 corMhapSensitivity=high -nanopore-raw your_reads.fasta canu -p r2 -d r2 -correct corOutCoverage=500 corMinCoverage=0 corMhapSensitivity=high -nanopore-raw r1/r1.correctedReads.fasta.gz canu -p r3 -d r3 -correct corOutCoverage=500 corMinCoverage=0 corMhapSensitivity=high -nanopore-raw r2/r2.correctedReads.fasta.gz canu -p r4 -d r4 -correct corOutCoverage=500 corMinCoverage=0 corMhapSensitivity=high -nanopore-raw r3/r3.correctedReads.fasta.gz canu -p r5 -d r5 -correct corOutCoverage=500 corMinCoverage=0 corMhapSensitivity=high -nanopore-raw r4/r4.correctedReads.fasta.gz
Then assemble the output of the last round, allowing up to 30% difference in overlaps:canu -p asm -d asm correctedErrorRate=0.3 utgGraphDeviation=50 -nanopore-corrected r5/r5.correctedReads.fasta.gz
- Nanopore R7 2D and Nanopore R9 1D
- The defaults were designed with these datasets in mind so they should work. Having very high coverage or very long Nanopore reads can slow down the assembly significantly. You can try the
overlapper=mhap utgReAlign=trueoption which is much faster but may produce less contiguous assemblies on large genomes.
- Nanopore R9 2D and PacBio P6
- Slightly decrease the maximum allowed difference in overlaps from the default of 14.4% to 12.0% with
- Early PacBio Sequel
- Based on one publicly released A. thaliana dataset, and a few more recent mammalian genomes, slightly increase the maximum allowed difference from the default of 4.5% to 6.5% with
- Nanopore R9 large genomes
- Due to some systematic errors, the identity estimate used by Canu for correction can be an over-estimate of true error, inflating runtime. For recent large genomes (>1gbp) with more than 30x coverage, we’ve used
'corMhapOptions=--threshold 0.8 --num-hashes 512 --ordered-sketch-size 1000 --ordered-kmer-size 14'. This is not needed for below 30x coverage.
Canu will likely mis-assemble, or completely fail to assemble, RNA data. It will do a reasonable job at generating corrected reads though. Reads are corrected using (local) best alignments to other reads, and alignments between different isoforms are usually obviously not ‘best’. Just like with DNA sequences, similar isoforms can get ‘mixed’ together. We’ve heard of reasonable success from users, but do not have any parameter suggestions to make.
Note that Canu will silently translate ‘U’ bases to ‘T’ bases on input, but NOT translate the output bases back to ‘U’.
The most important determinant for assembly quality is sequence length, followed by the repeat complexity/heterozygosity of your sample. The first thing to check is the amount of corrected bases output by the correction step. This is logged in the stdout of Canu or in canu-scripts/canu.*.out if you are running in a grid environment. For example on a haploid H. sapiens sample:-- BEGIN TRIMMING -- ... -- In gatekeeper store 'chm1/trimming/asm.gkpStore': -- Found 5459105 reads. -- Found 91697412754 bases (29.57 times coverage). ...
Canu tries to correct the longest 40X of data. Some loss is normal but having output coverage below 20-25X is a sign that correction did not work well (assuming you have more input coverage than that). If that is the case, re-running with
corMhapSensitivity=normalif you have >50X or
corMhapSensitivity=high corMinCoverage=0otherwise can help. You can also increase the target coverage to correct
corOutCoverage=100to get more correct sequences for assembly. If there are sufficient corrected reads, the poor assembly is likely due to either repeats in the genome being greater than read lengths or a high heterozygosity in the sample. Stay tuned for mor information on tuning unitigging in those instances.
For all stages:
rawErrorRateis the maximum expected difference in an alignment of two _uncorrected_ reads. It is a meta-parameter that sets other parameters.
correctedErrorRateis the maximum expected difference in an alignment of two _corrected_ reads. It is a meta-parameter that sets other parameters. (If you’re used to the
errorRateparameter, multiply that by 3 and use it here.)
minOverlapLength. The defaults are to discard reads shorter than 1000bp and to not look for overlaps shorter than 500bp. Increasing
minReadLengthcan improve run time, and increasing
minOverlapLengthcan improve assembly quality by removing false overlaps. However, increasing either too much will quickly degrade assemblies by either omitting valuable reads or missing true overlaps.
corOutCoveragecontrols how much coverage in corrected reads is generated. The default is to target 40X, but, for various reasons, this results in 30X to 35X of reads being generated.
corMinCoverage, loosely, controls the quality of the corrected reads. It is the coverage in evidence reads that is needed before a (portion of a) corrected read is reported. Corrected reads are generated as a consensus of other reads; this is just the minimum coverage needed for the consensus sequence to be reported. The default is based on input read coverage: 0x coverage for less than 30X input coverage, and 4x coverage for more than that.
utgOvlErrorRateis essentially a speed optimization. Overlaps above this error rate are not computed. Setting it too high generally just wastes compute time, while setting it too low will degrade assemblies by missing true overlaps between lower quality reads.
utgRepeatDeviationwhat quality of overlaps are used in contig construction or in breaking contigs at false repeat joins, respectively. Both are in terms of a deviation from the mean error rate in the longest overlaps.
utgRepeatConfusedBPcontrols how similar a true overlap (between two reads in the same contig) and a false overlap (between two reads in different contigs) need to be before the contig is split. When this occurs, it isn’t clear which overlap is ‘true’ - the longer one or the slightly shorter one - and the contig is split to avoid misassemblies.
For polyploid genomes:
Generally, there’s a couple of ways of dealing with the ploidy.
Avoid collapsing the genome so you end up with double (assuming diploid) the genome size as long as your divergence is above about 2% (for PacBio data). Below this divergence, you’d end up collapsing the variations. We’ve used the following parameters for polyploid populations (PacBio data):
corOutCoverage=200 "batOptions=-dg 3 -db 3 -dr 1 -ca 500 -cp 50"
This will output more corrected reads (than the default 40x). The latter option will be more conservative at picking the error rate to use for the assembly to try to maintain haplotype separation. If it works, you’ll end up with an assembly >= 2x your haploid genome size. Post-processing using gene information or other synteny information is required to remove redundancy from this assembly.
Smash haplotypes together and then do phasing using another approach (like HapCUT2 or whatshap or others). In that case you want to do the opposite, increase the error rates used for finding overlaps:
When trimming, reads will be trimmed using other reads in the same chromosome (and probably some reads from other chromosomes). When assembling, overlaps well outside the observed error rate distribution are discarded.
The basic idea is to use all data for assembly rather than just the longest as default. The parameters we’ve used recently are:
corOutCoverage=10000 corMhapSensitivity=high corMinCoverage=0 redMemory=32 oeaMemory=32 batMemory=200
For low coverage:
- For less than 30X coverage, increase the alllowed difference in overlaps by a few percent (from 4.5% to 8.5% (or more) with
correctedErrorRate=0.105for PacBio and from 14.4% to 16% (or more) with
correctedErrorRate=0.16for Nanopore), to adjust for inferior read correction. Canu will automatically reduce
corMinCoverageto zero to correct as many reads as possible.
For high coverage:
- For more than 60X coverage, decrease the allowed difference in overlaps (from 4.5% to 4.0% with
correctedErrorRate=0.040for PacBio, from 14.4% to 12% with
correctedErrorRate=0.12for Nanopore), so that only the better corrected reads are used. This is primarily an optimization for speed and generally does not change assembly continuity.
Canu creates three assembled sequence output files:
<prefix>.unassembled.fasta, where contigs are the primary output, unitigs are the primary output split at alternate paths, and unassembled are the leftover pieces.
The contigFilter parameter sets several parameters that control how small or low coverage initial contigs are handled. By default, initial contigs with more than 50% of the length at less than 3X coverage will be classified as ‘unassembled’ and removed from the assembly, that is,
contigFilter="2 0 1.0 0.5 3". The filtering can be disabled by changing the last number from ‘3’ to ‘0’ (meaning, filter if 50% of the contig is less than 0X coverage).
In Canu v1.6 and earlier only the longest 40X of data (based on the specified genome size) is used for correction. Datasets with uneven coverage or small plasmids can fail to generate enough corrected reads to give enough coverage for assembly, resulting in gaps in the genome or even no reads for small plasmids. Set
corOutCoverage=1000(or any value greater than your total input coverage) to correct all input data.
An alternate approach is to correct all reads (
-correct corOutCoverage=1000) then assemble 40X of reads picked at random from the
More recent Canu versions dynamically select poorly represented sequences to avoid missing short plasmids so this should no longer happen.
Some reads are trimmed during correction due to being chimeric or because there wasn’t enough evidence to generate a quality corrected sequence. Typically, this results in a 25% loss. Setting
corMinCoverage=0will report all bases, even low those of low quality. Canu will trim these in its ‘trimming’ phase before assembly.
For eukaryotic genomes, coverage more than 20X is enough to outperform current hybrid methods. Below that, you will likely not assemble the full genome.
This is expected for any circular elements. They can overlap by up to a read length due to how Canu constructs contigs. Canu provides an alignment string in the GFA output which can be converted to an alignment to identify the trimming points.
An alternative is to run MUMmer to get self-alignments on the contig and use those trim points. For example, assuming the circular element is in
tig00000099.fa. Run:nucmer -maxmatch -nosimplify tig00000099.fa tig00000099.fa show-coords -lrcTH out.delta
to find the end overlaps in the tig. The output would be something like:1 1895 48502 50400 1895 1899 99.37 50400 50400 3.76 3.77 tig00000001 tig00000001 48502 50400 1 1895 1899 1895 99.37 50400 50400 3.77 3.76 tig00000001 tig00000001
means trim to 1 to 48502. There is also an alternate writeup.
My genome is AT (or GC) rich, do I need to adjust parameters? What about highly repetitive genomes?¶
On bacterial genomes, no adjustment of parameters is (usually) needed. See the next question.
On repetitive genomes with with a significantly skewed AT/GC ratio, the Jaccard estimate used by MHAP is biased. Setting
corMaxEvidenceErate=0.15is sufficient to correct for the bias in our testing.
In general, with high coverage repetitive genomes (such as plants) it can be beneficial to set the above parameter anyway, as it will eliminate repetitive matches, speed up the assembly, and sometime improve unitigs.
FTP to ftp://ftp.cbcb.umd.edu/incoming/sergek. This is a write-only location that only the Canu developers can see.
Here is a quick walk-through using a command-line ftp client (should be available on most Linux and OSX installations). Say we want to transfer a file named
reads.fastq. First, run
ftp ftp.cbcb.umd.edu, specify
anonymousas the user name and hit return for password (blank). Then
put reads.fastq, and
That’s it, you won’t be able to see the file but we can download it.