Flye was extensively tested on various whole genome PacBio and ONT datasets. In particular, we used Flye to assemble PacBio's P5C3, P6C4, Sequel and Sequel II; ONT's R7-R9 basecalled with Albacore, Guppy and Flipflop (fast mode). We typically use raw reads, however error-corrected input is also supported (for example, for PacBio CCS). Flye is designed for all kinds of genomes, for various bacteria to mammalians. We are also testing Flye on metagenomic datasets (mock and real). You can also check the table with the Flye benchmarks in the Usage file.
We have NOT extensively tested Flye on targeted sequencing (not whole meta/genome).
Also, support of very short sequences, such as viruses, phages, mitochondria
is limited. On the other hand, assembly of short plasmids IS supported
through the --plasmids option.
Currently Flye does not explicitly support diploid assemblies. If heterzygosity is low, it should not be a problem for contiguity; however similar alternative haplotypes could be collapsed. Likewise, SNPs and short indels between the alternative haplotypes will not be captured. If heterozygosity is high, Flye will likely recover alternative haplotypes, but will not phase them. Because we do not attempt to reconstruct pseudo-haplotypes (like FALCON), this will also reduce the overall contiguity.
Yes, use the --meta option. This option also should be applied if you
expect highly non-uniform read coverage in your dataset.
A disadvantage of --meta mode is that it could be
slower and consume more RAM. Also, graph simplifications will be less aggressive,
which can hurt the contiguity of a single genome assembly.
In theory - yes, but RAM usage is the limit. We have tested Flye on many human assemblies, which typically require ~700Gb of RAM. Memory consumption grows linearly with genome size and reads coverage. Thus, genomes beyond ~5Gb is size are unlikely to be supported, currently.
Yes, use the --pacbio-corr option. Note that this is currently
not the primary focus of our developments. More improvements likely
could be made in the future as more datasets of this kind become available.
No, support of very short sequences, such as viruses, phages, mitochondria
is limited. On the other hand, assembly of short plasmids IS supported
through the --plasmids option.
For a typical bacterial assembly with ~100x read coverage,
Flye needs <10 Gb of RAM and finishes within an hour
using ~30 threads. This will scale linearly with the increase in
read coverage. If you coverage is above 100x, consider use
--asm-coverage 100 to use the longest 100x reads for disjointig
assembly - this should speed things up.
Mid-size eukaryotes (like C. elegans or D. melanogaster) with coverage around 50x might require 3-4 hours and 30-50Gb RAM to assemble (30 threads).
Mammalian assemblies with 40x coverage need ~700Gb of RAM and typically finish within 3-4 days using 30 threads.
Various benchmarks are also given in the Usage file. The time and memory usage usually scales linearly with genome size and read coverage. However, highly repetitive genomes might require more memory and be slower to assemble. Most of the Flye stages run in parallel, so the more threads you use, the faster it will be. We typically use 30-50 threads on our hardware.
Note that you can also use --asm-coverage option to
reduce the memory usage by sampling the longest reads
for the initial disjointig assembly.
One can typically get satisfying assembly contiguity with 40x+ PacBio / ONT reads, if the read length is sufficient. You might need higher coverage to improve the consensus quality.
Depending on the technology and read length distribution, you might have success with 20-30x long reads. Assembly of datasets with coverage below 10x is not recommended.
The genome size estimate is used for solid k-mer selection in the initial disjointig assembly stage. Flye is not very sensitive to this parameter, and the estimate could be rough. It is ok if the parameter is within 0.5x-2x of the actual genome size. If the final assembly size is very different from the initial guess, consider re-running the pipeline with an updated estimate for better results.
An alternative option is to run Flye in --meta mode, which uses
a different approach for solid k-mer selection. This
mode is almost independent from the genome size parameter
(you still need to provide an estimate for the selection of
some other parameters). When assembly is completed,
you can re-run in the normal mode with the inferred genome size.
First, make sure that your dataset type is supported (see above), and the parameters are set correctly. In particular, make sure you are correctly using either raw or corrected reads mode, and the genome size parameter is reasonable.
Secondly, make sure that coverage and read length is sufficient. Flye generally expects coverage to be more than 10x, and reads N90 over 1kb (5kb+ recommended).
Another problem with disjointig assembly could be that solid
k-mers are not properly selected because of some bias / contamination.
For example, a lot of nonsense (artificial) reads may confuse k-mer
selection. In this case, you can try the following. First, if you have
sufficient read coverage, try to use the --asm-coverage option
to subsample the longest reads, which typically have better quality.
Secondly, try the --meta option that represents an alternative
k-mer selection strategy.
We designed Flye to work on a wide range of datasets using the default parameters. We thus do not expose most of the technical parameters to the user. This also ensures the reproducibility of Flye assemblies in different environments.
If the quality of your assembly worse than expected, first make sure that all required parameters are set correctly (e.g. check the FAQ questions above). Make sure that input reads have sufficient quality, coverage and length.
A notable exception is the --min-overlap parameter. Intuitively,
we want keep it as high as possible (e.g. 5kb) to reduce the complexity
of a repeat graph. However, if the read length is not sufficient,
this might lead to gaps in assembly. Flye automatically
selects this parameter based on the read length distribution,
and for the most of datasets the selected value works well.
In some rare cases, this parameter needs to be adjusted manually,
for example if the read length distribution is skewd.
Currently, the minimum overlap is automatically selected with the 1kb-5kb range. For some datasets (such as ultra-long ONT reads) it makes sense to manually increase minimum overlap to 10k. We will likely include this as an automatic rule in the future releases.
You can do this as follows: first, run the pipeline with all your reads
in the --pacbio-raw mode (you can specify multiple files, no need to
merge all you reads into one). Also add --iterations 0 to stop the pipeline
before polishing.
Once the assembly finishes, run polishing using either PacBio or ONT reads only.
Use the same assembly options, but add --resume-from polishing. Here is an
example of a script that should do the job (thanks to @jvhaarst):
flye --pacbio-raw $PBREADS $ONTREADS --iterations 0 --out-dir $OUTPUTDIR --genome-size $SIZE --threads $THREADS
flye --pacbio-raw $PBREADS --resume-from polishing --out-dir $OUTPUTDIR --genome-size $SIZE --threads $THREADS
It is a somewhat difficult question to answer. Flye does include polishing step, and it producing high quality consensus on bacterial PacBio datasets with high coverage. For example, see this recent evaluation by Ryan Wick. On the other hand, PacBio has specialized Quiver/Arrow tools that are more advanced, since they use PacBio-specific signal information. It is possible, that you can get a bit of improvement after applying these tools.
For ONT data, Flye still has ~1% errors in consensus due to the systematic error pattern of ONT reads. One might need an external polisher (such as nanopolish/Medaka) to get higher consensus quality.
Illumina correction can fix many of the remaining errors and improve the assembly quality for both PacBio and ONT, for example, using Pilon or Racon. But it should be applied with caution to prevent over-correction of repetitive regions. Also see Watson and Warr paper for a discussion on the assembly quality.
Flye was designed and tested mainly using raw reads, so it is currently the recommended option.
No, usually it is not necessary. Flye automatically filters out
chimeric reads or reads with bad ends. If the read coverage is very high,
you can use the built-in --asm-coverage option for subsampling the longest ones.
Note that in PacBio mode, Flye assumes that the input files represent PacBio subreads, e.g. adaptors and scraps are removed and multiple passes of the same insertion sequence are separated. This is typically handled by PacBio instruments/toolchains, however we saw examples of problemmatic raw -> fastq conversions, which resulted into incorrectl subreads. In this case, consider using pbclip to fix your Fasta/q reads.
Currently, cluster environments are not supported. Flye was designed to run on a single high-memory node, and it will be difficult to make it run in a distributed environment. Note that Flye pipeline has multiple consecutive stages, that are could be resumed and run on different machines, if desired.
Yes, you can use the --polish-target option. Here is an example of
polishing using PacBio reads:
flye --polish-target SEQ_TO_POLISH --pacbio-raw READS --iterations NUM_ITER --out-dir OUTPUTDIR --threads THREADS
Please post your question to the issue tracker.
In case you prefer personal communcation, you can contact Mikhail at [email protected].
If you reporting a problem, please include the flye.log file and provide some
details about your dataset (if possible).