A minimap2 SMRT wrapper for PacBio data: native PacBio data in ⇨ native PacBio BAM out.
pbmm2 is a SMRT C++ wrapper for minimap2's C API. Its purpose is to support native PacBio in- and output, provide sets of recommended parameters, generate sorted output on-the-fly, and postprocess alignments. Sorted output can be used directly for polishing using GenomicConsensus, if BAM has been used as input to pbmm2. Benchmarks show that pbmm2 outperforms BLASR in sequence identity, number of mapped bases, and especially runtime. pbmm2 is the official replacement for BLASR.
Latest version can be installed via bioconda package pbmm2.
Please refer to our official pbbioconda page for information on Installation, Support, License, Copyright, and Disclaimer.
Version 1.16.0: Full changelog here
pbmm2 offers following tools
Tools:
index Index reference and store as .mmi file
align Align PacBio reads to reference sequences
A. Generate index file for reference and reuse it to align reads
$ pbmm2 index ref.fasta ref.mmi --preset SUBREAD
$ pbmm2 align ref.mmi movie.subreads.bam ref.movie.bam --preset SUBREAD
B. Align reads and sort on-the-fly, with 4 alignment and 2 sort threads
$ pbmm2 align ref.fasta movie.subreads.bam ref.movie.bam --preset SUBREAD --sort -j 4 -J 2
C. Align reads, sort on-the-fly, and create PBI
$ pbmm2 align ref.fasta movie.subreadset.xml ref.movie.alignmentset.xml --preset SUBREAD --sort
D. Omit output file and stream BAM output to stdout
$ pbmm2 align hg38.mmi movie1.subreadset.xml --preset SUBREAD | samtools sort > hg38.movie1.sorted.bam
E. Align CCS fastq input and sort output
$ pbmm2 align ref.fasta movie.Q20.fastq ref.movie.bam --sort --rg '@RG\tID:myid\tSM:mysample'
Indexing is optional, but recommended if you use the same reference with the same --preset multiple times.
Usage: pbmm2 index [options] <ref.fa|xml> <out.mmi>
Notes:
- If you use an index file, you can't override parameters
-k,-w, nor-uinpbmm2 align! - Minimap2 parameter
-H(homopolymer-compressed k-mer) is always on for SUBREAD and UNROLLED presets and can be disabled with-u. - You can also use existing minimap2
.mmifiles inpbmm2 align.
The output argument is optional. If not provided, BAM output is streamed to stdout.
Usage: pbmm2 align [options] <ref.fa|xml|mmi> <in.bam|xml|fa|fq> [out.aligned.bam|xml]
The number of alignment threads can be specified with -j,--num-threads.
If not specified, the maximum number of threads will be used, minus one thread for BAM IO
and minus the number of threads specified for sorting.
Sorted output can be generated using --sort.
Percentage: By default, 25% of threads specified with -j, maximum 8, are used for sorting. Example: --sort -j 12, 9 threads for alignment, 3 threads for sorting.
Manual override: To override the default percentage, -J,--sort-threads defines the explicit number of threads
used for on-the-fly sorting. Example: --sort -j 12 -J 4, 12 threads for alignment, 4 threads for sorting.
The memory allocated per sort thread can be defined with -m,--sort-memory, accepting suffixes M,G.
Temporary files during sorting are stored in the current working directory,
unless explicitly defined with environment variable TMPDIR.
The path used for temporary files is also printed if --log-level DEBUG is set.
Benchmarks on human data have shown that 4 sort threads are recommended, but no more than 8 threads can be effectively leveraged, even with 70 cores used for alignment. It is recommended to provide more memory to each of a few sort threads, to avoid disk IO pressure, than providing less memory to each of many sort threads.
Following compatibility table shows allowed input file types, output file types,
compatibility with GenomicConsensus, and recommended --preset choice.
More info about our dataset XML specification.
| Input | Output | GC | Preset |
|---|---|---|---|
.bam (aligned or unaliged) |
.bam |
Y | |
.fasta / .fa / .fasta.gz / .fa.gz |
.bam |
N | |
.fastq / .fq / .fastq.gz / .fq.gz |
.bam |
N | |
.Q20.fastq / Q20.fastq.gz |
.bam |
N | CCS |
bam.fofn |
.bam |
N | |
fasta.fofn |
.bam |
N | |
fastq.fofn |
.bam |
N | |
.subreadset.xml |
.bam \ .alignmentset.xml |
Y | |
.consensusreadset.xml |
.bam \ .consensusalignmentset.xml |
Y | CCS |
.transcriptset.xml |
.bam \ .transcriptalignmentset.xml |
Y | ISOSEQ |
In addition to native PacBio BAM input, reads can also be provided in FASTA and FASTQ formats, as shown above.
With FASTA/Q input, option --rg sets the read group. Example call:
pbmm2 align hg38.fasta movie.Q20.fastq hg38.movie.bam --rg '@RG\tID:myid\tSM:mysample'
All three reference file formats .fasta, .referenceset.xml, and .mmi can be combined with FASTA/Q input.
pbmm2 supports the .fofn file type (File Of File Names), containing the same
datatype. Supported are .fofn files with FASTA, FASTQ, or BAM.
Examples:
echo "m64001_190131_212703.Q20.fastq.gz" > myfiles.fofn
echo "m64001_190228_200412.Q20.fastq.gz" >> myfiles.fofn
pbmm2 align hg38.fasta myfiles.fofn hg38.myfiles.bam --rg '@RG\tID:myid\tSM:mysample'
ls *.subreads.bam > mymovies.fofn
pbmm2 align hg38.fasta mymovies.fofn hg38.mymovies.bam
- pbmm2 ≥v1.13.0: minimap2 v2.26
- pbmm2 ≥v1.13.0: minimap2 v2.15
Whenever the output is of type xml, a pbi file is being generated.
For sorted output via --sort, a bai file is being generated per default.
You can switch to csi for larger genomes with --bam-index CSI or skip
index generation completely with --bam-index NONE.
Per default, pbmm2 uses recommended parameter sets to simplify the plethora of possible combinations. For this, we currently offer:
Alignment modes of --preset:
- "SUBREAD" -k 19 -w 19 -o 5 -O 56 -e 4 -E 1 -A 2 -B 5 -z 400 -Z 50 -r 2000 -g 5000
- "CCS" or "HIFI" -k 19 -w 19 -u -o 6 -O 26 -e 2 -E 1 -A 1 -B 4 -z 400 -Z 50 -r 2000 -g 5000
- "ISOSEQ" -k 15 -w 5 -u -o 2 -O 32 -e 1 -E 0 -A 1 -B 2 -z 200 -Z 100 -r 200000 -g 2000 -C 5 -G 200000
- "UNROLLED" -k 15 -w 15 -o 2 -O 32 -e 1 -E 0 -A 1 -B 2 -z 200 -Z 100 -r 2000 -g 10000
Default ["CCS"]
If you want to override any of the parameters of your chosen set, please use the respective options:
-k k-mer size (no larger than 28). [-1]
-w Minimizer window size. [-1]
-u Disable homopolymer-compressed k-mer (compression is active for SUBREAD & UNROLLED presets).
-A Matching score. [-1]
-B Mismatch penalty. [-1]
-z Z-drop score. [-1]
-Z Z-drop inversion score. [-1]
-r Bandwidth used in chaining and DP-based alignment. [-1]
-g Stop chain enlongation if there are no minimizers in N bp. [-1]
For the piece-wise linear gap penalties, use the following overrides, whereas a k-long gap costs min{o+ke,O+kE}:
-o,--gap-open-1 Gap open penalty 1. [-1]
-O,--gap-open-2 Gap open penalty 2. [-1]
-e,--gap-extend-1 Gap extension penalty 1. [-1]
-E,--gap-extend-2 Gap extension penalty 2. [-1]
-L,--lj-min-ratio Long join flank ratio. [-1]
For ISOSEQ, you can override additional parameters:
-G Max intron length (changes -r). [-1]
-C Cost for a non-canonical GT-AG splicing. [-1]
--no-splice-flank Do not prefer splice flanks GT-AG.
If you have suggestions for our default parameters or ideas for a new parameter set, please open a GitHub issue!
To achieve similar alignment behavior like blasr, we implicitly use following minimap2 parameters:
- soft clipping with
-Y - long cigars for tag
CGwith-L X/=cigars instead ofMwith--eqx- no overlapping query intervals with repeated matches trimming
- no secondary alignments are produced with
--secondary=no
The idea of removing spurious or low-quality alignments is straightforward,
but the exact definition of a threshold is tricky and
varies between tools and applications. More on sequence identity
from Heng Li.
pbmm2 offers following filters:
--min-concordance-perc, legacy mapped concordance filter, inherited from its predecessor BLASR (hidden option)--min-id-perc, a sequence identity percentage filter defined as the BLAST identity (hidden option)--min-gap-comp-id-perc, a gap-compressed sequence identity filter accounting insertions and deletions as single events only (default)
By default, (3) is set to 70%, (1) and (2) are deactivated.
The problem with (1) the mapped concordance filter is that it also removes
biological structural variations, such as true insertions and deletions
w.r.t. used reference; it is only appropriate if applied to resequencing
data of haploid organisms.
The (2) sequence identity is the BLAST identity, a very natural metric for filtering.
The (3) gap-compressed sequence identity filter is very similar to (2),
but accounts insertions and deletions as single events only and
is the fairest metric when it comes to assess the actual error rate.
All three filters are combined with AND, meaning an alignment has to pass all
three thresholds.
The --min-concordance-perc option, whereas concordance is defined as
100 - 100 * (#Deletions + #Insertions + #Mismatches) / (AlignEndInRead - AlignStartInRead)
will remove alignments that do not pass the provided threshold in percent.
You can deactivate this filter with --min-concordance-perc 0.
The --min-id-perc option, whereas sequence identity is defined as the BLAST identity
100 * #Matches / (#Matches + #Mismatches + #Deletions + #Insertions)
will remove alignments that do not pass the provided threshold in percent.
You can deactivate this filter with --min-id-perc 0.
The --min-gap-comp-id-perc, -y option, whereas gap-compressed identity is defined as
100 * #Matches / (#Matches + #Mismatches + #DeletionEvents + #InsertionEvents)
will remove alignments that do not pass the provided threshold in percent.
This is the default filter.
You can deactivate this filter with --min-gap-comp-id-perc 0.
A repeated match is, when the same query interval is shared between a primary
and supplementary alignment. This can happen for translocations, where breakends
share the same flanking sequence:
And sometimes, when a LINE gets inserted, the flanks are/get duplicated leading
to complicated alignments, where we see a split read sharing a duplication.
The inserted region itself, mapping to a random other LINE in the reference
genome, may also share sequence similarity to the flanks:
To get the best alignments, minimap2 decides that two alignments may use up to 50% (default) of the same query bases. This does not work for PacBio, because we see pbmm2 as a blasr replacement and require that a single base may never be aligned twice. Minimap2 offers a feature to enforce a query interval overlap to 0%. What happens now if a query interval gets used in two alignments, one or both get flagged as secondary and get filtered. This leads to yield loss and more importantly missing SVs in the alignment.
Papers like this present dynamic programming approaches to find the optimal split to uniquely map query intervals, while maximizing alignment scores. We don't have per base alignment scores available, thus our approach is much simpler. We align the read, find overlapping query intervals, and trim non-primary alignments in order as provided by minimap2; trimming here means that pbmm2 rewrites the cigar and the reference coordinates on-the-fly. This allows us to increase number of mapped bases, slightly reduce identity, but boost SV recall rate.
pbmm2 adds following tags to each aligned record:
mc, stores mapped concordance percentage between 0.0 and 100.0, if the filter was usedmg, stores gap-compressed sequence identity percentage between 0.0 and 100.0, if the filter was usedmi, stores sequence identity percentage between 0.0 and 100.0, if the filter was usedrm, is set to1if an alignment has been manipulated by repeated matches trimming
As for any two alignments of the same data with different mappers, alignments will differ. This is because of many reasons, but mainly a combination of different scoring functions and seeding techniques.
We integrated samtools sort code into pbmm2 to use it as on-the-fly sorting.
This allows pbmm2 to skip writing unaligned BAM as output and thus save
one round-trip of writing and reading unaligned BAM to disk, minimizing disk IO
pressure.
This highly depends on your filesystem. Our tests are showing that there is no clear winner; runtimes differ up to 10% in either directions, depending on read length distribution, genome length and complexity, disk IO pressure, and possibly further unknown factors. For very small genomes post-alignment sorting is faster, but for larger genomes like rice or human on-the-fly sorting is faster. Keep in mind, scalability is not only about runtime, but also disk IO pressure.
We recommend to use on-the-fly sorting via pbmm2 align --sort.
If you use --log-level INFO, after alignment is done, you get following
alignment metrics:
Mapped Reads: 1529671
Alignments: 3087717
Mapped Bases: 28020786811
Mean Sequence Identity: 88.4%
Max Mapped Read Length : 122989
Mean Mapped Read Length : 35597.9
If you use --log-level INFO, after alignment is done, you get following
timing and memory information:
Index Build/Read Time: 22s 327ms
Alignment Time: 5s 523ms
Sort Merge Time: 344ms 927us
BAI Generation Time: 150ms
PBI Generation Time: 161ms 120us
Run Time: 28s 392ms
CPU Time: 39s 653ms
Peak RSS: 12.5847 GB
If you use --log-level DEBUG, you will following reports:
#Reads, #Aln, #RPM: 1462688, 2941000, 37393
#Reads, #Aln, #RPM: 1465877, 2948000, 37379
#Reads, #Aln, #RPM: 1469103, 2955000, 37350
That is:
- number of reads processed,
- number of alignments generated,
- reads per minute processed.
If you are interested in unrolled alignments that is, align the full-length
ZMW read or the HQ region of a ZMW against an unrolled template, please use
--zmw or --hqregion with *.subreadset.xml as input that contains
one *.subreads.bam and one *.scraps.bam file. Keep in mind, to unroll the
reference on your own.
This is beta feature and still in development.
You can override the sample name (SM field in RG tag) for all read groups
with --sample.
If not provided, sample names derive from the dataset input with order of
precedence: SM field in input read group, biosample name, well sample name, UnnamedSample.
If the input is a BAM file and --sample has not been used, the SM field will
be populated with UnnamedSample.
Yes, --split-by-sample generates one output BAM file per sample name, with
the sample name as file name infix, if there is more than one aligned sample name.
Yes, --strip removes following extraneous tags if the input is BAM,
but the resulting output BAM file cannot be used as input into GenomicConsensus:
dq, dt, ip, iq, mq, pa, pc, pd, pe, pg, pm, pq, pt, pv, pw, px, sf, sq, st
Per default, unmapped reads are omitted. You can add them to the output BAM file
with --unmapped.
Use -N, --best-n. If set to 0, default, maximum filtering is disabled.
Using --median-filter, only the subread closest to the median subread length
per ZMW is being aligned.
Preferably, full-length subreads flanked by adapters are chosen.
The idea behind --collapse-homopolymers is to collapse any two or more
consecutive bases of the same type. In this mode, the reference is collapsed and
written to disk with the same prefix as your output alignment and appended
with suffix .ref.collapsed.fasta. In addition, each read is collapsed
before alignment. This mode cannot be combined with .mmi input.
Due to multithreading the ouput alignment ordering can differ between multiple runs
with the same input parameters. The same can occur even with option --sort for
records that align to the same target sequence, the same position within that target,
and in the same orientation, which are the only fields that samtools sort uses.
-
1.16.0
- SMRT Link 25.1 release
- Do not warn about aligned input for minimizer sorted unaligned BAM input
- There are currently no plans to update the source code on GitHub, but binary releases will continue regularly
-
1.14.99
- Remove minimum DP length threshold for HiFi preset
-
1.14.0:
- Add option
--include-fail-readsfor datasets
- Add option
-
1.13.1:
- Documentation changes, included in SMRT Link v13.0
-
1.13.0:
- Update minimap2 to version 2.26
-
1.12.0:
- Set
--preset CCSas default - Change repeated matches trimming to adhere to minimap2 alignment ordering
- Set
-
1.11.0:
- Strip HiFi kinetics tags
-
1.10.0:
- Allow reverse-complemented unaligned records as input
- Allow infix, but not flanking, spaces in sample name
- Do not allow overwriting input files
- Store number of mismatches, tag NM
-
1.9.0:
- Print dependency versions
- Set unaligned MAPQ to 0
-
1.8.0:
- Add support for *.fsa files
-
1.7.0:
- Set
TLEN, for information only - Trim insertions, deletions, and mismatches from the alignment flanks
- Set
-
1.6.0:
SAtag contains full cigar; use--short-sa-cigarto use legacy version- Sanitize bio sample names
-
1.5.0:
- Hide
--min-concordance-percand--min-id-perc - Change default identity filter to
--min-gap-comp-id-perc
- Hide
-
1.4.0:
- Official SMRT Link v10 release
- Case-insensitive
--preset - Read groups without
SMtag are labelled asUnnamedSample
-
1.3.0:
- New internal features for HiFi assembly
- htslib 1.10 support
-
1.2.1:
- Abort if input fofn contains non-existing files
- Add new filters
--min-id-percand--min-gap-comp-id-perc - Updated CLI UX
- Add
-gto control minimap2'smax_gap - Add
--bam-index
-
1.1.0:
- Add support for gzipped FASTA and FASTQ
- Allow multiple input files via
.fofn - Add
--collapse-homopolymers - Use
TMPDIRenv variable to set path for temporary files - Minor memory leak fix, if you used the API directly
-
1.0.0:
- First stable release, included in SMRT Link v7.0
- Minor documentation changes
-
0.12.0:
- Enable
--unmappedto add unmapped records to output - Add repeated matches trimming
- Add BAI for sorted output
- Allow
0value overrides - Abort if insufficient memory is available for sorting
- Enable
-
0.11.0:
- Change input argument order
- Library API access
- Add fasta/q input support
- Add
--lj-min-ratio,--rg,--split-by-sample,--strip - Fix
SAtag - Fix BAM header for idempotence
-
0.10.1:
- Idempotence. Alignment of alignments results in identical alignments
- Use different technique to get tmpfile pipe
- Median filter does not log to DEBUG
-
0.10.0:
- Add
--preset CCS - Allow disabling of homopolymer-compressed k-mer
-u - Adjust concordance metric to be identical to SMRT Link
- Add reference fasta to dataset output
- Output run timings and peak memory
- Change CLI UX
- No overlapping query intervals
- Use BioSample or WellSample name from input dataset
- Drop fake @SQ checksum
- Add
SAtag
- Add
-
0.9.0:
- Add
--sort - Add
--preset ISOSEQ - Add
--median-filter
- Add
Many thanks to Heng Li for a pleasant API experience and to Lance Hepler for the initial idea and code.
THIS WEBSITE AND CONTENT AND ALL SITE-RELATED SERVICES, INCLUDING ANY DATA, ARE PROVIDED "AS IS," WITH ALL FAULTS, WITH NO REPRESENTATIONS OR WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTIES OF MERCHANTABILITY, SATISFACTORY QUALITY, NON-INFRINGEMENT OR FITNESS FOR A PARTICULAR PURPOSE. YOU ASSUME TOTAL RESPONSIBILITY AND RISK FOR YOUR USE OF THIS SITE, ALL SITE-RELATED SERVICES, AND ANY THIRD PARTY WEBSITES OR APPLICATIONS. NO ORAL OR WRITTEN INFORMATION OR ADVICE SHALL CREATE A WARRANTY OF ANY KIND. ANY REFERENCES TO SPECIFIC PRODUCTS OR SERVICES ON THE WEBSITES DO NOT CONSTITUTE OR IMPLY A RECOMMENDATION OR ENDORSEMENT BY PACIFIC BIOSCIENCES.