Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


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.

Latest Version

Version 1.3.0: Full changelog here


pbmm2 offers following tools

    index      Index reference and store as .mmi file
    align      Align PacBio reads to reference sequences

Typical workflows

A. Generate index file for reference and reuse it to align reads
  $ pbmm2 index ref.fasta ref.mmi
  $ pbmm2 align ref.mmi movie.subreads.bam

B. Align reads and sort on-the-fly, with 4 alignment and 2 sort threads
  $ pbmm2 align ref.fasta movie.subreads.bam --sort -j 4 -J 2

C. Align reads, sort on-the-fly, and create PBI
  $ pbmm2 align ref.fasta movie.subreadset.xml --sort

D. Omit output file and stream BAM output to stdout
  $ pbmm2 align hg38.mmi movie1.subreadset.xml | samtools sort > hg38.movie1.sorted.bam

E. Align CCS fastq input and sort output
  $ pbmm2 align ref.fasta movie.Q20.fastq --preset CCS --sort --rg '@RG\tID:myid\tSM:mysample'


Indexing is optional, but recommended it you use the same reference with the same --preset multiple times.

Usage: pbmm2 index [options] <ref.fa|xml> <out.mmi>


  • If you use an index file, you can't override parameters -k, -w, nor -u in pbmm2 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 .mmi files in pbmm2 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]

Alignment Parallelization

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.

Input file types

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

FASTA/Q input

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 --preset CCS --rg '@RG\tID:myid\tSM:mysample'

All three reference file formats .fasta, .referenceset.xml, and .mmi can be combined with FASTA/Q input.

Multiple input files

pbmm2 supports the .fofn file type (File Of File Names), containing the same datatype. Supported are .fofn files with FASTA, FASTQ, or BAM.


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 --preset CCS --rg '@RG\tID:myid\tSM:mysample'
ls *.subreads.bam > mymovies.fofn
pbmm2 align hg38.fasta mymovies.fofn hg38.mymovies.bam


Which minimap2 version is used?

Minimap2 version 2.15 is used, to be specific, SHA1 c404f49.

When are pbi files created?

Whenever the output is of type xml, a pbi file is being generated.

When are BAM index files created?

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.

What are parameter sets and how can I override them?

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 10    -o 5 -O 56 -e 4 -E 1 -A 2 -B 5 -z 400 -Z 50  -r 2000   -L 0.5 -g 5000
    - "CCS" or "HIFI" -k 19 -w 10 -u -o 5 -O 56 -e 4 -E 1 -A 2 -B 5 -z 400 -Z 50  -r 2000   -L 0.5 -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 -L 0.5 -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   -L 0.5 -g 10000
  Default ["SUBREAD"]

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!

What other special parameters are used implicitly?

To achieve similar alignment behavior like blasr, we implicitly use following minimap2 parameters:

  • soft clipping with -Y
  • long cigars for tag CG with -L
  • X/= cigars instead of M with --eqx
  • no overlapping query intervals with repeated matches trimming
  • no secondary alignments are produced with --secondary=no

What sequence identity filters does pbmm2 offer?

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:

  1. --min-concordance-perc, legacy mapped concordance filter, inherited from its predecessor BLASR (hidden option)
  2. --min-id-perc, a sequence identity percentage filter defined as the BLAST identity (hidden option)
  3. --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.

How do you define mapped concordance?

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.

How do you define identity?

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.

How do you define gap-compressed identity?

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.

What is repeated matches trimming?

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 will be much simpler. We align the read, find overlapping query intervals, determine one alignment to be maximal reference spanning, and all others get trimmed; by trimming, I mean 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.

What SAM tags are added by pbmm2?

pbmm2 adds following tags to each aligned record:

Why is the output different from BLASR?

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.

How does sorting work?

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.

Is pbmm2 unsorted + samtools sort faster than pbmm2 --sort?

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.

Can I get alignment statistics?

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

Is there any benchmark information, like timings and peak memory consumption?

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

Can I get progress output?

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.

Can I perform unrolled alignment?

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.

How can I set the sample name?

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.

Can I split output by sample name?

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.

Can I remove all those extra per base and pulse tags?

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

Where are the unmapped reads?

Per default, unmapped reads are omitted. You can add them to the output BAM file with --unmapped.

Can I output at maximum the N best alignments per read?

Use -N, --best-n. If set to 0, default, maximum filtering is disabled.

Is there a way to only align one subread per ZMW?

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.

What is --collapse-homopolymers?

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.

Full Changelog

  • 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
  • 1.6.0:

    • SA tag contains full cigar; use --short-sa-cigar to use legacy version
    • Sanitize bio sample names
  • 1.5.0:

    • Hide --min-concordance-perc and --min-id-perc
    • Change default identity filter to --min-gap-comp-id-perc
  • 1.4.0:

    • Official SMRT Link v10 release
    • Case-insensitive --preset
    • Read groups without SM tag are labelled as UnnamedSample
  • 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-perc and --min-gap-comp-id-perc
    • Updated CLI UX
    • Add -g to control minimap2's max_gap
    • Add --bam-index
  • 1.1.0:

    • Add support for gzipped FASTA and FASTQ
    • Allow multiple input files via .fofn
    • Add --collapse-homopolymers
    • Use TMPDIR env 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 --unmapped to add unmapped records to output
    • Add repeated matches trimming
    • Add BAI for sorted output
    • Allow 0 value overrides
    • Abort if insufficient memory is available for sorting
  • 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 SA tag
    • 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 SA tag
  • 0.9.0:

    • Add --sort
    • Add --preset ISOSEQ
    • Add --median-filter


Many thanks to Heng Li for a pleasant API experience and to Lance Hepler for the initial idea and code.