minimap2-fpga: FPGA-accelerated Minimap2

This work presents an end-to-end integrated FPGA-accelerated solution for the gold-standard third generation genomics sequence analysis tool - Minimap2. The solution carefully integrates an FPGA-based hardware accelerator designed for the chaining step of the tool (top hotspot of the tool for most of the configurations), with the rest of the software so that a speed-up in tool’s end-to-end time is achieved.

Getting Started

1. Create an Amazon EC2 F1 instance loaded with FPGA Developer AMI (AWS f1.2xlarge instance loaded with FPGA Developer AMI Version 1.12.1 was used in this work).

2. Use the commands below to download the AWS EC2 FPGA Development Kit and setup the runtime environment.

git clone https://github.com/aws/aws-fpga.git $AWS_FPGA_REPO_DIR
source $AWS_FPGA_REPO_DIR/vitis_runtime_setup.sh

2. Use the commands below to download this work's GitHub repo and change to project directory.

git clone https://github.com/kisarur/minimap2-fpga.git
cd minimap2-fpga

3. [OPTIONAL] If you use Xilinx UltraScale+ VU9P based FPGA board available on AWS EC2 F1 instance, for which our accelerator implementation (on this branch) is optimized, it is recommended that you use the already built FPGA hardware binary (also called AFI - Amazon FPGA Image) included with this repo at bin/minimap2_opencl.awsxclbin. However, if you want to build this AFI from source (in device/minimap2_opencl.cl), you can use the guide available at https://github.com/aws/aws-fpga/tree/master/Vitis (as recommended on the guide, using a non-F1 EC2 compute instance for this time-consuming hardware build step will help minimize costs).

4. If minimap2-fpga is to be used for the first time in the system, the parameters used in splitting the chaining tasks for hardware and software executions, need to be first trained. A small (~500K reads) representative dataset (<query>) corresponding to the target minimap2 configuration¹ (<minimap2_configuration>) should be used with the reference human genome (<reference>) for this training process. Note that this one-time process can take couple of hours to complete depending on the size of the training dataset.

   1. Use the commands below to train the parameters (requires Python 3 with numpy and scikit-learn packages installed).

   cd hw_sw_split
   ./find_hw_sw_split_params.sh <minimap2_configuration> <reference> <query>
   (e.g. ./find_hw_sw_split_params.sh map-ont hg38noAlt.fa reads.fastq)
   
   cd ../
   

   Given below is a sample output of the above command, that contains the hardware-software split parameters.

   ----------------------
   HW/SW split parameters
   ----------------------
   K1_HW   : 7.8272969079034e-06
   K2_HW   : 4.216579210613027e-05
   C_HW    : 0.5789597220291505
   K_SW    : 8.14078022400557e-06
   C_SW    : -3.3955657611056456
   

   2. The parameters in chain_hardware.h for the relavent dataset type (i.e. ONT, PacBio CCS) should now be replaced with the parameters obtained above. For example, if the parameter training was performed for ONT datasets, the values for ONT_K1_HW, ONT_K2_HW, ONT_C_HW, ONT_K_SW, ONT_C_SW in chain_hardware.h should be replaced with values obtained above for K1_HW, K2_HW, C_HW, K_SW, C_SW respectively.

5. Use the commands below to build the host application and run FPGA-accelerated Minimap2 with the relative arguments¹.

make host
./minimap2-fpga [minimap2 arguments]
(e.g. ./minimap2-fpga -x map-ont <reference> <query>)

Notes

1. Currently, only map-ont and asm20 configurations of Minimap2 are supported. Please contact us if you want more configurations to be added.

---

The README content of original minimap2 software is copied below.

Getting Started

git clone https://github.com/lh3/minimap2
cd minimap2 && make
# long sequences against a reference genome
./minimap2 -a test/MT-human.fa test/MT-orang.fa > test.sam
# create an index first and then map
./minimap2 -x map-ont -d MT-human-ont.mmi test/MT-human.fa
./minimap2 -a MT-human-ont.mmi test/MT-orang.fa > test.sam
# use presets (no test data)
./minimap2 -ax map-pb ref.fa pacbio.fq.gz > aln.sam       # PacBio genomic reads
./minimap2 -ax map-ont ref.fa ont.fq.gz > aln.sam         # Oxford Nanopore genomic reads
./minimap2 -ax asm20 ref.fa pacbio-ccs.fq.gz > aln.sam    # PacBio CCS genomic reads
./minimap2 -ax sr ref.fa read1.fa read2.fa > aln.sam      # short genomic paired-end reads
./minimap2 -ax splice ref.fa rna-reads.fa > aln.sam       # spliced long reads (strand unknown)
./minimap2 -ax splice -uf -k14 ref.fa reads.fa > aln.sam  # noisy Nanopore Direct RNA-seq
./minimap2 -ax splice:hq -uf ref.fa query.fa > aln.sam    # Final PacBio Iso-seq or traditional cDNA
./minimap2 -ax splice --junc-bed anno.bed12 ref.fa query.fa > aln.sam  # prioritize on annotated junctions
./minimap2 -cx asm5 asm1.fa asm2.fa > aln.paf             # intra-species asm-to-asm alignment
./minimap2 -x ava-pb reads.fa reads.fa > overlaps.paf     # PacBio read overlap
./minimap2 -x ava-ont reads.fa reads.fa > overlaps.paf    # Nanopore read overlap
# man page for detailed command line options
man ./minimap2.1

Unimap is recommended for aligning long contigs against a reference genome. It often takes less wall-clock time and is much more sensitive to long insertions and deletions.

Table of Contents

• Getting Started
• Users' Guide
  • Installation
  • General usage
  • Use cases
    • Map long noisy genomic reads
    • Map long mRNA/cDNA reads
    • Find overlaps between long reads
    • Map short accurate genomic reads
    • Full genome/assembly alignment
  • Advanced features
    • Working with >65535 CIGAR operations
    • The cs optional tag
    • Working with the PAF format
  • Algorithm overview
  • Getting help
  • Citing minimap2
• Developers' Guide
• Limitations

Users' Guide

Minimap2 is a versatile sequence alignment program that aligns DNA or mRNA sequences against a large reference database. Typical use cases include: (1) mapping PacBio or Oxford Nanopore genomic reads to the human genome; (2) finding overlaps between long reads with error rate up to ~15%; (3) splice-aware alignment of PacBio Iso-Seq or Nanopore cDNA or Direct RNA reads against a reference genome; (4) aligning Illumina single- or paired-end reads; (5) assembly-to-assembly alignment; (6) full-genome alignment between two closely related species with divergence below ~15%.

For ~10kb noisy reads sequences, minimap2 is tens of times faster than mainstream long-read mappers such as BLASR, BWA-MEM, NGMLR and GMAP. It is more accurate on simulated long reads and produces biologically meaningful alignment ready for downstream analyses. For >100bp Illumina short reads, minimap2 is three times as fast as BWA-MEM and Bowtie2, and as accurate on simulated data. Detailed evaluations are available from the minimap2 paper or the preprint.

Installation

Minimap2 is optimized for x86-64 CPUs. You can acquire precompiled binaries from the release page with:

curl -L https://github.com/lh3/minimap2/releases/download/v2.18/minimap2-2.18_x64-linux.tar.bz2 | tar -jxvf -
./minimap2-2.18_x64-linux/minimap2

If you want to compile from the source, you need to have a C compiler, GNU make and zlib development files installed. Then type make in the source code directory to compile. If you see compilation errors, try make sse2only=1 to disable SSE4 code, which will make minimap2 slightly slower.

Minimap2 also works with ARM CPUs supporting the NEON instruction sets. To compile for 32 bit ARM architectures (such as ARMv7), use make arm_neon=1. To compile for for 64 bit ARM architectures (such as ARMv8), use make arm_neon=1 aarch64=1.

Minimap2 can use [SIMD Everywhere (SIMDe)][simde] library for porting implementation to the different SIMD instruction sets. To compile using SIMDe, use make -f Makefile.simde. To compile for ARM CPUs, use Makefile.simde with the ARM related command lines given above.

General usage

Without any options, minimap2 takes a reference database and a query sequence file as input and produce approximate mapping, without base-level alignment (i.e. no CIGAR), in the PAF format:

minimap2 ref.fa query.fq > approx-mapping.paf

You can ask minimap2 to generate CIGAR at the cg tag of PAF with:

minimap2 -c ref.fa query.fq > alignment.paf

or to output alignments in the SAM format:

minimap2 -a ref.fa query.fq > alignment.sam

Minimap2 seamlessly works with gzip'd FASTA and FASTQ formats as input. You don't need to convert between FASTA and FASTQ or decompress gzip'd files first.

For the human reference genome, minimap2 takes a few minutes to generate a minimizer index for the reference before mapping. To reduce indexing time, you can optionally save the index with option -d and replace the reference sequence file with the index file on the minimap2 command line:

minimap2 -d ref.mmi ref.fa                     # indexing
minimap2 -a ref.mmi reads.fq > alignment.sam   # alignment

Importantly, it should be noted that once you build the index, indexing parameters such as -k, -w, -H and -I can't be changed during mapping. If you are running minimap2 for different data types, you will probably need to keep multiple indexes generated with different parameters. This makes minimap2 different from BWA which always uses the same index regardless of query data types.

Use cases

Minimap2 uses the same base algorithm for all applications. However, due to the different data types it supports (e.g. short vs long reads; DNA vs mRNA reads), minimap2 needs to be tuned for optimal performance and accuracy. It is usually recommended to choose a preset with option -x, which sets multiple parameters at the same time. The default setting is the same as map-ont.

Map long noisy genomic reads

minimap2 -ax map-pb  ref.fa pacbio-reads.fq > aln.sam   # for PacBio subreads
minimap2 -ax map-ont ref.fa ont-reads.fq > aln.sam      # for Oxford Nanopore reads

The difference between map-pb and map-ont is that map-pb uses homopolymer-compressed (HPC) minimizers as seeds, while map-ont uses ordinary minimizers as seeds. Emperical evaluation suggests HPC minimizers improve performance and sensitivity when aligning PacBio reads, but hurt when aligning Nanopore reads.

Map long mRNA/cDNA reads

minimap2 -ax splice:hq -uf ref.fa iso-seq.fq > aln.sam       # PacBio Iso-seq/traditional cDNA
minimap2 -ax splice ref.fa nanopore-cdna.fa > aln.sam        # Nanopore 2D cDNA-seq
minimap2 -ax splice -uf -k14 ref.fa direct-rna.fq > aln.sam  # Nanopore Direct RNA-seq
minimap2 -ax splice --splice-flank=no SIRV.fa SIRV-seq.fa    # mapping against SIRV control

There are different long-read RNA-seq technologies, including tranditional full-length cDNA, EST, PacBio Iso-seq, Nanopore 2D cDNA-seq and Direct RNA-seq. They produce data of varying quality and properties. By default, -x splice assumes the read orientation relative to the transcript strand is unknown. It tries two rounds of alignment to infer the orientation and write the strand to the ts SAM/PAF tag if possible. For Iso-seq, Direct RNA-seq and tranditional full-length cDNAs, it would be desired to apply -u f to force minimap2 to consider the forward transcript strand only. This speeds up alignment with slight improvement to accuracy. For noisy Nanopore Direct RNA-seq reads, it is recommended to use a smaller k-mer size for increased sensitivity to the first or the last exons.

Minimap2 rates an alignment by the score of the max-scoring sub-segment, excluding introns, and marks the best alignment as primary in SAM. When a spliced gene also has unspliced pseudogenes, minimap2 does not intentionally prefer spliced alignment, though in practice it more often marks the spliced alignment as the primary. By default, minimap2 outputs up to five secondary alignments (i.e. likely pseudogenes in the context of RNA-seq mapping). This can be tuned with option -N.

For long RNA-seq reads, minimap2 may produce chimeric alignments potentially caused by gene fusions/structural variations or by an intron longer than the max intron length -G (200k by default). For now, it is not recommended to apply an excessively large -G as this slows down minimap2 and sometimes leads to false alignments.

It is worth noting that by default -x splice prefers GT[A/G]..[C/T]AG over GT[C/T]..[A/G]AG, and then over other splicing signals. Considering one additional base improves the junction accuracy for noisy reads, but reduces the accuracy when aligning against the widely used SIRV control data. This is because SIRV does not honor the evolutionarily conservative splicing signal. If you are studying SIRV, you may apply --splice-flank=no to let minimap2 only model GT..AG, ignoring the additional base.

Since v2.17, minimap2 can optionally take annotated genes as input and prioritize on annotated splice junctions. To use this feature, you can

paftools.js gff2bed anno.gff > anno.bed
minimap2 -ax splice --junc-bed anno.bed ref.fa query.fa > aln.sam

Here, anno.gff is the gene annotation in the GTF or GFF3 format (gff2bed automatically tests the format). The output of gff2bed is in the 12-column BED format, or the BED12 format. With the --junc-bed option, minimap2 adds a bonus score (tuned by --junc-bonus) if an aligned junction matches a junction in the annotation. Option --junc-bed also takes 5-column BED, including the strand field. In this case, each line indicates an oriented junction.

Find overlaps between long reads

minimap2 -x ava-pb  reads.fq reads.fq > ovlp.paf    # PacBio read overlap
minimap2 -x ava-ont reads.fq reads.fq > ovlp.paf    # Oxford Nanopore read overlap

Similarly, ava-pb uses HPC minimizers while ava-ont uses ordinary minimizers. It is usually not recommended to perform base-level alignment in the overlapping mode because it is slow and may produce false positive overlaps. However, if performance is not a concern, you may try to add -a or -c anyway.

Map short accurate genomic reads

minimap2 -ax sr ref.fa reads-se.fq > aln.sam           # single-end alignment
minimap2 -ax sr ref.fa read1.fq read2.fq > aln.sam     # paired-end alignment
minimap2 -ax sr ref.fa reads-interleaved.fq > aln.sam  # paired-end alignment

When two read files are specified, minimap2 reads from each file in turn and merge them into an interleaved stream internally. Two reads are considered to be paired if they are adjacent in the input stream and have the same name (with the /[0-9] suffix trimmed if present). Single- and paired-end reads can be mixed.

Minimap2 does not work well with short spliced reads. There are many capable RNA-seq mappers for short reads.

Full genome/assembly alignment

minimap2 -ax asm5 ref.fa asm.fa > aln.sam       # assembly to assembly/ref alignment

For cross-species full-genome alignment, the scoring system needs to be tuned according to the sequence divergence.

Advanced features

Working with >65535 CIGAR operations

Due to a design flaw, BAM does not work with CIGAR strings with >65535 operations (SAM and CRAM work). However, for ultra-long nanopore reads minimap2 may align ~1% of read bases with long CIGARs beyond the capability of BAM. If you convert such SAM/CRAM to BAM, Picard and recent samtools will throw an error and abort. Older samtools and other tools may create corrupted BAM.

To avoid this issue, you can add option -L at the minimap2 command line. This option moves a long CIGAR to the CG tag and leaves a fully clipped CIGAR at the SAM CIGAR column. Current tools that don't read CIGAR (e.g. merging and sorting) still work with such BAM records; tools that read CIGAR will effectively ignore these records. It has been decided that future tools will will seamlessly recognize long-cigar records generated by option -L.

TL;DR: if you work with ultra-long reads and use tools that only process BAM files, please add option -L.

The cs optional tag

The cs SAM/PAF tag encodes bases at mismatches and INDELs. It matches regular expression /(:[0-9]+|\*[a-z][a-z]|[=\+\-][A-Za-z]+)+/. Like CIGAR, cs consists of series of operations. Each leading character specifies the operation; the following sequence is the one involved in the operation.

The cs tag is enabled by command line option --cs. The following alignment, for example:

CGATCGATAAATAGAGTAG---GAATAGCA
||||||   ||||||||||   |||| |||
CGATCG---AATAGAGTAGGTCGAATtGCA

is represented as :6-ata:10+gtc:4*at:3, where :[0-9]+ represents an identical block, -ata represents a deltion, +gtc an insertion and *at indicates reference base a is substituted with a query base t. It is similar to the MD SAM tag but is standalone and easier to parse.

If --cs=long is used, the cs string also contains identical sequences in the alignment. The above example will become =CGATCG-ata=AATAGAGTAG+gtc=GAAT*at=GCA. The long form of cs encodes both reference and query sequences in one string. The cs tag also encodes intron positions and splicing signals (see the minimap2 manpage for details).

Working with the PAF format

Minimap2 also comes with a (java)script paftools.js that processes alignments in the PAF format. It calls variants from assembly-to-reference alignment, lifts over BED files based on alignment, converts between formats and provides utilities for various evaluations. For details, please see misc/README.md.

Algorithm overview

In the following, minimap2 command line options have a dash ahead and are highlighted in bold. The description may help to tune minimap2 parameters.

1. Read -I [=4G] reference bases, extract (-k,-w)-minimizers and index them in a hash table.

2. Read -K [=200M] query bases. For each query sequence, do step 3 through 7:

3. For each (-k,-w)-minimizer on the query, check against the reference index. If a reference minimizer is not among the top -f [=2e-4] most frequent, collect its the occurrences in the reference, which are called seeds.

4. Sort seeds by position in the reference. Chain them with dynamic programming. Each chain represents a potential mapping. For read overlapping, report all chains and then go to step 8. For reference mapping, do step 5 through 7:

5. Let P be the set of primary mappings, which is an empty set initially. For each chain from the best to the worst according to their chaining scores: if on the query, the chain overlaps with a chain in P by --mask-level [=0.5] or higher fraction of the shorter chain, mark the chain as secondary to the chain in P; otherwise, add the chain to P.

6. Retain all primary mappings. Also retain up to -N [=5] top secondary mappings if their chaining scores are higher than -p [=0.8] of their corresponding primary mappings.

7. If alignment is requested, filter out an internal seed if it potentially leads to both a long insertion and a long deletion. Extend from the left-most seed. Perform global alignments between internal seeds. Split the chain if the accumulative score along the global alignment drops by -z [=400], disregarding long gaps. Extend from the right-most seed. Output chains and their alignments.

8. If there are more query sequences in the input, go to step 2 until no more queries are left.

9. If there are more reference sequences, reopen the query file from the start and go to step 1; otherwise stop.

Getting help

Manpage minimap2.1 provides detailed description of minimap2 command line options and optional tags. The FAQ page answers several frequently asked questions. If you encounter bugs or have further questions or requests, you can raise an issue at the issue page. There is not a specific mailing list for the time being.

Citing minimap2

If you use minimap2 in your work, please cite:

Li, H. (2018). Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics, 34:3094-3100. doi:10.1093/bioinformatics/bty191

Developers' Guide

Minimap2 is not only a command line tool, but also a programming library. It provides C APIs to build/load index and to align sequences against the index. File example.c demonstrates typical uses of C APIs. Header file minimap.h gives more detailed API documentation. Minimap2 aims to keep APIs in this header stable. File mmpriv.h contains additional private APIs which may be subjected to changes frequently.

This repository also provides Python bindings to a subset of C APIs. File python/README.rst gives the full documentation; python/minimap2.py shows an example. This Python extension, mappy, is also available from PyPI via pip install mappy or from BioConda via conda install -c bioconda mappy.

Limitations

• Minimap2 may produce suboptimal alignments through long low-complexity regions where seed positions may be suboptimal. This should not be a big concern because even the optimal alignment may be wrong in such regions.

• Minimap2 requires SSE2 instructions on x86 CPUs or NEON on ARM CPUs. It is possible to add non-SIMD support, but it would make minimap2 slower by several times.

• Minimap2 does not work with a single query or database sequence ~2 billion bases or longer (2,147,483,647 to be exact). The total length of all sequences can well exceed this threshold.

• Minimap2 often misses small exons.