PBSIM2

Simulate PacBio and Oxford Nanopore long reads from a reference genome using an HMM model of quality scores.

License: GPL-2.0. Source code.

What it simulates

PBSIM2 turns a reference FASTA into long reads that look like PacBio or Oxford Nanopore output. It draws read lengths from a gamma-type length model, cuts them from random positions on the reference, and adds errors. A hidden Markov model (HMM) trained on real runs sets the quality codes, which sets where the errors fall. You give it a reference, a chemistry model and a depth. It returns a FASTQ, a MAF alignment of each read to the reference, and a copy of the reference.

Key assumptions: read length follows the mean and standard deviation you set, and the mix of substitutions, insertions and deletions follows the ratio you set. The ratio is a global fixed mix. It does not depend on sequence context.

When to use it

Test an assembler, mapper or variant caller on long reads with a known truth (the MAF file lists every error).
Compare how P6C4 (PacBio) and R103 (Nanopore) error profiles change an alignment.
Plan coverage: set --depth and count the reads you get.

Alternatives: PBSIM3 adds multi-pass reads, transcript and template modes and newer ONT models. SimLoRD models PacBio errors that fall with the number of passes. CycSim makes errors depend on the surrounding k-mer. Use wgsim or DWGSIM for short reads.

Install it yourself

PBSIM2 2.0.1 is on bioconda. The recipe builds it for linux/amd64 only. A dry-run resolve on linux-64 picked build h9948957_4 (378 kB), so on Apple Silicon run it through emulation.

conda create -n pbsim2 -c conda-forge -c bioconda pbsim2=2.0.1
conda activate pbsim2

The six HMM models (P4C2, P5C3, P6C4, R94, R95, R103) are in $CONDA_PREFIX/data. Minimal command:

pbsim --hmm_model $CONDA_PREFIX/data/R103.model --depth 5 reference.fa

Defaults from the program's usage text: prefix sd, depth 20, mean length 9000, SD 7000, minimum length 100, maximum 1,000,000, mean accuracy 0.85, ratio 6:50:54 (substitution:insertion:deletion), seed taken from the clock.

Run the tested example

This recipe builds PBSIM2 in a container and runs the example below. It needs Docker and nothing else. It last ran on 2026-10-12 in 0.2 s. PBSIM2 has no ARM build, so on Apple silicon Docker runs the x86 image under emulation, which is slower.

Download pbsim2-recipe.tar.gz (Dockerfile, example reference.fa, run.sh).

Unpack it and run the script:

tar -xzf pbsim2-recipe.tar.gz
cd pbsim2
sh run.sh

What run.sh does

docker build --platform linux/amd64 -t biosimhub/pbsim2:2.0.1 .
mkdir -p out && chmod a+w out
docker run --rm --platform linux/amd64 \
  --user "$(id -u):$(id -g)" -e HOME=/tmp \
  -v "$PWD/in:/work/in:ro" -v "$PWD/out:/work/out" -w /work/out \
  biosimhub/pbsim2:2.0.1 pbsim --hmm_model /opt/conda/data/R103.model \
    --difference-ratio 23:31:46 --depth 5 --length-mean 3000 \
    --length-sd 1000 --length-min 100 --accuracy-mean 0.85 \
    --seed 42 --prefix sim /work/in/reference.fa > out/pbsim_log.txt

Inputs are read from in/ and results land in out/. Change the values after each option to run your own case; the parameter table lists them.

Expected output

File	Size
`pbsim_log.txt`	0 B
`sim_0001.fastq`	587.4 KB
`sim_0001.maf`	630.0 KB
`sim_0001.ref`	59.4 KB

First lines of sim_0001.fastq

@S1_1
ACGACAAAATGTTCAAGTAAGGCGTGCAAGGCGTAAGCCCTCACCTGGTTAATAGAGTTAAACGCCAAGGCGCCAATCCGGAGTGCTGCGTTTTCACATAGTTCTCAGTTCACAGGGCAATCAACCCAATCCGTAAGTAAGATGGGTGGGCAGCAAGCGTACCATTACTCCTGCCCATTACACTTAAAGGGCCAACTAAT

First lines of sim_0001.maf

a
s ref 53253 3688 + 60000 A-GAGAAAAT-ATTCAAGTAAGGCGGTGCAAGGCGTAAGC-CATACACCTGGTTAATAGAGTTAAACGCCAAGTGCGCCAATCCGGAGTGCTGCGTTCTTCACATAGTTCTCAGTTCCACACC-GCCAATCAACCCAAT-CGTCAAGTAAGCATGAGGTGGGCAGCCAGCCGTAC

First lines of sim_0001.ref

>synthetic_ref
GCTAAAGACAATTACATAACATACACGTCAGCACGAAACTTGTTGGCCCAGTGTGAATCGCTTAAGGGTT
AAGTAAGTGTGATGCATACGCCTTTACTTGCTGTGTCCACCCCATCGGACTGGCATTTTTATTACACTCA
GAAACAGAACTCGGGTAATTTTGACAGGTCACGCAGAGGCGCGCCCTCCTGAAGTGCGTGGACACTCGCT
ATGAATCTCTGATTTACCCACTCTGCCAAACTCCAGCGCGGTCAGTTCCATCACCCTAAGTAACCGAATA
ATGCGTTCGCTCTATTGACTACGACGCGCTCATTCCCTTGTCGGAGAGTTATGGAACAAGGACGCTGTCT
GAGACTAGAAGACAGATAGTGCACACGACCGGCGTCGGAGAAACTCTATTTGCCGCCTGACAAGTCAATG
CGATCCGTAGGGGCAGCGCAGTATGCCAAGACTATAGGCACTGTCGCATCACAAACGATTAACTGATAAA
TGAGCCCTTTATGACACGGGCATATGACTGGTTTACGATAGTATGTCCAACGGCGAGCTTTACATTTGCT
GTGAGAGGTACAGGGATTAGTGAGAAGCCGTGCGTATCAATTCGTACCTTGGGGGTCGTTACCACTCTGT
TCCCACGAGCGGCATTTCTGGATGGCCAGCTTTTGACATTTAATTTCACCCATAAACCAGCGTAAAGCTG
CAAGTGGCTCCATGAACTTAGCTGCTAGTGTCAGACTCGCCTCGGATCCTTACTACACTAACTTGAACGC

Parameters

Option	Meaning	Example value	Range
`/opt/conda/data/….model`	Quality HMM modelP* are PacBio chemistries, R* are Oxford Nanopore pore versions	R103	P4C2, P5C3, P6C4, R94, R95, R103
`--difference-ratio`	Substitution:insertion:deletion ratio6:50:54 is the PacBio default, 23:31:46 suits Nanopore	23:31:46	6:50:54, 23:31:46, 10:60:30
`--depth`	Coverage depth	5	0.1 to 50
`--length-mean`	Mean read length	3000	100 to 100000
`--length-sd`	Read length standard deviation	1000	0 to 50000
`--length-min`	Minimum read length	100	10 to 10000
`--accuracy-mean`	Mean read accuracy	0.85	0.6 to 0.99
`--seed`	Seed	42	1 to 2147483647

Ranges are the limits used for the example recipe; the program itself may accept more.

Reading the output

The example simulates 5x on a 60,000 bp reference with the R103 Nanopore model and ratio 23:31:46. pbsim_log.txt is empty because the example saves only stdout. The run summary goes to stderr.

sim_0001.fastq (601,470 bytes): 93 reads, 300,000 bases in total, exactly 5 x 60,000. Mean length is 3,226 bp for a requested mean of 3,000. IDs run @S1_1, @S1_2 and so on.
sim_0001.maf (645,111 bytes): one a block per read. The first s line is the reference (ref 53253 3688 + 60000: start, aligned length, strand, source size). The second is the read (S1_1 0 3643 + 3643). Gaps (-) mark insertions and deletions.
sim_0001.ref: the input sequence, named synthetic_ref.

Counting columns in the MAF gives 319,919 alignment columns with 86.5% identity. Of the errors, 23.1% are substitutions, 30.9% insertions and 46.0% deletions. That matches the 23:31:46 ratio you asked for.

Your own runs

All runs below use the recipe's image and the example reference with seed 42.

PacBio model, deeper. P6C4, ratio 6:50:54, depth 20, mean length 9000, SD 7000. Result: 138 reads, 1,200,000 bases, mean length 8,696, longest read 41,394. Substitutions fall to 5.4% of errors, insertions 45.3%, deletions 49.3%.
Higher accuracy. Same as the example but --accuracy-mean 0.95. Result: 100 reads, identity 96.0% (86.5% before), so the error rate drops from 13.5% to 4.0% of columns.
Seeds. Seed 42 twice gives byte-identical FASTQ files. Seed 7 gives 104 reads with 86.0% identity.

With your own genome, mount the FASTA and keep depth near 20 to start. Multi-FASTA input produces sim_0001, sim_0002 and so on, one set per sequence.

Pitfalls

Ratio defaults suit PacBio. The default 6:50:54 gives insertion-heavy reads. For Nanopore the authors recommend about 23:31:46. Set it with --difference-ratio.
Accuracy is a target for the model. The example asked for 0.85 and got 86.5% identity. Check the MAF before you rely on a number.
Length defaults are long. The default mean is 9,000 with SD 7,000. On a 60 kb reference that gives few, very long reads. Set --length-mean and --length-sd for your genome.
Model path. --hmm_model needs a file path. The models sit under $CONDA_PREFIX/data, so pass the full path.
Unseeded runs differ. Without --seed the clock is the seed, so two runs give different reads.