A*PA & A*PA2: A* Pairwise Aligner

A*PA is a global pairwise sequence aligner for edit distance using A*, co-authored by @pesho-ivanov and @RagnarGrootKoerkamp.

A*PA2 is an improvement of A*PA that uses a DP-based approach instead of plain A*. It achieves up to 20x speedup over other exact aligners and is competitive with approximate aligners.

An alignment of two sequences of length 500 with 30% error rate using A*PA:

An alignment of two sequences of length 10’000 with 15% error rate using A*PA2:

Papers

For A*PA, we recommend reading the bioRxiv version which directly includes the supplement and has better formatting. But please cite the Bioinformatics version!

• A*PA BioRxiv preprint:

  Ragnar Groot Koerkamp, Pesho Ivanov. “Exact global alignment using A* with chaining seed heuristic and match pruning”. bioRxiv (2024). 10.1101/2022.09.19.508631

• A*PA OUP Bioinformatics paper:

  Ragnar Groot Koerkamp, Pesho Ivanov. “Exact global alignment using A* with chaining seed heuristic and match pruning”. Bioinformatics (2024). 10.1093/bioinformatics/btae032

• A*PA2 BioRxiv preprint:

  Ragnar Groot Koerkamp. “A*PA2: up to 20 times faster exact global alignment”. bioRxiv (2024). 10.1101/2024.03.24.586481

Links

• Twitter: @curious_coding, @peshotrie,
• Matrix: @curious_coding:matrix.org,
• Blog: https://curiouscoding.nl

Usage

If you run into any kind of problem or unclarity, please (please 🥺) make an issue or reach out on twitter or matrix.

Rust API

To call A*PA2 from another Rust crate, simply add the astarpa[2] crate in this repo as a git dependency.

For A*PA2, use astarpa2_simple(a, b) or astarpa2_full(a, b) in the ~astarpa2~ crate, or customize parameters with e.g.

let mut params = astarpa2::AstarPa2Params::full();
params.front.incremental_doubling = false;
let mut aligner = params.make_aligner(true);
let (cost, cigar) = aligner.align(a, b);

The astarpa crate is the main entrypoint for A*PA. See the docs there. Use astarpa::astarpa(a, b) for alignment with default settings or astarpa::astarpa_gcsh(a,b,r,k,end_pruning) to use GCSH+DT with custom parameters.

More complex usage examples can be found in pa-bin/examples.

C API

The astarpa-c crate contains simple C-bindings for the astarpa::{astarpa,astarpa_gcsh} and astarpa2::astarpa2_{simple,full} functions and an example with makefile. More should not be needed for simple usage. To run the resulting binary, make sure to export LD_LIBRARY_PATH=/path/to/astarpa/target/release.

Command line application

pa-bin is a small command line application that takes as input consecutive pairs of sequences from a .fasta, .seq, or .txt file (or can generate random input) and outputs costs and alignments to a .csv.

Install pa-bin to ~/.local/share/cargo/bin/pa-bin using the following (cloning this repo is not needed):

cargo install --git https://github.com/RagnarGrootKoerkamp/astar-pairwise-aligner pa-bin

This requires cargo Rust nightly. To get both, first install rustup. Then enable nightly: rustup install nightly; rustup default nightly.

To run from the repository: clone and cargo run --release -- <pa-bin flags>.

cargo run --release -- -h

Globally align pairs of sequences using A*PA

Usage: pa-bin [OPTIONS] <--input <INPUT>|--length <LENGTH>>

Options:
  -i, --input <INPUT>      A .seq, .txt, or Fasta file with sequence pairs to align
  -o, --output <OUTPUT>    Write a .csv of `{cost},{cigar}` lines
      --aligner <ALIGNER>  The aligner to use [default: astarpa2-full] [possible values: astarpa,
                           astarpa2-simple, astarpa2-full]
  -h, --help               Print help (see more with '--help')

Generated input:
  -n, --length <LENGTH>          Target length of each generated sequence [default: 1000]
  -e, --error-rate <ERROR_RATE>  Error rate between sequences [default: 0.05]

Visualization

The Rust API supports generating visualizations using the sdl2 library and ttf fonts. If this gives errors, install sdl2: e.g. using apt-get install libsdl2-ttf-dev.

Here are some sample videos. The first five correspond to figure 1 of the A*PA paper. Timings are not comparable due to differences in visualization strategies (cell vs layer updates).

Dijkstra	Ukkonen’s exponential search (Edlib)
Diagonal transition (WFA)	DT + Divide & Conquer (BiWFA)
A*PA (GCSH+DT)	A*PA2-full (8-bit words; block size 32)

Paper artefacts

Figures

Paper figures are generated using the example binaries at pa-bin/examples/astarpa-figures and pa-bin/examples/astarpa2-figures.

Evals

Benchmarking code, evals, and datasets can be found in the pa-bench repo. For A*PA, results can be found in this notebook and reproduced using this makefile. For A*PA2, results can be found in this notebook and reproduced using this justfile. Dataset downloads are in this release.

Tests

Code is tested for correctness in various tests (astarpa/src/tests.rs) against triple-accel. The benchmark tool pa-bench also checks correctness automatically.

Crate structure

Code is spread out over multiple crates. From low to high:

• pa-types: Basic types such as Seq, Pos, Cigar, and Cost, hosted in the pairwise-alignment org.
• pa-affine-types: Types for affine edit graphs such as State = (Pos, Layer), AffineCigar, and CostModel. Not used by A*PA, but other algorithms and the visualizer support it.
• pa-heuristic: Code for
  • finding matches
  • computing contours (fast and bruteforce)
  • heuristics themselves
  • wrapper/bruteforce heuristics for debugging
• pa-vis-types: Trait definition of the visualizer callbacks, and the empty NoVis visualizer.
• astarpa: Main A*PA API entrypoint containing the astar and astar_dt functions, the bucket_queue data structure, and the astarpa(a,b) entrypoint.
• astarpa-c: C-bindings for astarpa
• pa-vis: The visualizer. Contains a Canvas trait implemented for the SDL2Canvas. The sdl2 feature is optional.
• pa-generate: Library and binary to generate different types of random sequences.
• pa-bin: Main command line interface to A*PA. Allows for input from file, generated input, visualizing, and customization of the A*PA parameters.
• pa-bitpacking: Implementation of Myers’ bitpacking algorithms and SIMD extensions.
• astarpa2: A*PA2 entrypoint containing astarpa2_simple and astarpa2_full functions.
• pa-base-algos: Re-implementations of Needleman-Wunsch/Edlib and Diagonal-transition/WFA/BiWFA for visualizations.
• astarpa-next: Some code for other new ideas such as path-pruning.
• pa-web: web-interface to A*PA by compiling to webassembly. Implements the Canvas trait for HTMLCanvas. (Not maintained.)

License

MPL-2.0