SIMD-accelerated library for computing global and X-drop affine gap penalty sequence-to-sequence or sequence-to-profile alignments using an adaptive block-based algorithm.
Preprint paper available here.
```rust use blockaligner::{cigar::*, scanblock::, scores::};
let minblocksize = 32; let maxblocksize = 256;
// A gap of length n will cost: open + extend * (n - 1) let gaps = Gaps { open: -2, extend: -1 };
// Note that PaddedBytes, Block, and Cigar can be initialized with sequence length
// and block size upper bounds and be reused later for shorter sequences, to avoid
// repeated allocations.
let r = PaddedBytes::frombytes::
// Align with traceback, but no X-drop threshold (global alignment).
let mut a = Block::
asserteq!(res, AlignResult { score: 7, queryidx: 24, reference_idx: 21 });
let mut cigar = Cigar::new(res.queryidx, res.referenceidx); // Compute traceback and resolve =/X (matches/mismatches). a.trace().cigareq(&q, &r, res.queryidx, res.reference_idx, &mut cigar);
asserteq!(cigar.tostring(), "2=6I16=3D"); ``` See the docs for detailed API information.
Block aligner provides a new efficient way to compute pairwise alignments on proteins, DNA sequences, and byte strings with dynamic programming. Block aligner also supports aligning sequences to profiles, which are position-specific scoring matrices and position-specific gap open costs.
It works by calculating scores in a small square block that is shifted down or right in a greedy manner, based on the scores at the edges of the block. This dynamic approach results in a much smaller calculated block area compared to previous approaches, though at the expense of some accuracy. The block can also go back to a previous best checkpoint and grow larger, to handle difficult regions with large gaps. The block size can also dynamically shrink when it detects that a large block is not needed. Both block growing and shrinking are based on heuristics.
By trading off some accuracy for speed, block aligner is able to efficiently handle a variety of scoring matrices and adapt to sequences of varying sequence identities. In practice, it is still very accurate on a variety of protein and nucleotide sequences.
Block aligner is designed to exploit SIMD parallelism on modern CPUs. Currently, SSE2 (128-bit vectors), AVX2 (256-bit vectors), Neon (128-bit vectors), and WASM SIMD (128-bit vectors) are supported. For score calculations, 16-bit score values (lanes) and 32-bit per block offsets are used.
Block aligner behaves similarly to an (adaptive) banded aligner when the minimum and maximum block size is set to the same value.
For long, noisy Nanopore reads, a min block size of ~1% sequence length and a max block size
of ~10% sequence length performs well (tested with reads up to ~50kbps).
For proteins, a min block size of 32 and a max block size of 256 performs well.
Using a minimum block size that is at least 32 is recommended for most applications.
Using a maximum block size greater than 2^14 = 16384 is not recommended.
If the alignment scores are saturating (score too large), then use a smaller block size.
Let me know how block aligner performs on your data!
This library can be used on both stable and nightly Rust channels. The nightly channel is needed for running tests and benchmarks. Additionally, the tests and benchmarks need to run on Linux or MacOS.
To use this as a crate in your Rust project, add the following to your Cargo.toml:
[dependencies]
block-aligner = { version = "0.4", features = ["simd_avx2"] }
Use the simd_sse2, simd_neon, or simd_wasm feature flag for x86 SSE2, ARM Neon, or WASM SIMD support, respectively.
It is your responsibility to ensure the correct feature to be enabled and supported by the
platform that runs the code because this library does not automatically detect the supported
SIMD instruction set. More information on specifying different features for different platforms
with the same dependency here.
Here's a simple example:
[target.'cfg(target_arch = "x86_64")'.dependencies]
block-aligner = { version = "0.4", features = ["simd_avx2"] }
[target.'cfg(target_arch = "aarch64")'.dependencies]
block-aligner = { version = "0.4", features = ["simd_neon"] }
For developing, testing, or using the C API, you should clone this repo and use Rust nightly. In general, when building, you need to specify the correct feature flags through the command line.
For x86 AVX2:
cargo build --features simd_avx2 --release
For x86 SSE2:
cargo build --features simd_sse2 --release
For ARM Neon:
cargo build --target=aarch64-unknown-linux-gnu --features simd_neon --release
For WASM SIMD:
cargo build --target=wasm32-wasi --features simd_wasm --release
To run WASM programs, you will need wasmtime
installed and on your $PATH.
There are C bindings for block aligner. More information on how to use them is located in
the C readme.
See the 3di branch for an example of using block aligner to do local alignment in C,
along with block aligner modifications to support aligning with amino acid 3D interaction (3Di) information.
Most of the instructions below are for benchmarking and testing block aligner.
Some Illumina/Nanopore (DNA), Uniclust30 (protein), and SCOP (protein profile) data are used in some tests and benchmarks. You will need to download them by following the instructions in the data readme.
Run scripts/test_avx2.sh or scripts/test_wasm.sh to run tests.
CI will run these tests when commits are pushed to this repo.
More testing and evaluating scripts are available in the scripts directory.
For debugging, there exists a debug feature flag that prints out a lot of
useful info about the internal state of the aligner while it runs.
There is another feature flag, debug_size, that prints the sizes of blocks after they grow.
To manually inspect alignments, run scripts/debug_avx2.sh with two sequences as arguments.
Run scripts/doc_avx2.sh or scripts/doc_wasm.sh to build the docs locally.
Run scripts/bench_avx2.sh or scripts/bench_wasm.sh for basic benchmarks.
See the scripts directory for more benchmark scripts on real data.
Use the Jupyter notebook in the vis/ directory to gather data and plot them. An easier way
to run the whole notebook is to run the vis/run_vis.sh script.
Use
brew install cargo-instruments
RUSTFLAGS="-g" cargo instruments --example profile --release --features simd_avx2 --open
Use
scripts/build_ir_asm.sh
to generate assembly output and run LLVM-MCA.
Use either scripts/build_ir_asm.sh, objdump -d on a binary (avoids recompiling code in
some cases), or a more advanced tool like Ghidra (has a decompiler, too).
Edits were made to Hajime Suzuki's adaptive banding benchmark code and difference recurrence benchmark code. These edits are available here and here, respectively. Go to those repos, then follow the instructions for installing and running the code.
If you run the scripts in those repos for comparing scores produced by different algorithms,
you should get .tsv generated files. Then, in this repo's directory, run
scripts/compare_avx2.sh /path/to/file.tsv 50
to get the comparisons. The X-drop threshold is specified after the path.
See the ideas file.