Introduction

Sfasta is a replacement for the FASTA/Q format with a goal of both saving space but also having very fast random-access for machine learning, even for large datasets (such as the nt database, 217Gb gzip compressed, 225Gb bgzip compressed, and 175Gb with sfasta, including an index).

Speed comes from assuming modern hardware, thus:

• Multiple compression threads
• I/O dedicated threads
• SIMD bitpacking support
• modern compression algorithms (ZSTD, as default).
• Dual level ID Index
• Everything stored as sequence streams (for better compression).

If you need different hardware support, open an issue. There are SIMD support for other architectures that could be implemented.

While the goals are random-access speed by ID query, and smaller size, I hope it can become a more general purpose format. It supports other compression algorithms, which could be used for archival purposes (such as xz compression). This is a work in progress, but ready for community feedback. Because the goals are not simple decompression, this part of the code is not-optimized yet, and is much slower than zcat and other tools. This will be remedied in the future.

Note

This has taken a few years of on again, off again development. FORMAT.md and other files are likely out of date.

Community Feedback Period

I'm hopeful folks will check this out, play around, break it, and give feedback.

FASTQ support

It's now trivial to add the scores as another compression stream, and the struct that puts it all back together already has an entries for scores. I'll get to it soon.

TODO

• Does Zstd dict support work for small block sizes (8kb, 16kb, 4kb)?
• less than 8kbp block sizes should use dicts by default
• Masking as bitvec or [Bool; N]?

Usage

Installation

cargo install sfasta Don't have cargo?

Usage

To compress a file:

sfa convert MyFile.fasta

You can also convert directly from gzipped files:

sfa convert MyFile.fasta.gz

You can use other compression schemes. The software automatically detects which is used and decompresses accordingly.

sfa convert --snappy MyFile.fasta sfa convert --xz MyFile.fasta

You can also change the block size. Units are in * 1024 bytes. Default is 4Mb (4096):

sfa convert --block-size 8192 MyFile.fasta

View a file:

sfa view MyFile.sfasta

Query a file by sequence ID:

sfa faidx MyFile.sfasta Chr1

For help:

sfa --help

Please note, not all subcommands are implemented yet. The following should work: convert, view, list, faidx.

Requirements

Should work anywhere that supports Rust. Tested on Ubuntu, Red Hat, and Windows 10. I suspect it will work on Mac OS. I also do not think it supports WASM at this time.

Comparisons

Features

Uniprot Random Access

Samtools index pre-built

Uniprot Compression Speed

Uniprot Size

Uncompressed: 282M

Nanopore Reads

As a FASTA file.

Nanopore Reads Compression Speed

Nanopore Reads Random Access

Samtools index pre-built

Nanopore Reads Size

Uncompressed Size: 8.8G

Genome

Genome Random Access Speed

Samtools index pre-built

Genome Compression Speed

Genome Size

Uncompressed: 2.7G

Future Plans

Implement NAF-like algorithm

NAF has an advantage with 4bit encoding. It's possible to implement this, and use 2bit when possible, to gain additional speed-ups. Further, there is some SIMD support for 2bit and 4bit DNA/RNA encoding.

Subsequence support for faidx CLI

Only open the block(s) that contain the subsequence. The index is already set up to support this and I've had it working before in the python bindings.

Additional Speed-ups

There is plenty of room for speed improvements, including adding more threads for specific tasks, CPU affinities, native compilation, and maybe more Cow.

Additional Compression

There is likely room to decrease size as well without hurting speed.

Command-line interface

As I've refactored much of the library, the CLI code withered and decayed. Need to fix this.

Quality Scores

To support FASTQ files

Adjust compression level and compression method

For other applications (such as long term storage)

C and Python bindings

To make it easier to use in other programs and in python/jupyter

Small file optimization

Sfasta is currently optimized for larger files.

GFA file format support

Graph genome file format is in dire need of an optimized format

Profile Guided Optimization

Never mind. This somehow doubled the time it takes to compress binaries. Enable PGO for additional speed-ups

Fuzzing

The following format parsers are fuzzed. To fuzz execute the following in the libsfasta directory:

cargo fuzz run parse_fastq  -- -max_len=8388608
cargo fuzz run parse_fasta  -- -max_len=8388608
cargo fuzz run parse_sfasta -- -detect_leaks=0 -rss_limit_mb=4096mb -max_len=8388608

FAQ

I get a strange symbol near the progress bar

You need to install a font that supports Unicode. I'll see if there is a way to auto-detect.

XZ compression is fast until about halfway, then slows to a crawl.

The buffers can store lots of sequence, but the compression algorithm takes longer.

Why so many dependencies?

Right it it works with a wide range of compression functions. Once some are determined to be the best others could be dropped from future versions. The file format itself has a version identifier so we could request people rollback to an older version if they need to.

Why samtools comparison?

I've got plenty of experiments trying to get a fast gzip compressed multi-threaded reader, but even when mounted on a ramdisk, it is too slow. Samtools is an awesome, handy tool that has the 'faidx' function, of which I am a huge fan. While the faidx is not it's main function, it is not optimized for large datasets, thus the test is a little unfair. Still, it's helpful to have something to compare to.

File Format

The best source is currently this file: conversion.rs. Masking is converted into instructions in a u32, see ml32bit.rs.

Header