James Bonfield, Wellcome Trust Sanger Institute 2011 - 2013.
This is the 4th variant of an experimental code to compress fastq data. It has a number of caveats.
The fastq must consist of [ACGTN], or [0123.]. Anything else is considered as an “N” or “.”.
Sequence and quality should be on one line only.
All bases called as N will be assumed to have quality 0. This can disagreements after decompression. Arguably this is a feature as how can you have anything other than zero confidence that “N” is the correct call?
Although the program usage and documentation refer to .fqz files, this is not compatible with the previous two variants using that format extension. (However the magic number has been amended to include the appropriate format version code so they can detect which files are which.)
It has been tested on Illumina, 454 data, PacBio and SOLiD data files. However there may be specific caveats with this formats as some break the assumptions listed above. See below for suggested compression parameters for each machine type.
The program reads from stdin and writes to stdout, so it can be used as part of a pipeline.
To compress: fqzcomp options < a.fasta > a.fqz
To uncompress fqzcomp -d < a.fqz > a.fasta
Additionally in v4.0 onwards it can read/write to files instead.
fqzcomp -s5 -q3 a.fasta a.fqz
fqzcomp -d a.fqz a.fastaCompression options are as follows.
-s
Specifying a "+" after level will use two context sizes (eg -s5+);
a small fixed size and the <level> specified. This greatly helps
compression of small files and also slightly helps compression of
larger files.
Defaults to -s3.-b Use both strands when updating sequence contexts. This has no memory issues, but it slows down the compression and has only a small gain in compression ratios. Off by default.
Do not use this with SOLiD data files.-e By default the sequence contexts use 8-bit counters for memory efficiency. This forces 16-bit counters. It has a very small gain in compression ratio, but doubles sequence compression memory usage.
-q
-q1
Uses the previous quality value and the maximum of the two
previous to that as a context for predicting the next value.
Combined this adds 12 bits of context.
-q2
In addition to -q1, this extends the context by adding a
single bit to indicate if the 2nd and 3rd previous qualities
are identical to each other, as well as using 3 bits of
context for the running-delta. (A measure of how variable a
string of quality values are.) This is the default level.
-q3
As per -q2, but also adds 4 bits worth of context holding the
position each the sequence.-Q
All quality values are stored within +/- <distance> from their
actual value, but not necessarily always the same distance.
Defaults to -Q0 (lossless).-n
-n1 uses a very simple string based compressor that takes into
account the previous byte on this name and the same byte position
in the previous name, with some basic logic to try and keep
current and previous names approximately in alignment.
-n2 breaks the name down into tokens based on non alpha-numeric
characters and attempts to encode using numeric deltas where
appropriate. It typically compresses better on regular names
produced by (for example) Illumina machines, but may behave poorly
on other data. You should test which performs best for your data
set.
Defaults to -n2.-P Disables all multi-threading. This has no impact on the data format produced.
The default threading used is very minor, typically using under
200% of cpu usage, and considerably less on high -s levels.To uncompress, simply use -d. The compression parameters are saved in the output file so the decompression program knows what options to repeat during decompression.
The decompression may not perfectly match the original data. Sometimes this can generate checksum failures too (intentionally), although these can be glossed over by using -X on the decompressor.
Known issues are:
@SRR1234 ACGTACGTACGTGGGA +SRR1234 %$&!&((&&&&&
becomes
@SRR1234 ACGTACGTACGTGGGA + %$&!&((&&&&&
(This does not generate a checksum failure.)
For example:
@SRR1234 ACGTACGTACGTNGGA +SRR1234 ^%%%$$"&
becomes
@SRR1234 ACGTACGTACGTNGGA +SRR1234 ^%%%!$"&
The quality for base 9 (A) has been changed from !(0) to "(1). The quality for base 13(N) has been changed from $ to !(0).
Many of the algorithms are derived from fqzcomp-3.0 and in turn from fqzcomp-2.1, although the primary parsing code has been completely rewritten since version 2. Each type of data gets encoded using its own model.
Names are compared against the previous name, with a certain amount of wobble to match up colons and spaces where the intervening text is of different length (essentially a poor mans name alignment). This essentially means the context for encoding a name symbol is the character at the same position within the previous name. There is also dedicated prefix and suffix matching to help speed up encoding and reduce space further.
Sequences are converted into 2-bit values and the last N bases are used as the context for encoding the next base. The length of this context is the primary affect of changing the level with -s1 to -s9. Note it is not expected that you will have sufficient memory to cope with -s9. Practically going beyond -s6 may pose difficulties.
N bases are replaced with the most likely of A,C,G,T - whichever takes the least space to encode - and are replaced by Ns during decoding by noting that the quality is zero. (This is enforced.)
Qualities use an Order-2 model based on the previous two quality values. This can be augmented (-q1) with an additional context holding the approximate amount of variance along the sequence. By summing the differences between each quality and the next we can arrive at a figure indicating whether this sequence has smooth behaviour - a gradual reduction in quality along the length of the read - or highly variable behaviour with lots of dips in quality.
It was found that this delta summation had better discrimination of quality distributions than purely the position along the sequence. However on very large data sets the approximate position along the sequence can also be added in to the context by specifying -q3. Note that this uses memory memory and is slower.
Normally quality is encoded losslessly, but the -Q option permits encoding approximate qualities. With -Q 2 we could encode quality 35 as 33, 34, 35, 36 or 37. We choose (with approximations for speed) the quality value from that set which requires the fewest bits to encode and emit that instead. This was found to be significantly better at reducing storage volumes than simply degrading the quality into a series of N-wide bins, while still closely approximating the shape of a quality profile and preserving any dips and spikes.
The entropy encoding is based on a range coder taken from coders6c2 from http://ctxmodel.net/rem.pl?-37. This range coder was written by Eugene Shelwien, but it looks probable that it derived from Dmitry Subbotin’s range coder in PPMd. The modelling code is API compatible with the copy from Eugene Shelwien’s coders6c2 (which looks in turn to be derived from Dmitry Shkarin’s PPMd code), but the functions themselves have been completely rewritten to better fit the stationary probability models used in this data.
Different platforms give very different profiles of qualities and sequences. While all the compression parameters are valid, you may find that attempting to overcompress gives no gains (or even marginally harms compression) while still incurring additional CPU time. The space/time tradeoff will therefore differ per platform.
Illumina: -n2 -s7+ -b -q3, or for small files -n2 -s5 -q2
The bulk of development was done against Illumina files as these
were the test set for the SequenceSqueeze competition. It should
work fine with the defaults.454: -n1 -s7+ -b -q2, or -n1 -s5 -b -q2 for smaller filers
Tested to work OK, but -n1 gives a better name compression ratio
than -n2. (I have more work to do with -n2 it seems.) For large
files try "-n1 -s7+ -b -q3", with smaller -s for small files.PacBio: -n2 -s4+ -q1, or -n2 -s3 -q1 for smaller files
Tested to work mostly fine, but various non-N bases get turned
into Ns due to them having quality 0. Solution: replace quality 0
by quality 1 to prevent this. (tr '!' '"') How to fix in this
code?
PacBIO doesn't compress too well due to significantly higher
variability in the quality values (close to 3.7bit per).
Increasing -q has no real gains and infact harms compression as it
takes longer to train the model, so stick with -q1 for speed. Also
high order sequence models do not do too well due to the high
error rate. If you use a high -s, be sure to mix it with a low
order model too by adding '+'.SOLiD: -S -n2 -s5+ -q1
SOLiD support is available in a rudamentary form. It is not
auto-detected so you must currently specify -S when encoding. (The
decoder does not need this option.)
I do correctly deal with the latest NCBI SRA download format which
has sequences starting with the last primer base [ACGT] followed
by a series of [0123.] characters. The quality is one character
shorter. Eg SRR402768 encodes and decodes correctly.
However I have seen SOLiD fastq formats where the quality string
includes a dummy character for the un-measured primer base, and
also seen fastq files that have no primer base listed at all or
have one listed at each end. These formats are not supported
currently as I cannot find any definitive statement on what the
SOLiD fastq format is. It appears to change with wind direction!
In terms of compression, SOLiD quality values are far more
variable and have only minor dependance on position or context. So
-q1 is the most sensible option.
At present the complementary strand detection has not been
implemented for SOLiD, so do not use the -b parameter.(Compatible with 4.5 and 4.4)
Cope with Fastq files containing windows style carriage-return newline rather than unix newline only.
No longer generates checksum errors when encoding/decoding files containing data after the “+” line (eg the read name).
Bug fix to the arithmetic coder, which occasionally caused crashes.
Now treats non IUPAC ambiguity codes as N.
Improvements to reading/writing from pipes; flushing blocks when they’re full (as we do from a file) rather than when the pipe stalls.
Improved memory prefetching, particularly of sequence models. This speeds up high order sequence encoding by 30-40%. Decoding speeds have only improved by less, but overall the program is around 20% faster.
Adjusted the sequence model to use 20% less memory.
Support for more SOLiD formats. We now cope with the primer base having a dummy quality (or not as the case may be). Standards? What are they?
Bug fix to name parsing containing long strings of numbers.
Added -X option to ignore checksum failures (eg due to N vs ! issues). It seems quite commmon to have broken fastq files with impossible quality values. Eg calling an A,C,G or T with quality 0 (meaning it’s 100% incorrect, so why call it?) or calling an N with a non-zero quality (meaning we believe the N call may actually be correct).
Checksum failures now properly return a valid exit code.
Basic support for SOLiD sequence files added. There appear to be multiple conflicting ideas of what SOLiD fastq files look like though so mileage may vary.
Code tidyup of argument passing.
Using -Q
Minor tweaks to -q2 quality contexts.
I have added a fast checksum system to spot any potential errors in decoding. It’s still more likely that a corrupted file will yield a crash in the decompression code than emitting data that looks like valid fastq, but this is for safety.
Further improvements to the sequence model (-s5 vs -s5+) by combining multiple models. This only really helps on small data sets.
Completely reimplemented the various order-0 models upon which the other models are built. This was previous Eugene Shelwien’s tidy up of code from Dmitry Shkarin’s PPMd (I think). This means the only code not entirely mine now is the rangecoder itself.
Implemented a new name encoder, -n2. This is typically smaller than the old -n1 encoding on Illumina data, but has not been extensively tested on other naming schemes yet.
Minor tweaks to the quality encoding on -q3.
Improved reporting of results. It now reports how many bytes are taken up by sequences, qualities and names. (Sequence lengths are also stored, but not reported. They are typically tiny.)
There are no globals or static variables now. All data used by the fqz class is held internal to it. I still need to improve on the I/O mechanisms so it can compress and decompress blocks to memory rather than stdin/stdout, but this is most of the way there.
Split -1 to -N compression levels into explicit -s and -q for sequence and quality compression metrics. Added a new quality compression level for large datasets where we use the position along sequence too. Default of -q2 was old behaviour. -q1 does less compression than before, but is faster.
Added additional levels of sequence compression, up to -s9 if you have the memory.
Changed from 16-bit to 8-bit counters for sequence compression by default. This is significant memory reduction and for some sizes a big speed up too if it makes the difference of in/out of cache.
The -e option switches back to 16-bit mode.
Reworking of the models. Sequence and binary models have been sped up considerably, exploiting their smaller alphabet. The more general purpose 8-bit and 7-bit variants are now one model controlled by templates for the alphabet size.
The format consists of a header and a series of data blocks. The size of the data blocks is controlled by BLK_SIZE in the source code. Increasing the size may give files that cannot be uncompressed with existing fqzcomp builds.
The header has 8 bytes.
0: ‘.’ (4 bytes of magic number) 1: ‘f’ 2: ‘q’ 3: ‘z’
4: MAJOR_VERS (version numbers, encoded in the source) 5: MINOR_VERS
6: 4-bits -s level, 2 bits -q level, 2 bits -n level. (slevel<<0) | (qlevel << 4) | (nlevel << 6).
7: flags: bit 0: both_strands bit 1: extreme_seq (16-bit seq models) bit 2: multi_seq_model bit 3: SOLiD mode bits 4+ are reserved for future use, set to zero for now.
Additionally, if SOLiD mode is set, the first base type is appended to the header (typically T), as this does not appear in the actual quality strings. (Sometimes - NCBI seem to prefer it that way at least.)
Each block consists of a series of little endian 32-bit words followed by the compressed data streams:
Byte Value 0- 3 Total compressed size of this block 4- 7 32-bit checksum 8-11 Uncompressed size 12-15 Number of sequences 16-19 Size of compressed length data 20-23 Size of compressed name data 24-27 Size of compressed seq data 28-31 Size of compressed qual data 32+ Length data ? Name data ? Seq data ? Qual data
Each of the 4 blocks of data - length, name, seq and qualities - are all compressed using their own range coders to permit compression within independent threads if desired.
Investigate the impact of resetting the fqz class per block. This simulates efficiency of allowing random access. What block size is needed? (Need to move rc from global to a parameter of the models.)
Once we reset compression per block, we can perform multi-threaded compression.
Alternatively with 4 threads simply slice file into 4 and have a flag to indicate block-continue vs block-reset.