count-fgs-sam 1.0.dev4

Count Functional Genomics Screen alignments in a SAM/BAM file with filtering options

count_fgs_sam

Count Functional Genomics Screen alignments in a SAM/BAM file with filtering

This program will count the alignments in a SAM/BAM file. The number of primary alignments that are mapped to each reference sequence are counted. The alignments are filtered for various criteria before counting, and (optionally) additional buckets of counts (for each reference) are produced.

This script uses the pysam module. It is designed for alignments produced by Bowtie 2, and relies on the AS and XS score tags.

See below for Bowtie 2 parameters to produce compatible alignments.

Command-line parameters

usage: count_fgs_sam [-h] [--version] [-v] [-p PERFECTSCORE]
                     [-l EXPECTEDLENGTH] [-a UNAMBIGUOUSDELTA]
                     [-c ACCEPTABLESCORE] [-m ACCEPTABLEMINLENGTH]
                     [-M ACCEPTABLEMAXLENGTH] [-o OUTPUTFILE]
                     [--show-unacceptable] [--show-ambiguous] [--show-length]
                     [-d]
                     inputfile

positional arguments:
  inputfile             Input SAM/BAM file, type determined by extension

optional arguments:
  -h, --help            show this help message and exit
  --version             show program's version number and exit
  -v, --verbose         Increase logging verbosity (levels 0 to 3) (default:
                        0)
  -p PERFECTSCORE, --perfectscore PERFECTSCORE
                        Cutoff for perfect alignments, must have alignment
                        score (AS) of at least this (AS>=cutoff) (default:
                        200)
  -l EXPECTEDLENGTH, --expectedlength EXPECTEDLENGTH
                        Expected length of reads, reads filtered by this
                        length (length==expectedlength) (default: 20)
  -a UNAMBIGUOUSDELTA, --unambiguousdelta UNAMBIGUOUSDELTA
                        Cutoff for unambiguous alignments, reads filtered for
                        unambiguous alignments that have alignment score (AS)
                        minus alternative score (XS) less than equal this (AS-
                        XS<=cutoff), XS assumed to be 0 if not present for an
                        alignment (default: 3)
  -c ACCEPTABLESCORE, --acceptablescore ACCEPTABLESCORE
                        Cutoff for acceptable alignments AS>=cutoff; if
                        specified and different from --perfectscore, counts of
                        acceptable reads are shown (default: None)
  -m ACCEPTABLEMINLENGTH, --acceptableminlength ACCEPTABLEMINLENGTH
                        Minimum length for acceptable alignments (length
                        >=minlength); if specified and different from
                        --expectedlength, counts of shorter reads are reported
                        (default: None)
  -M ACCEPTABLEMAXLENGTH, --acceptablemaxlength ACCEPTABLEMAXLENGTH
                        Maximum length for acceptable alignments (length
                        <=maxlength); if specified and different from
                        --expectedlength, counts of longer reads are reported
                        (default: None)

optional output options:
  Options to included additional count buckets in output

  -o OUTPUTFILE, --outputfile OUTPUTFILE
                        Output TSV file, output to stdout if absent (default:
                        None)
  --show-unacceptable   Show counts of ambiguous alignments (default: False)
  --show-ambiguous      Show counts of unacceptable alignments (default:
                        False)
  --show-length         Show overall counts of reads by length classes, not
                        filtered by any other criteria; counts outside
                        specified lengths counted as 'length-out-of-range'
                        (default: False)
  -d, --detailed        Turn on all optional count buckets (--show* options
                        enabled, as well as any buckets enabled with
                        additional threshold parameters). Note that these
                        optional count buckets may all be zero if
                        corresponding parameters are not used. (default:
                        False)

Detail on scoring parameters

The following are parameters for scoring the alignments:

• Alignment score: The alignment score is read from the AS tag in the alignment entry in the SAM/BAM file.

• Read length: The read length is inferred from the CIGAR alignment, or if the read is unmapped, the read length is obtained from the sequence in the SAM/BAM file.

• Alternative alignment score: The alternative alignment score is read from the XS tag in the alignment entry in the SAM/BAM file.

• Alignment score delta: This is calculated as the alignment score (AS) minus the alternative alignment score (XS) (i.e. delta = AS-XS). It represents how much better the primary alignment is from the next best alignment, i.e. a larger delta indicates the primary alignment is clearly better than any alternatives, while a small delta may indicate that the 'best' alignment is ambiguous.

This script expects an input file consisting of unpaired reads aligned to the forward strand of the references and only primary alignments. Reads/alignments with various SAM flags are processed as follows:

• Only primary alignments are processed by this script. All secondary or supplementary alignments (SAM flags 0x100 or 0x800) are ignored, with a warning produced by the script.

• Any reads with the QC failed (0x200) or PCR/optical duplicate (0x400) flags are ignored. No warning is produced but counts of these are logged.

• A warning is produced if paired reads (0x1) or alignments to the reverse strand (0x10) are detected, although any such alignments are processed normally.

Filtering criteria

Based on the scoring parameters, the following filtering criteria are applied prior to counting the alignments:

• Perfect alignment score (-p or --perfectscore): Alignment score must be at least (>=) this value for the alignment to be considered perfect. In addition, an alignment must be of expected-length (below) to be considered perfect. Therefore, any alignments with score >= this value that are not of expected-length can be at best considered acceptable (below). Any read/alignment that is not perfect or acceptable is considered unacceptable. This includes any unmapped reads. By default, only perfect alignments are counted and reported.

• Expected length (l or --expectedlength): The read length must be equal (==) to this length for an alignment resulting from the read to be considered as being of expected-length. An alignment must be of expected-length to be considered perfect, although an alignment that is not of expected-length can still be considered to be acceptable if it meets the shorter or longer criteria (below). Any alignment that is outside expected or acceptable lengths is considered unacceptable. By default, only alignments that are perfect and therefore of expected-length are counted and reported.

• Unambiguous delta (-a or --unambiguousdelta): The alignment score delta must be greater (>) than this value for the alignment to be considered unambiguous. Any alignment with a delta less than or equal (<=) to this value is considered ambiguous. By default, only alignments that are unambiguous are counted and reported. Care should be used when choosing this value based on the scoring scheme used. If set to too high, it is possible that all alignments to certain references may be considered ambiguous and not counted.

  If any reads with perfect alignments scores are found to be ambiguous, a warning will be displayed since these will always be excluded from the counts. It is likely that this is a configuration error. This may occur if there are replicated sequences among the references (which should be consolidated into a single reference sequence, and the reads realigned to this new set of references), or the unambiguous delta is set too high.

• Acceptable alignment score (-c or --acceptablescore): Any non-perfect alignments may be considered acceptable if their scores are at least (>=) this value. By default, no acceptable are allowed. Providing an acceptable alignment score will enable counting and reporting of acceptable alignments, these still have to be of expected-length unless acceptable minimum or maximum lengths are also provided.

• Acceptable minimum length (-m or --acceptableminlength): An alignment from a read that is below (<) the expected length but greater than or equal (>=) the acceptable minimum length will be considered shorter. Providing an acceptable minimum length will enable counting and reporting of shorter alignments. Alignments that are shorter will be considered acceptable if they meet the acceptable alignment score threshold. If no acceptable alignment score is provided, any shorter alignment that meets the perfect score threshold will be considered acceptable, but in typical usage a shorter alignment could not have a perfect alignment score and this option should be combined with an acceptable alignment score.

• Acceptable maximum length (-M or --acceptablemaxlength): An alignment from a read that is above (>) the expected length but less than or equal (<=) the acceptable maximum length will be considered longer. Providing an acceptable maximum length will enable counting and reporting of longer alignments. Alignments that are longer will be considered acceptable if they meet the acceptable alignment score threshold. If no acceptable alignment score is provided, any longer alignment that meets the perfect score threshold will be considered acceptable.

These filtering criteria determine which count buckets each read/alignment goes into.

Count buckets

Various buckets of counts are produced for each reference sequence. Each reference sequence is a row in the output file, with the reference sequence name (ref-name) and length (ref-length) reported. The buckets in the output are the columns (see further below for example output). In addition to the reference sequences, there are rows in the output file to count unmapped reads if they are included in the criteria for that count bucket (row has ref-name of *unmapped), the sum of all counted reads (including any counted unmapped reads; **total_counted), any excluded alignment/reads that don't meet the bucket criteria (**excluded), and the grand total of all reads (***grand_total). Note that for the *unmapped row, unmapped reads are automatically considered unacceptable and therefore excluded for most buckets (i.e. only counted under **excluded rather than being counted under *unmapped), apart for unacceptable buckets. The buckets that can be produced are given below:

• perfect/unambiguous/expected-length (default)

• acceptable/unambiguous/expected-length (enabled if an acceptable alignment score is provided)

• acceptable/unambiguous/shorter (enabled if an acceptable minimum length is provided)

• acceptable/unambiguous/longer (enabled if an acceptable maximum length is provided)

• perfect/ambiguous/expected-length (enabled with --show-ambiguous)

• acceptable/ambiguous/expected-length (enabled with --show-ambiguous and if an acceptable alignment score is provided)

• acceptable/ambiguous/shorter (enabled with --show-ambiguous and if an acceptable minimum length is provided)

• acceptable/ambiguous/longer (enabled with --show-ambiguous and if an acceptable maximum length is provided)

• unacceptable/expected-length (enabled with --show-unacceptable)

• unacceptable/shorter (enabled with --show-unacceptable and if an acceptable minimum length is provided)

• unacceptable/longer (enabled with --show-unacceptable and if an acceptable maximum length is provided)

• unacceptable/length-out-of-range (enabled with --show-unacceptable)

All of the above buckets are mutually exclusive. A read cannot be in more than one of the above buckets (excluding the total rows). This means, for example, that downstream analyses should sum perfect/unambiguous/expected-length and acceptable/unambiguous/expected-length counts if both perfect and acceptable counts are to be included. Buckets for ambiguous and unacceptable counts are primarily meant to be used for diagnostic purposes.

The following buckets for overall (total) counts can also be produced:

• expected-length: counts all reads that are of expected-length, regardless of score or delta criteria (enabled with --show-length)

• shorter: counts all reads that are shorter, regardless of score or delta criteria (enabled with --show-length and if an acceptable minimum length is provided)

• longer: counts all reads that are longer, regardless of score or delta criteria (enabled with --show-length and if an acceptable maximum length is provided)

• length-out-of-range: counts all reads that are not in any of the above length categories, including any shorter/longer reads where acceptable minimum or maximum lengths are not provided (enabled with --show-length)

• any: counts all reads regardless of score, delta or length criteria (default)

By default, only perfect/unambiguous/expected-length and any counts are produced.

Example output (formatted with spaces for presentation, actual output is tab-separated):

ref-name                ref-length  perfect/unambiguous/expected-length  any
Non-Targeting___76442   20          21                                   25
Non-Targeting___76443   20          22                                   22
...
*unmapped               nan         0                                    4803
**total_counted         nan         17078                                23015
**excluded              nan         5937                                 0
***grand_total          nan         23015                                23015

Suggested Bowtie 2 parameters

The suggested parameters for Bowtie 2 are as follows:

bowtie2 --ma=10 --mp=-4,-6 --np=-6 --rdg=0,1 --rfg=0,1 --score-min=C,150,0 --n-ceil=C,2,0 --local --norc --gbar 1 -D 20 -R 1 -N 0 -L 10 -i L,1,0

This uses local alignment with bonus scoring for matches (match bonus --ma=10), (lower) bonus scoring for mismatches/Ns (mismatch penalty --mp=-4,-6, N penalty --np=-6; negative values turn this into a bonus) and small gap penalties (read and reference gap penalties --rdg=0,1 --rfg=0,1; gaps at ends of the read are not penalised during local alignment, but all gaps are also penalised by a lack of match/mismatch bonus, i.e -10 compared to a match).

The minimum alignment score can be set leniently (--score-min=C,150,0) as this script will do additional filtering of alignments prior to counting. The remaining options set a maximum of 2 Ns (--n-ceil=C,2,0) and the optimal parameters for producing to most mapped reads (determined mostly by trial and error --local --norc --gbar 1 -D 20 -R 1 -N 0 -L 10 -i L,1,0).

For reads with length of 20 bp, these options will produce a perfect alignment score of 200, and alignment score of at least 194 with one mismatch, and a score of 189 to 190 with one gap (depending of whether it is internal or at ends). For reads of 19 bp, a maximum score of 189 to 190 is possible for alignments with no mismatches depending on the position of the gap. For reads of 21 bp, a maximum score of 200 is possible even with the additional unmatched nucleotide at the read ends. Ambiguous alignments can be detected based on the alternative alignment score, an additional mismatch will reduce the score by at least 4 and an additional gap will reduce the score by at least 10. Changes in the position of the gap or an character of a mismatch will alter score by only 0, -1 or -2.

With these Bowtie 2 parameters, a perfect alignment score of 200, expected length of 20 and an unambiguous delta of 3 will count perfect, unambiguous alignments of the expected length (default parameters of --perfectscore 200 --expectedlength 20 --unambiguousdelta 3). An acceptable alignment score of 194 can be used to count alignments with one mismatch (additional parameter --acceptablescore 194). An acceptable alignment score of 189, and acceptable minimum and maximum lengths of 19 and 21 can also be included to detect unambiguous alignments with one mismatch or one gap, including those from reads 1 nucleotide shorter or longer than expected (additional parameters --acceptablescore 189 --acceptableminlength 19 --acceptablemaxlength 21).

The input SAM/BAM file does not need to be sorted or indexed prior to use in this script, by sorting generally reduces the size of a BAM file.

Performance

This script is implemented in Python but has been through basic optimisations to increase performance. On a 3.8 Ghz Ryzen 9, this script produces detailed counts for about 6M reads in about 80 seconds. No multi-threading is implemented, but if multiple input files are to be processed, processing of each file could be performed in parallel.

---

Author	Tet Woo Lee
Created	2020-03-12
Copyright	© 2020 Tet Woo Lee
License	GPLv3
Dependencies	pysam, tested with v0.15.4

Change log

• version 1.0.dev4 2020-03-28
  Minor improvements & additional checks/tests

  • added different levels of verbosity
  • warnings now produced using Python warnings module
  • added warning for length of reference(s) not equal expected length
  • unit tests for warnings: reference length issues, ignored alignments, ambiguous perfect alignments
  • unit test input file now has alternative alignment scores
  • added processing additional flags: warning if paired reads detected, ignore/report QC fail or duplicates, warning if reverse alignments detected
  • specifying --acceptablescore to be the same as --perfectscore, and --acceptableminlength or --acceptablemaxlength to be the same as --expectedlength produces same results as not using these parameters

• version 1.0.dev3 2020-03-26
  First production version

  • added unit tests
  • PyPi and conda packaging
  • add additional output options

• version 1.0.dev2 2020-03-17
  Modifications to improve performance, total 5.5x speedup:

  • 468 s to 350 s (6.4M reads) by switching from Flag to IntFlag
  • 350 s to 134 s by modifying to use int rather than IntFlag in add_counts
  • 134 s to 85 s by modifying to use int with flag addition or operations
  • Further improvements possible, e.g. 70 s possible by Cythonizing as-is but current performance should suffice.

• version 1.0.dev1 2020-03-15
  First working version