Page tree

Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

Expand
titleAnswer
Recall that these are 100 bp reads and we did not remove adapter contamination. There will be a distribution of fragment sizes – some will be short – and those short fragments may not align without adapter removal (fastx_trimmer or cutadapt).

Exercise #2: Bowtie2 - Vibrio Cholerae RNA-seq

While we have focused on aligning eukaryotic data, the same tools can be used to perform identical functions with prokaryotic data.  The major differences are less about the underlying data and much more about the external/public databases established to store and distribute reference data.  For example, the Illumina iGenome resource provides pre-processed and uniform reference data, designed to be out-of-the-box compatible with aligners like bowtie2 and bwa.  However, the limited number of available species are heavily biased towards model eukaryotes. If we wanted to study a prokaryote, the reference data must be downloaded from a resource like GenBank, and processed/indexed similarly to the procedure for mirbase.  

...

After the header lines, each feature in the genome is represented by a line that gives chromosome, start, stop, strand, and other information.  Features are things like "mRNA," "CDS," and "EXON."  As you would expect in a prokaryotic genome it is frequently the case that the gene, mRNA, CDS, and exon annotations are identical, meaning they share coordinate information. You could parse these files further using commands like grep  and awk  to extract, say, all exons from the full file or to remove the header lines that begin with #.

Building the bowtie2 vibCho index

To build the reference index for alignment, we actually only need the FASTA file, since annotations are often not necessary for alignment. (This is not always true - extensively spliced transcriptomes requires splice junction annotations to align RNA-seq data properly, but for now we will only use the FASTA file.)  We build the reference files exactly as we did with mirbase, only swapping out the FASTA files as follows:

Code Block
mkdir -p $WORK/core_ngs/references/bt2/vibCho
mv $WORK/core_ngs/references/vibCho.O395.fa $WORK/core_ngs/references/fasta
cd $WORK/core_ngs/references/bt2/vibCho
ln -s -f ../../fasta/vibCho.O395.fa
ls -la

Now build the index using bowtie2-build like we did for the mirbase index:

Code Block
bowtie2-build vibCho.O395.fa vibCho.O395

This should also go pretty fast: you can see the resulting files using ls like before.

Performing the bowtie2 alignment

Now we will go back to our scratch area to do the alignment, and set up symbolic links to the index in the work area to simplify the alignment command:

...

 

Introducing bowtie2

 

Go ahead and load the bowtie2 module so we can examine some help pages and options. To do that, you must first load the perl module, and then the a specific version of bowtie2

 

Code Block
module load perl
module load bowtie/2.2.0

 

Now that it's loaded, check out the options. There are a lot of them! In fact for the full range of options and their meaning, Google "Bowtie2 manual" and bring up that page. The Table of Contents is several pages long! Ouch!

 

This is the key to using bowtie2 - it allows you to control almost everything about its behavior, but that also makes it is much more challenging to use than bwa – and it's easier to screw things up too!

Building the bowtie2 vibCho index

To build the reference index for alignment, we actually only need the FASTA file, since annotations are often not necessary for alignment. (This is not always true - extensively spliced transcriptomes requires splice junction annotations to align RNA-seq data properly, but for now we will only use the FASTA file.)  We build the reference files exactly as we did with mirbase, only swapping out the FASTA files as follows:

Code Block
mkdir -p $WORK/core_ngs/references/bt2/vibCho
mv $WORK/core_ngs/references/vibCho.O395.fa $WORK/core_ngs/references/fasta
cd $WORK/core_ngs/references/bt2/vibCho
ln -s -f ../../fasta/vibCho

...

.O395.fa
ls -la

Now build the index using bowtie2-build like we did for the mirbase index:

Code Block
bowtie2-build vibCho.O395.fa vibCho.O395

This should also go pretty fast: you can see the resulting files using ls like before.

Performing the bowtie2 alignment

Now we will go back to our scratch area to do the alignment, and set up symbolic links to the index in the work area to simplify the alignment command:

Code Block
cd $SCRATCH/core_ngs/alignment
ln -s -f $WORK/core_ngs/references/bt2/vibCho vibCho

In the previous exercise, we set specific local alignment parameters for bowtie2 that took into account the fact that the dataset was derived from small RNA.  However, in this case the data is from standard mRNA sequencing, meaning that the DNA fragments are typically longer than the reads.  There is likely to be very little contamination that would require using a local rather than global alignment, or many other pre-processing steps (e.g. adapter trimming).  Thus, we will run bowtie2 with default parameters, omitting any options other than the input, output, and reference index.  The command would look like this:

...

When the job is complete you should have a cholera_rnaseq.sam file that you can examine using whatever commands you like.  In the past two exercises, we have explored a few different sorts of commands that you might run on this file to get different sorts of information.  Those commands will provide similar information here, since this file is also in SAM/BAM format.  In the last exercise, we will come back to this SAM file (and the others) to explore ways to compress them effectively and to extract basic quality statistics.

Exercise #4:

...

Bowtie2 and Local Alignment - Human

...

microRNA-seq

After bowtie2 came out with a local alignment option, it wasn't long before bwa developed its own local alignment algorithm called BWA-MEM (for Maximal Exact Matches), implemented by the bwa mem command. bwa mem has the following advantages:

  • It incorporates a lot of the simplicity of using bwa with the complexities of local alignment, enabling straightforward alignment of datasets like the mirbase data we just examined
  • It can align different portions of a read to different locations on the genome
    • In a long RNA-seq experiment, reads will (at some frequency) span a splice junction themselves, or a pair of reads in a paired-end library will fall on either side of a splice junction. We want to be able to align reads that do this for many reasons, from accurate transcript quantification to novel fusion transcript discovery.

Thus, our last exercise will be the alignment of a human long RNA-seq dataset composed (by design) almost exclusively of reads that cross splice junctions.

bwa mem was made available when we loaded the bwa module, so take a look at its usage information. The most important parameters, similar to those we've manipulated in the past two sections, are the following:

OptionEffect
-kControls the minimum seed length (default = 19)
-wControls the "gap bandwidth", or the length of a maximum gap. This is particularly relevant for MEM, since it can determine whether a read is split into two separate alignments or is reported as one long alignment with a long gap in the middle (default = 100)
-rControls how long an alignment must be relative to its seed before it is re-seeded to try to find a best-fit local match (default = 1.5, e.g. the value of -k multiplied by 1.5)
-cControls how many matches a MEM must have in the genome before it is discarded (default = 10000)
-tControls the number of threads to use

There are many more parameters to control the scoring scheme and other details, but these are the most essential ones to use to get anything of value at all.

The test file we will be working with is just the R1 file from a paired-end total RNA-seq experiment, meaning it is (for our purposes) single-end. Go ahead and take a look at it, and find out how many reads are in the file.

Expand
titleHint:
Code Block
languagebash
cd $SCRATCH/core_ngs/alignment
ls fastq
gunzip -c fastq/human_rnaseq.fastq.gz | echo $((`wc -l`/4))

RNA-seq alignment with bwa aln

Now, try aligning it with bwa aln like we did in Example #1, but first link to the hg19 bwa index directory.  In this case, due to the size of the hg19 index, we are linking to Anna's scratch area INSTEAD of our own work area containing indexes that we built ourselves.

Code Block
languagebash
titleLink to BWA hg19 index directory
cd $SCRATCH/core_ngs/alignment
ln -s -f /scratch/01063/abattenh/ref_genome/bwa/bwtsw/hg19
ls hg19

You should see a set of files analogous to the yeast files we created earlier, except that their universal prefix is hg19.fa.  

Go ahead and try to do a single-end alignment of the file to the human genome using bwa aln like we did in Exercise #1, saving intermediate files with the prefix human_rnaseq_bwa. Go ahead and just execute on the command lineNow we're going to switch over to a different aligner that was originally designed for very short reads and is frequently used for RNA-seq data. Accordingly, we have prepared another test microRNA-seq dataset for you to experiment with (not the same one you used cutadapt on). This data is derived from a human H1 embryonic stem cell (H1-hESC) small RNA dataset generated by the ENCODE Consortium – its about a half million reads.

However, there is a problem!  We don't know (or, well, you don't know) what the adapter structure or sequences were. So, you have a bunch of 36 base pair reads, but many of those reads will include extra sequence that can impede alignment – and we don't know where! We need an aligner that can find subsections of the read that do align, and discard (or "soft-clip") the rest away – an aligner with a local alignment mode. Bowtie2 is just such an aligner.

Overview miRNA alignment workflow with bowtie2

If the adapter structure were known, the normal workflow would be to first remove the adapter sequences with cutadapt. Since we can't do that, we will instead perform a local alignment of the single-end miRNA sequences using bowtie2. This workflow has the following steps:

  1. Prepare the mirbase v20 reference index for bowtie2 (one time) using bowtie2-build 
  2. Perform local alignment of the R1 reads with bowtie2, producing a SAM file directly
  3. Convert the SAM file to a BAM file (samtools view)
  4. Sort the BAM file by genomic location (samtools sort)
  5. Index the BAM file (samtools index)
  6. Gather simple alignment statistics (samtools flagstat and samtools idxstat)

This looks so much simpler than bwa – only one alignment step instead of three! We'll see the price for this "simplicity" in a moment...

As before, we will just do the alignment steps leave samtools for the next section.

Mirbase is a collection of all known microRNAs in all species (and many speculative miRNAs). We will use the human subset of that database as our alignment reference.  This has the advantage of being significantly smaller than the human genome, while likely containing almost all sequences likely to be detected in a miRNA sequencing run.

Expand
titleIf it's simpler and faster, would one ever want to align a miRNA dataset to hg19 rather than mirbase? If so, why?
  1. Due to natural variation, sequencing errors, and processing issues, variation between reference sequence and sample sequence is always possible. Alignment to the human genome allows a putative "microRNA" read the opportunity to find a better alignment in a region of the genome that is not an annotated microRNA. If this occurs, we might think that a read represents a microRNA (since it aligned in the mirbase alignment), when it is actually more likely to have come from a non-miRNA area of the genome. This is a major complication involved when determining, for example, whether a potential miRNA is produced from a repetitive region.
  2. If we suspect our library contained other RNA species, we may want to quantify the level of "contamination". Aligning to the human genome will allow rRNA, tRNA, snoRNA, etc to align. We can then use programs such as bedtools, coupled with appropriate genome annotation files, to quantify these "off-target" hits. This is particularly plausible if, after a miRNA sequencing run, the alignment rate to mirbase is very low.

These are the four reference genomes we will be using today, with some information about them (and here is information about many more genomes):

 

Building the bowtie2 mirbase index

Before the alignment, of course, we've got to build a mirbase index using bowtie2-build (go ahead and check out its options). Unlike for the aligner itself, we only need to worry about a few things here:

Code Block
bowtie2-build <reference_in> <bt2_index_base>
  • reference_in file is just the FASTA file containing mirbase v20 sequences
  • bt2_index_base is the prefix of where we want the files to go

 Following what we did earlier for BWA indexing, namely move our FASTA into place, create the index directory, and establish our symbolic links.

We are looking at mapping quality values for both aligned and un-aligned records, but mapping quality only makes sense for aligned reads. This expression does not distinguish between mapping quality = 0 because the read mapped to multiple locations, and mapping quality = 0 because the sequence did not align.

The proper solution will await the use of samtools to filter out unmapped reads.
Code Block
Expand
titleAnswer
languagebash
titleCommands to bwa aln RNA-seq data
bwa aln hg19/hg19.fa fastq/human_rnaseq.fastq.gz > human_rnaseq_bwa.sai
bwa samse hg19/hg19.fa human_rnaseq_bwa.sai fastq/human_rnaseq.fastq.gz > human_rnaseq_bwa.sam

Once this is complete use less to take a look at the contents of the SAM file, using the space bar to leaf through them. You'll notice a lot of alignments look basically like this:

Code Block
HWI-ST1097:228:C21WMACXX:8:1316:10989:88190     4       *       0       0       *       *       0       0
  AAATTGCTTCCTGTCCTCATCCTTCCTGTCAGCCATCTTCCTTCGTTTGATCTCAGGGAAGTTCAGGTCTTCCAGCCGCTCTTTGCCACTGATCTCCAGCT
  CCCFFFFFHHHHHIJJJJIJJJJIJJJJHJJJJJJJJJJJJJJIIIJJJIGHHIJIJIJIJHBHIJJIIHIEGHIIHGFFDDEEEDDCDDD@CDEDDDCDD

Notice that the contig name (field 3) is just an asterisk ( * ) and the alignment flags value is a 4 (field 2), meaning the read did not align (decimal 4 = hex 0x4 = read did not map).

Essentially, nothing (with a few exceptions) aligned. Why?

Expand
titleAnswer

Because this file was generated exclusively from reads in a larger dataset that cross at least one splice junction. The sequences as they exists in most of the reads do not correspond to a single location in the genome. However subsections of each read do exist somewhere in the genome.

So, we need an aligner that is capable aligning different parts of the read to different genomic loci.

RNA-seq alignment with bwa mem

Exercise: use bwa mem to align the same data

Based on the following syntax and the above reference path, use bwa mem to align the same file, saving output files with the prefix human_rnaseq_mem. Go ahead and just execute on the command line.

Code Block
bwa mem <ref.fa> <reads.fq> > outfile.sam
Expand
titleHint:
Code Block
languagebash
bwa mem hg19/hg19.fa fastq/human_rnaseq.fastq.gz > human_rnaseq_mem.sam

Check the length of the SAM file you generated with wc -l. Since there is one alignment per line, there must be 586266 alignments (minus no more than 100 header lines), which is more than the number of sequences in the FASTQ file. This is bwa mem can report multiple alignment records for the same read, hopefully on either side of a splice junction. These alignments can still be tied together because they have the same read ID.

Expand
titleMore advanced piping...

To get an idea of how often each read aligned, and what the 'real' alignment rate is, use the following commands:

Code Block
languagebash
# This gives you a view where each read is listed next to the number of entries it has in the SAM file
cut -f 1 human_rnaseq_mem.sam | sort | uniq -c | less

#This gives essentially a histogram of the number of times each read aligned - a plurality of reads aligned twice, which seems reasonable since these are all reads crossing a junction, but plenty aligned more or less
cut -f 1 human_rnaseq_mem.sam | sort | uniq -c | awk '{print $1}' | sort | uniq -c | less	

#This gives a better idea of the alignment rate, which is how many reads aligned at least once to the genome.  Divided by the number of reads in the original file, the real alignment rate is much higher.
cut -f 1 human_rnaseq_mem.sam | sort | uniq | wc -l	

# NOTE: Some of these one-liners are only reasonably fast if the files are relatively small (around a million reads or less). For bigger files, there are better ways to get this information, mostly using samtools.

This alignment rate is pretty good, but it could get better by playing around with the finer details of bwa mem.

Tip

Be aware that some downstream tools (for example the Picard suite, often used before SNP calling) do not like it when a read name appears more than once in the SAM file. To mark the extra alignment records as secondary, specify the bwa mem -M option. This option leaves the best (longest) alignment for a read as -is but marks additional alignments for the read as secondary (the 0x100 BAM flag). This designation also allows you to easily filter the secondary reads with samtools if desired.

BWA-MEM vs Tophat

Another approach to aligning long RNA-seq data is to use an aligner that is more explicitly concerned with sensitivity to splice sites, namely a program like Tophat. Tophat uses either bowtie (tophat) or bowtie2 (tophat2) as the actual aligner, but performs the following steps:

  • aligns reads to the genome
  • reads that do not align to the genome are aligned against a transcriptome, if provided
    • if they align, the transcriptome coordinates are converted back to genomic coordinates, with gaps represented in the CIGAR string, for example as 196N
  • reads that do not align to the transcriptome are split into smaller pieces, each of which Tophat attempts to map to the genome

Note that Tophat also reports secondary alignments, but they have a different meaning. Tophat always reports spliced alignments as one alignment records with the N CIGAR string operator indicating the gaps. Secondary alignments for Tophat (marked with the 0x100 BAM flag) represent alternate places in the genome where a  read (spliced or not) may have mapped.

As you can imagine from this series of steps, Tophat is very computationally intensive and takes much longer than bwa mem – very large alignments (hundreds of millions of reads) may not complete in stampede's 48 hour maximum job time!

Exercise #4: Bowtie2 and Local Alignment - Human microRNA-seq

Now we're going to switch over to a different aligner that was originally designed for very short reads and is frequently used for RNA-seq data. Accordingly, we have prepared another test microRNA-seq dataset for you to experiment with (not the same one you used cutadapt on). This data is derived from a human H1 embryonic stem cell (H1-hESC) small RNA dataset generated by the ENCODE Consortium – its about a half million reads.

However, there is a problem!  We don't know (or, well, you don't know) what the adapter structure or sequences were. So, you have a bunch of 36 base pair reads, but many of those reads will include extra sequence that can impede alignment – and we don't know where! We need an aligner that can find subsections of the read that do align, and discard (or "soft-clip") the rest away – an aligner with a local alignment mode. Bowtie2 is just such an aligner.

Overview miRNA alignment workflow with bowtie2

If the adapter structure were known, the normal workflow would be to first remove the adapter sequences with cutadapt. Since we can't do that, we will instead perform a local alignment of the single-end miRNA sequences using bowtie2. This workflow has the following steps:

  1. Prepare the mirbase v20 reference index for bowtie2 (one time) using bowtie2-build 
  2. Perform local alignment of the R1 reads with bowtie2, producing a SAM file directly
  3. Convert the SAM file to a BAM file (samtools view)
  4. Sort the BAM file by genomic location (samtools sort)
  5. Index the BAM file (samtools index)
  6. Gather simple alignment statistics (samtools flagstat and samtools idxstat)

This looks so much simpler than bwa – only one alignment step instead of three! We'll see the price for this "simplicity" in a moment...

As before, we will just do the alignment steps leave samtools for the next section.

Mirbase is a collection of all known microRNAs in all species (and many speculative miRNAs). We will use the human subset of that database as our alignment reference.  This has the advantage of being significantly smaller than the human genome, while likely containing almost all sequences likely to be detected in a miRNA sequencing run.

Expand
titleIf it's simpler and faster, would one ever want to align a miRNA dataset to hg19 rather than mirbase? If so, why?
  1. Due to natural variation, sequencing errors, and processing issues, variation between reference sequence and sample sequence is always possible. Alignment to the human genome allows a putative "microRNA" read the opportunity to find a better alignment in a region of the genome that is not an annotated microRNA. If this occurs, we might think that a read represents a microRNA (since it aligned in the mirbase alignment), when it is actually more likely to have come from a non-miRNA area of the genome. This is a major complication involved when determining, for example, whether a potential miRNA is produced from a repetitive region.
  2. If we suspect our library contained other RNA species, we may want to quantify the level of "contamination". Aligning to the human genome will allow rRNA, tRNA, snoRNA, etc to align. We can then use programs such as bedtools, coupled with appropriate genome annotation files, to quantify these "off-target" hits. This is particularly plausible if, after a miRNA sequencing run, the alignment rate to mirbase is very low.

These are the four reference genomes we will be using today, with some information about them (and here is information about many more genomes):

Introducing bowtie2

Go ahead and load the bowtie2 module so we can examine some help pages and options. To do that, you must first load the perl module, and then the a specific version of bowtie2

Code Block
module load perl
module load bowtie/2.2.0

Now that it's loaded, check out the options. There are a lot of them! In fact for the full range of options and their meaning, Google "Bowtie2 manual" and bring up that page. The Table of Contents is several pages long! Ouch!

This is the key to using bowtie2 - it allows you to control almost everything about its behavior, but that also makes it is much more challenging to use than bwa – and it's easier to screw things up too!

Building the bowtie2 mirbase index

Before the alignment, of course, we've got to build a mirbase index using bowtie2-build (go ahead and check out its options). Unlike for the aligner itself, we only need to worry about a few things here:

Code Block
bowtie2-build <reference_in> <bt2_index_base>
  • reference_in file is just the FASTA file containing mirbase v20 sequences
  • bt2_index_base is the prefix of where we want the files to go

 Following what we did earlier for BWA indexing, namely move our FASTA into place, create the index directory, and establish our symbolic links.

Code Block
languagebash
titlePrepare Bowtie2 index directory for mirbase
mkdir -p $WORK/core_ngs/references/bt2/mirbase.v20
mv $WORK/core_ngs/references/hairpin_cDNA_hsa.fa $WORK/core_ngs/references/fasta
cd $WORK/core_ngs/references/bt2/mirbase.v20
ln -s -f ../../fasta/hairpin_cDNA_hsa.fa
ls -la 

Now build the index with bowtie2-build:

Code Block
languagebash
titlePrepare Bowtie2 index directory for mirbase
bowtie2-build hairpin_cDNA_hsa.fa hairpin_cDNA_hsa.fa

That was very fast!  It's because the mirbase reference genome is so small compared to what programs like this are used to dealing with, which is the human genome (or bigger).  You should see the following files:

Code Block
titlebowtie2 index files for miRNAs
hairpin_cDNA_hsa.fa
hairpin_cDNA_hsa.fa.1.bt2
hairpin_cDNA_hsa.fa.2.bt2
hairpin_cDNA_hsa.fa.3.bt2
hairpin_cDNA_hsa.fa.4.bt2
hairpin_cDNA_hsa.fa.rev.1.bt2
hairpin_cDNA_hsa.fa.rev.2.bt2

Performing the bowtie2 local alignment

Now, we're ready to actually try to do the alignment.  Remember, unlike BWA, we actually need to set some options depending on what we're after. Some of the important options for bowtie2 are:

OptionEffect
--end-to-end or --localControls whether the entire read must align to the reference, or whether soft-clipping the ends is allowed to find internal alignments. Default --end-to-end
-LControls the length of seed substrings generated from each read (default = 22)
-NControls the number of mismatches allowable in the seed of each alignment (default = 0)
-iInterval between extracted seeds. Default is a function of read length and alignment mode.
--score-minMinimum alignment score for reporting alignments. Default is a function of read length and alignment mode.

To decide how we want to go about doing our alignment, check out the file we're aligning with less:

Code Block
languagebash
titleExamine miRNA sequences to be aligned
cd $SCRATCH/core_ngs/alignment
less fastq/human_mirnaseq.fastq.gz

Lots of reads have long strings of A's, which must be an adapter or protocol artifact.  Even though we see how we might be able to fix it using some tools we've talked about, what if we had no idea what the adapter sequence was, or couldn't use cutadapt or other programs to prepare the reads?

 In that case, we need a local alignment where the seed length smaller than the expected insert size. Here, we are interested in finding any sections of any reads that align well to a microRNA, which are between 16 and 24 bases long, with most 20-22. So an acceptable alignment should have at least 16 matching bases, but could have more.

If we're also interested in detecting miRNA SNPs, we might want to allow a mismatch in the seed. So, a good set of options might look something like this:

Code Block
-N 1 -L 16 --local

 

Expand
titleIf our reads were longer

Because these are short reads we do not have to adjust parameters like inter-seed distance (-i) or minimum alignment score (--min-score) that are a function of read length. If we were processing longer reads, we might need to use parameters like this, to force bowtie2 to "pretend" the read is a short, constant length:

-i C,1,0
--score-min=C,32,0

Yes, that looks complicated, and it kind of is. It's basically saying "slide the seed down the read one base at a time", and "report alignments as long as they have a minimum alignment score of 32 (16 matching bases x 2 points per match, minimum).

See the bowtie2 manual (after you have had a good stiff drink) for a full explanation.

As you can tell from looking at the bowtie2 help message, the general syntax looks like this:

Code Block
bowtie2 [options]* -x <bt2-idx> {-1 <m1> -2 <m2> | -U <r>} [-S <sam>]

Let's make a link to the mirbase index directory to make our command line simpler:

Code Block
languagebash
titleLink to mirbase index for bowtie2
cd $SCRATCH/core_ngs/alignment
ln -s -f $WORK/core_ngs/references/bt2/mirbase.v20 mb20

Putting this all together we have a command line that looks like this.

Code Block
languagebash
titleLocal bowti2 alignment of miRNA data
bowtie2 --local -N 1 -L 16 -x mb20/hairpin_cDNA_hsa.fa -U fastq/human_mirnaseq.fastq.gz -S human_mirnaseq.sam
Expand
titleWhat's going on?

Parameters are:

  • --local – local alignment mode
  • -L 16 – seed length 16
  • -N 1 – allow 1 mismatch in the seed
  • -x  mb20/hairpin_cDNA_hsa.fa – prefix path of index files
  • -U fastq/human_mirnaseq.fastq.gz – FASTQ file for single-end (Unpaired) alignment
  • -S human_mirnaseq.sam – tells bowtie2 to report alignments in SAM format to the specified file

Create a commands file called bt2.cmds with this task definition then generate and submit a batch job for it (time 1 hour, normal queue).

Use nano to create the bt2.cmds file. Then:

Expand
titleWhat's going on?
Code Block
languagebash
titleLocal bowti2 alignment of miRNA data
launcher_creator.py -n bt2 -j bt2.cmds -t 01:00:00 -q normal
sbatch bt2.slurm; showq -u

When the job is complete you should have a human_mirnaseq.sam file that you can examine using whatever commands you like. An example alignment looks like this.

Code Block
titleExample miRNA alignment record
TUPAC_0037_FC62EE7AAXX:2:1:2607:1430#0/1  0  hsa-mir-302b  50  22 3S20M13S * 0 0
    TACGTGCTTCCATGTTTTANTAGAAAAAAAAAAAAG  ZZFQV]Z[\IacaWc]RZIBVGSHL_b[XQQcXQcc
    AS:i:37 XN:i:0  XM:i:1  XO:i:0  XG:i:0  NM:i:1  MD:Z:16G3       YT:Z:UU

Notes:

  • This is one alignment record, although it has been broken up below for readability.
  • This read mapped to the mature microRNA sequence hsa-mir-302b, starting at base 50 in that contig.
  • Notice the CIGAR string is 3S20M13S, meaning that 3 bases were soft clipped from one end (3S), and 13 from the other (13S).
    • If we did the same alignment using either bowtie2 --end-to-end mode, or using bwa aln as in Exercise #1, very little of this file would have aligned.
  • The 20M part of the CIGAR string says there was a block of 20 read bases that mapped to the reference.
    • If we had not lowered the seed parameter of bowtie2 from its default of 22, we would not have found many of the alignments like this one that only matched for 20 bases.

Such is the nature of bowtie2 – it it can be a powerful tool to sift out the alignments you want from a messy dataset with limited information, but doing so requires careful tuning of the parameters, which can take quite a few trials to figure out.

Exercise: About how many records in human_mirnaseq.sam represent aligned reads?

Expand
titleSolution

We can use our cut / grep trick from Exercise #1, but on the human_mirnaseq.sam file. Since all read names in this file start with TUPAC, we'll use that pattern to select non-header lines.

Code Block
languagebash
titleCount aligned SAM records
grep -P '^TUPAC' human_mirnaseq.sam | cut -f 3 | grep -v '*' | wc -l

This expressions returns 221086.

Use sort and uniq to create a histogram of mapping qualities

The mapping quality score is in field 5 of the human_mirnaseq.sam file. We can do this to pull out only that field:

Code Block
languagebash
titleCut mapping quality SAM field
grep -P '^TUPAC' human_mirnaseq.sam | cut -f 5 | head

We will use the uniq create a histogram of these values. The first part of the --help for uniq says:

Code Block
titleWhat uniq does
Usage: uniq [OPTION]... [INPUT [OUTPUT]]
Filter adjacent matching lines from INPUT (or standard input),
writing to OUTPUT (or standard output).

With no options, matching lines are merged to the first occurrence.
Mandatory arguments to long options are mandatory for short options too.
  -c, --count           prefix lines by the number of occurrences

To create a histogram, we want to organize all equal mapping quality score lines into an adjacent block, then use uniq -c option to count them. The sort -n command does the sorting into blocks (-n means numerical sort). So putting it all together, and piping the output to a pager just in case, we get:

Code Block
languagebash
titleCut mapping quality SAM field
grep -P '^TUPAC' human_mirnaseq.sam | cut -f 5 | sort -n | uniq -c | more

Exercise: What is the flaw in this "program"?

Prepare Bowtie2 index directory for mirbase
mkdir -p $WORK/core_ngs/references/bt2/mirbase.v20
mv $WORK/core_ngs/references/hairpin_cDNA_hsa.fa $WORK/core_ngs/references/fasta
cd $WORK/core_ngs/references/bt2/mirbase.v20
ln -s -f ../../fasta/hairpin_cDNA_hsa.fa
ls -la 

Now build the index with bowtie2-build:

Code Block
languagebash
titlePrepare Bowtie2 index directory for mirbase
bowtie2-build hairpin_cDNA_hsa.fa hairpin_cDNA_hsa.fa

That was very fast!  It's because the mirbase reference genome is so small compared to what programs like this are used to dealing with, which is the human genome (or bigger).  You should see the following files:

Code Block
titlebowtie2 index files for miRNAs
hairpin_cDNA_hsa.fa
hairpin_cDNA_hsa.fa.1.bt2
hairpin_cDNA_hsa.fa.2.bt2
hairpin_cDNA_hsa.fa.3.bt2
hairpin_cDNA_hsa.fa.4.bt2
hairpin_cDNA_hsa.fa.rev.1.bt2
hairpin_cDNA_hsa.fa.rev.2.bt2

Performing the bowtie2 local alignment

Now, we're ready to actually try to do the alignment.  Remember, unlike BWA, we actually need to set some options depending on what we're after. Some of the important options for bowtie2 are:

OptionEffect
--end-to-end or --localControls whether the entire read must align to the reference, or whether soft-clipping the ends is allowed to find internal alignments. Default --end-to-end
-LControls the length of seed substrings generated from each read (default = 22)
-NControls the number of mismatches allowable in the seed of each alignment (default = 0)
-iInterval between extracted seeds. Default is a function of read length and alignment mode.
--score-minMinimum alignment score for reporting alignments. Default is a function of read length and alignment mode.

To decide how we want to go about doing our alignment, check out the file we're aligning with less:

Code Block
languagebash
titleExamine miRNA sequences to be aligned
cd $SCRATCH/core_ngs/alignment
less fastq/human_mirnaseq.fastq.gz

Lots of reads have long strings of A's, which must be an adapter or protocol artifact.  Even though we see how we might be able to fix it using some tools we've talked about, what if we had no idea what the adapter sequence was, or couldn't use cutadapt or other programs to prepare the reads?

 In that case, we need a local alignment where the seed length smaller than the expected insert size. Here, we are interested in finding any sections of any reads that align well to a microRNA, which are between 16 and 24 bases long, with most 20-22. So an acceptable alignment should have at least 16 matching bases, but could have more.

If we're also interested in detecting miRNA SNPs, we might want to allow a mismatch in the seed. So, a good set of options might look something like this:

Code Block
-N 1 -L 16 --local

 

Expand
titleIf our reads were longer

Because these are short reads we do not have to adjust parameters like inter-seed distance (-i) or minimum alignment score (--min-score) that are a function of read length. If we were processing longer reads, we might need to use parameters like this, to force bowtie2 to "pretend" the read is a short, constant length:

-i C,1,0
--score-min=C,32,0

Yes, that looks complicated, and it kind of is. It's basically saying "slide the seed down the read one base at a time", and "report alignments as long as they have a minimum alignment score of 32 (16 matching bases x 2 points per match, minimum).

See the bowtie2 manual (after you have had a good stiff drink) for a full explanation.

As you can tell from looking at the bowtie2 help message, the general syntax looks like this:

Code Block
bowtie2 [options]* -x <bt2-idx> {-1 <m1> -2 <m2> | -U <r>} [-S <sam>]

Let's make a link to the mirbase index directory to make our command line simpler:

Code Block
languagebash
titleLink to mirbase index for bowtie2
cd $SCRATCH/core_ngs/alignment
ln -s -f $WORK/core_ngs/references/bt2/mirbase.v20 mb20

Putting this all together we have a command line that looks like this.

Code Block
languagebash
titleLocal bowti2 alignment of miRNA data
bowtie2 --local -N 1 -L 16 -x mb20/hairpin_cDNA_hsa.fa -U fastq/human_mirnaseq.fastq.gz -S human_mirnaseq.sam
Expand
titleWhat's going on?

Parameters are:

  • --local – local alignment mode
  • -L 16 – seed length 16
  • -N 1 – allow 1 mismatch in the seed
  • -x  mb20/hairpin_cDNA_hsa.fa – prefix path of index files
  • -U fastq/human_mirnaseq.fastq.gz – FASTQ file for single-end (Unpaired) alignment
  • -S human_mirnaseq.sam – tells bowtie2 to report alignments in SAM format to the specified file

Create a commands file called bt2.cmds with this task definition then generate and submit a batch job for it (time 1 hour, normal queue).

Use nano to create the bt2.cmds file. Then:

Expand
titleWhat's going on?
Code Block
languagebash
titleLocal bowti2 alignment of miRNA data
launcher_creator.py -n bt2 -j bt2.cmds -t 01:00:00 -q normal
sbatch bt2.slurm; showq -u

When the job is complete you should have a human_mirnaseq.sam file that you can examine using whatever commands you like. An example alignment looks like this.

Code Block
titleExample miRNA alignment record
TUPAC_0037_FC62EE7AAXX:2:1:2607:1430#0/1  0  hsa-mir-302b  50  22 3S20M13S * 0 0
    TACGTGCTTCCATGTTTTANTAGAAAAAAAAAAAAG  ZZFQV]Z[\IacaWc]RZIBVGSHL_b[XQQcXQcc
    AS:i:37 XN:i:0  XM:i:1  XO:i:0  XG:i:0  NM:i:1  MD:Z:16G3       YT:Z:UU

Notes:

  • This is one alignment record, although it has been broken up below for readability.
  • This read mapped to the mature microRNA sequence hsa-mir-302b, starting at base 50 in that contig.
  • Notice the CIGAR string is 3S20M13S, meaning that 3 bases were soft clipped from one end (3S), and 13 from the other (13S).
    • If we did the same alignment using either bowtie2 --end-to-end mode, or using bwa aln as in Exercise #1, very little of this file would have aligned.
  • The 20M part of the CIGAR string says there was a block of 20 read bases that mapped to the reference.
    • If we had not lowered the seed parameter of bowtie2 from its default of 22, we would not have found many of the alignments like this one that only matched for 20 bases.

Such is the nature of bowtie2 – it it can be a powerful tool to sift out the alignments you want from a messy dataset with limited information, but doing so requires careful tuning of the parameters, which can take quite a few trials to figure out.

Exercise: About how many records in human_mirnaseq.sam represent aligned reads?

Expand
titleSolution

We can use our cut / grep trick from Exercise #1, but on the human_mirnaseq.sam file. Since all read names in this file start with TUPAC, we'll use that pattern to select non-header lines.

Code Block
languagebash
titleCount aligned SAM records
grep -P '^TUPAC' human_mirnaseq.sam | cut -f 3 | grep -v '*' | wc -l

This expressions returns 221086.

Use sort and uniq to create a histogram of mapping qualities

The mapping quality score is in field 5 of the human_mirnaseq.sam file. We can do this to pull out only that field:

Code Block
languagebash
titleCut mapping quality SAM field
grep -P '^TUPAC' human_mirnaseq.sam | cut -f 5 | head

We will use the uniq create a histogram of these values. The first part of the --help for uniq says:

Code Block
titleWhat uniq does
Usage: uniq [OPTION]... [INPUT [OUTPUT]]
Filter adjacent matching lines from INPUT (or standard input),
writing to OUTPUT (or standard output).

With no options, matching lines are merged to the first occurrence.
Mandatory arguments to long options are mandatory for short options too.
  -c, --count           prefix lines by the number of occurrences

To create a histogram, we want to organize all equal mapping quality score lines into an adjacent block, then use uniq -c option to count them. The sort -n command does the sorting into blocks (-n means numerical sort). So putting it all together, and piping the output to a pager just in case, we get:

Code Block
languagebash
titleCut mapping quality SAM field
grep -P '^TUPAC' human_mirnaseq.sam | cut -f 5 | sort -n | uniq -c | more

Exercise: What is the flaw in this "program"?

Expand
titleAnswer

We are looking at mapping quality values for both aligned and un-aligned records, but mapping quality only makes sense for aligned reads. This expression does not distinguish between mapping quality = 0 because the read mapped to multiple locations, and mapping quality = 0 because the sequence did not align.

The proper solution will await the use of samtools to filter out unmapped reads.

Exercise #4: BWA-MEM - Human mRNA-seq

After bowtie2 came out with a local alignment option, it wasn't long before bwa developed its own local alignment algorithm called BWA-MEM (for Maximal Exact Matches), implemented by the bwa mem command. bwa mem has the following advantages:

  • It incorporates a lot of the simplicity of using bwa with the complexities of local alignment, enabling straightforward alignment of datasets like the mirbase data we just examined
  • It can align different portions of a read to different locations on the genome
    • In a long RNA-seq experiment, reads will (at some frequency) span a splice junction themselves, or a pair of reads in a paired-end library will fall on either side of a splice junction. We want to be able to align reads that do this for many reasons, from accurate transcript quantification to novel fusion transcript discovery.

Thus, our last exercise will be the alignment of a human long RNA-seq dataset composed (by design) almost exclusively of reads that cross splice junctions.

bwa mem was made available when we loaded the bwa module, so take a look at its usage information. The most important parameters, similar to those we've manipulated in the past two sections, are the following:

OptionEffect
-kControls the minimum seed length (default = 19)
-wControls the "gap bandwidth", or the length of a maximum gap. This is particularly relevant for MEM, since it can determine whether a read is split into two separate alignments or is reported as one long alignment with a long gap in the middle (default = 100)
-rControls how long an alignment must be relative to its seed before it is re-seeded to try to find a best-fit local match (default = 1.5, e.g. the value of -k multiplied by 1.5)
-cControls how many matches a MEM must have in the genome before it is discarded (default = 10000)
-tControls the number of threads to use

There are many more parameters to control the scoring scheme and other details, but these are the most essential ones to use to get anything of value at all.

The test file we will be working with is just the R1 file from a paired-end total RNA-seq experiment, meaning it is (for our purposes) single-end. Go ahead and take a look at it, and find out how many reads are in the file.

Expand
titleHint:
Code Block
languagebash
cd $SCRATCH/core_ngs/alignment
ls fastq
gunzip -c fastq/human_rnaseq.fastq.gz | echo $((`wc -l`/4))

RNA-seq alignment with bwa aln

Now, try aligning it with bwa aln like we did in Example #1, but first link to the hg19 bwa index directory.  In this case, due to the size of the hg19 index, we are linking to Anna's scratch area INSTEAD of our own work area containing indexes that we built ourselves.

Code Block
languagebash
titleLink to BWA hg19 index directory
cd $SCRATCH/core_ngs/alignment
ln -s -f /scratch/01063/abattenh/ref_genome/bwa/bwtsw/hg19
ls hg19

You should see a set of files analogous to the yeast files we created earlier, except that their universal prefix is hg19.fa.  

Go ahead and try to do a single-end alignment of the file to the human genome using bwa aln like we did in Exercise #1, saving intermediate files with the prefix human_rnaseq_bwa. Go ahead and just execute on the command line.

Code Block
languagebash
titleCommands to bwa aln RNA-seq data
bwa aln hg19/hg19.fa fastq/human_rnaseq.fastq.gz > human_rnaseq_bwa.sai
bwa samse hg19/hg19.fa human_rnaseq_bwa.sai fastq/human_rnaseq.fastq.gz > human_rnaseq_bwa.sam

Once this is complete use less to take a look at the contents of the SAM file, using the space bar to leaf through them. You'll notice a lot of alignments look basically like this:

Code Block
HWI-ST1097:228:C21WMACXX:8:1316:10989:88190     4       *       0       0       *       *       0       0
  AAATTGCTTCCTGTCCTCATCCTTCCTGTCAGCCATCTTCCTTCGTTTGATCTCAGGGAAGTTCAGGTCTTCCAGCCGCTCTTTGCCACTGATCTCCAGCT
  CCCFFFFFHHHHHIJJJJIJJJJIJJJJHJJJJJJJJJJJJJJIIIJJJIGHHIJIJIJIJHBHIJJIIHIEGHIIHGFFDDEEEDDCDDD@CDEDDDCDD

Notice that the contig name (field 3) is just an asterisk ( * ) and the alignment flags value is a 4 (field 2), meaning the read did not align (decimal 4 = hex 0x4 = read did not map).

Essentially, nothing (with a few exceptions) aligned. Why?

Expand
titleAnswer

Because this file was generated exclusively from reads in a larger dataset that cross at least one splice junction. The sequences as they exists in most of the reads do not correspond to a single location in the genome. However subsections of each read do exist somewhere in the genome.

So, we need an aligner that is capable aligning different parts of the read to different genomic loci.

RNA-seq alignment with bwa mem

Exercise: use bwa mem to align the same data

Based on the following syntax and the above reference path, use bwa mem to align the same file, saving output files with the prefix human_rnaseq_mem. Go ahead and just execute on the command line.

Code Block
bwa mem <ref.fa> <reads.fq> > outfile.sam
Expand
titleHint:
Code Block
languagebash
bwa mem hg19/hg19.fa fastq/human_rnaseq.fastq.gz > human_rnaseq_mem.sam

Check the length of the SAM file you generated with wc -l. Since there is one alignment per line, there must be 586266 alignments (minus no more than 100 header lines), which is more than the number of sequences in the FASTQ file. This is bwa mem can report multiple alignment records for the same read, hopefully on either side of a splice junction. These alignments can still be tied together because they have the same read ID.

Expand
titleMore advanced piping...

To get an idea of how often each read aligned, and what the 'real' alignment rate is, use the following commands:

Code Block
languagebash
# This gives you a view where each read is listed next to the number of entries it has in the SAM file
cut -f 1 human_rnaseq_mem.sam | sort | uniq -c | less

#This gives essentially a histogram of the number of times each read aligned - a plurality of reads aligned twice, which seems reasonable since these are all reads crossing a junction, but plenty aligned more or less
cut -f 1 human_rnaseq_mem.sam | sort | uniq -c | awk '{print $1}' | sort | uniq -c | less	

#This gives a better idea of the alignment rate, which is how many reads aligned at least once to the genome.  Divided by the number of reads in the original file, the real alignment rate is much higher.
cut -f 1 human_rnaseq_mem.sam | sort | uniq | wc -l	

# NOTE: Some of these one-liners are only reasonably fast if the files are relatively small (around a million reads or less). For bigger files, there are better ways to get this information, mostly using samtools.

This alignment rate is pretty good, but it could get better by playing around with the finer details of bwa mem.

Tip

Be aware that some downstream tools (for example the Picard suite, often used before SNP calling) do not like it when a read name appears more than once in the SAM file. To mark the extra alignment records as secondary, specify the bwa mem -M option. This option leaves the best (longest) alignment for a read as -is but marks additional alignments for the read as secondary (the 0x100 BAM flag). This designation also allows you to easily filter the secondary reads with samtools if desired.

BWA-MEM vs Tophat

Another approach to aligning long RNA-seq data is to use an aligner that is more explicitly concerned with sensitivity to splice sites, namely a program like Tophat. Tophat uses either bowtie (tophat) or bowtie2 (tophat2) as the actual aligner, but performs the following steps:

  • aligns reads to the genome
  • reads that do not align to the genome are aligned against a transcriptome, if provided
    • if they align, the transcriptome coordinates are converted back to genomic coordinates, with gaps represented in the CIGAR string, for example as 196N
  • reads that do not align to the transcriptome are split into smaller pieces, each of which Tophat attempts to map to the genome

Note that Tophat also reports secondary alignments, but they have a different meaning. Tophat always reports spliced alignments as one alignment records with the N CIGAR string operator indicating the gaps. Secondary alignments for Tophat (marked with the 0x100 BAM flag) represent alternate places in the genome where a  read (spliced or not) may have mapped.

As you can imagine from this series of steps, Tophat is very computationally intensive and takes much longer than bwa mem – very large alignments (hundreds of millions of reads) may not complete in stampede's 48 hour maximum job time!

Exercise #5: Simple SAMtools Utilities

...