Table of Contents
Overview and Objectives
Once raw sequence files are generated (in FASTQ format) and quality-checked, the next step in most NGS pipelines is mapping to a reference genome. For individual sequences of interest, it is common to use a tool like BLAST to identify genes or species of origin. However, a typical example NGS dataset may have tens to hundreds of millions of reads, which BLAST and similar tools are not designed to handle.
Thus, a large set of computational tools have been developed to quickly, and with sufficient (but not absolute) accuracy align each read to its best location, if any, in a reference. Even though many mapping tools exist, a few individual programs have a dominant "market share" of the NGS world. These programs vary widely in their design, inputs, outputs, and applications. In this section, we will primarily focus on two of the most versatile mappers: BWA and Bowtie2, the latter being part of the Tuxedo suite (e.g. transcriptome-aware Tophat2).
You have already worked with a paired-end yeast ChIP-seq dataset, which we will continue to use here. The paired end data should already be located at:
We will also use two additional RNA-seq datasets, which are located at:
|Sample_Yeast_L005_R1.cat.fastq.gz||Paired-end Illumina, First of pair, FASTQ||Yeast ChIP-seq|
|Sample_Yeast_L005_R2.cat.fastq.gz||Paired-end Illumina, Second of pair, FASTQ||Yeast ChIP-seq|
|human_rnaseq.fastq.gz||Paired-end Illumina, First of pair only, FASTQ||Human RNA-seq|
|human_mirnaseq.fastq.gz||Single-end Illumina, FASTQ||Human microRNA-seq|
First copy the two human datasets to your $SCRATCH/core_ngs/fastq_prep directory.
Create a $SCRATCH/core_ngs/align directory and make a link to the fastq_prep directory.
Before we get to alignment, we need a genome to align to. We will use three different references here:
- the human genome (hg19)
- the yeast genome (sacCer3)
- and mirbase (v20), human subset
Mirbase is a collection of all known microRNAs in all species. 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 containing all the sequences we are actually interested in.
- 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.
- 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.
These are the three reference genomes we will be using today, with some information about them (and here is information about many more genomes):
|Reference||Species||Base Length||Contig Number||Source||Download|
|hg19||Human||3.1 Gbp||25 (really 93)||UCSC||UCSC GoldenPath|
|sacCer3||Yeast||12.2 Mbp||17||UCSC||UCSC GoldenPath|
|mirbase V20||Human||160 Kbp||1908||Mirbase||Mirbase Downloads|
Searching genomes is hard work and takes a long time if done on an un-indexed, linear genomic sequence. So aligners require that references first be indexed for quick access The aligners we are using each require a different index, but use the same method (the Burrows-Wheeler Transform) to get the job done. This requires taking a FASTA file as input, with each chromosome (or contig) as a separate entry, and producing some aligner-specific set of files as output. Those index files are then used by the aligner when performing the sequence alignment.
hg19 is way too big for us to index here, so we're not going to do it. Instead, all hg19 index files are located at:
We will grab the FASTA files for the other two references and build each index right before we use. These two references are located at:
First stage the yeast and mirbase reference FASTA files in your work archive area in a directory called references.
With that, we're ready to get started on the first exercise.
Exercise #1: BWA – Yeast ChIP-seq
Overview ChIP-seq alignment workflow with BWA
We will perform a global alignment of the paired-end Yeast ChIP-seq sequences using bwa. This workflow generally has the following steps:
- Trim the FASTQ sequences down to 50 with fastx_clipper
- this removes most of any 5' adapter contamination without the fuss of specific adapter trimming w/cutadapt
- Prepare the sacCer3 reference index for bwa (one time) using bwa index
- Perform a global bwa alignment on the R1 reads (bwa aln) producing a BWA-specific binary .sai intermediate file
- Perform a global bwa alignment on the R2 reads (bwa aln) producing a BWA-specific binary .sai intermediate file
- Perform pairing of the separately aligned reads and report the alignments in SAM format using bwa sampe
- Convert the SAM file to a BAM file (samtools view)
- Sort the BAM file by genomic location (samtools sort)
- Index the BAM file (samtools index)
- Gather simple alignment statistics (samtools flagstat and samtools idxstat)
We're going to skip the trimming step for now and see how it goes. We'll perform steps 2 - 5 now and leave samtools for the next course section, since those steps (6 - 10) are common to nearly all post-alignment workflows.
Building the BWA sacCer3 index
Like other tools you've worked with so far, you first need to load bwa using the module system. Go ahead and do that now, and then enter bwa with no arguments to view the top-level help page (many NGS tools will provide some help when called with no arguments).
Program: bwa (alignment via Burrows-Wheeler transformation)
Contact: Heng Li <firstname.lastname@example.org>
Usage: bwa <command> [options]
Command: index index sequences in the FASTA format
mem BWA-MEM algorithm
fastmap identify super-maximal exact matches
pemerge merge overlapping paired ends (EXPERIMENTAL)
aln gapped/ungapped alignment
samse generate alignment (single ended)
sampe generate alignment (paired ended)
bwasw BWA-SW for long queries
fa2pac convert FASTA to PAC format
pac2bwt generate BWT from PAC
pac2bwtgen alternative algorithm for generating BWT
bwtupdate update .bwt to the new format
bwt2sa generate SA from BWT and Occ
Note: To use BWA, you need to first index the genome with `bwa index'.
There are three alignment algorithms in BWA: `mem', `bwasw', and
`aln/samse/sampe'. If you are not sure which to use, try `bwa mem'
first. Please `man ./bwa.1' for the manual.
As you can see, bwa offers a number of sub-commands one can use with to do different things. We're going to index the genome with the index command. To learn what this sub-command needs in the way of options and arguments, enter bwa index with no arguments.
Here, we only need to specify two things:
- the name of the FASTA file
- whether to use the bwtsw or is algorithm for indexing
Since sacCer3 is relative large (~12 Mbp) we will specify bwtsw as the indexing option, and the name of the FASTA file is sacCer3.fa.
Importantly, the output of this command is a group of files that are all required together as the index. So, within the references directory, we will create another directory called bwa/sacCer3, make a symbolic link to the yeast FASTA there, and run the index command in that directory.
Now execute the bwa index command.
Since the yeast genome is not large when compared to human, this should not take long to execute (otherwise we would do it as a batch job). When it is comple you should see a set of index files like this:
Exploring the FASTA with grep
A common question is what contigs are in a given FASTA file. You'll usually want to know this before you start the alignment so that you're familiar with the contig naming convention – and to verify that it's the one you expect.
We saw that a FASTA consists of a number of contig entries, each one starting with a name line of the form below, followed by many lines of bases.
How do we dig out just the lines that have the contig names and ignore all the sequences? Well, the contig name lines all follow the pattern above, and since the > character is not a valid base, it will never appear on a sequence line.
We've discovered a pattern (also known as a regular expression) to use in searching, and the command line tool that does regular expression matching is grep.
Regular expressions are so powerful that nearly every modern computer language includes a "regex" module of some sort. There are many online tutorials for regular expressions (and a few different flavors of them). But the most common is the Perl style (http://perldoc.perl.org/perlretut.html). We're only going to use the most simple of regular expressions here, but learning more about them will pay handsome dividends for you in the future.
Here's how to execute grep to list contig names in a FASTA file.
- The -P option tells grep to use Perl-style regular expression patterns.
- This makes including special characters like Tab ( \t ), Carriage Return ( \r ) or Linefeed ( \n ) much easier that the default Posix paterns.
- While it is not really required here, it generally doesn't hurt to include this option.
'^>' is the regular expression describing the pattern we're looking for (described below)
- sacCer3.fa is the file to search. Lines with text that match our pattern will be written to standard output; non matching lines will be omitted.
- We pipe to more just in case there are a lot of contig names.
Now down to the nuts and bolts of our pattern, '^>'
First, the single quotes around the pattern – they are only a signal for the bash shell. As part of its friendly command line parsing and evaluation, the shell will often look for special characters on the command line that mean something to it (for example, the $ in front of an environment variable name, like in $SCRATCH). Well, regular expressions treat the $ specially too – but in a completely different way! Those single quotes tell the shell "don't look inside here for special characters – treat this as a literal string and pass it to the program". The shell will obey, will strip the single quotes off the string, and will pass the actual pattern, ^>, to the grep program. (Aside: We've see that the shell does look inside double quotes ( " ) for certain special signals, such as looking for environment variable names to evaluate.)
So what does ^> mean to grep? Well, from our contig name format description above we see that contig name lines always start with a > character, so > is a literal for grep to use in its pattern match.
We might be able to get away with just using this literal alone as our regex, specifying '>' as the command line argument. But for grep, the more specific the pattern, the better. So we constrain where the > can appear on the line. The special carat ( ^ ) character represents "beginning of line". So ^> means "beginning of a line followed by a > character, followed by anything. (Aside: the dollar sign ( $ ) character represents "end of line" in a regex. There are many other special characters, including period ( . ), question mark ( ? ), pipe ( | ), parentheses ( ( ) ), and brackets ( [ ] ), to name the most common.)
Exercise: How many contigs are there in the sacCer3 reference?
grep -P '^>' sacCer3.fa | wc -l
or use grep's -c option that says "just count the line matches"
grep -P -c '^>' sacCer3.fa
There are 17 contigs.
Performing the alignment
Now, we're ready to execute the actual alignment, with the goal of initially producing a SAM file from the input FASTQ files and reference. First go to the align directory, and link to the sacCer3 reference directory (this will make our commands more readable).
As our workflow indicated, we first use bwa aln on the R1 and R2 FASTQs, producing a BWA-specific .sai intermediate binary files. Since these alignments are completely independent, we can execute them in parallel in a batch job.
What does bwa aln needs in the way of arguments?
There are lots of options, but here is a summary of the most important ones. BWA, is a lot more complex than the options let on. If you look at the BWA manual on the web for the aln sub-command, you'll see numerous options that can increase the alignment rate (as well as decrease it), and all sorts of other things.
|-l||Controls the length of the seed (default = 32)|
|-k||Controls the number of mismatches allowable in the seed of each alignment (default = 2)|
|-n||Controls the number of mismatches (or fraction of bases in a given alignment that can be mismatches) in the entire alignment (including the seed) (default = 0.04)|
|-t||Controls the number of threads|
The rest of the options control the details of how much a mismatch or gap is penalized, limits on the number of acceptable hits per read, and so on. Much more information can be accessed at the BWA manual page.
For a simple alignment like this, we can just go with the default alignment parameters, with one exception. At TACC, we want to optimize our alignment speed by allocating more than one thread (-t) to the alignment. We want to run 2 tasks, and will use a minimum of one 16-core node. So we can assign 8 cores to each alignment by specifying -t 8.
Also note that bwa writes its (binary) output to standard output by default, so we need to redirect that to a .sai file. And of course we need to redirect standard error to a log file, one per file.
Create an aln.cmds file (using nano) with the following lines:
Create the batch submission script specifying a wayness of 8 (8 tasks per node) on the normal queue and a time of 1 hour, then submit the job and monitor the queue:
Since you have directed standard error to log files, you can use a neat trick to monitor the progress of the alignment: tail -f. The -f means "follow" the tail, so new lines at the end of the file are displayed as they are added to the file.
When it's done you should see two .sai files. Next we use the bwa sampe command to pair the reads and output SAM format data. For this command you provide the same reference prefix as for bwa aln, along with the two .sai files and the two original FASTQ files.
Again bwa writes its output to standard output, so redirect that to a .sam file. (Note that bwa sampe is "single threaded" – it does not have an option to use more than one processor for its work.) We'll just execute this at the command line – not in a batch job.
You did it! You should now have a SAM file that contains the alignments. It's just a text file, so take a look with head, more, less, tail, or whatever you feel like. In the next section, with samtools, you'll learn some additional ways to analyze the data once you create a BAM file.
Exercise: What kind of information is in the first lines of the SAM file?
Exercise: How many alignment records (not header records) are in the SAM file?
This looks for the pattern '^HWI' which is the start of every read name (which starts every alignment record).
Remember -c says just count the records, don't display them.
grep -P -c '^@' yeast_pairedend.sam
Or use the -v (invert) option to tell grep to print all lines that don't match a particular pattern, here the header lines starting with @.
grep -P -v -c '^HWI' yeast_pairedend.sam
Exercise: How many sequences were in the R1 and R2 FASTQ files combined?
gunzip -c fq/Sample_Yeast_L005_R1.cat.fastq.gz | echo $((`wc -l` / 2))
- Do both R1 and R2 reads have separate alignment records?
- Does the SAM file contain both aligned and un-aligned reads?
- What is the order of the alignment records in this SAM file?
- Do both R1 and R2 reads have separate alignment records?
- yes, they must, because there were 1,184,360 R1+R2 reads and an equal number of alignment records
- Does the SAM file contain both aligned and un-aligned reads?
- yes, it must, because there were 1,184,360 R1+R2 reads and an equal number of alignment records
- What is the order of the alignment records in this SAM file?
- the names occur in the exact same order as they did in the FASTQ, except that they come in pairs
- the R1 read comes first, then its corresponding R2
- this ordering is called read name ordering
Exercise #2: 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 a test microRNA-seq dataset for you to experiment with. 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 will impede alignment! We need an aligner that can find subsections of the read that align, and discard (or "soft-clip") the rest away. Bowtie2 is just such an aligner. Go ahead and load it up so we can examine some help pages and options. To do that, you must first load the "perl" module, and then the bowtie2 module designated "bowtie/2.2.0".
Now that it's loaded, check out the options. There are a LOT of them! This is the key to using Bowtie2 - it allows you to control almost everything about its behavior, but that also makes it easier to screw things up. Before getting to using the tool, though, we've got to build a mirbase index. This requires using the command "bowtie2-build" - go ahead and check out its options. Unlike for the aligner itself, we only need to worry about a few things here:
Here, the reference_in file is just our FASTA file containing mirbase sequences, and the bt2_index_base is the prefix of where we want the files to go. Following what we did earlier for BWA indexing:
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). Now, your $SCRATCH/references/mirbase directory should be filled with the following files:
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. These are the most important options when using Bowtie2:
|-N||Controls the number of mismatches allowable in the seed of each alignment (default = 0)|
|-L||Controls the length of seed substrings generated from each read (default = 22)|
|--end-to-end or --local||Controls whether the entire read must align to the reference, or whether soft-clipping the ends is allowed to find internal alignments|
|-ma||Controls the alignment score contribution of a matching base (0 for --end-to-end, 2 for --local|
To decide how we want to go about doing our alignment, check out the file we're aligning with 'less'.
Lots of those 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 is the small boundary of the acceptable internal alignments. Here, we are interested in finding any sections of any reads that align well to a microRNA. These sequences are between 16 and 22 bases long, so any good alignment should have at least 16 matching bases, but could have more. Also, maybe we want to allow a mismatch or two in the seed, since we might be interested in miRNA SNPs. So, a good set of options might look something like this:
This leaves the default scoring method as "-ma 2", meaning that a 16 base pair alignment will have a score of 32, and so on. This is VERY different from the alignment scores assigned by other aligners, so it's worth remembering.
As you can tell from looking at the Bowtie2 help message, the syntax looks like this:
As such, our alignment command (now that we have the FASTQ file and the reference sequence ready) could be this (make sure you are located in your scratch directory!):
Now, you should have a human_mirnaseq.sam file in your alignments directory, that you can check out using whatever commands you like. An example alignment looks like this:
Note how the CIGAR string is 3S20M13S, meaning that 13 bases were soft clipped from one end, and 3 from the other. If we did the same alignment using either --end-to-end mode, or using BWA in the same way as we did in Exercise #1, very little of this file would have aligned. However, 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 the one shown above, because the read only matched for 20 bases - a matching 22 base seed does not exist. Such is the nature of Bowtie2 - 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 in itself take a lot of time to perfect.
Exercise #3: BWA-MEM (and Tophat2) - Human mRNA-seq
After Bowtie2 came out with a local alignment option, it wasn't long before BWA generated their own local-aligner called BWA-MEM (for Maximal Exact Matches). This aligner is very, very nice because it incorporates a lot of the simplicity of using BWA with the complexities of local alignment. This functionality, while enabling the alignment of datasets like the mirbase data we just examined, also permits more complex alignments, such as that of spliced mRNAs. 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 should have been loaded when we loaded the BWA module, so to look at the details of MEM alignment, just enter "bwa mem" to get the help menu with the options list. The most important parameters, similar to those we've manipulated in the past two sections, are the following:
|-k||Controls the minimum seed length (default = 19)|
|-w||Controls 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 one long alignment with a long gap in the middle (default = 100)|
|-r||Controls 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)|
|-c||Controls how many matches a MEM must have in the genome before it is discarded (default = 10000)|
|-t||Controls 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 parameters 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.
Now, try aligning it with BWA like we did in example 1. This won't take very long, but you'll need to use our pre-indexed hg19 reference. It is located at:
if you look at the contents of the 'bwtsw' directory in the above path, you'll see a set of files that are analogous to the yeast files we created earlier, except that their universal prefix is 'hg19.fa'. Now, using that index, go ahead and try to do a single-end alignment of the file to the human genome like we did in Exercise #1, and save the contents to intermediate files with the prefix 'human_rnaseq_bwa'.
Now, 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:
meaning they are not alignments at all. Essentially, nothing (with a few exceptions) aligned. That's because this file was generated exclusively from reads in a larger dataset that cross at least one splice junction (surprise!), meaning that they sequence as it exists in most of the reads does not exist anywhere in the genome, but some subsections of each read do exist somewhere in the genome. So, we need an aligner that is capable of finding regions of each read (above some length cutoff) that align to the genome. Based on the following syntax, and using the reference path, we noted earlier, try to use BWA MEM to align the same file, single-end, saving all output files with the prefix 'human_rnaseq_mem':
Now, check the length of the SAM file you generated with 'wc -l'. Since there is one alignment per line, there must be 586266 alignments, which is more than the number of sequences in the FASTQ file. This is because many reads will align twice or more, hopefully on either side of a splice junction. These alignments can still be associated because they will have the same read ID, but are reflected in more than one line. To get an idea of how often each read aligned, and what the 'real' alignment rate is, use the following commands:
If you want to be able to generate a file with only the best (for BWA MEM, that means the longest) alignment for each read, so that every read can only be represented by one line, the option -M will add a flag to the end of each alignment entry that designates any alignments that are not the longest for their originating read as secondary. Then, you can filter all reads that do not have that flag.
This alignment rate is pretty good, but it could get better by playing around with the finer details of BWA MEM. It is also a bit higher if you use an aligner that is more explicitly concerned with sensitivity to splice sites, namely a program like Tophat2. All Tophat programs use Bowtie or Bowtie2 as the actual aligner, but split up each read into smaller pieces to align individually, then tries to reconstruct reasonable overall alignments from the pieces. In fact, these reads can all be aligned by using Tophat2 with the right parameters.