Genome Assembly (velvet)

Overview

Velvet is a De Bruijn graph assembler works fairly rapidly on short (microbial) genomes. In this tutorial we will use velvet to assemble an E. coli genome from simulated Illumina reads.

Learning Objectives

  • Run velvet to perform de novo assembly on fragment, paired-end, and mate-paired data.
  • Use contig_stats.pl to display assembly statistics.
  • Find proteins of interest in an assembly using Blast.

Table of Contents

Data

First, let's copy the fastq read files.

Move to scratch, copy the raw data, and change into this directory for the tutorial
cds
mkdir velvet_tutorial
cp $BI/ngs_course/velvet/data/*/* velvet_tutorial
cd velvet_tutorial

Now we have a bunch of Illumina reads. These are simulated reads. If you'd ever like to simulate some on your own, you might try using Mason.

Files in the tutorial directory
login1$ ls
paired_end_2x100_ins_1500_c_20.fastq  paired_end_2x100_ins_400_c_20.fastq  single_end_100_c_50.fastq
paired_end_2x100_ins_3000_c_20.fastq  paired_end_2x100_ins_400_c_25.fastq
paired_end_2x100_ins_3000_c_25.fastq  paired_end_2x100_ins_400_c_50.fastq

There are 4 sets of simulated reads:

 

Set 1

Set 2

Set 3

Set 4

Read Size

100

100

100

100

Paired/Single Reads

Single

Paired

Paired

Paired

Gap Sizes

NA

400

400, 3000

400, 3000, 1500

Coverage

50

50

25 for each subset

20 for each subset

Number of Subsets

1

1

2

3

Note that these fastq files are "interleaved", with each read pair together one-after-the-other in the file. The #/1 and #/2 in the read names indicate the pairs.

Interleaved fastq
login1$ head paired_end_2x100_ins_1500_c_20.fastq
@READ-1/1
TTTCACCGTTGACCAGCACCCAGTTCAGCGCCGCGCGACCACGATATTTTGGTAACAGCGAACCATGCAGATTAAATGCACCTGCGGGAGCGAGCTGCAA
+
*@A+<at:var at:name="55G" />T@@I&+@A+@@<at:var at:name="II" />G@+++A++GG++@++I@+@+G&/+I+GD+II@++G@@I?<at:var at:name="I" />@<at:var at:name="IIGGI" /><at:var at:name="A4" />6@A,+AT=<at:var at:name="G" />+@AA+GAG++@
@READ-1/2
TTAACACCGGGCTATAAGTACAATCTGACCGATATTAACGCCGCGATTGCCCTGACACAGTTAGTCAAATTAGAGCACCTCAACACCCGTCGGCGCGAAA
+
I@@H+A+@G+&+@AG+I>G+I@+CAIA++$+T<at:var at:name="GG" />@+++1+<at:var at:name="GI" />+ICI+A+@<at:var at:name="I" />++A+@@A.@<G@@+)GCGC%I@IIAA++++G+A;@+++@@@@6

Often your read pairs will be "separate" with the corresponding paired reads at the same index in two different files (each with exactly the same number of reads).

Velvet Assembly

Now let's use Velvet to assemble the reads.

First, you need to load Velvet via module.

 I need to see the command...
Load the velvet module
 module load velvet

Using velvet consists of a sequence of two commands:

  1. velveth - analyzes kmers in the reads in preparation for assembly
  2. velvetg - constructs the assembly and filters contigs from the graph

Look at the help for each program.

The <hash_length> parameter of velveth is the kmer value that is key to the assembly process. Choosing it controls the tradeoff between sensitivity (lower hash_length, more reads included, longer contigs) and specificity (higher hash length, less chance of misassembly, more reads ignored, shorter contigs)). There is more discussion about choosing an appropriate kmer value in the Velvet manual and in this blog post.

Velvet has an option of specifying the insertion size of a paired read file (-ins_length). If no size is given, Velvet will guess the insertion size. We'll just have Velvet guess the size.

Velvet also has an option to specify the expected coverage of the genome, which helps it choose how to resolve repeated sequences (-exp_cov). We set this parameter to estimate this from the data. A common problem with using Velvet is that you have many very short contigs and the last line of output tells you that it used 0 of your reads. This is caused by not setting this parameter. The default is NOT auto.

We'll need to create a commands file and submit it to TACC. Let's make the commands file say:

For a "commands" file - to run four velvet assemblies in parallel. If you copy and paste, be sure that there are ONLY four lines in your file.
velveth single_out 61 -fastq single_end_100_c_50.fastq && velvetg single_out -exp_cov auto -amos_file yes
velveth pairedc20_out 61 -fastq -shortPaired paired_end_2x100_ins_3000_c_20.fastq paired_end_2x100_ins_1500_c_20.fastq paired_end_2x100_ins_400_c_20.fastq && velvetg pairedc20_out -exp_cov auto -amos_file yes
velveth pairedc25_out 61 -fastq -shortPaired paired_end_2x100_ins_3000_c_25.fastq paired_end_2x100_ins_400_c_25.fastq && velvetg pairedc25_out -exp_cov auto -amos_file yes
velveth pairedc50_out 61 -fastq -shortPaired paired_end_2x100_ins_400_c_50.fastq && velvetg pairedc50_out -exp_cov auto -amos_file yes

Now use launcher_creator.py to make a *.sge for the commands file and qsub it.

Use the correct "wayness"

Velvet and other assemblers usually take large amounts of RAM to complete. Running these 4 commands on a single node will use more RAM than is available on a single node so it's necessary to change the number of commands per node (wayness) from the default of 12 to 1. When you use launcher_creator.py, you set the "wayness" (number of commands per node) using the -w flag.

You can set the allotted time for this job to just 10 minutes.

 I need some help with my launcher_creator.py command
launcher_creator.py -w 1 -n velvet -q normal -j commands -t 00:10:00

If you are assembling large genomes or have high coverage depth data in the future, you will probably need to submit your jobs to the "largemem" queue.

If you find yourself waiting a long time for the assembly process to run, you can also start an idev session and try running some of the velveth and velvetg commands interactively. Each one takes a few minutes to complete.

Velvet Output

Explore each output directory that was created by Velvet.

Checking the tail of the Log files in each of the output folders, we see lines like the following:

 The results...
Single Set
Median coverage depth = 2.657895
Final graph has 9748 nodes and n50 of 191, max 1427, total 1865207, using 281499/2314900 reads
Set with one group of reads at 50 coverage
Median coverage depth = 11.131337
Final graph has 265 nodes and n50 of 127102, max 397974, total 4558511, using 1464201/2314900 reads
Set with 2 groups of reads both at 25 coverage each
Median coverage depth = 11.109244
Final graph has 203 nodes and n50 of 698134, max 1032531, total 4585717, using 1465818/2314900 reads
Set with 3 groups of reads all at 20 coverage each
Median coverage depth = 13.353287
Final graph has 202 nodes and n50 of 698626, max 1139610, total 4602729, using 1758595/2777880 reads

With better read pairs that link more distant locations in the genome, there are fewer contigs, and contigs are are longer, giving us a more complete picture of linkage across the genome.

The complete E. coli genome is about 4.6 Mb. Why weren't we able to assemble it, even with this "perfect" data?

  Possibilities...
  1. Sometimes errors in reads lead to dead-ends in the graphs that are trimmed when they should not be.
  2. There are 7 nearly identical ribosomal RNA operons in E. coli spaced throughout the chromosome. Since each is >3000 bases, contigs cannot be connected across them using this data.

More assembly statistics: contig_stats.pl

The output file stats.txt contains information about every contig in the assembly, but it isn't sorted and can be a bit overwhelming.

You can generate some summary stats and graphs about each assembly using the contig_stats.pl script that we have installed under $BI/bin. Try figuring out how to do this on your own. You probably want to change into the directory of results for a specific assembly before running this command, since it generates several output files in the current working directory. You will need to copy PNG output back to your computer to view it.

 Example command
contig_stats.pl -plot -One_pdf -tool Velvet contigs.fa

What do I do now?

Many choices:

  1. Get a better assembly: maybe add a different library size, or go into a detailed genome completion project (commonly called "finishing") using a sequence assembly editor like consed or gap4 or AMOS. (Be careful though, the amount of data in NGS data sets can be very difficult for these programs to deal with, since many were designed for Sanger sequencing reads.) You may have a lot of PCR products to make to close gaps and/or to order and orient scaffolds. consed in particular makes this pretty easy, but it may still consume a lot more time and money than the initial shotgun assembly. You can identify some misassemblies by mapping the original reads to the assembly and then viewing them in IGV to look for discordant mate pairs, for example.
  2. Look for things: If you're just after a few homologs, an operon, etc. you're probably done. Most assemblers will be able to take 2x100 data and give you full gene sequences since these are non-repetitive and so assemble well. You can turn the contigs.fa into a blast database (formatdb or makeblastdb depending on which version of blast you have) and start blasting away.

Exercise: Finding proteins in the assembly

Change into one of the Velvet result directories for the set of contigs that you want to analyze and load the blast module.

Find your favorite bacterial protein and see if it exists within the assembly.

 Hint...

Try NCBI's "Protein" database - search for Escherichia.

 If you don't want to find them, use these proteins
Copy over a few E. coli protein sequences for the next blast
cp /corral-repl/utexas/BioITeam/ngs_course/velvet/test.fa .

Once you have some query sequences you'd like to find in your assembly, turn your assembly into a local blast database and search for them:

Make a blast database of the contigs and try to find a gene (might have to load the blast module first)
makeblastdb -in contigs.fa -dbtype nucl
tblastn -query test.fa -db contigs.fa -evalue 1e-50 | more
 You could also do the reverse search...
Reverse search - all the contigs blasted against a small protein database
makeblastdb -in test.fa -dbtype prot
blastx -query contigs.fa -db test.fa -outfmt "6 qseqid sallseqid bitscore frames ppos sallacc" -evalue 1e-10 | more

Other paths from here:

  1. Predict genes/annotate the genome with de novo gene prediction tool like glimmer, maker, or one of the online gene prediction tools available at NCBI or JCVI.
  2. Use a variety of assembly evaluation tools (a rapidly growing field) - we'll talk about that more on the next page.

Advanced Exercise: A5 Haploid Microbial Genome Assembler

Another approach to try for haploid microbial genomes is the A5 pipeline.

It includes several additional quality control steps (like filtering adaptor contamination and low quality score portions of reads) that make sense when using a De Bruijn graph assembler and is optimized for haploid genomes. They're described in this paper.

Install the a5 pipeline by downloading from here and either 1) moving all of the items in the bin directory into your local bin directory or adding the location of the a5/bin directory to your $PATH.

 Sample command line for using the A5 assembler...
For an interleaved paired end input file
a5_pipeline.pl paired_end_2x100_ins_3000_c_25.fastq output

For more information, read the directions.

Further Reading