This is a question that I have been asked many times. I think that the opposite question should also be asked — which genome assembler gives you the worst genome assembly? — but people seem less interested in asking this. By the end of this blog post, I will try to answer both questions.

Earlier this week, I wrote about the fact that I get to look at lots of genome assemblies that people have asked me to run CEGMA on. The reason I run CEGMA is to calculate what percentage of a set of 248 Core Eukaryotic Genes (CEGs) are present in each genome (or transcriptome) assembly. This figure can be used as a proxy for the much larger set of 'all genes'.

I can also calculate the N50 length of each assembly, and if you plot both statistics for each assembly you capture a tiny slice of that nebulous characteristic known as 'assembly quality'. Here's what such a plot looks like for 32 different genome assemblies (from various species, assembled by various genome assemblers):

Figure 1. CEGMA results vs N50 length for 32 genome assemblies. Click to enlarge.

Three key observations can be made from this figure:

There's a lot of variation in the percentage of CEGs present (52.4–98.8%).
There's a lot of variation in N50 length (997 bp all the way to 3.9 million bp).
There is almost no correlation between the two measures (r = 0.04)

Let's dwell on that last point by highlighting a couple of the extreme data points:

Figure 2: CEGMA results vs N50 length — red data points highlight assemblies with highest values of %248 CEGs or N50 length. Click to enlarge.

The highlighted point at the top of the graph represents the assembly with the best CEGMA result (245 out of 248 CEGs present). However, this assembly ranks 13th for N50 length. The data point on the far right of the graph represents a genome assembly with the highest N50 length (almost 4 million bp). But this assembly only ranks 27th for its CEGMA results. Such inconsistency was exactly what we saw in the Assemblathon 2 paper (but with even more metrics involved).

Can we shed any light as to which particular genome assemblers excel in either of these statistics? Well, as I now ask people who submit CEGMA jobs to me what was the principle assembly tool used?, I can overlay this information on the plot:

Figure 3: CEGMA results vs N50 length — colors refer to different assemblers used as the principle software tool in the assembly process. Click to enlarge.

It might not be clear but there are more data points for the Velvet assembler than any other (12/32). You can see that Velvet assemblies produce a relatively wide range in CEGMA results. Assemblies made by SOAPdenovo produce an even wider range of CEGMA results (not to mention a wide range of N50 results). The truth is that there is no consistent pattern of quality in the results of any one assembler (and remember we are only measuring 'quality' by just two paltry metrics).

To answer the questions raised at the start of this post:

All assemblers can be used to make terrible genome assemblies
Some assemblers can (probably) be used to make great genome assemblies

There is no magic bullet in genome assembly and there are so many parameters that can affect the quality of your final assembly (repeat content of genome, sequencing technology biases, amount of heterozygosity in genome, quality of input DNA, quality of sample preparation steps, suitable mix of libraries with different insert sizes, use of most suitable assembler options for your genome of interest, amount of coffee drunk by person running the assembler, etc. etc.).

Don't even ask the question which genome assembler gives you the best genome assembly? if you are not prepared to define what you mean by 'best' (and please don't define it just on the basis of two simple metrics like %248 CEGs present and N50 length).