Okay, I guess a more accurate four word summary would be 'genome assemblies sometimes suck'.

The paper that I'm referring to, by Denton et al, looks at the problem of fragmentation (of genes) in many draft genome assemblies:

Extensive Error in the Number of Genes Inferred from Draft Genome Assemblies

As they note in their introduction:

Our results suggest that low-quality assemblies can result in huge numbers of both added and missing genes, and that most of the additional genes are due to genome fragmentation (“cleaved” gene models).

One section of this paper looks at the quality of different versions of the chicken genome and CEGMA is one of the tools they use in this analysis. I am a co-author of CEGMA, and reading this paper brought back some memories of when we were also looking at similar issues.

In our 2nd CEGMA paper we tried to find out why 36 core genes were not present in the v2.1 version of the chicken genome (6.6x coverage). Turns out that there were ESTs available for 29 of those genes, indicating that they are not absent from the genome, just from the genome assembly. This led us to find pieces of these missing genes in the unanchored set of sequences that were included as part of the set of genome sequences (these often appear as a 'ChrUn' FASTA file in genome sequence releases).

Something else that we reported on in our CEGMA paper is that, sometimes, a newer version of a genome assembly can actually be worse than what it replaces (at least in terms of genic content). Version 1.95 of the Ciona intestinalis genome contained several core genes that subsequently disappeared in the v2.0 release.

In conclusion — and echoing some of the findings of this new paper by Denton et al.:

Many genomes are poorly assembled
Many genomes are poorly annotated (often a result of the poor assembly)
Newer versions of genome assemblies are not always better