Sunday, July 02, 2017

Confusion about the number of genes

My last post was about confusion over the sizes of the human and mouse genomes based on a recent paper by Breschi et al. (2017). Their statements about the number of genes in those species are also confusing. Here's what they say about the human genome.
[According to Ensembl86] the human genome encodes 58,037 genes, of which approximately one-third are protein-coding (19,950), and yields 198,093 transcripts. By comparison, the mouse genome encodes 48,709 genes, of which half are protein-coding (22,018 genes), and yields 118,925 transcripts overall.
The very latest Ensembl estimates (April 2017) for Homo sapiens and Mus Musculus are similar. The difference in gene numbers between mouse and human is not significant according to the authors ...
The discrepancy in total number of annotated genes between the two species is unlikely to reflect differences in underlying biology, and can be attributed to the less advanced state of the mouse annotation.
This is correct but it doesn't explain the other numbers. There's general agreement on the number of protein-coding genes in mammals. They all have about 20,000 genes. There is no agreement on the number of genes for functional noncoding RNAs. In its latest build, Ensemble says there are 14,727 lncRNA genes, 5,362 genes for small noncoding RNAs, and 2,222 other genes for nocoding RNAs. The total number of non-protein-coding genes is 22,311.

There is no solid evidence to support this claim. It's true there are many transcripts resembling functional noncoding RNAs but claiming these identify true genes requires evidence that they have a biological function. It would be okay to call them "potential" genes or "possible" genes but the annotators are going beyond the data when they decide that these are actually genes.

Breschi et al. mention the number of transcripts. I don't know what method Ensembl uses to identify a functional transcript. Are these splice variants of protein-coding genes?

The rest of the review discusses the similarities between human and mouse genes. They point out, correctly, that about 16,000 protein-coding genes are orthologous. With respect to lncRNAs they discuss all the problems in comparing human and mouse lncRNA and conclude that "... the current catalogues of orthologous lncRNAs are still highly incomplete and inaccurate." There are several studies suggesting that only 1,000-2,000 lncRNAs are orthologous. Unfortunately, there's very little overlap between the two most comprehensive studies (189 lncRNAs in common).

There are two obvious possibilities. First, it's possible that these RNAs are just due to transcriptional noise and that's why the ones in the mouse and human genomes are different. Second, all these RNAs are functional but the genes have arisen separately in the two lineages. This means that about 10,000 genes for biologically functional lncRNAs have arisen in each of the genomes over the past 100 million years.

Breschi et al. don't discuss the first possibility.

Breschi, A., Gingeras, T.R., and Guigó, R. (2017) Comparative transcriptomics in human and mouse. Nature Reviews Genetics [doi: 10.1038/nrg.2017.19]


  1. "I don't know what method Ensembl uses to identify a functional transcript. Are these splice variants of protein-coding genes? "
    GENCODE (the manual annotation arm of Ensembl) has 128,000 transcripts annotated in mouse now (, so either the paper is a bit out of date or there is a discrepancy between Ensembl and GENCODE for some reason.
    The transcripts are always predictions (the correct terminology should be transcript models) and based on the balance of the available evidence. Obviously some transcript models have better evidence than others. They aren't all coding transcripts or from coding genes, just 55,000 of the models are predicted as coding. The full (long) list of annotated transcript types is at the bottom of the page.

  2. You will find the different procedures ENSEMBL uses for genome annotation under

    1. I don't find that very helpful. Do you? If you can figure out how they decided there are 14,727 lncRNA genes then please share it with me.

    2. Try this link. As you can see from the last paragraph (lincRNA) they are predicted using cDNA alignments and chromatin-state maps:

    3. That link is also useless. At best it helps us understand a bit about which RNAs they are going to consider but it says nothing about how they decide which ones have a function (= gene) and which ones are spurious transcripts.

    4. Larry, you asked how they decided there were 14,727 lncRNA genes (now > 15,000). From the link: "lincRNA (Long intergenic non-coding RNAs) Ensembl gene annotation, cDNA alignments and chromatin-state map data from the Ensembl regulatory build are used to predict lincRNAs for human and mouse. We do not import the lincRNAs identified by Guttman et al [1], but their publication guided us to our current approach for automatically annotating lincRNAs. First, regions of chromatin methylation (H3K4me3 and H3K36me3) outside known protein-coding loci are identified. Next, cDNAs which overlap with H3K4me3 or H3K36me3 features are identified as candidate lincRNAs. A final evaluation step investigates if each candidate lincRNA has any protein-coding potential. Any candidate lincRNA containing a substantial open reading frame (ORF) covering 35% or more of its length and containing PFAM/tigrfam protein domains will be rejected. Candidate lincRNAs that pass the final evaluation step are included in the human or mouse gene set as lincRNA genes."

      All the lincRNA are predictions. Once they make their way into the annotation they can be validated by various means such as RT-PCRseq (Howald, 2012). At present 4,609 of the lincRNA transcripts (not genes) are tagged as validated experimentally.

      How many of these have a function is whole new question. And, especially if you are talking about known function, I suspect the answer is very few.

  3. "This means that ... genes ... have arisen in each of the genomes over the past [...] 100 million years."

    That 100 million years assumes unchanging or slowly changing clocks. There is much confusion over the difference between generation-to-generation mutation rate and long-term genetic clock, about how fast the latter changes, and what makes it change. What are your views about the differences and the variability?