The main problem is counting the number of genes that produce functional RNA molecules. The latest Ensembl results are based on build CRch37 from February 2009 and the GENCODE annotation from last year (GENCODE 19) [see Human assembly and gene annotation and Harrow et al., 2014]
The most recent estimates are 20,807 protein-encoding genes, 9,096 genes for short RNAs, and 13,870 genes for long RNAs. This gives 43,773 genes. Nobody knows for sure how many of the putative genes for RNAs actually exist. They may only be a few thousand functional genes in this category.
It's a lot easier to figure out whether a gene really encodes a functional protein so most of the annotation effort is focused on those genes. I want to draw your attention to a recent paper by Ezkurdia et al. (2014) that discusses this issue. The authors begin with a bit of history ...
Multiple evidence suggests that there may be as few as 19 000 human protein-coding genesI wish scientists would stop repeating this myth about estimates of gene number. The fact is that many knowledgeable experts predicted that there would be fewer than 30,000 genes and they turned out to be right. By and large, the only people who were predicting much larger numbers were people not familiar with the literature on the subject [False History and the Number of Genes] [Facts and Myths Concerning the Historical Estimates of the Number of Genes in the Human Genome].
The actual number of protein-coding genes that make up the human genome has long been a source of discussion. Before the first draft of the human genome came out, many researchers believed that the final number of human protein-coding genes would fall somewhere between 40 000 and 100 000 (1). The initial sequencing of the human genome revised that figure drastically downwards by suggesting that the final number would fall somewhere between 26 000 (2) and 30 000 (3) genes. With the publication of the final draft of the Human Genome Project (4), the number of protein-coding genes was revised downwards again to between 20 000 and 25 000. Most recently, Clamp and co-workers (5) used evolutionary comparisons to suggest that the most likely figure for the protein-coding genes would be at the lower end of this continuum, just 20 500 genes.
Setting aside this little slip-up, the paper is actually very good.
The best test of whether a given DNA sequence actually encodes a polypeptide is to identify the protein in vivo. This may not be conclusive since several pseudogenes produce aberrant, non-functional, proteins but that's not the main problem. The real problem is confirming gene predictions where the predicted protein is not conserved or otherwise identifiable. It's now common for newly published sequences to over-predict the number of genes and, invariably, the number drops as the annotators get to work and eliminate all the errors.
This procedure applies equally well to predictions of genes for functional RNAs but there it's much more difficult to confirm those genes. The task of identifying protein-encoding genes provides us with a case study that should make us skeptical of the predictions for other genes. We should also be skeptical of ALL gene predictions for most published genomes. As this paper points out, there are a huge number of scientists working on the annotation of the human genome and we still don't know for sure how many genes we have. (The authors of this paper are part of the ENCODE/GENCODE project.)
Modern technology has made it possible to identify tiny amounts of a given protein as long as you have a reliable genome sequence. The procedure works like this. You isolate all the proteins from a given tissue and digest them with an enzyme that cleaves proteins at a particular site. Typically, you use trypsin, an enzyme that cuts proteins at the site of arginine and lysine residues. If you know the sequence of the gene you can predict the sequence of the peptides (pieces of protein) that are produced by treating the protein with trypsin.
If you know the sequence of every single protein-encoding gene then you can predict every peptide that can possibly be produced by treating cellular proteins with trypsin. Each one of these peptides has a particular mass that depends on the sequence of amino acids. For example, the peptide AAFTECCQAADK has a mass of exactly 1262.63 so if you can detect a peptide with that mass then you know that the gene is producing the protein.
Genomics, Proteomics and Mass Spectrometry].
This technique has become so common that we do an experiment like this in our third-year undergraduate laboratory.
In theory, you could isolate proteins from every human tissue at every stage of development and identify all the proteins that are ever encoded by real protein-encoding genes. In practice, there are several limitations that make the task more difficult. The paper discusses many of these limitations and that's why it's going to be assigned to the students in the laboratory class.
The authors of this paper looked at the combined results of seven large-scale analyses and concluded that expression of only 11,840 protein-encoding genes could be confirmed. That's only about half of the number of predicted protein-encoding genes. One of the most severe limitations of the technique is that membrane proteins are often resistant to digestion by trypsin even though they have lysine and arginine residues. The authors address this issue by looking at known membrane proteins and they confirm that these are under-represented in the databases of peptides. There are lots of other problems that could explain why half the genes aren't confirmed.
Recall that the issue is whether all predicted genes actually exist. It looks like you can't conclude that a prediction is false just because you don't find a protein so you have too use other criteria. One of them might be conservation. If a predicted gene is only found in humans and not in other species then it's called a "orphan" gene. It could be a gene that just recently evolved in the human lineage since humans and chimpanzees diverged but it is also a prime candidate for a false prediction. (Most of the "genes" that have already been eliminated by GENCODE annotators fall into this category.)
You might predict that the probability of detection decreases as the level of conservation declines because it's more likely that predicted genes that are not well-conserved may not be genes at all. That's what the authors found as shown in the figure below.
If you look at the far right, you'll see that predicted genes that are only found in primates are almost never detected in the mass spec experiments. We all know that such genes, if real, are most likely to be expressed rarely so it may be difficult to detect them. However, we also know that there's a good chance that the genes aren't real and that's why they aren't confirmed.
In addition to lack of conservation, you can apply a number of other criteria to determine whether a gene is likely to be real. For example, you can look at the predicted amino acid sequences. Real proteins have certain characteristics that aren't found in random sequences. Real protein-encoding genes will produce a transcript that corresponds to the open reading frame. The authors looked at 19 different features to determine whether a gene was likely to encode a real functional protein.
They identified a set of 2001 predicted genes that were highly questionable. The set includes 248 out of the 308 orphan genes that remain in the annotated human genome. These 248 predicted genes will be eliminated in the next version of GENCODE leaving only 60 potential genes that might have arisen de novo in the human lineage.
A total of 394 genes were removed in the latest version of GENCODE. Of these, 349 were in the group that this paper identifies as questionable indicating that the criteria used here is a pretty good predictor of false genes. Nine of the genes that were removed have now been found to produce protein so they will be reinstated.
There were 651 new protein-encoding genes added to the last version of GENCODE. The authors state that, "... we found that 596 of the newly annotated genes had features that suggested they were not protein-coding ...." They will likely be removed in the next version of GENCODE.
The bottom line is that 1867 of the 20,719 protein-encoding genes are probably not genes. On the other hand, the authors found evidence for 58 new genes that are currently not annotated. The total number of protein-encoding genes is now estimated to be 18,910 or about 19,000.
Nobody knows how to apply this type of rigorous analysis to all those predicted genes that might make a functional RNA product. Based on our knowledge of the annotation process for protein-encoding genes, we might expect that a large percentage of those putative RNA genes will turn out to be false predictions.
You may think that this is a pretty esoteric exercise that has few consequences for researchers but that's not true. Falsely annotated genes can lead to false predictions about protein domains and gene families and these propagate in the databases. As the authors put it,
Many of the 2001 genes in the potential non-coding set may turn out not to code for proteins under any circumstances. Unfortunately, genes labelled as protein coding at the gene annotation level can have complications for downstream services and research groups that are sometimes difficult to undo. The Pfam functional domain database, for example, has a recent surge in the numbers of newly defined protein functional domains, and many of these have almost certainly been defined on the back of ‘protein-coding’ genes, some of which may turn out not to code for proteins. Overestimating the numbers of protein-coding genes can also hinder experiments such as large-scale proteomics projects and biomedical projects, such as the mapping of cancer or disease-related variations to human genes.Keep in mind that the human genome is being subjected to much more intent scrutiny than any other genome. You need to take gene predictions in other genomes with a large grain of salt.
The human genome is still in the process of being annotated, and the Ensembl/GENCODE merge of the human genome is in constant flux as the annotators withdraw, redefine gene models and add new genes.
It's also worth keeping in mind that there are huge numbers of so-called "orphan" genes in most newly published genomes. It's very likely that these are not genes at all. This is not widely appreciated. See, for example, the Wikipedia article on orphan genes or some of the creationist blogs: [Orphan Genes: A Guide for the Perplexed] [Newly Discovered 'Orphan Genes' Defy Evolution] [Orphan Genes - Powerful Evidence for Intelligent Design and Creation]
Image Credit: Moran, L.A., Horton, H.R., Scrimgeour, K.G., and Perry, M.D. (2012) Principles of Biochemistry 5th ed., Pearson Education Inc. page 175 [Pearson: Principles of Biochemistry 5/E] [1,177 human orphan genes removed by evolutionists from databases].
Ezkurdia, I., Juan, D., Rodriguez, J. M., Frankish, A., Diekhans, M., Harrow, J., Vazquez, J., Valencia, A. and Tress, M. L. (2014) Multiple evidence strands suggest that there may be as few as 19 000 human protein-coding genes. Human molecular genetics, ddu309, advanced access July 1, 2014. [doi: 10.1093/hmg/ddu309]
Harrow, J., Frankish, A., Gonzalez, J. M., Tapanari, E., Diekhans, M., Kokocinski, F., Aken, B. L., Barrell, D., Zadissa, A., Searle, S. et al. (2012) GENCODE: the reference human genome annotation for The ENCODE Project. Genome research 22, 1760-1774. [doi: 10.1101/gr.135350.111 ]