For the characterization of the functionally crucial consensus we assumed: (i) that all ORs have a common ancestral origin 710and hence the most conserved residues in a multiple alignment comparison would be those that are subjected to the strongest functional constraints; (ii) that residue conservation likely represents functional or structural aspects such as signal transduction processes or disulphide bridges shared by all ORs. Accordingly, we analyzed OR genes of dog and mouse, two macrosmatic mammals that still heavily rely on their sense of smell for survival. Intact OR genes in these species are likely to be under strong evolutionary pressure which would tend to eliminate deleterious mutations. We identified residues that were significantly more conserved than expected (P < 0.05) both in orthologs and in paralogs (Fig. 1). This was aimed to eliminate positions conserved only in orthologous pairs and not among paralogs, which are candidate contact residues for odorant ligands 1920. Also eliminated are residues that are highly conserved in only one of the species (species-specific conservation).

Based on this analysis (Fig. 1), 65 positions were found to be highly conserved both among orthologs and paralogs, hence constructing a conservation core which might be related to functionalities shared by all or most ORs. Inspection of the relative position of these conserved residues along the inferred transmembrane helix topology of the OR protein 19 revealed a trend towards an intracellular localization (Fig. 2). This finding supports other studies indicating higher amino acid conservation in the intracellular portion of OR genes 921, related to interaction with downstream transduction molecules.

To distinguish between tolerant and deleterious amino acids in the conservation core, we employed the SIFT algorithm 18, a sequence homology-based tool that predicts whether an amino acid substitution would have a phenotypic effect. The same dog and mouse intact OR dataset was used, and SIFT was applied separately to class I and class II ORs 68 to accommodate class-specific amino acid preferences. Consequently, a class-related SIFT matrix for OR genes was constructed (Fig. 3).

To validate the SIFT matrix results, we examined its agreement with the above pairwise conservation core analysis. For each of the 65 conserved positions, SIFT-intolerant mutations were counted in both mouse and dog intact ORs and the results were compared to their corresponding pairwise conservation scores. Overall, a good agreement was seen (r = 0.73, P < 10^-5) (Fig. 4A). In five positions we observed large differences in the intolerant mutation frequency between the two methods (Fig. 4A) and hence conservatively removed these positions from further analysis.

A key hypothesis of the present study is that a protein's cumulative SIFT intolerance score is a predictor of its pseudogene status. We used the set of 245 non-redundant (< 80% similarity) full-length human OR pseudogenes to obtain support for this notion. For each OR pseudogene the total number of SIFT-intolerant amino acids was compared to the number of frame disruptions (nonsense and in/del mutations). We observed high correlation between these two types of deleterious sequence alterations (r = 0.75, P < 10^-5) (Fig. 4B), suggesting that the cumulative SIFT deviation is an adequate indicator for pseudogenizing sequence disruptions. We also investigated the effect of the position of the coding frame disruption. For this, we reanalyzed the data with only considering frame disruptions between TM1-TM7, as the amino and carboxy termini are the most likely to be dispensable in shortened open reading frame (Fig. 4B). The results show that the observed correlation is not significantly changed.

The central goal of this study is to predict whether an OR is functional or not according to its deviation form a protein consensus sequence. For that we used the resulted SIFT matrix to generate a binary vector for each OR gene, indicating a tolerant (0) or intolerant (1) amino acids along the 60 conserved position. To achieve the best separation between presumed active and inactive OR genes, these vectors were subjected to a logistic regression analysis 22, a supervised learning-based classification routine.

The training step of the logistic regression procedure is aimed at computing weighting coefficients for the predictor variables that would subsequently produce likelihood score for belonging to one of the two classes. Here, functional genes were represented by a non-redundant set (< 80% similarity) of 598 mouse intact OR genes, and non-functional genes by 295 human OR pseudogenes. We considered genes with a pseudogene likelihood score ψ_L ≤ 0.5 as functional and genes with ψ_L > 0.5 as pseudogenes. The resulting logistic model correctly characterizing 96.3% of the learning set intact ORs and 69.8% of its human OR pseudogenes (χ² = 541.5, P < 10^-8). We further performed a cross-validation jackknifing analysis, whereby the original training set of 893 genes (both intact and inactive) was divided into 8 equal groups, and learning was performed for different 7/8 subsets, followed by testing on the remained 1/8. Correct identification was obtained for 65% ± 7% of the human pseudogenes and 95% ± 3% of mouse intact genes (P < 10^-8) (see Additional file 2).

Subsequently, we used the derived logistic model weight parameters to assess the propensity of human OR genes, nominally classified as functional, to encode non-functional proteins. The test set constituted 384 human OR genes with full length open reading frame (> 280 amino acids and including all 7TM regions) labeled as intact in the HORDE database 23. The procedure resulted in the re-classification of 98 purportedly intact OR genes as non-functional (see Additional file 2). Taking into account the false positive (~5%) and false negative (~35%) values of our logistic model in correctly identifying true pseudogenes (see above), the extrapolated number of intact human ORs that likely encode non-functional proteins is expected to be ~135. This brings the total potential number of human OR pseudogenes to ~70%. Recently, Gilad et al 24 have examined the evolution of OR genes in primates by a genomic comparisons of human to chimpanzee (Pan troglodytes). Among other things, they assessed the rate at which neutral gene disruptions accumulate in human OR genes. This led to a subsequent estimation that ~135 human intact OR genes evolve under no selective constraints 24 and Y. Gilad, private communication), which is in agreement with our inference. Thus, both analyses suggest that while a large group of pseudogenes have both missense functional disruptions as well as in-frame stop codons, a significant number of genes (which is in agreement with statistically-based expectations) have only missense-induced loss of function, without coding frame disruptions.

We also utilized CORP to analyze the recently published chimpanzee OR subgenome 24. Consequently, of 279 chimpanzee full-length intact OR genes, 59 were predicted to be non-functional. The fraction of putatively inactive ORs with intact open reading frame in chimpanzee (0.204) was slightly smaller than observed in human (0.255), as expected according to the similar difference in the frame-disrupted pseudogenes of these species 24. Such differences suggest that the chimpanzee OR genes evolved under stronger selective constrains than the human OR genes since the speciation of these two higher apes 5–10 million years ago.

CORP could also be utilized to predict the functional status of allelic variants in OR genes. In this realm, nonsynonymous SNPs that exchange between tolerant and intolerant amino acids in a highly conserved position may result in segregation between active and inactive alleles. To examine this, we performed a database search for missense SNPs at the 60 highly conserved sequence positions of the OR conservation core of all intact genes. A total of 91 such SNPs were identified in 71 intact OR genes. To assess the potential functional impact of the polymorphisms in these genes, we applied all possible allelic states of these ORs to CORP. Consequently, we found 30 human OR genes segregated between active and inactive states in the human population, arising from mutation-like variations at highly conserved positions. Another 16 ORs that were among the 98 ORs already annotated as non-functional according to other fixed intolerant amino acid substitutions did not change their functional status due to these SNPs. Conversely, 25 ORs were predicted as functional in all their allelic states despite the potentially damaging SNPs in their sequences. Notably, due to the false negative rate of pseudogene identification of our algorithm (~0.3), it is likely that it failed in detecting functional segregation in additional 8 genes. Thereby, it brings the current number of missense segregating OR genes to 38 which more than doubling the known count of such important human genetic variation loci 16.

The initial data set comprised 1913 full-length (>280 amino acids) OR protein sequences. These included 753 human genes from build 40 of the HORDE database 23, 1039 mouse 19 and 121 dog 6 OR genes with full-length open reading frame with up to two frame disruptions. Each of these sequences was aligned to a well-curated multiple alignment of mouse and human ORs 19, and the amino acid positions enumerated from 1 to 301 as per the original multiple alignment. Pseudogenes were regarded as genes with at least one frame disrupting mutation between the initiating Methionine and amino acid number 280. The alignments were constructed using Clustal X 37 in a profile alignment mode, with default parameters. Sequences too disrupted to be aligned were discarded. The final data set contained an alignment of 1039, 753 and 83, mouse, human and dog sequences respectively.

We used a previous definition based on mutual best hit 6 for identifying 83 pairs of mouse-dog full-length orthologous OR sequences. The 433 mouse OR paralog pairs were retrieved from 19.

The positional conservation scores were calculated separately for pairwise alignments of 83 mouse-dog OR orthologs, and for a similar alignments of 433 mouse OR paralogous pairs. Only residues between TM1 and TM7 (positions 32–301) of the OR sequences were considered, due to poor alignment in the N-terminal and C-terminal ends of the protein. The conservation scores at each residue position i was calculated as:

C ( i ) = − log ⁡ m ( i ) n ( i )       ( 1 ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGdbWqcqGGOaakcqWGPbqAcqGGPaqkcqGH9aqpcqGHsislcyGGSbaBcqGGVbWBcqGGNbWzdaWcaaqaaiabd2gaTjabcIcaOiabdMgaPjabcMcaPaqaaiabd6gaUjabcIcaOiabdMgaPjabcMcaPaaacaWLjaGaaCzcamaabmGabaGaeGymaedacaGLOaGaayzkaaaaaa@438A@

where m(i) is the number of mismatches between pairs at position i and n(i) is the total number of comparisons at the same positions. The statistical significance of the positional conservation score (C(i)) was assessed by a one-sided chi-square test with one degree of freedom, using the dataset average conservation score as a reference. Finally, we applied a false detection rate (FDR) analysis 38 to eliminate possible false positives due to multiple tests.

To distinguish between tolerant and deleterious amino acids along the OR protein sequence we applied the SIFT algorithm 18 to the dog and mouse OR datasets. Class I and class II ORs were subjected to SIFT separately to eliminate spurious classifications of amino acid substitutions due to the relatively high divergence of these OR classes 710.

Logistic regression was utilized to distinguish between functional and non-functional OR genes according to their deviations from a functionally crucial consensus. This statistical approach, that attempts to distinguish between classes using the most parsimonious model, employs a set of variables with a known class-identifier to create a model that includes the weighted predictor variables, so as to provide the best separation between classes. Logistic regression was preferred to other statistical methods (e.g. linear discrimination analysis) because it makes no assumptions about the distribution of the independent variables. All the logistic regression analyses in this research were processed by the JMP logistic regression package 39 with its default parameters.

To build the logistic model i.e. to assign the appropriate weight to each conserved position, we used a non-redundant training set (< 80% similarity) of 598 mouse intact OR genes and 295 human OR pseudogenes representing active and inactive OR proteins respectively. According to the resulting model, the probability of belonging to one of the two classes was indicated by a pseudogene likelihood score (ψ_L) ranging between 0 (active) and 1 (inactive). Consequently, we considered genes with ψ_L ≤ 0.5 as functional and genes with ψ_L >0.5 as non-functional.

Missense SNPs in the 60 highly conserved positions were taken from the HORDE databases 23 for all human intact OR genes. SNP information in HORDE is extracted from the NCBI dbSNP database 40 in reference to the OR genomic location. Currently HORDE contains 3395 SNPs of which, 1089 are missense SNPs in intact ORs and hence have a potential functional consequence.

For each of the human and mouse ORs we matched the closest ortholog and paralog according to sequence identity. Then we inferred the ancestral sequence of the gene according to the sequence consensus of this triplet. Subsequently, a value of Ka/Ks was calculated for each gene based on the sequence divergence between the gene and its inferred ancestral sequence. This procedure was performed using GCG package 41.