To assess whether SNP A/S ratios from the current large-scale SNP data set provide a good measure for selective constraints, we first compared them with Ka/Ks ratios derived from inter-species alignments. We collected 9,759 human proteins with both validated coding-region SNPs and available human-mouse Ka/Ks data from Ensemble 13, binned them by their Ka/Ks values, and measured the SNP A/S ratios for each group. There is a strong positive correlation between these two measure (Figure 1a; Kendall's rank correlation 14 τ = 0.50, p-value < 1e-04), which is in agreement with the neutral theory of molecular evolution. Analysis of data from chimpanzee and Old World monkey (Macaca mulatta) led to similar conclusions, although the Ka/Ks values may need to be corrected to subtract the contribution of SNPs due to relatively short evolutionary distance.

We next investigated whether the conservation in protein sequences correlates with the SNP A/S ratio under the assumption that both the conservation at the protein sequence level and the SNP A/S ratio at the nucleotide level are indications for selective constraints. Using the position-specific alignment entropy (a measure for conservation) from PSI-BLAST profiles 15, we calculated A/S ratios for residues with different conservation scores. We indeed observed a monotonic decrease of the A/S ratio with an increase in protein sequence conservation (Figure 1b). The residues with the conservation range of 0-0.5 have a ratio of 0.33, while those having conservation scores bigger than 3.5 have an A/S ratio of 0.06.

Many studies have been published addressing the correlation between evolutionary constraints and other variables, most of which were based on relatively small data sets. Having established the SNP A/S ratio as a good measure for selective constraints, we attempted to use the large-scale human SNP data set to revisit some of the features in the earlier studies, and also to investigate several protein properties that had not been examined before.

Until a few years ago, the prevalent theory in molecular evolution was that evolutionary rate is largely dependent on structural and functional constraints. Recently, increasingly more evidence suggests that there is a strong correlation between evolutionary rate and gene expression. It has been observed that highly expressed genes evolve slowly in bacteria 16, yeast 17, and mammals 18. In yeast, it has been shown by principal component regression that the number of translation events is the dominant determinant of evolutionary rate among several other functional attributes 19, leading to the increasingly popular 'translational robustness' hypothesis 20. However, a later study suggested that the dominant effect may result from the noise in biological data that confounded the analysis 21. Studies of human mRNA expression data showed that the breadth of expression (that is, the number of tissues in which a gene is expressed) also correlates with evolutionary rate 2223; it is still debatable whether the breadth or the rate of expression is the stronger predictor 18. We obtained mRNA expression data for 10,885 genes in our data set that are available from a published microarray experiment (Gene Expression Atlas) 24 and investigated the correlation between selective constraints and four gene expression parameters examined previously: peak expression level, mean expression level, expression breadth, and tissue specificity. Overall, this set of genes with available mRNA expression data has an SNP A/S ratio of 0.25, lower than that of our entire data set (0.28). We indeed observed that highly expressed genes tend to have low A/S ratios (Figure 2a,b): both mean and peak expression rate negatively correlate with the SNP A/S ratio (τ = -0.178 and -0.160, respectively; Table S1 in Additional data file 1). Genes with the lowest mean expression levels have an A/S ratio of 0.38, about twice as high as the ratio in the highest expression group (Figure 2a). The SNP A/S ratio also correlates well with the breadth of expression (Figure 2c; τ = -0.213, p-value < 1e-04), but only marginally with tissue specificity (Figure 2d; τ = 0.047, p-value = 0.003). Since these four expression parameters correlate strongly with each other, we carried out partial correlation analysis 14 to identify the stronger predictors for evolutionary rates. The correlation between tissue specificity and the A/S ratio disappeared entirely after controlling for mean expression level (τ = 0.0107, p-value = 0.499; Table S1 in Additional data file 1) or expression breadth (τ = 0.0084, p-value = 0.596; Table S1 in Additional data file 1). Expression breadth and mean expression level both remain significantly correlated with the A/S ratio when controlling one for the other (τ = -0.096 and -0.064, p-values < 1e-04 and 7e-04, respectively; Table S1 in Additional data file 1). Peak expression level is highly correlated with mean expression level and its partial correlation patterns largely resemble those of mean expression level. It has recently been recognized that it is critical to control for expression when studying the statistical relevance of other variables as predictors for evolutionary rates, since many previously reported correlations became insignificant after this control. As expression breadth appeared to have the strongest correlation with the SNP A/S ratio in our data set among the four parameters, we chose to control for it in the following correlation analysis between selective constraints and other variables. The results did not change qualitatively when controlling for mean expression level instead.

Consistent with the hypothesis that gene duplications are an important source of new protein function, it has been observed that duplicated genes evolve under weaker purifying selection than unduplicated ones 2526. We collected 12,460 human genes without paralogs and 167 genes with paralogs according to the HomoloGene database 2728, and found that the A/S ratio is markedly higher for genes with paralogs (0.46 versus 0.27, p-value < 1e-04; Figure 3a, dark gray bars). To control for expression breadth, we analyzed the subset of genes with mRNA expression data from the Gene Expression Atlas 24. The two groups of genes do not differ in their distribution of expression breadth (Kolmogorov-Smirnov test, p-value = 0.507). The difference in the A/S ratio did not change significantly when the expression breadth was controlled by Monte Carlo sampling (Figure 3a, light gray bars and white bars). We then examined whether the higher rate could be solely explained by additional copies of paralogs while keeping one copy stable. When we selected the fastest evolving genes from each homology group, they have an A/S ratio of 0.55 compared with 0.36 for the batch of the slowest-evolving genes from each homology group. Both numbers are higher than the A/S ratio for genes without paralogs (0.27), suggesting that both duplicated copies are evolving faster than unduplicated genes. The much bigger variation in the with-paralog group (95th percentile confidence interval = [0.38, 0.58]) reflects the small number of genes in that particular group.

To determine whether the SNP A/S ratio correlates with the age of proteins, we classified each protein into one of seven age groups according to their most ancient homologs. It appears that young proteins (for example, those found in human or primates only) have the highest A/S ratios (0.76 for human and 0.66 for primates), whereas proteins traceable to all animals or other eukaryotes have much lower ratios of about 0.25 (Figure 3b). This is consistent with a previous finding that proteins that arose earlier in evolution tend to have a larger proportion of sites subjected to negative selection 29, although there was some debate about whether the observation was an artifact resulting from the inability of BLAST to detect homology for the fastest-evolving genes 3031. We examined the functions of proteins in each group by their Gene Ontology (GO) 32 annotation of biological process. The human-specific group is the least well annotated, with only 6% having GO annotation compared with 62% overall and 84% for proteins conserved in both eukaryotes and prokaryotes (the 'universal' group). Among the proteins with GO annotation of biological process, we observed the enrichment of 'epidermis development', 'defense response to bacterium', and 'spermatogenesis' in the human and primate groups, whereas 'amino acid metabolic process', 'glycolysis', and 'fatty acid metabolic process' are overrepresented in the 'universal' group.

As an example of the many conflicting reports in the literature about correlations with evolutionary rates, for a variable as simple as protein length, it was shown that there was positive correlation 33, negative correlation 3435, or no correlation 36. In addition, there was a study based on protein sequence alignments that showed that less conserved proteins are shorter than more conserved ones on average 37. In our data set, we observed a negative correlation between protein length and SNP A/S ratio (Kendall's τ = -0.137, p-value < 1e-04). The correlation did not change upon controlling for expression breadth. Our analysis also showed that this correlation is only prominent for proteins shorter than 500 residues, and disappears for longer proteins (Figure 4a).

Solvent accessibility measures the degree of an amino acid residue's exposure to the surrounding solvent. There have been a number of studies about the effect of mutations on solvent accessibility and its implication in human diseases; most of them were based on relatively small collections of SNPs in known protein structures. The general consensus was that buried residues are less likely to vary and their mutations are more likely to cause disease 3839. We obtained solvent accessibility predictions for all proteins in our dataset using PROFacc 40, and compared the SNP A/S ratios. Exposed residues have an A/S ratio of 0.31, significantly higher than that of 0.24 for the buried residues (Figure 4b). The p-value for this difference is smaller than 1e-04 according to bootstrap analysis. Similar results were obtained when using three-state prediction (buried, intermediate, and exposed) or numeric relative accessibility values. This underscores higher selective constraints on buried residues, possibly due to their importance in maintaining protein stability.

We also investigated selective constraints upon different protein structure conformations. We first grouped all residues into different secondary structure conformations (alpha-helix, beta-strand, or loop) according to predictions by PSIPRED 41. Significantly higher A/S ratios were observed for residues in the loop conformation (Figure 4c), suggesting relaxed selective pressure on these residues. There is no difference between residues in alpha-helices and beta-strands. We next examined natively disordered proteins, a class of structurally flexible proteins that have recently gained traction because of their potential important roles in dynamic molecular recognition of macromolecules 42. It has been estimated that one-third of eukaryotic proteins contains disordered regions 43, and that they are more likely to be involved in regulatory functions and protein-protein interactions 4445. We obtained disorder predictions using DISOPRED2 43 and retained only the disordered regions longer than 30 residues. Interestingly, while proteins with disordered regions have a lower A/S ratio (Figure 4d; Figure S2b in Additional data file 1), the residues in disordered regions have a much higher A/S ratio than other residues (0.38 versus 0.22; Figure 4d). This seems to suggest that disordered proteins as a class are under stronger selective pressure, but the disordered residues are allowed to evolve much faster to explore different ways to interact with other molecules. Since disordered regions are often characterized by low sequence complexity 4244, we also examined the selective constraints on low complexity regions as defined by SEG 46. Not surprisingly, low complexity regions have a higher A/S ratio, but the profile is different from that of the disordered regions (Figure 4e), confirming that disorder and low complexity are related but different sequence features.

Subcellular localization is an important aspect of protein function. There have been conflicting reports about the correlation between protein subcellular localization and evolutionary rate. While a previous survey of human SNPs in 2002 did not find a significant correlation of selective pressure against deleterious non-synonymous SNPs with localization 47, a more recent study of mammalian sequences found that secreted proteins evolve much faster than cytoplasmic proteins (Ka/Ks 0.27 versus 0.12), and that membrane segments are under higher selective pressure than non-membrane segments (0.07 versus 0.15) 48. We attempted to address this issue by examining A/S ratios from several subcellular localization assignment methods. When we divide our data set into 3,064 secreted proteins and 10,622 non-secreted proteins according to SignalP 49 predictions, there is a small and insignificant difference between these two classes, but the residues within the signal peptides appear under much less selective pressure (A/S ratios of 0.42 versus 0.29; Figure 5a). Interestingly, when only the subset of genes that have mRNA expression data was examined (both before and after controlling for expression), secreted proteins had significantly higher A/S ratios than non-secreted proteins (p-value < 1e-04; Figure S3a in Additional data file 1). There is no difference between membrane proteins and non-membrane proteins, membrane segments and non-membrane segments according to TMHMM 50 predictions (Figure 5b; Figure S3b in Additional data file 1). We also obtained predictions of subcellular localizations for non-membrane proteins by LOCtree 51, a hierarchical prediction system mimicking cellular sorting mechanisms. Predicted extracellular proteins have an A/S ratio of 0.34 on average, significantly higher than nuclear and cytoplasmic proteins (Figure 5c). Lastly, we examined A/S ratios of 6,228 proteins that have unambiguous GO cellular component assignments. We observed the same trend as for the LOCtree predictions, although the absolute numbers are slightly lower (Figure 5d). This may be explained by the fact that more conserved proteins are more likely to get GO annotation through sequence homology. The selective constraints acted upon membrane proteins seem to fall between the extracellular and cytoplasmic proteins according to the GO annotations (Figure 5d). The results from both LOCtree predictions and GO annotation did not change qualitatively when controlling for expression breadth (Figure S3c,d in Additional data file 1). Overall, our analysis suggests that extracellular proteins are indeed under more relaxed selection than cytoplasmic and nuclear proteins, but the difference is not as dramatic as previously reported. The absence of difference between membrane and non-membrane proteins according to TMHMM predictions may result from the lack of distinction between the extracellular and cytoplasmic/nuclear proteins.

We next studied the variation in SNP distribution of functional categories based on GO annotations. A/S ratios were calculated for 176 GO biological process categories and 152 molecular function categories that have at least 20 genes in our data set. As expected, there are dramatic differences in selective constraints among different categories: A/S ratios range from 0.72 for 'sensory perception of smell' to 0.07 for 'protein kinase C activation' (Table 1). We compared our results with a comparative genomic study of human and chimpanzee 12. Seven of the top ten categories with highest divergence rates between human and chimpanzee are not present in our entire set of 176 categories due to differences in gene sets and the availability of SNP data. Among the three that are present, all show elevated A/S ratios, and two of them are also in our top ten list (GO:0007608 sensory perception of smell and GO:0007565 female pregnancy). When GO terms were mapped to a small set of high level terms according to Gene Ontology Annotation 52 (GOA slim), the biological process category with the most relaxed selective constraint was 'response to stimulus', which has a significantly higher A/S ratio of 0.33 compared with 'multicellular organismal development', 'transport', 'macromolecule metabolic process', and 'cell differentiation' (Figure 6a). In terms of molecular function, the least variable groups are 'protein transporter activity' and 'motor activity', and the opposite groups are 'receptor activity' and 'isomerase activity' (Figure 6b).

We also sought to quantify the selective pressure on protein families. Of the 13,686 proteins in our data set, 10,629 can be assigned to at least one Pfam 53 family using the HMMER program. Among the 190 Pfam families that have at least 20 members, the families with the lowest A/S ratios include protein kinase C-terminal domain family (PF00433) and core histones (PF00125); on the high end there are mammalian taste receptors (PF05296), the rhodopsin family (PF00001), and glutathione S-transferases (PF02798 and PF00043) (Table 2). We took a closer look at the G protein-coupled receptor (GPCR) family. GPCRs comprise a large protein family of seven transmembrane receptors that play important roles in sensing environmental signals. They are the targets of more than 40% of all modern drugs. There are five Pfam GPCR families that have more than 20 proteins in our data set. Mammalian taste receptor proteins (PF05296) and rhodopsin family (PF00001) are among the most variable protein families, with an A/S ratio of 0.49. The other three (PF00002 secretin family, PF00003 metabotropic glutamate family, and PF01461 7TM chemoreceptor) have A/S ratios of around 0.25, similar to the overall A/S ratio of 0.28 in our entire dataset. There are 558 proteins that belong to the rhodopsin family, including 286 olfactory receptors. The elevated A/S ratio in the family can be largely attributed to olfactory receptors (A/S = 0.73): the non-olfactory receptors in this family have an A/S ratio of 0.30. Therefore, it appears that among GPCRs, only olfactory and taste receptors have extraordinarily high variations, while other proteins behave like average human proteins.

Knowledge about the degree of selection for disease-related genes can help us understand the etiology of human diseases. An early study found that human disease genes evolve faster at both synonymous and non-synonymous sites than non-disease genes, and Ka/Ks ratios of disease genes are 24% higher 54. Although the elevated Ks has subsequently been confirmed by others, later studies reported no difference in Ka/Ks between disease genes and non-disease genes 55 or lower Ka for disease genes 56. It has also been shown that significant differences exist between the Ka/Ks ratio for different pathophysiological classes: genes related to neurological diseases evolve much slower than those associated with immune, hematological and pulmonary diseases 55. We investigated the SNP distribution of human disease genes using two cancer-related gene collections (243 genes from Cancer Gene Census (CGC) 57, and 3,103 genes from the Catalogue of Somatic Mutations in Cancer (COSMIC) 58) and the catalog of heritable human disease genes from Online Mendelian Inheritance in Man (OMIM; 2,334 genes) 27. These three data sets represent 4,649 unique human genes, and 139 genes are common to all three sets. Our analysis of the SNP data shows that disease related genes indeed have a higher synonymous SNP density (OMIM, 5.14; COSMIC, 4.41; CGC, 4.73; non-disease, 4.19, per 1,000 synonymous sites). However, the numbers of non-synonymous SNPs per site for disease genes are lower than that for non-disease genes, resulting in significantly lower A/S ratios in disease genes (p-value < 1e-04; Figure 7). The difference between our analysis and some previous studies could be explained by two factors. First, our data sets are substantially bigger than what were used in previous studies. For example, the Smith and Eyre-Walker study 54 analyzed only 392 genes in the disease set and 2,038 genes in the non-disease set, and the Huang et al. study 55 included 1,178 human disease genes. The other possibility is that the evolution of disease-related genes has different patterns in the human lineage, leading to the difference in SNP A/S ratios and Ka/Ks ratios from human-rodent alignments. It has also been suggested that when non-disease genes are partitioned into housekeeping genes and others, the evolutionary rates of disease genes lie between them 59. This is consistent with our data: the SNP A/S ratio for OMIM is 0.24, indeed higher than housekeeping genes (genes with the broadest expression patterns, A/S = 0.19; Figure 2c). Moreover, when controlling for expression breadth (so that different groups have the same distribution of expression breadth, and thus the same proportion of housekeeping genes), non-disease genes still showed significantly higher A/S ratios than genes in the OMIM and COSMIC sets, while the confidence interval of A/S ratios for genes in the CGC set slightly overlapped with that for non-disease genes (Figure S2c in Additional data file 1), mostly due to large variance in the CGC set resulting from a smaller number of genes.

It is still debatable whether there is any correlation between protein-protein interaction and selective pressure. Most studies so far have been based on data from the budding yeast Saccharomyces cerevisiae. After an initial report that yeast proteins with more interaction partners evolve slowly 60, several studies suggested that the correlation is dependent on interaction data sets 61, or that it may be a secondary effect due to protein abundance 62. The latest and most conclusive study in yeast suggested that there is no correlation between connectivity and evolutionary rate in a higher quality literature curated interaction data set, while negative correlations observed in some high-throughput data sets even after controlling for expression could be artifacts of the data sets 63. We obtained human protein-protein interaction data from the IntAct database 64 and examined how SNP A/S ratios are correlated with the connectivity of proteins in the protein-protein interaction network. When all types of interactions were included, proteins with more than five interaction partners appear to have significantly lower A/S ratios than proteins with no more than one partner (Figure 8a, gray bars). We also noticed that proteins with more interaction partners tend to have higher mRNA expression (Figure 8b, gray bars). The Kendall's rank correlation between connectivity and the SNP A/S ratio was -0.131 (p-value < 1e-04), and it dropped to -0.106 (p-value < 1e-04) after controlling for both mean expression level and expression breadth. The correlation between protein abundance and high connectivity in the interaction network could be either a real biological phenomenon or experimental bias; for example, mass spectrometry-based protein complex pulldown experiments are more likely to identify interaction partners for abundant proteins. When we included only yeast two-hybrid interactions in our analysis, which are supposedly less biased with respect to intrinsic expression levels, the correlation between connectivity and abundance largely disappeared, except for the proteins with no interaction partners in the database (Figure 8b, white bars); at the same time, the difference in A/S ratios between proteins with only one partner and those with more than one became smaller and lost statistical significance in some cases according to bootstrap analysis (Figure 8a, white bars). For yeast two-hybrid interactions only, the correlation between connectivity and the SNP A/S ratio was -0.100, and it dropped only slightly to -0.090 (p-value = 0.007) after controlling for expression. Nevertheless, the partial correlations were still statistically significant in both the yeast two-hybrid interaction set and the all interaction set. Our analysis supports the idea that the correlation between evolutionary rate and connectivity in the interaction network can, in part, be explained by protein abundance and that some of the correlation may result from experimental bias. Similar to all the conflicting studies in yeast, it is likely that this result is inconclusive and may vary from data set to data set.

A recent study suggested that protein evolution is strongly affected by mRNA splicing, in addition to the biology of the protein 65. Based on Ka and Ka/Ks from the human-mouse comparison, it was reported that the proportion of sequence near intron-exon boundaries is a strong predictor of evolutionary rates in human, in part due to splice enhancers located close to intron-exon junctions. We were able to confirm this result using SNP data. Codons within 70 bp of the intron-exon boundaries have a SNP A/S ratio of 0.22, significantly lower than 0.30 for codons that are far from the junction. At the protein level, proteins with more than 80% of the sequences within 70 bp of the boundaries have an SNP A/S ratio of 0.20, much lower than those with only 5% close to the boundaries (Figure 9). The Kendall's rank correlation between the proportion of sequence near intron-exon boundaries and SNP A/S ratio is -0.163, comparable to the correlation between mean mRNA expression levels and the SNP A/S ratios. After controlling for expression breadth, the correlation remained significant (τ = -0.147, p-value < 1e-04).

The average SNP density in our data set is about 2 SNPs per 1,000 nucleotides. Although we have limited our analysis to proteins with at least one validated coding SNP, many proteins in our set still have no non-synonymous SNPs or synonymous SNPs. Therefore, it is neither practical nor reliable to measure the selective constraints on individual proteins using the SNP A/S ratio. Nevertheless, we have demonstrated that when a group of proteins (or residues) are measured together, the measure can be quite robust and often in good agreement with Ka/Ks for divergence. For example, the SNP A/S ratio for proteins with paralogs in our data set is 0.47, very close to the Ka/Ks ratio of 0.45 for duplicated mammalian genes reported earlier 25. Although this may not seem surprising given that it is an expected prediction from the neutral theory of evolution, our analysis showed that the SNP A/S ratio can indeed be a practical and robust measure given the current volume of SNP data in public databases.

There are two unique advantages of using SNP A/S ratios to measure selective constraints. First, for human-specific genes and genes whose ortholog relationships can not be determined reliably, inter-species Ka/Ks ratios can not be estimated. For example, among the 13,686 genes in our data set, human-mouse Ka/Ks ratios for 3,937 genes are not available through Ensemble. In those cases, SNP A/S ratios provide an alternative way of measuring selective pressure by the same evolutionary principle. Second, the SNP A/S ratio is a direct measure of the selective constraints specific to the human lineage, which can not be obtained from Ka/Ks for species divergence.

In comparison with other simpler measures used in some earlier SNP studies (for example, 38), such as density of synonymous and non-synonymous SNPs or the fraction of SNPs in different protein classes, the SNP A/S ratio offers several advantages: its interpretation is clearer from evolutionary theories; it is not subject to data selection bias arising from the popularity of genes in sequencing efforts, which could be a problem when using SNP density; and it is normalized by the number of synonymous and non-synonymous sites, so the numbers are comparable across different protein classes or different studies.

To address the issue of determinants of molecular evolution, many studies have been published examining the correlation between variables that characterize the evolution, expression and function of genes. However, as noted in the Results section, there are many controversies among those reports. It is clear that the validity and significance of correlations depend on many factors, including feature variables, data sets, and the correlation measure. Here, we provide the first large-scale study of SNP A/S ratios, and reported correlations between the ratio and a number of variables; not surprisingly, some are not in agreement with previous studies. Although further studies are still needed to draw definitive conclusions about the major determinants, we agree with many others that it is likely that many of these variables correlate with each other in some way, and some of them are secondary effects rather than primary determinants 66. Several statistical methods, such as partial correlation and principle component regression, have been used to attempt to dissect these complex and rich connections; yet it remains an important open challenge in molecular evolution.

We first filtered the human entries in the NCBI's Entrez Gene database 27 by the presence of RefSeq 27 records, and limited our analysis to genes that have at least one validated coding SNP according to dbSNP 6 build 127. For each gene, we chose one transcript (and its corresponding protein product) by selecting the transcript with the most validated coding SNPs, most advanced RefSeq annotation status, and the longest sequence. We also discarded proteins that are longer than 5,000 residues or shorter than 25 residues. Our final data set has 13,686 proteins with 45,538 coding SNPs.

mRNA expression data were obtained from Gene Expression Atlas 2467. Only normal adult samples were included in the analysis. Samples were sorted into 54 non-redundant tissue types. The expression level of each probe set in a given tissue was calculated as the mean of log (base 10) MAS5 signal intensities of all samples in that tissue. The 'mean expression level' of a probe set was defined as the mean across all tissues, while 'peak expression level' was defined as the maximum among all tissues. The tissue specificity of a probe set was defined as the heterogeneity of its expression level across all tissues. It was calculated according to 68 as:

∑ j = 1 n ( 1 − log ⁡ S j log ⁡ S max ⁡ ) n − 1 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8GiVeY=Pipec8Eeeu0xXdbba9frFj0xb9Lqpepeea0xd9q8qiYRWxGi6xij=hbbc9s8aq0=yqpe0xbbG8A8frFve9Fve9Fj0dmeaabaqaciGacaGaaeqabaqabeGadaaakeaajuaGdaWcaaqaamaaqadabaGaaiikaiaaigdacqGHsisldaWcaaqaaiGacYgacaGGVbGaai4zaiaadofadaWgaaqaaiaadQgaaeqaaaqaaiGacYgacaGGVbGaai4zaiaadofaciGGTbGaaiyyaiaacIhaaaGaaiykaaqaaiaadQgacqGH9aqpcaaIXaaabaGaamOBaaGaeyyeIuoaaeaacaWGUbGaeyOeI0IaaGymaaaaaaa@4750@

where n = 54 is the number of human tissues examined here, S_j is the expression level in each tissue, and S_max is the highest expression level of the probe set across all tissues. When a gene has multiple probe sets, its expression levels and tissue specificity were represented by the probe set with the highest mean expression level.

Affymetrix present/absent calls were used to calculate the breadth of the expression. A probe set was considered 'present' in a tissue if it had 'present' calls in no less than half of the samples in that tissue, and the expression breadth of a probe set was defined as the number of tissues in which the probe set was 'present'. When a gene has multiple probe sets, its expression breadth was represented by the probe set with the highest value of breadth.

Ka/Ks ratios from human-mouse alignments were downloaded from Ensembl 13. The information about human paralogs was extracted from HomoloGene database 2728 release 57; the existence of paralogs is indicated by the presence of other human proteins in the same homology group. To investigate the degree of conservation throughout the evolutionary history, that is, the age of a protein, we performed BLAST searches 15 of each protein sequence against NCBI's RefSeq database, and collected all hits with an e-value < 1e-10 and a protein length difference smaller than 30% as potential homologs. The query protein was then classified into one of the seven age groups (human, primate, mammal, vertebrate, animal, eukaryote, or universal) according to its most ancient homolog. The use of different e-values and length difference cutoffs did not change the results qualitatively. We also obtained the conservation score for each residue in all proteins in our data set by running PSI-BLAST 15 against NCBI's nr database (parameters: '-j 3 -h 5e-3 -F F') and taking the values in the 'information per position' (a measure of alignment entropy) column from the ASCII format of PSI-BLAST profiles.

We obtained protein structure features by the following computational methods using their default parameters: two-state (exposed or buried) solvent accessibility by PROFacc 40, signal peptides by SignalP version 3.0 49, transmembrane helices by TMHMM 2.0 50, secondary structures by PSIPRED 41, sequence complexity by SEG 46, and natively disordered proteins by Disopred2 43. Disopred2 predictions were subsequently filtered to retain only the disordered regions longer than 30 residues.

GO annotations 32 were extracted from the GenBank record of each gene, and the GO terms were subsequently mapped to a selection of high-level terms according to GOA slim 5269. For cellular component annotations, the GOA slim terms were further collapsed into four categories (extracellular region, nucleus, cytoplasm, and membrane) where appropriate. Proteins assigned to more than one of these four categories were excluded from the analysis of subcellular localization by GO annotation. Subcellular localization prediction for non-membrane proteins (that is, no transmembrane helix predictions from TMHMM) were also obtained by using LOCtree 51. To assign proteins to protein families, we ran hmmpfam from the HMMER package (version 2.3) against Pfam_ls 53 models (release 21.0) and obtained Pfam family hits with an e-value < 0.01. Numbers of protein-protein interaction partners were obtained from the IntAct database 6470 (28 September 2007 release). We also obtained disease related genes from CGC (243 genes) 5771, COSMIC 5872 (3,103 genes), and OMIM databases 2773 on 28 September 2007. For OMIM, only those genes with the 'confirmed' status (2,334 genes) were included in the analysis.

Sequences within 70 bp of an exon-intron junction were collected in the same way as described in 65. Briefly, transcripts with less than three exons were excluded from the analysis. All internal exons were trimmed so that the first base was the first base of the first complete codon, and the last base the last of the final complete codon. The first and last codons were then removed from each exon, and remaining codons within 70 bp of the intron-exon boundary were defined as sequences close to the boundary.

The features we considered can also be divided into two classes: protein-level features and residue-level features. A protein-level feature describes the property of an entire protein, for example, whether a protein has a transmembrane helix or not; in contrast, a residue-level feature describes the property of a subset of residues within a protein, for example, whether a residue resides in the transmembrane helix or not.

The SNP A/S ratio, also known as the Ka/Ks ratio for polymorphism, is defined as the ratio of the number of non-synonymous SNPs per non-synonymous site to the number of synonymous SNPs per synonymous site. The numbers of synonymous sites and non-synonymous sites were calculated using the method of Miyata and Yasunaga 74. The ratio for a set of proteins (or residues) was calculated by summing the number of SNPs and the number of sites to obtain A and S for the concatenated set before taking the ratio.

We performed bootstrap re-sampling analysis to assess the statistical significance of the differences observed in A/S ratios between different groups. We obtained 10,000 bootstrap replicates by re-sampling with replacement from the original data set. A/S ratios for different groups were calculated for each replicate, and confidence intervals and p-values of the differences were obtained from those 10,000 sets of A/S ratios. For protein-level feature groups, re-samplings were performed within each group, so that those re-sampled data sets had the same number of proteins in each group as the original data set had. For residue-level feature groups, the re-sampled data sets were constructed by re-sampling the 13,686 proteins in our entire data set.

Monte Carlo samplings were used to assess the differences in A/S ratios between groups when controlling for expression parameters. Briefly, the distribution of expression breadth (or expression level) for the group with the smallest number of genes was set as the target distribution; genes in all other groups were then sampled without replacement using Monte Carlo simulation so that all groups had the same distribution of the expression breadth (or level). The effectiveness of Monte Carlo samplings was confirmed by Kolmogorov-Smirnov tests. An example is shown in Figure S1 in Additional data file 1. For each group, 100 samplings were performed, and the A/S ratio for the group was taken as the mean of A/S ratios from the 100 Monte Carlo samples.

Since the average SNP density in our data set is about 2 SNPs per 1,000 nucleotides, and many proteins in our data set have either no non-synonymous SNPs or no synonymous SNPs, it is not possible to reliably calculate the correlation between the SNP A/S ratio and other continuous variables using each protein as a data point. We chose to randomly group every six proteins together as a data point so that, on average, each data point had roughly the same number of SNPs as the reported number of single-nucleotide substitutions (1.23%) between the human and chimpanzee genomes 12. Non-parametric Kendall's τ rank correlation coefficients 14 and two-tailed p-values were used throughout the study. When controlling for expression parameters, Kendall's partial correlation between x and y controlling for z was calculated as:

τ x y . z = τ x y − τ x z τ y z ( 1 − τ x z 2 ) ( 1 − τ y z 2 ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8GiVeY=Pipec8Eeeu0xXdbba9frFj0xb9Lqpepeea0xd9q8qiYRWxGi6xij=hbbc9s8aq0=yqpe0xbbG8A8frFve9Fve9Fj0dmeaabaqaciGacaGaaeqabaqabeGadaaakeaacqaHepaDdaWgaaWcbaGaamiEaiaadMhacaGGUaGaamOEaaqabaGccqGH9aqpjuaGdaWcaaqaaiabes8a0naaBaaaleaacaWG4bGaamyEaaqcfayabaGaeyOeI0IaeqiXdq3aaSbaaSqaaiaadIhacaWG6baajuaGbeaacqaHepaDdaWgaaWcbaGaamyEaiaadQhaaKqbagqaaaqaamaakaaabaGaaiikaiaaigdacqGHsislcqaHepaDlmaaDaaabaGaamiEaiaadQhaaeaacaaIYaaaaKqbakaacMcacaGGOaGaaGymaiabgkHiTiabes8a0TWaa0baaeaacaWG5bGaamOEaaqaaiaaikdaaaqcfaOaaiykaaqabaaaaaaa@5671@

The random grouping was performed 100 times. The correlation and partial correlation coefficients were computed from these 100 samples, and the medians of those 100 sets of coefficients were reported.