Posts Tagged ‘disease associated genetic variation’

Investigating Functional Compensation by Human Paralogous Proteins

Larry H. Bernstein, MD, FCAP, Curator




Using Disease-Associated Coding Sequence Variation to Investigate Functional Compensation by Human Paralogous Proteins

Evolutionary Bioinformatics 2015:11 245-251    http://dx.doi.org:/10.4137/EBO.S30594


In this article, we examined the functional compensation among paralogs as a general phenomenon through an analysis of disease-associated genetic variation in humans.23–26 In contrast to expectations under the functional compensation hypothesis, we found that multigene families have a greater tendency to harbor dSNVs than singleton proteins. We proposed that differences in functional constraints (evolutionary constraint hypothesis) explain the observed pattern to a large degree.


Gene duplication enables the functional diversification in species. It is thought that duplicated genes may be able to compensate if the function of one of the gene copies is disrupted. This possibility is extensively debated with some studies reporting proteome-wide compensation, whereas others suggest functional compensation among only recent gene duplicates or no compensation at all. We report results from a systematic molecular evolutionary analysis to test the predictions of the functional compensation hypothesis. We contrasted the density of Mendelian disease-associated single nucleotide variants (dSNVs) in proteins with no discernable paralogs (singletons) with the dSNV density in proteins found in multigene families. Under the functional compensation hypothesis, we expected to find greater numbers of dSNVs in singletons due to the lack of any compensating partners. Our analyses produced an opposite pattern; paralogs have over 35% higher dSNV density than singletons. We found that these patterns are concordant with similar differences in the rates of amino acid evolution (ie, functional constraints), as the proteins with paralogs have evolved 33% slower than singletons. Our evolutionary constraint explanation is robust to differences in family sizes, ages (young vs. old duplicates), and degrees of amino acid sequence similarities among paralogs. Therefore, disease-associated human variation does not exhibit significant signals of functional compensation among paralogous proteins, but rather an evolutionary constraint hypothesis provides a better explanation for the observed patterns of disease-associated and neutral polymorphisms in the human genome.



Gene duplication is an important mechanism for the origin of novelty in evolution.1–3 When a gene is duplicated, one of the duplicate copies usually decays within a few million years due to an accumulation of deleterious mutations.4 However, duplicates may be retained if they become functionally important to the organism.5–7 It has been suggested that duplicate genes may be able to carry out the original gene function, which means that paralogs may compensate for each other.8,9 Gene knockout/knockdown experiments have been conducted in multiple species to examine the degree of functional redundancy in gene families. The results suggest that the loss of function in genes with paralogs is associated with higher organismal survival than the loss of function in genes without any known paralogs (singletons), supporting the functional compensation hypothesis.10–16 However, Liao and Zhang17 reported that duplicates rarely compensate for each other in mice, which has been debated.18–22 Overall, experimental data have not yet provided definitive evidence about whether paralogous genes do compensate for each other in most instances.

The predictions of functional compensation can be tested computationally by analyzing the disease-associated genetic variation in humans. These variants are currently experiencing negative selection in the human populations, which means that they constitute data of functional impact in nature. If functional compensation among gene family members is substantial, it is expected that fewer significant statistical associations between variants and disease phenotypes will be detected for proteins in multigene families than for singletons. Using this idea, Dickerson and Robertson23 tested the predictions of functional compensation and found no difference between the proportion of singletons and para logs implicated in diseases (2% difference), supporting the conclusions of Liao and Zhang.17 However, they and others have suggested that recently diverged paralogs are less likely to be disease-associated than singletons and proteins with distantly related paralogs.23–26 These results suggest functional redundancy among young gene duplicates.

However, the abovementioned computational studies have not accounted for many potentially confounding factors. First, disease-associated single nucleotide variants (dSNVs) are found preferentially at slowly evolving amino acid positions27; thus, we expect to observe a higher frequency of dSNVs in more conserved proteins. This could distort comparisons between singletons and multigene family proteins if the distributions of amino acid evolutionary rates are not the same for these two classes. Second, the numbers of dSNVs found in different proteins are not expected to be the same because the numbers of amino acids in proteins vary by an order of magnitude. This means that commonly used metrics, such as the relative fractions of disease and nondisease proteins in different protein classes, are too coarse. Metrics that take into account the number of amino acids in proteins (sequence length) are necessary for more robust hypothesis testing.

In the following section, we tested the hypothesis of functional compensation by considering the abovementioned factors to better understand the genome-wide pattern of functional evolution in gene families, which is vital for understanding genome evolution and predicting disruptive effects of the mutations of proteins that have paralogs.

We obtained a set of 15,485 human proteins and their homologs from 46 diverse species from the UCSC genome browser (see Material and Methods). For each protein, we also obtained a list of paralogs from the HOVERGEN database.28 Our set of proteins is representative of the whole human gene set because about half (52%) of these proteins have at least one paralog, a fraction that is similar to the overall fraction of proteins with paralogs in the human genome (49% in HOVERGEN database28). For each human protein, we computed the average rate of amino acid substitution (number of substitutions per site per billion years) using the interspecific amino acid sequence alignments (see Material and Methods). Figure 1 shows the distributions of evolutionary rates in singleton and multigene family proteins. Overall, singletons are less conserved than multigene family proteins, with a ∼20% mean and ∼30% median difference (P < 0.01 by two-sample Kolmogorov–Smirnov test; Fig. 1A). Similar patterns are observed when considering paralogs belonging to small (2–5) and large (.5) multigene families (P < 0.01; Fig. 1B).


Figure 1. Distributions of evolutionary rates of singleton (broken line) and multigene family proteins (solid or dotted line). (A) Evolutionary rates are in the units of the number of amino acid substitutions per amino acid site per billion years. the mean and median of these distributions are 1.05 and 0.89, respectively, for singletons, and 0.80 and 0.61, respectively, for proteins in multigene families. these distributions are significantly different (two-sample Kolmogorov– smirnov test; P < 0.01). (b) multigene family proteins were separated into those with two to five paralogs (small family; solid line) and greater than five paralogs (large family; dotted line). The mean and median of these distributions are 0.75 and 0.60, respectively, for the proteins from the small multigene families (two to five paralogs) and 0.87 and 0.63, respectively, for the proteins from the large multigene families (greater than five paralogs). These distributions are significantly different from the distribution for singletons (P < 0.01).


dsNVs in singletons and multigene families. We analyzed all available SNVs associated with Mendelian diseases in singleton and multigene family proteins. There were a total of 47,382 dSNVs in 2,589 proteins. In these data, the proportion of proteins with at least one dSNV was slightly lower (2.2%) for singletons than that of proteins with paralogs, which is consistent with the recent reports.23,29 However, the number of dSNVs in proteins varied extensively and was found to be positively correlated with the protein length (P < 0.05 for multigene family and singletons; Fig. 2). This is reasonable because longer proteins should have a greater chance of accumulating random mutations and are, therefore, more likely to be classified as disease genes. Thus, we normalized the number of dSNVs by protein length to avoid any bias due to length differences in subsequent analyses.


Figure 2. Distributions of the number of dsnvs. (A) a frequency diagram showing the number of proteins with at least one dsnv. (b) the average number of dsnvs per protein for proteins at different length thresholds at 100 amino acids intervals. the average number of dsnvs per protein is positively correlated with the average protein length for both multigene family (correlation = 0.005; P < 0.01) and singleton proteins (correlation = 0.002; P = 0.04).


We compared the number of dSNVs per 100 amino acid positions (dSNV density) between multigene family and singleton proteins. Multigene family proteins have 1.6 times higher density of dSNVs than detected in singleton proteins (0.66 and 0.42, respectively). We can statistically reject the null hypothesis of equal dSNV densities in singletons and multigene family proteins (P < 0.01). However, the direction of effect is opposite to the predictions of functional compensation from paralogous genes in multigene families, as the multigene family proteins contained significantly more dSNVs than singletons. We tested the influence of outliers on this result by excluding all proteins with .0.5 dSNVs per amino acid. This reduced the number of proteins slightly (131 proteins were excluded), but the ratio of multigene family and singleton protein dSNV densities remained unchanged (1.6; P < 0.01). We, nevertheless, excluded all proteins in which the number of dSNVs per position was .0.5 in all subsequent analyses to remove the influence of proteins with unusually high dSNV density when comparing the patterns between different classes of proteins. We also tested if the observed patterns reflect the mutations of specific amino acids (eg, arginine) that comprise a major fraction of the dataset of dSNVs (16%). Arginine codons contain a CpG dinucleotide in the first two positions and are, thus, more prone to transitional mutations, leading to amino acid variation.30 We computed the dSNV densities using only the arginine positions in proteins and found the dSNV density in multigene family proteins to be 1.5 times greater than observed in singletons (0.09 and 0.06, respectively; P < 0.01). A similar pattern was observed for glycine (replacement of glycine residues occurs for 12% of dSNVs in this dataset). The dSNV density in multigene family proteins was twice than observed in singletons (0.08 and 0.04, respectively; P < 0.01).

Finally, we looked for the signatures of functional compensation using dSNVs that are expected to be the most severe, with the rationale that functional compensation may be easier to detect, as ameliorating severe phenotypic effects will have greater relative effect on individual fitness. We designated a dSNV to be severe if the predicted functional impact score for the variant was in the top 5% of all dSNVs (see Material and Methods). For these data, the multigene family proteins have a dSNV density 2.3 times higher than that observed for singletons (0.034 and 0.015, respectively; P < 0.01), which does not support the functional compensation hypothesis. Therefore, the patterns of greater abundance of dSNVs in multigene families are robust to the predicted effect sizes of dSNVs analyzed and the amino acid composition bias of the variation dataset.

Relationship of evolutionary conservation and dsNVs.

We examined if protein conservation difference between singletons and multigene family proteins can explain the above mentioned pattern because it is now well established that highly conserved proteins are significantly more likely to contain dSNVs.27,31 Because the protein evolutionary rate distributions are neither normal nor symmetrical (Fig. 1), we compared medians (0.61 and 0.89, respectively) and found a ratio of 0.69 between the multigene family and singleton proteins. The inverse of this ratio (1.5) is only slightly different from the ratio of dSNV densities (1.6). This similarity suggests that the higher rate of dSNVs in multigene family proteins is mostly explained by the degree of functional constraint on proteins in multigene families versus singleton proteins. Based on this observation, we propose the evolutionary constraints hypothesis, which posits that the differences in dSNV densities among different classes of proteins (eg, singleton vs. multigene) are primarily a result of the differences in the degree of natural selection acting upon them. If true, this would be consistent with the neutral theory of molecular evolution.32 Evolutionary constraint hypothesis does not preclude the existence of functional compensation (among other factors) in some proteins or positions, but it does claim that differences in the intensity of purifying selection will be the primary cause of observed differences in the preponderance of SNVs in different groups of proteins.

We tested the prediction of the evolutionary constraint hypothesis in an analysis of 12,952 common neutral SNVs (nSNVs) obtained from the 1000 Genomes Project.33 These common nSNVs are complementary in nature to dSNVs, as common nSNVs persist in the human population and have risen to moderate frequencies (.5%) because their impact on fitness is effectively neutral (opposite of dSNVs that cause disease). Therefore, if functional constraints and, thus, the conservation level of human protein sequence explain the observed differences in dSNV density, we should also observe fewer nSNVs in multigene family proteins, as these proteins evolve more slowly and are expected to be subject to more severe purifying selection.34 Indeed, the nSNV density (number of nSNVs per 100 amino acids) in multigene family proteins was lower than that of singletons (ratio = 0.82; 0.13 and 0.16, respectively; P < 0.01). This ratio (0.82) is again similar to the ratio of the evolutionary rates (0.69) for these two classes of proteins. These results suggest that the occurrence of dSNVs and nSNVs in proteins is largely concordant with the degree of functional constraint on proteins, which is captured in their evolutionary rates.

Disease sNV prevalence in proteins with young and old paralogs.

Next, we tested the hypothesis that functional compensation is more common in proteins with younger paralogs.23,24 If functional compensation generally occurs only for a brief period after the gene duplication event, then the most recently diverged paralogs will provide the most powerful signal to detect functional compensation. We first identified the closest paralog for each protein within a given gene family by selecting the paralog with the smallest nucleotide divergence in their codons (third positions only). To estimate the relative antiquity of the duplicate event, we used the protein-specific human–mouse third positions in codons to normalize each closest paralog divergence across gene families (see Materials and Methods). This normalized value yields an approximate gene duplication time when it is scaled using the human–mouse divergence time (92.3 million years ago35). This approximation is reasonable, as third positions in codons evolve relatively neutrally and because we use divergence times primarily for identifying and sorting young paralogs for hypothesis testing.

Density of dSNV for duplicates that have diverged from their paralogs in the last 200 million years shows a tendency to increase with estimated duplicate age (Fig. 3A). The same pattern is observed for the positions of arginine and glycine and those with predicted severe functional impacts (Fig. 3B–D). Also, the dSNV densities for the youngest duplicates are lower than those for singletons (triangle in Fig. 3). We found that the evolutionary rate of proteins is negatively correlated with time since duplication, and the youngest duplicates have higher evolutionary rates than singletons (Fig. 4A). These patterns do not support the functional compensation hypothesis23 and are consistent with our evolutionary constraint hypothesis. These trends are confirmed in the analysis of nSNV densities that showed expected complementary patterns (Fig. 4B).


Figure 3. the dsnv density in duplicates over time. Each point shows the dsnv density of all proteins with duplication age less than or equal to a threshold time (x-axis; 10 million year intervals). the dsnv density of singletons is shown with a triangle. Panels show patterns obtained for all dsnvs (A), arginine dsnvs (b), and glycine dsnvs (C). Panel D shows patterns for dsnvs with severe impact predicted by EvoD.46


Figure 4. the average evolutionary rates (A) and nsnv densities (b) of all proteins with duplication age less than or equal to a threshold time (x-axis; 10 million year intervals). the decreasing trend for evolutionary rate (A) is opposite to that observed for dsnvs, but it is similar to that observed for nsnvs (b). in each panel, triangle shows the value from singletons.


Disease sNV prevalence in proteins with very similar paralogs.

We also tested the functional compensation hypothesis in proteins that show high amino acid sequence similarities with their paralogs, as studied by Hsiao and Vitkup.24 We found that paralogs with the highest amino acid sequence similarities (.95%) actually have higher dSNV densities than other paralogs (0.98 vs. 0.57; P < 0.01). This is inconsistent with the functional compensation hypothesis but agrees with our evolutionary constraint hypothesis because the evolutionary rates were lower in paralogs with .95% similarity (0.59 and 0.78 substitutions/site/billion years; P < 0.01). Therefore, differences in the degree of functional constraint (measured using evolutionary rates) account for the observed patterns of dSNV densities.

Next, we compared nSNV densities in paralogs with .95% sequence similarity to those with #95% similarity. For this comparison, we needed to be cognizant of the fact that variant calls are difficult when the paralogs have very similar DNA sequences.36–39 This is the case for paralogs with .95% amino acid sequence similarity because most of these proteins also showed small divergences at the third positions in codons between paralogs (#0.2 substitutions per site). To accommodate the variant call errors, we used proteins with #0.2 distance (third positions) for comparison between paralogs for two groups of proteins (225 and 69 proteins). The nSNV density was 0.30 and 0.52 for proteins that have paralogs with .95% and #95% sequence similarity, respectively (P < 0.01). The former proteins are more conserved (rate = 0.89) than the latter (rate = 1.97; P < 0.01), and so the result is consistent with the evolutionary constraint hypothesis.


In this article, we examined the functional compensation among paralogs as a general phenomenon through an analysis of disease-associated genetic variation in humans.23–26 In contrast to expectations under the functional compensation hypothesis, we found that multigene families have a greater tendency to harbor dSNVs than singleton proteins. We proposed that differences in functional constraints (evolutionary constraint hypothesis) explain the observed pattern to a large degree. We confirmed that singleton proteins show lower functional constraint than proteins with identifiable duplicates in the genome, which explains the increased detection of disease-associated variation observed in multigene families.

Some recent theoretical and empirical studies suggest that functional compensation can lead to enhanced purifying selection and, therefore, may actually be associated with slower evolutionary rates.14,40 Other studies indicate that the youngest duplicates are evolving under relaxed selection pressures, which would cause an increase in evolutionary rates for a few million years.4 Such short-term and localized rate changes (faster or slower) will not have significant impact on the estimates of very long-term evolutionary rates that we have used to quantify the functional constraint. We have calculated the evolutionary rates using sequence differences in proteins that have accumulated changes for hundreds of millions of years across major groups of vertebrates. There is no evidence that pervasive functional compensation exists across the phylogenetic breadth and genomic scale reflected in our analyses. We expect our major conclusions to hold true in general, while acknowledging that functional compensation may occur in some multigene families and some amino acid positions. In summary, we suggest that there is a need to fully consider differences in the evolutionary conservation of proteins when studying the patterns of sequence variation and variant–phenotype associations.





Read Full Post »

%d bloggers like this: