Feature Identification of Compensatory Gene Pairs without Sequence Homology in Yeast

2.1. Dataset of Synthetic Lethal Genes and mRNA Expression Data

With systematic generation of double mutant strains of Saccharomyces cerevisiae [32–34], two approaches, synthetic genetic arrays (SGA) and diploid-based synthetic lethality analysis on microarrays (dSLAM), have been developed to identify genome-wide synthetic lethal interactions. We collected all experimentally verified SSL interactions from BioGrid database (http://www.thebiogrid.org/). Recent data sets of high-density epistatic miniarray profiles (E-MAPs) were also included [35, 36]. Temporal mRNA expression data for 6359 S. cerevisiae genes in 40 natural and perturbed conditions were obtained from ExpressDB [37]. Genome-wide responses to 259 single-gene mutants were also collected [38]. All expression profiles of the genes in each condition were standardized with respect to the mean and variance.

2.2. Calculation of Sequence Similarity and Assignment of Functional Module

For each pair of synthetic lethal genes, the corresponding protein sequences were downloaded from NCBI RefSeq (release 24). We defined nonhomologous protein pairs as two sequences that, by BLASTP with standard parameters, share less than 30% identity. Of 9237 SSL pairs, 6186 were nonhomologous. Tong et al. reported that 98% of SSL gene pairs were nonhomologous [34], but our criterion was stricter: only 67% of SSL pairs were retained for further functional analyses.

Functional modules applied on our dataset were collected from the work by Petti and Church [39]. They defined 72 functional modules from the S. cerevisiae genome database at MIPS (Munich Information Center for Protein Sequences) [40].

2.3. Biological Network Construction

We constructed two types of biological networks, the molecular functional network and the gene coexpression network, to estimate the functional similarity between two genes.

The 40 expression profiles under natural and perturbed conditions were used to construct the coexpression network. To evaluate the degree of coexpression between each pair of genes, Pearson correlation coefficients (PCCs) under different conditions were calculated and genes with PCC > 0.7 (upper 5th percentiles) were connected.

As for the functional network, we collected protein-protein and genetic interaction data from the BioGRID (http://www.thebiogrid.org/) and BOND (http://bond.unleashedinformatics.com/) databases.

2.4. Promoter Regulatory Elements Analyses

Promoter sequences of all nonhomologous SSL genes with functional similarity were retrieved from NCBI RefSeq (release 24) (http://www.ncbi.nlm.nih.gov/RefSeq/). A set of 103 yeast regulatory elements (motifs) and their gene assignments were collected from the TRANSFAC database (version 11.3) (http://www.gene-regulation.com/pub/databases.html), which records all experimentally verified transcription factors of S. cerevisiae and their target genes.

For each pair of nonhomologous SSL genes with functional similarity, a pattern search program in TRANSFAC (pMatch) was applied to match both promoter sequences individually with all experimentally determined regulatory motifs. Then, the motif-content overlap (MCO) score [6] was calculated as follows:

where r ₁ and r ₂ are sets of matched regulatory elements in promoter one and promoter two, respectively. |r _x| denotes the number of regulatory elements in gene x.

2.5. MES and PCoR Analyses

MES and PCoR defined by Kafri et al. stand for “mean expression similarity” and “partial coregulation”, respectively [6]. For each pair of the nonhomologous SSL pairs with high functional similarity, Pearson correlation coefficients of mRNA expression profiles under 40 different conditions were calculated to compute these scores.

2.6. Functional Relatedness Analyses

For all gene pairs, we defined the CN (common neighbor) score as a measure for the fraction of shared partners. The score was defined as

where n ₁ and n ₂ denote the number of neighbors for one and the other synthetic lethal counterpart, respectively, and n ₁₂ denotes the number of common neighbors shared between the two SSL genes. If the two SSL genes are connected directly, the value of n ₁₂ is two. The significance of the CN score was estimated by a P value, (Pr(CN_random≧CN)), computed by randomly sampling sets of 10⁵ pairs of genes.

3.1. Sequence Similarity and Functional Module Analysis

The central issue is whether transcriptional compensation plays a significant role for nonhomologous genes with similar functions. To exclude all homologous SSL gene pairs, we first evaluated a set of SSL pairs of S. cerevisiae through sequence similarity. Of 9237 SSL pairs, 6186 were nonhomologous and retained for further analyses. Next, we identified 1771 nonhomologous SSL pairs that are categorized into either one of 72 functional modules [39], which are defined as a group of genes or proteins involved in a common cellular process in the gene ontology.

3.2. Gene Expression Analyses

Kafri et al. showed that the compensatory capability of a gene pair is optimal when its mean expression similarity (MES) falls from 0 to ~0.2 and the standard deviation of gene expression correlations (PCoR) higher than 0.4 [6]. To examine whether this feature also exists in the selected 1771 gene pairs, we plotted their scores on a plane spanned by MES and PCoR (blue circles in Figure 1) against the remaining nonhomologous SSL pairs that were not mapped into the same functional modules (red circles in Figure 1). The two-sample t-test revealed that the difference between the two sets was not statistically significant (P values are 0.59 and 0.91 for MES and PCoR, resp.). We also analyzed the distribution of both PCoR and MES for homologous SSL genes. In this case, the P values of the two-sample t-test for MES and PCoR were 0.000038304 and 0.000176, respectively. Our results demonstrated that the measurements of PCoR and MES did not fully delineate compensatory capability in nonhomologous genes.

3.3. Network-Based Modeling

Paralogs with backup capacity have high propensity to be coclustered in the same protein complexes and share common interacting partners [41]. To further filter nonhomologous pairs with high functional similarity, we defined the CN (common neighbor) score (see Section 2).

We used two complementary biological networks: the molecular functional network and the gene coexpression network. Both types of networks were constructed to estimate the functional similarity between two genes (see Section 2).

Of the 1771 nonhomologous SSL pairs in the same functional module, 171 pairs had significant CN scores (P≦0.01) in both the molecular functional network and the coexpression network. Figure 2 shows the distribution of these 171 SSL pairs on the MES-PCoR plane. About 70% of these functionally overlapping nonhomologous SSL pairs had MES values between 0 and 0.5 as well as PCoR values from 0.4 to 0.6. The above MES and PCoR values suggest that expression patterns of nonhomologous SSL pairs are conditionally coexpressed. The distribution of MES and PCoR values for the 171 nonhomologous pairs showed a trend similar to that observed for paralogous backup genes [6]. Based on the gene expression profiles corresponding to diverse mutations [38], we selected a set of genes as potential compensatory genes, which show expression alteration after mutation or deletion of their SSL partner. Figure 2 illustrates the distribution of PCoR and MES for the 171 nonhomologous genes (in blue color) and selected potential compensatory genes from (in red color). The values of the PCoR and MES are more clustered in the intervals [0.4, 0.6] and [0, 0.5], respectively. To further investigate the distribution of PCoR and MES, both measurements were also estimated on CN-significant and non-CN-significant nonhomologous genes in the same functional module. In this case, PCoR and MES were not associated with the CN score (P values of 2-sample t-test are 0.07 and 0.97 for MES and PCoR, resp.). On the other hand, for homologous SSL pairs in the same functional module, MES and PCoR are statistically significant between CN-significant and non-CN-significant pairs (P values of 2-sample t-test are 0.002279 and 0.0197 for MES and PCoR, resp.). Again, these two measurements failed to find cofunctional compensatory pairs when they are not homologous.

3.4. Motif-Sharing Analysis of Nonhomologous SSL Pairs

Because the maximal duplicate-associated compensatory capability might coincide with intermediate levels of motif sharing in the promoter regions of the backup gene pair, partial similarity of regulatory controls may form the basis of transcriptional reprogramming in response to the loss of one paralogous partner [6]. Hence, we next also investigated the regulatory elements in the promoter sequences of the functionally overlapping nonhomologous SSL pairs and calculated the MCO score for each pair.

We used the aforementioned mRNA expression profiles of single-gene mutants to study the relationship between sharing regulatory elements and the compensation capability of nonhomologous SSL pairs. Of the 259 knockouts in this expression dataset, 66 nonhomologous genes with functional similarity had synthetic lethal interactions. Figure 3 presents the relationship between MCO scores and the corresponding probabilities of functional compensation. The greatest proportion of compensatory genes was observed to have MCO scores between 30% and 40%. Similar to compensation by duplicate genes, backup capability was obtained when two genes share only part of their regulatory elements. It should be noted that highly similar regulatory controls in a gene pair result in strongly correlated expression profiles. In this scenario, both genes might be expressed at similar levels and might be expected to be required simultaneously. This may be why nonhomologous SSL pairs with overlapping functions lack the potential for compensation capability when their MCO score is higher than 0.5. Although the analysis of the mRNA expression data is not comprehensive, this analysis provides an evidence to show that partial overlap of the sharing regulatory element is one of the features of compensatory genes.

Taken together, we proposed that the aforementioned ratio of common neighbors in the biological network and the score of the motif-content overlap are critical determinants of the compensatory capability for nonhomologous genes.

3.5. Identification of Functionally Compensatory Genes

As described above, three features were investigated for nonhomologous compensatory genes: (i) synthetic lethality; (ii) the ratio of common neighbors in the biological network; (iii) partial overlap of the regulatory elements of two genes. With these parameters, we selected nonhomologous functionally compensatory genes (see the supplementary table). 89.3% of the selected compensation pairs were over-expressed (log₂ fold change > 3) when its partner was mutated [38] (see the supplementary table in Supplementary Material available online at doi:10.1155/2012/653174). Consistent with the previous study [30], transcriptional compensation between SSL gene pairs may only appear in a small portion. Many of them were associated with signal transduction, metabolic processes, ribosomal proteins, and posttranslational protein modification (see the supplementary table for the detail function of each gene). The compensation mechanisms between genes seemed to be highly divergent [42]. We took two candidate pairs, kar2/sil1 and kar2/lhs1, as examples. The former, kar2 and sil1, shares common regulatory elements in their promoters [43] and exhibits synthetic lethal phenotype [44]. Similarly, lhs1 and sil1 also exhibit SSL interaction [45]. Both SIL1 and LHS1 are nucleotide exchange factors of KAR2 and were proposed to bind KAR2 in a mutually exclusive manner [46]. The promoter of lhs1 was observed to be transcriptionally induced when the mutation occurs in kar2 [47, 48]. On the other hand, at least in the function for the protein translocation, overexpressed SIL1 can partially compensate for LHS1 during loss of lhs1. However, this compensation seems not originating from modulation through kar2 [45]. The mechanism of robustness among these three genes therefore cannot be simply explained by transcriptional reprogramming. Another example of a nonhomologous compensatory gene pair is provided by two members of the RAD2 nuclease family, rad27 and exo1. Functional overlap of RAD27 and EXO1 was observed from identical structure-specific endonuclease and 5′ exonuclease activities. The overexpression of EXO1 results in the suppression of multiple rad27 null mutation-associated phenotypes. Interestingly, similar compensatory behaviors were found in RAD2, but in complementing a different type of mutation, that is, base excision repair. EXO1 and RAD2 complement the defects of the rad27 mutant to different extents. These results suggest that compensation of RAD27 can be achieved in an alternative way [49]. Other examples, such as the chitin synthase gene chs3 and the β-1, 3-glucan synthase gene fks1: mutations in fks1 results in upregulation of CHS3 [29, 50]. As the above examples suggest, the compensation mechanism of nonhomologous genes might occur beyond the level of transcriptional reprogramming.