Strategy for the comparison of different processes in different organisms

The strategy we use to establish how similar a given process is in two different organisms is as follows. First, we identify orthologs (i.e. genes in different species that evolved from a common ancestral gene by speciation) between the genome of the potential model organism and that of the target organism(s). Then, we attribute function to the different genes in the organisms under comparison and assign each gene to specific biological processes, using biological ontologies [21]. Specifically, we use:

1. The Gene Ontology (GO) [22], which has been widely used for annotating function and localization of genes at a coarse level in many organisms [23], [24], [25], [26], [27], and
2. The pathways that regulate and execute the processes that one is interested in studying, as defined in KEGG [28] (one can also use MetaCYC [29]).

Finally, we compare the sets of genes responsible for the different processes that are present in each organism. Such an approach predicts if two organisms are likely to be comparable with respect to specific processes of interest, by establishing whether the elements that are a part of the molecular circuits executing the relevant processes are analogous between the organisms (see methods for further details).

Because this is a pilot study, we focus on an organism that is widely used and well characterized, S. cerevisiae. We have attributed function to each of the proteins in S. cerevisiae, according to the information derived from GO and KEGG. This allowed us to create a functional classification of the proteins with respect to the biological processes that they are involved in. Details about this classification are given in Figure S1 and Tables S1–S3 materials. With the functional classification of proteins in place, we can compare the different molecular circuits and processes of yeast to their analogs in 704 other organisms.

To compare these molecular circuits and biological processes between S. cerevisiae and other organisms, we created clusters of orthologs (ScCOGs: S. cerevisiae Clusters of Orthologs), homologues (ScCHGs: S. cerevisiae Clusters of Homologues) and absent proteins (ScCAGs: S. cerevisiae Clusters of Absent Genes) for each S. cerevisiae protein with respect to the translated genome of each of the other 704 organisms. Hereafter we only discuss the results for ScCOGs, because these are consistent with those for ScCHGs. The results for each organism are summarized in Table S1. The detailed clusters are provided as Text S1 and Text S2. We are also preparing a server where these results can be further explored and the method can be applied to other organisms.

Each cluster was associated with the functional terms corresponding to its S. cerevisiae protein. To analyze the differences between S. cerevisiae and a specific organism with respect to a given process, we compare the fraction of proteins that are annotated as functioning in that process in both organisms. We investigate if orthologs or homologues for each of these proteins are simultaneously present in both organisms or not. Then, we rank organisms with respect to the differences in the set of proteins responsible for each process, analyzing for ScCOGs, ScCHGs and ScCAGs at the level of domain, kingdom and phyla for all the 704 organisms (summarized in Tables S2 and S3).

Functional comparison of the full S. cerevisiae protein complement to that of archaea, bacteria and eukaryotes

We compared how well the proteins in the different ScCOGs, ScCHGs and ScCAGs are conserved between S. cerevisiae and various classes of organisms. This allowed us to predict if S. cerevisiae can be a good model for specific processes in different classes of organisms, rather than in individual species. The details of the analysis are presented in appendix S1. No S. cerevisiae protein has orthologs in all 704 organisms. Furthermore, no S. cerevisiae protein has homologues in all the Prokaryotes (Archaea & Bacteria together). In addition, 2642 (45%) S. cerevisiae proteins are absent in all the Prokaryotes (for more details see Tables S2 and S3).

ARCHAEA DOMAIN

We analyzed 48 species of Archaea. About 20% (1158) of all S. cerevisiae proteins generate ScCOGs that contain Archaea sequences. However, only 2% (103) of all yeast proteins generate ScCOGs that contain at least a sequence from each sequenced species of Archaea. An additional 18 (0.3%) S. cerevisiae proteins have homologues in all Archaea. 3672 (62%) S. cerevisiae proteins are absent in all Archaea. Most of these have unknown function. Overall, there is no group of organisms for which the networks of proteins responsible for a large fraction of biological processes in S. cerevisiae are similar to their counterparts in Archaea. However, some biological processes are predicted to be similar between S. cerevisae and some Archaea (see below).

BACTERIA DOMAIN

We analyzed 598 species of Bacteria. 1612 (27%) of all S. cerevisiae proteins generate ScCOGs that contain bacteria sequences. However, no ScCOG or ScCHG contains a sequence from each bacterial species. Furthermore, 2881 (49%) S. cerevisiae genes are absent from all Bacteria, a smaller percentage than that for Archaea. As was the case in archaea, overall, there is no group of organisms for which the networks of proteins responsible for a large fraction of biological processes in S. cerevisiae are similar to their counterparts in Bacteria. However, some biological processes are predicted to be similar between S. cerevisae and some Bacteria (see below).

EUKARYOTA DOMAIN

Overall, there are 59 species of Eukaryotes in our dataset. About 4.5% (263) of all ScCOGs contain sequences from each of these organisms. Between 40% and 60% of all S. cerevisiae proteins involved in “MAPK signaling pathways”, “Signal transduction” biological process, and “Helicase activity” molecular functions have orthologs in all 59 species. Furthermore, between 60% and 80% of all proteins involved in “Microtubule organizing center” of S. cerevisiae are also found in all 59 sequenced eukaryotes. Overall, the networks of proteins responsible for a large fraction of biological processes in S. cerevisiae are similar to their counterparts in ascomycetes. Furthermore, several biological processes are predicted to be similar between S. cerevisae and other Eukaryotes.

A more detailed analysis of the three domains is given in Appendix S1.

Functional comparison of biological processes and pathways between S. cerevisiae and other organisms

After getting such a bird's eye view of the similarities and differences between S. cerevisiae and different clades of organisms with respect to different biological processes, we now focus on individual organisms. To obtain an approximate estimation of how close a given biological process is between S. cerevisiae and another organism we build a matrix of 704×5880 entries. In this matrix, a row represents an organism, while a column represents a ScCOG. The matrix entries are 0 if no sequence from the corresponding organism is found in the appropriate ScCOG and 1 otherwise.

Then, we build a secondary set of four additional matrices containing information about KEGG pathways, biological processes, molecular activity and cellular localization. In each matrix, the rows represent the organisms and the columns represent the biological process, the cellular localization, the molecular function, or the KEGG pathway. Each entry in one of these matrices is a vector with a variable number of elements that is constant for each column of a matrix. The number of elements in the vector is equal to the number of different proteins that is associated to the specific biological process or pathway corresponding to the column (See methods for details).

Subsequently, we calculate the Normalized Hamming Distance (NHD) between the vector of proteins in one entry of the matrix and the corresponding vector for S. cerevisiae from that same column. This NHD is a metric based on the number of elements that are different between the two vectors. The smaller the NHD, the more similar the two vectors are and the more similar the set of proteins executing a specific process in both organisms is. Consequently, the more likely it is that S. cerevisiae is a good model to study the relevant process and generalize the results to the other organism. Using this metric we have clustered the organisms in the matrix according to growing overall NHD with respect to S. cerevisiae.

KEGG Pathways

Figure 1 summarizes the results for KEGG pathways (see Figure S2 for a complete analysis). “Benzoate degradation via hydroxylation”, “Geraniol degradation”, “Propanoate metabolism”, “Valine, leucine and isoleucine biosynthesis”, “Glycolysis/Gluconeogenesis”, “methane metabolism”, “Glycolysis/Gluconeogenesis” and “Aminoacyl-t-RNA biosynthesis” are pathways that appear to be similar to those of S. cerevisiae in a large fraction of organisms. Pathways such as S. cerevisiae's “RNA polymerase” (29 genes), “Lysosome” (14 genes), “Endocytosis” (33 genes), “Oxidative phosphorylation” (76 genes), “Ribosome” (142 genes), “MAPK signaling pathway - yeast” (55 genes), “DNA replication” (30 genes), and “Ubiquitin mediated proteolysis” (44 genes) and “Nucleotide excision repair” (34 genes) are much more similar to those from other eukaryotes than to the corresponding prokaryotic pathways (when they exist). Among the pathways that are central for life, the one that appears to be more unique to S. cerevisiae and other Saccharomycetes is cell cycle (115 genes), because only a small fraction of its proteins have orthologs in other eukaryotes. Thus, these results suggest that extrapolating cell cycle studies in S. cerevisiae to other organisms outside of the Saccharomycetes clade should be done only at the level of basic principles, if at all (see for example [30], [31]). A more detailed analysis of these pathways and their similarity between S. cerevisiae and the other 704 organisms can be found in the appendix and in Figure S2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Details of a heat-map representation showing how distant each organisms is from S. cerevisiae with respect to each individual KEGG pathway.

A green square indicates a high level of coincidence between the set of proteins involved in the specific pathway (column) in a given organism (row) and the set of proteins for the same pathway in S. cerevisiae. A red square indicates complete absence of the set of proteins involved in the specific pathway (column) in a given organism (row) with respect to the same pathway in S. cerevisiae. Intermediate colors indicate intermediate degrees of coincidence between the set of proteins in the target organism and that in S. cerevisiae. The complete heat-map can be seen in Figure S1.

https://doi.org/10.1371/journal.pone.0016015.g001

An encouraging observation for the use of S. cerevisiae as a model organism for mammals is that most of the studied mammals (humans, dogs, mice, cows and rats) are among the non-fungal organisms that have biological processes with protein sets that are similar to the corresponding sets of S. cerevisiae. Specifically, the sets of S. cerevisiae proteins that are associated to “Mismatch repair” (18 genes), “Ubiquitin and other terpenoid-quinone biosynthesis” (5 genes), “Inositol phosphate metabolism” (15 genes), “Steroid biosynthesis” (15 genes), “Ubiquitin mediated proteolysis” (44 genes), “DNA replication” (30 genes), “Ribosome” (142 genes), “Proteasome” (35 genes), “Mismatch repair” (18 genes), “Galactose metabolism” (23 genes), “One carbon pool by folate” (14 genes) and “Glycolysis/gluconeogenesis” (48 genes) are those that appear to be more similar to the corresponding sets of proteins in man. A more thorough analysis is given in the Appendix S1.

GO Biological Processes, Cellular Component and Molecular Function

Figures 2 and 3 summarize the results for the comparisons between S. cerevisiae and the other organisms using the GO categories classification. Details can be further analyzed in Figures S3, S4 and S5. The results are similar to those described for Figure 1 (or those reported in Figure S2), which suggests that these functional classifications are, to a large extent, equivalent, in spite of all problems that they might have (see discussion). S. cerevisiae metabolic activities like “Cellular amino acid and derivative metabolic process”, “Cellular aromatic compound metabolic process”, “Heterocycle metabolic process”, “Cofactor metabolic process” and “Vitamin metabolic process” are the ones that are more conserved in all organisms. In contrast, “cytoskeleton organization”, “Transcription”, “Anatomical structure morphogenesis”, “Transposition”, “conjugation”, “Cell budding”, and “Protein modification process” appear to be conserved mostly in eukaryotes. Conservation of the “Cell wall organization” pathway is restricted to fungi.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Details of a heat-map representation showing how distant each organisms is from S. cerevisiae with respect to each biological process from the GOSLIM classification.

A green square indicates a high level of coincidence between the set of proteins involved in the specific biological process (column) in a given organism (row) and the set of proteins for the same pathway in S. cerevisiae. A red square indicates complete absence of the set of proteins involved in the specific pathway (column) in a given organism (row) with respect to the same biological process in S. cerevisiae. Intermediate colors indicate intermediate degrees of coincidence between the set of proteins in the target organism and that in S. cerevisiae. The complete heat-map can be seen in Figure S2.

https://doi.org/10.1371/journal.pone.0016015.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Details of a heat-map representation showing how distant each organisms is from S. cerevisiae with respect to each molecular function from the GOSLIM classification.

A green square indicates a high level of coincidence between the set of proteins involved in the specific molecular function (column) in a given organism (row) and the set of proteins for the same pathway in S. cerevisiae. A red square indicates complete absence of the set of proteins involved in the specific pathway (column) in a given organism (row) with respect to the same molecular function in S. cerevisiae. Intermediate colors indicate intermediate degrees of coincidence between the set of proteins in the target organism and that in S. cerevisiae. The complete heat-map can be seen in Figure S3.

https://doi.org/10.1371/journal.pone.0016015.g003

Validating the predictions

The analysis described above and the results given in Figures 1, 2 and 3 and in Figures S2-S5 ranks the difference between the protein set responsible for a given biological process in each organism and the corresponding set in S. cerevisiae. If our earlier arguments are correct, one would expect that the similarity between the adaptive responses that involve a given process in other organisms and the same responses in S. cerevisiae is directly correlated to the similarity between the protein sets that regulate and execute that process.

In other words, we define a static metric of closeness of processes between organisms that is based solely on the similarity between the sets of proteins involved in those processes in both organisms. Can we assume that such a metric is also a good measure of closeness between physiological and adaptive responses of the pathways regulating the processes in the organisms being compared, even though it does not include any kinetic or regulatory information?

To answer this question we selected pathways for which dynamic, regulatory, and/or phenotypic information was available for S. cerevisiae and for a scope of different organisms. This selection was based on a careful analysis of Figure S3. We systematically identified pathways or processes with more than 4 genes and then searched the literature for comparable studies of the dynamical and adaptive behavior of these processes in different organisms that belong to our dataset. We were able to identify twelve cases that could be used to answer the question from the previous paragraph.

The results are summarized in Table S4. They show that the phenotypic adaptations and dynamical behavior of a given pathway is more similar to that of S. cerevisiae in organisms that are found to be closer to S. cerevisiae according to our analysis than in more distant organisms. Thus, even if the method we propose is based on static information, the results of the analysis appear to be adequate for pinpointing an appropriate model organism from which to study and extrapolate the dynamical and adaptive behavior of specific biological processes.

The rational choice of model organisms and its technical limitations

In this work we ask the question “How can one chose an appropriate model organism in which to study a specific biological process in such a way that the results may be extrapolated to another organism?” We propose a systematic way to answer this question that involves comparing the similarity between the set of proteins that participate in the biological process of interest in the organism to the equivalent set of proteins in the organism to which we want to extrapolate the results. The closer the set of proteins is between the two, the more likely it is that the results from one organism can be extrapolated to the other.To compare the sets of proteins between organisms, we propose a procedure that involves: a) associating a protein to a process or pathway, for example using GO categories or the KEGG pathways, and b) compare the sets of proteins associated to the process between the relevant organisms. This method offers a proxy for establishing probable equivalency of processes between organisms, but it has some drawbacks.

First, more often than not, there will be little functional information associated to the proteins of a given organism. To overcome such a problem, we propose choosing an initial subject organism that is well studied and functionally well characterized at the molecular level. As our method relies on ortholog identification and functional annotation, it requires that this annotation be continuously improved even in well studied organisms. By choosing S. cerevisiae as an example we use the eukaryotic organism that we believe has the best overall functional annotation. It must also be emphasized that, when comparing the set of proteins that participate in a given process in different organisms, one must consider the “super set” of proteins participating in that process and compare the differences. In other words, for example when comparing KEGG pathways, one can consider the pathway that includes all possible EC numbers and then compare the two organisms in this context. This was also done here. Otherwise, one may find a situation where two organisms are predicted as being good models with respect to a given process when the proteins in one organism are a small subset of those in the other.

Second, using sequence similarity to establish functional orthology also has its drawbacks. On one hand, sometimes functional orthology exists even in the absence of sequence orthology and vice versa. Comparing the structures of proteins as well as their amino acid motifs and active centers provides some assistance in tackling this problem. However, at the current stage of development in bioinformatics, sequence comparison is still the most efficient and accurate way to make such predictions on the scale that we made them for this work. On the other hand, sometimes, due to gene duplication and domain shuffling, proteins that are unique in one organism may have several close sequence homologues in another. We address this problem by proposing a procedure that takes several similarity factors between sequences into account before deciding which of the homologues is the more likely to be orthologous to the query protein. These factors include e-value score, similarity of the sequences and the fraction of the two proteins that is comparable. Nevertheless, if one also analyzes homologues separately, as we also do here, one stands a better chance of controlling for false negative orthologs.

Third, by comparing only the set of proteins associated with a given biological process in different organisms, we are disregarding regulatory and dynamic information that could be important for the comparison. This shortcoming may not be problematic. On one hand our method is a good way to eliminate processes and organisms for which the reference organism is not a good model. If the sets of proteins that execute a given process are very dissimilar, then the dynamics are not even an issue because other model organisms need to be chosen. On the other hand, having a more similar set of proteins associated to a specific process makes it more likely that the adaptive and regulatory responses of the process be similar. This claim can be supported by comparing the physiological responses of different organisms to that of the model organism (see below).

Fourth, sometimes the logic used to define the proteins associated to specific biological pathways or processes is questionable. This is a very important factor and a successful general application of the method described here requires that the annotation of genomes and ontologies/pathways keeps on improving. Poorly characterized biological processes will lead to greater errors in the comparisons. There is little we can do with respect to this limitation at this time. One of the actions that can be taken to minimize this problem is to choose as a model an organism that is one of the best annotated worldwide. We did so by choosing S. cerevisiae as a model for the study. This organism has the additional advantages of being well characterized at the molecular level and used to study many biological processes that are important in other organisms. To further ameliorate this problem we carefully curated both the KEGG and GO associations of yeast.

S. cerevisiae as a model organism

We apply our method to a pilot study of S. cerevisiae as a model organism, by comparing it to 704 other organisms. The results are presented in detail in Figures 1, 2 and 3, Figures S1–S5 and Tables S1–S4. In S. cerevisiae 4571 proteins are not associated to any pathway in the KEGG database. Analyzing the approximately 1000 proteins that have such a functional association, we find that, as expected, in many cases evolutionary closeness goes on par with similarity between sets of proteins that are associated to a specific biological process.

As mentioned above, our inference of closeness between S. cerevisiae and the other organisms is based upon an analysis of similarity between the sets of proteins involved in a specific process in both organisms. This analysis does not include any information about the physiological responses and the dynamic or regulatory aspects of the biological processes and pathways being compared between organisms. To understand if this limitation is in general important we selected pathways for which dynamic, regulatory, and phenotypic information was available for S. cerevisiae and for a scope of other different organisms. We then compare the behavior of those pathways in yeast and in the other organisms. In this comparison, organisms that are predicted to be closer for a specific pathway or process also have more similar adaptive responses (Table S4). Furthermore, recent work that uses orthology between human genes and those in other organisms to find models for human diseases support these results [32], [33], [34]. Together, this suggests that our method is adequate both for eliminating unsuitable model organisms and for choosing an appropriate model organism from which to study and extrapolate the dynamical and adaptive behavior of specific biological processes.

Conclusion

Our results support the use of S. cerevisiae as a model organism to study different biological processes and pathways in specific organisms, while pinpointing specific processes in this yeast that may not be readily generalizable to other organisms. We conclude that using a single proteome as a reference and applying a methodology such as the one suggested here, one can in general appropriately select model organisms to study the dynamic and adaptive responses of a given biological process, as long as the proteins that participate in that process are known.

Selection of genome sequences

The complete proteome of Saccharomyces cerevisiae (5880 proteins) was downloaded from NCBI (December 2009). The complete sequences for the full protein complement of 704 organisms with fully sequenced genomes was downloaded from the KEGG database (December 2009) and cross-referenced to that provided the NCBI database.

Homology analysis

We downloaded BLAST version 2.2.18 from NCBI and used it locally. All genome and protein sequences were formatted using FormatDB. A pipeline for selecting orthologous proteins, homologous proteins and proteins of the S. cerevisiae that are absent in each of the other organisms was developed and implemented in PERL.

Orthology analysis

The collection of all proteins in a target genome that blasted against a specific protein of S. cerevisiae with an e-value ≤10⁻¹⁰ was analyzed. Manually and through the comparison of the S. cerevisiae proteome to that of two organisms from each class, we setup a cutoff value for separating orthologs from homologues. Pairs of proteins with e-value between 10⁻¹⁰ and 10⁻³⁶ and identity score below 30% are considered as homologues. If the alignment spans over 85% of either sequence and either the e-value of the blast search is bellow 10⁻³⁶ or the identity score is higher than 30%, both proteins are considered as belonging to the same family of orthologs [35]. When more than one protein in a target genome meets these conditions with respect to the same S. cerevisiae protein we calculate an orthology score function, F. The protein with the highest F-score function is considered to be the most likely ortholog with respect to the S. cerevisiae protein, while the remaining proteins are flagged as in-paralogs of that ortholog. F is defined as follows:(Eq.1)

Factor F1 is calculated as follows.(Eq.2)

In Eq. 2, S represents the similarity score and I represent the identity score of the alignment. F1 is always between 0 and 1. The more similar two sequences are, the closer to 1 will F1 be.

Factor F2 is calculated as follows.(Eq.3)

In Eq. 3, AL represents the length of the alignment, and PL is the total length of the query sequence. F2 is always between 0 and 1. The larger the fraction of the query sequence that aligns with the target sequence is, the more similar the two proteins will be and the closer to 1 will F2 be.

Finally, factor F3 is calculated as follows.(Eq.4)

In Eq. 4, G1 represents the number of gaps within the aligned region of the query sequence, L1 represents the length of the query sequence, G2 represents the number of gaps within the aligned region of the target sequence, and L2 represents the full length of the target sequence. The closer to zero F3 is the more similar will the two sequences be.

Theoretically, -∞≤F≤2. However, in practice, we found that F typically assumes values between 0 and 2. The higher F is, the more likely it is that the query and target sequence are orthologs.

The whole process is summarized in Figure 4. At the end of the analysis we obtain clusters of orthologs (ScCOGs) and homologues (ScCHGs) for all the S. cerevisiae genes with respect to the other 704 organisms. We also obtain a third family of clusters (ScCAGs), that of proteins from S. cerevisiae that are absent from the target genomes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Summary of the process used to build ScCOGs, ScCHGs and ScCAGs.

The full proteome of S. cerevisiae was compared to the full proteome of each of 704 different organisms using BLAST. See methods for details.

https://doi.org/10.1371/journal.pone.0016015.g004

Classification of clusters according to pathways and biological processes

In order to attribute biological function to the ScCOGs, ScCHGs and ScCAGs, we implemented the following procedure. On one hand, we used the GOSLIM classification of gene function for S. cerevisiae from SGD [36], [37] to attribute biological function, molecular functions and cellular localization to each cluster. On the other, we downloaded data from KEGG that associates genes to KEGG metabolic circuits in fully sequenced genomes [38] and attribute pathways terms to each of the clusters.

Calculation of the Hamming distance

The Hamming Distance (HD) between the vector,. of protein functions associated to a specific process, localization or pathway in S. cerevisiae and the vector, . of corresponding protein functions in another organism gives a measure of how different the two vectors are. It is calculated using the formula . where, is the Kronecker delta, is 1 if the elements in position i of both vectors are orthologs and 0 otherwise. The smaller the distance, the more similar the two vectors are and the more similar is the set of genes executing a specific process in both organisms. HD can be normalized (NHD) by dividing it by the maximum HD between corresponding vectors of all organisms. Consequently, the smaller NHD is, the more likely that S. cerevisiae is a good model to study the relevant process or pathways and generalize the results for the other organism. The vectors we define for each pathway include all proteins that could participate in that pathway in all organisms in the KEGG database. This ensures that the comparison we are making accounts for differences between the pathway in S. cerevisiae and that in the other organism and vice-versa. All calculations were performed using Mathematica [39].