Predicting Protein Phenotypes Based on Protein-Protein Interaction Network

Background Identifying associated phenotypes of proteins is a challenge of the modern genetics since the multifactorial trait often results from contributions of many proteins. Besides the high-through phenotype assays, the computational methods are alternative ways to identify the phenotypes of proteins. Methodology/Principal Findings Here, we proposed a new method for predicting protein phenotypes in yeast based on protein-protein interaction network. Instead of only the most likely phenotype, a series of possible phenotypes for the query protein were generated and ranked acording to the tethering potential score. As a result, the first order prediction accuracy of our method achieved 65.4% evaluated by Jackknife test of 1,267 proteins in budding yeast, much higher than the success rate (15.4%) of a random guess. And the likelihood of the first 3 predicted phenotypes including all the real phenotypes of the proteins was 70.6%. Conclusions/Significance The candidate phenotypes predicted by our method provided useful clues for the further validation. In addition, the method can be easily applied to the prediction of protein associated phenotypes in other organisms.


Introduction
Identifying phenotypes of proteins is a central challenge of the modern genetics in post-genome era.The study on phenotypes always involves many major diseases, such as HIV [1,2,3,4,5], different kinds of cancers [6,7,8,9], chronic liver diseases [10], Gaucher disease [11].The high-throughput phenotype assays [12,13] combining with gene perturbation technology [14,15] provide fast identification for gene active in a response [16].For example, yeast mutant strain collections has become increasingly used to identify the phenotypes [17].However, these assays are often trapped in the high false negative rates [18].On the other hand, the study on phenotypes is highly complex for the multifactorial trait often results from contributions of many proteins.Consequently, using experimental approaches alone is insufficient, and the computational methods should be applied for the identification of protein phenotypes [18].
In principle, there are two kinds of computational methods: the sequence-based methods and network-based methods.A sequence-based method is often designed on a benchmark dataset, sequence features such as amino acid composition [19], pseudo amino acid [20] (PseAAC), are used to represent the data (e.g. protein sequence), then a prediction model can be built according to the machine learning algorithm (e.g.nearest neighbor algorithm).In the past decade, a series of predictors have been designed for phenotype prediction.For example, Resch W et al. used a neural network model to identify the phenotype of HIV type 1 from loop 3 sequences [21].Pillai S et al. proposed a classifier based on support vector machine for V3 phenotype prediction [22].Recently, Onuki R et al. also employed a support vector machine method for predicting phenotype from genotype data [23].With the ever-increasing build-up of high-throughput techniques, biological data acquisition has never increased more rapidly.More and more biological networks, such as gene-regulatory networks and metabolic networks are constructed from multi data sources (e.g.microarrays, literature mining, and protein-protein interaction).Consequently, many network-based methods are proposed to contribute to various aspects of biology, including phenotype prediction.For instance, Keleta C et al. implemented the prediction of the 16 different growth phenotypes in E.coli based on regulated metabolic networks [24].McGary KL et al. demonstrate that the loss-of function Saccharomyces cerevisiae phenotypes are predictable in the functional gene network, and the proposed network-based method succeeded in the identification of yeast orthologs of human disease genes.
In this research, we presented a new network-based method for predicting budding yeast protein phenotypes.Unlike previous methods, our method can rank the possible phenotypes associated with the query protein and shows a more comprehensive view of the protein's biological effects.With the results, we also demonstrated that using protein-protein network is effective for predicting protein phenotypes.Owing to many protein-protein network of other organisms are available, we suggest that this method will be widely applied.

Data Set
Because of the complexity of phenotype research, we selected the budding yeast Saccharomyces cerevisiae (a well studied model organism [25,26]) as a model system.The protein data used here was taken from CYGD [27] (the MIPS Comprehensive Yeast Genome Database, ftp://ftpmips.gsf.de/yeast/), which dedicated to information on the molecular structure and functional network of the budding yeast.Among the 6,732 proteins of the yeast proteome, only those with both sequence and phenotypic annotations were selected.Thus we obtained 1,460 such proteins belonging to 11 phenotypic categories (see Table S1).The number of proteins in each category was listed in the Table 1, from which we can easily find that the total number of proteins (2,397) in 11 phenotypic categories is much larger than the total number of proteins (1,460).That is because many proteins exhibit more than one phenotype and this is the reason why we developed this method to predict the possible phenotypes with ranked scores, rather than only one predicted phenotype like previous tools.
The yeast protein-protein interaction (PPI) network used here was retrieved from STRING [28] (http://string.embl.de/),whose primary mission is to provide researchers with both physical (direct) and functional (indirect) interactions.For each species, a PPI network is constructed by integrating huge information derived from numerous sources such as experimental repositories, computational methods, and text-mining methods.In the functional protein association network, the interaction unit consists of two nodes (proteins) and an edge between them.The interaction confidence score is used as the edge weight to represent the likelihood that a predicted association exists between two nodes.Weight confidence limits are as follows: low confidence 215% (or better), medium confidence 240%, high confidence 270%, highest confidence 290%.In this research, we chose the highest confidence limit 290% to obtain reliable yeast PPI network (see Table S2), which contains 32,513 functional linkages among 4,209 yeast proteins.
Among the 1,460 proteins with phenotypic annotations, 1,267 proteins could be mapped to the yeast PPI network downloaded from STRING.Thus, the nodes in the network could be grouped into two kinds: those with phenotypic information, others without phenotypic information.Here, we called the protein with phenotypic annotation in the PPI network ''seed protein'', and the dataset consisting of 1,267 seed proteins ''seed set'', which were then used to test the network-based method.
The availability of using the PPI network to predict protein phenotypes In the functional network, PPI contains both physical (direct) and functional (indirect) interactions.Physically interacting proteins exist in the same complex, while functional interacting proteins tend to participate in the same pathway or cellular process.Here, we investigated the relationships between complex/ pathway and phenotype to explain the availability of using the PPI network to predict protein phenotypes.In order to analyze the relationship conveniently, we selected the proteins with single phenotype.The complex annotation of proteins was also downloaded from CYGD [27], and the pathway annotation of proteins was retrieved from KEGG [29] (Kyoto Encyclopedia of Genes and Genomes) (see Table S3 and Table S4).Totally, these proteins belonged to 733 complexes and 86 pathways.Each protein was coded by the vectors: V complex ~½v 1 ,v 2 ,:::,v i ,:::v 733 V pathway ~½v 1 ,v 2 ,:::,v i ,:::v 86 ð1Þ where v i ~1 if the protein belonged to the i-th complex/pathway, otherwise v i ~0.Then m-th phenotype can be represented by the protein complex/pathway information as the vector: where n is the number of proteins that had the m-th phenotype.
The similarity between any two phenotypes was calculated as: C~V p : V q jjV p jj : jjV q jj (p~1,2,:::,11; q~1,2,:::,11; p=q) where V p : V q is the vectors' inner product, jjV jj is the module of vector.Generally, two phenotypes are difficult to discriminated from each other using the complex/pathway if the value of the similarity of them is larger than 0.5.Using the protein complex information, the distribution of the similarities of 11 phenotypes was shown in Figure 1.Clearly, all the 55 similarities are smaller than 0.5.Because the proteins with the phenotype of sensitivity to immunosuppressants lacked the pathway annotation, the similarities of other 10 phenotypes were calculated using the protein pathway information.The distribution of the similarities of 10 phenotypes was shown in Figure 2, where two thirds of the similarities are smaller than 0.5.The results indicate the phenotypes can be classified by using protein complex/pathway information.Therefore, protein phenotypes can be predicted by using the functional PPI network.

Network-based Method
In the PPI network, when we were to predict the phenotypes of a node (protein), just like the weighted vote, not only the number of its neighbor nodes, but also the strengths of interactions (i.e., the edge weights) were considered by the method.The phenotypic categories of each protein in the network can be predicted as following.
First, let us consider the PPI network consisting of n proteins fP 1 ,P 2 ,:::,P n g, in which seed proteins belonged to 11 phenotypic categories (T~½T 1 ,T 2 ,:::,T 11 ), where T 1 represents the ''Conditional phenotypes'' category, T 2 the ''Cell cycle defects'', T 3 the ''Mating and sporulation defects'', and so forth (cf.Table 1).And the phenotypes of the i-th protein in the network can be denoted by T(P i )~½t i,1 ,t i,2 ,:::,t i,j ,:::,t i,11 T (i~1,2,:::,n; j~1,2,:::,11) ð4Þ where Towards a query protein P k , its interaction weights with m seed proteins can be defined as follows W (P k )~½w k,1 ,w k,2 ,:::,w k,i ,:::,w k,m T (i~1,2,:::m) ð6Þ where w k,i is the interaction weight (confidence score [28]) between P k and the i-th protein in the seed set.If there is no edge between them, w k,i ~0.Since we did not consider the selfinteraction of protein, w k,i ~0 when k~i.Subsequently, we proposed a new concept called ''tethering potential'' of protein P k to the j-th phenotype to reflect the potential of protein P k belonging to the j-th phenotype, which can be calculated as follows From this equation, we know that the proteins in seed set without association with the query protein do not contribute to the score of S(P k [j).Thus the tethering potential of protein P k to the j-th phenotype can be also described as the sum of interaction weights of it with neighbor proteins of the j-th phenotype in seed set.Obviously, the larger the value of S(P k [j) is, the more likely the protein P k belongs to the j-th phenotypic category.Therefore, the most likely phenotype of the query protein P k can be predicted to belong to the m-th phenotypic category as follows m~arg max j S(P k [j) j j~1,2,:::,11 where m stands for the argument of j that maximizes the value of S(P k [j).However, many proteins in yeast give rise to more than one phenotype; the prediction result with only the most likely candidate phenotype is insufficient.In view of this, to make the method able to handle the proteins with multiple phenotypes and benefit biologists with more flexible information in prioritizing candidate phenotypes, we introduced a 11-D (dimensional) vector to reflect the likelihood that the query protein may give rise to each of the 11 phenotypes, which can be formulated as follows  [5), …, then that the query protein P k gives rise to the 1 st phenotype (Conditional phenotypes) will have the maximum likelihood, that P k gives rise to the 7 th phenotype (Carbohydrate and lipid biosynthesis defects) will have the second maximum likelihood, that P k gives rise to the 5 th phenotype (Cell morphology and organelle mutants) will have the third maximum likelihood, and so forth (cf.Table 1).In rare cases, when more than one element of the vector in Eq.6 has the same value, the order will be randomly sorted.Based on the descending order of Eq.6, the predicted results are respectively called the 1 st -order predicted result, the 2 nd -order predicted result, the 3 rd -order predicted result, and so forth.

Jackknife Cross-validation and Evaluation
In statistical prediction, three cross-validation methods are often used to examine the prediction quality: subsampling (K-fold) test, independent dataset test and jackknife test [30].
Among the three methods, jackknife test is regarded as the most objective as discussed in Chou's work [31,32] and has been used more and more frequently to test and evaluate various predictors [33,34,35,36,37,38,39,40,41,42].In this research, the jackknife cross-validation was also applied to test the network-based method.During the validation, each protein in the seed set is in turn knocked out as a query protein sample, and the remaining proteins of the seed set in the PPI network are used for prediction by the network-based method.Thus, the i-th order prediction accuracy Q i can be calculated as follows Where M ij is the number of correctly predicted proteins of the jth phenotypic category in the seed set, and N is the total number of proteins in the seed set.Finally, the 11-order prediction accuracies are obtained to evaluate the network-based method.The large Q i with a small i and the small Q i with a large i imply a good performance of the method.
The average number of phenotypes that each protein in the network exhibits can be calculated as follows Therefore, another evaluation for the network-based method was proposed as the likelihood that the first r-order predicted results include all the phenotypes of proteins, which can be calculated as follows A large P r accompanied with a small r also implies a good performance of the method for the protein phenotype prediction.

Prediction
Besides the seed proteins, there are also 2,942 proteins in the PPI network.The tethering potential of such protein to the each phenotype can be calculated according to Eq. ( 7) and then ranked in descending order.In this manner, the phenotypes of these proteins can be predicted by the network-based method.

Performance of Network-based method
Through leave-one-out cross-validation, the overall 11-order success rates by the network-based method on the aforementioned 1,267 seed proteins are listed in Table 2.As we can see from the table, the most likely (first-order) prediction accuracy is 65.4%, and the least likely (last-order) one is 3.39%.The former minus the latter equals 61%.Based on the prediction criteria, the bigger the difference value is, the better the method performs.According to Table 2, a downward-slope curve is drawn in the Figure 3, showing that higher-order phenotype prediction is better than the lower-order one.This is the exact phenomenon that we want to see, and it may imply that the predicted phenotypic categories of proteins are well arranged by the method according to the prediction criteria.
The average number of phenotypes that each seed protein has is 1.7 according to Eq. ( 11).The chance that a random guess of a protein phenotype will succeed is 1.7/11 = 15.4%,much lower than the first order prediction success rate.As is shown in the Table 2, the first 3 prediction accuracies are larger than the success rates of random guess.And the likelihood of the first 3order predicted results including the phenotypic categories of the proteins in seed set is 70.6% according to the Eq. ( 12).These results may imply that our method performs well in the prediction of protein phenotypes in budding yeast.
In genetics, mutations that cause the same phenotype are inferred to functionally associated, and vice versa [18].Phenotype is a multifactorial trait that often results from the contribution of many proteins.Because the interacting proteins are often in the same complex or pathway, it is rational to expect that interacting proteins often share the common phenotypes.For example, the interactions of seed protein YBR039W with the other seed proteins are listed in Table 3.The complex information about those proteins is retrieved from CYGD [27].We can easily see that protein YBR039W and its neighbors YBL099W, YDL004W, YDR298C, YLR295C, YML081C-A, YPL078C, YPL271W are members of the same F0/F1 ATP synthase (complex V) complex.Additionally, proteins YDR298C and YPL078C are also members of complex in [43], and protein YPR024W is component of Yme1 protease complex.And these proteins share the common phenotype auxotrophies, carbon and nitrogen utilization defects.Therefore, when protein YBR039W is predicted as a test sample by the method, the first candidate phenotype will be assigned its real phenotype.For another example, the interactions of seed protein YDL028C with the other seed proteins are listed in Table 4.The information of pathways that yeast proteins participate in is retrieved from Kyoto Encyclopedia of Genes and Genomes [44] (KEGG).Except proteins YDR168W, YKL042W, YPL209C with no pathway annotation, proteins YDL028C, YBL084C, YGL116W, YGR113W, YGR188C, YIL106W, YKL022C, YMR055C, YOR026W involve in the same pathway sce04111 (Cell cycle in budding yeast).The loss-of-function of any one of these 9 proteins likely disrupts the mitotic cell cycle progression and lead to cell cycle defects.Based on the interactions listed in the table, we can arrange the first, second candidate phenotype of protein YDL028C as the cell cycle defects, cell morphology and organelle mutants respectively according to the prediction criteria.The correct phenotype predictions of proteins YBR039W and YDL028C support the hypothesis that the functional associated proteins often share the same phenotypes.Therefore, the protein phenotypes can be predicted from the phenotypes of its interacting proteins by the method.
Table 2.The leave-one-out cross-validation (Jackknife test) success rates by a random guess and the network-based method.Protein phenotype prediction with inactivating its interacting protein

Most likely category
Here, we discuss the robustness of our method by applying the method to the proteins whose interacting proteins are inactivated.First, we chose a protein and took away one of its interacting proteins from the PPI network.Then the phenotype of the protein was predicted by the method based on the broken PPI network.In this way, the phenotypes of 6 proteins were predicted, as shown in Table 5.The phenotypes predicted from the unbroken network and the recent phenotype studies focusing on these proteins are also listed in Table 5.We found that the phenotypes predicted from the broken network were different from the phenotypes predicted from the unbroken network, while the proteins were verified to have these new phenotypes predicted from broken network in the recent studies.For example, with protein YOR196C in the network, the 1 st order predicted phenotype of protein YER178W by the method is ''auxotrophies, carbon and nitrogen utilization defects'', which is the same as the annotation from CYGD [27].After inactivating protein YOR196C, the phenotype of protein YER178W is predicted as the ''conditional phenotypes''.In the study [45], protein YER178W was reported to have the phenotype-''Heat sensitivity: increased'', which is one kind of ''conditional phenotypes'' according to the phenotype classification in CYGD.In the table, the new phenotypes of other proteins predicted from the broken network can also be supported by the literatures [13,46,47,48,49,50].The examples listed in the table indicate that our method may provide new phenotypes for proteins and serve as a complementary tool for the existing resources.

Application and improvement
As is discussed above, the first 3-order predicted results (approximately double the average number of phenotypes 1.7) can be considered as the candidate phenotypes of the proteins concerned by the biologists.Genetic experiments can focus on these candidate phenotypes of the proteins, which may accelerate the research progress and decrease the cost.At least, the last three predicted phenotypes can be excluded because the last 3-order prediction accuracies are lower than 5% (See Table 2).
The effectiveness of the functional network for predicting phenotypes of proteins in yeast suggests the possibility of application to other species.The method is based on the functional protein association network.Besides an abundance of such networks in STRING [28] (Version 8.0 of STRING covered 630 networks of different organisms), the PPI networks can also derived from worm PPI database [51], fly database [52], human PPI database [53,54,55], and so on.When in possession a series of proteins with known phenotypes, one can predict the possible phenotypes of other proteins in the networks.Therefore, the method can be easily applied to the prediction of protein phenotypes in other organisms, especially model organisms.
The performance of our method can be improved if the following problems are solved.First, increase the quality of PPI network and exclude the false positive interaction; currently we used high confidence score cutoff to filter the network (See section Data Set).Second, proteins in the same complex or pathway may exert opposite effects on a phenotype, playing as actors or repressors [18].If the network can discriminate the positive or negative regulation, our method can be modified and the performance will be improved.Third, the performance of the network-based method depends on the number of seed proteins.This problem can be solved in future when the phenotypes of more proteins are investigated.In summary, identification of protein phenotypes is an extremely complicated work and there is a long way to go.

Conclusion
In this research, we proposed a multi-target model [40] to predict phenotypes of proteins in budding yeast based on the protein-protein network.Because some proteins can give rise to more than one phenotype, rather than the most likely phenotype, a series of candidate phenotypes are predicted for each protein.
With the performance of the method, it is anticipated that the promising approach may serve as a useful tool for annotating the phenotypes for uncharacterized protein sequences.

Table 1 .
Breakdown of 1,460 budding yeast proteins according to their 11 phenotypes.
9 Sensitivity to amino acid analogs and other drugs 124 10 Sensitivity to antibiotics 43 11 Sensitivity to immunosuppressants 14 See the texts of the paper for further explanation.doi:10.1371/journal.pone.0017668.t001

Table 3 .
Interactions of protein YBR039W with its neighbor proteins.

Table 4 .
Interactions of protein YDL028C with its neighbor proteins.