Learn more: PMC Disclaimer | PMC Copyright Notice
Prediction of Human-Plasmodium vivax Protein Associations From Heterogeneous Network Structures Based on Machine-Learning Approach
Associated Data
Abstract
Malaria caused by Plasmodium vivax can lead to severe morbidity and death. In addition, resistance has been reported to existing drugs in treating this malaria. Therefore, the identification of new human proteins associated with malaria is urgently needed for the development of additional drugs. In this study, we established an analysis framework to predict human-P. vivax protein associations using network topological profiles from a heterogeneous network structure of human and P. vivax, machine-learning techniques and statistical analysis. Novel associations were predicted and ranked to determine the importance of human proteins associated with malaria. With the best-ranking score, 411 human proteins were identified as promising proteins. Their regulations and functions were statistically analyzed, which led to the identification of proteins involved in the regulation of membrane and vesicle formation, and proteasome complexes as potential targets for the treatment of P. vivax malaria. In conclusion, by integrating related data, our analysis was efficient in identifying potential targets providing an insight into human-parasite protein associations. Furthermore, generalizing this model could allow researchers to gain further insights into other diseases and enhance the field of biomedical science.
Introduction
Plasmodium is a parasite that has proven to be difficult to eradicate. Plasmodium vivax is 1 of the 5 species of the parasite group Plasmodium that infects humans. 1 P. vivax has the ability to confer virulence to humans and survive in human hosts and has been categorized as a benign infection. At present, P. vivax malaria is recognized as a cause of severe morbidity and mortality. 2 Approximately, 14.3 million cases of P. vivax infection are recorded annually. 3 Although the global incidence of P. vivax malaria infection has decreased by 42% since 2000, the disease burden has increased in the Middle East and South America since 2013. 4 In addition, P. vivax is able to evolve its strategy to interact with the host, which has led to the development of drug-resistant parasites. The first-line treatment drug for P. vivax is chloroquine to treat blood-stage parasitemia together with primaquine to eradicate persistent liver-stage infection. 3 However, P. vivax parasites resistant to their respective first-line therapies have been found in Southeast Asia. 5 Recently, tafenoquine, a promising new drug, has been highlighted as a radical cure for P. vivax infection. Results have shown that it resulted in a significantly lower risk of P. vivax recurrence than placebo in patients with normal glucose-6-phosphate dehydrogenase (G6PD) activity. 6 However, tafenoquine causes hemolysis in patients with G6PD deficiency. Therefore, there is a need for testing G6PD activity before prescription of tafenoquine.7-9 The Plasmodium parasite has the ability to evade the human immune system, recruit host responses to regulate its life cycle, and adapt to the host environment. 10 Specifically, P. vivax invades erythrocytes during blood-stage growth in humans. Duffy antigen receptor for chemokines (DARC), which is a host receptor, is recognized by a critical invasion ligand, P. vivax Duffy Binding Proteins (DBP), for the invasion of immature red blood cells. 11 Therefore, DBP has been highlighted as a leading vaccine candidate against P. vivax malaria. 12 To control this parasite, we require a better understanding of host-parasite interactions which is crucial in the development and design of therapeutic approaches for this infectious disease.
Although recent technological advances in high-throughput techniques have enabled the characterization of proteins that may be involved in the parasitic invasion of target cells, maintaining a continuous in vitro culture for P. vivax is still very difficult to standardize. 13 This is the main obstacle to the development of a new effective vaccine. However, computational methods can be employed to solve this problem. One of the most widely used methods is a network-based approach that focuses on protein-protein interaction (PPI) networks. The analysis of a PPI network has been widely studied in several organisms.14-17 In Plasmodium, several studies have investigated the PPI networks with the aim of revealing many important aspects of protein interactions.10,18-24 Most studies of PPI networks have applied the calculation of degree and centralities, focusing on a single organism in their analyses. In addition, PPI networks have also been used to study the associations between proteins and diseases14,25-27 and host-parasite protein associations.10,18,19,24,28 Saha et al 24 investigated the characteristics of a host-pathogen protein interaction network based on interconnectivity and centrality properties. They analyzed the significance of central, peripheral, hub and non-hub protein nodes in the infection process of malaria. They also found few topologically unimportant but biologically significant proteins between humans and malaria. Notably, most such studies have been performed for Plasmodium falciparum. Several studies have used ortholog-based methods to predict the association of proteins across species.29-33 Specifically, Cuesta-Astroz et al 34 developed a method based on orthologous proteins to identify a transferred interaction between host and parasite proteins. They identified common and specific mechanisms of parasitic infection and survival in 15 human parasites. They also intensively analyzed the human-Schistosoma mansoni protein interaction network and revealed biological processes, pathways, and tissue-specific interactions that may be essential in the life cycle of the parasites. Lee et al 29 predicted PPIs between P. falciparum calmodulin and H. sapiens proteins based on orthologous pairs. From the associations between host and parasite, they found that P. falciparum may use calcium-modulating proteins in the host cell to maintain the Ca2+ levels. Recently, a heterogeneous network has been developed to propagate interaction information from the human PPI network and the P. vivax PPI network to infer new associations between human and P. vivax proteins. 19 This method was based on protein interactions that were considered to globally represent of these 2 networks. The study used protein similarities between human and parasite proteins to establish their associations; the idea behind this is that a malaria protein that is homologous to a human protein may interact or work together with human proteins to maintain their lives in the host and be related to the same set of cooperative proteins in humans. Thus, the study of the relationship between similar proteins in humans and malarial parasites is of great interest to investigate their network topology in PPI networks. Similar proteins may also have the same level of importance in the PPI, as the centrality measures reflect the essentiality of a protein in terms of the network topology and connections under a specific aspect of the measure. For example, the betweenness centrality provides an insight into a node that may be involved with the paths of communication of any pairs of nodes in the network.17,35,36 Therefore, the integration of these network topologies for the recognition of human-parasite protein associations via machine learning has the potential to provide important insights and reveal new associations and protein targets in human hosts.
In this study, alternative properties based on local network topology features and machine-learning techniques were used to elucidate new associations between human and P. vivax proteins. The associations presented in this study indicate the existence of functional interactions between human and P. vivax proteins, implying that these proteins cooperate to perform a task in the underlying mechanisms. A ranking technique was also developed to predict potential protein targets in humans which may be important for the treatment of P. vivax malaria. Clustering analysis was performed using information from the heterogeneous network analysis to identify groups of related proteins and functional proteins. Finally, a list of human proteins that are crucial for the cellular mechanisms of P. vivax was reported and validated via a literature search. This list may be useful in further studies that wish to develop drugs for the treatment of P. vivax.
Materials and Methods
Overview of the analysis framework
The analysis framework was initiated with the network reconstruction process as shown in Figure 1. First, PPI networks for humans and malarial parasites were constructed based on the interaction information obtained from the STRING database. 37 Each protein node in each network was then extracted for its network topological features such as the degree and the betweenness centrality. Subsequently, both networks were linked together to form a heterogeneous network based on their protein sequence similarity. Then, the topological features of a pair of human and malaria proteins were compared and evaluated to obtain the strength of the differences and to build a similarity profile of the human-parasite protein pairs. The protein sequence similarities obtained from BlastP searches (E-value ⩽ 1e−05) were then used as an initial class label of a pair of human and P. vivax proteins. The complete profile was then applied to various machine-learning techniques (naïve Bayes, neural network, random forest, and support vector machine). Cross-validations were performed for each technique, and the performances were measured using the receiver operator characteristic (ROC) curve. The top classifiers from the best technique were selected as models to predict new potential associations. Finally, the human proteins in the list of predicted associations were ranked to identify potential protein targets for malaria invasion in the human host.
Network construction and topology features
Our analysis was performed on PPI networks of human proteins and P. vivax proteins. The networks were obtained from the STRING database (version 11.0). 37 To ensure that only reliable interactions were obtained, interactions with a high confidence score (>900) were retained. A total of 12 038 human proteins with 313 359 interactions and 1787 P. vivax proteins with 11 477 interactions were obtained. Subsequently, a heterogeneous network was constructed by connecting human-human protein interactions and P. vivax-P. vivax protein interactions with the human-P. vivax protein associations.
The network topology features of all proteins were extracted based on centrality measurements. Several studies have shown that a relationship exists between gene essentiality and network centrality in PPI networks.38-40 Thus, we further investigated 5 topological features: betweenness centrality, closeness centrality, degree, eccentricity, and Kleinberg’s hub centrality. Each of these features explained different aspects of the measurement. Betweenness centrality reflects an important node in term of overloading paths passing through it in the communication of the network.35,36 Closeness centrality measures how close a given node is to the other nodes in the network.35,36 The degree represents the level of the local connections of a given node.35,36 Eccentricity calculates the local density of the connections among neighboring nodes of a given node. The Kleinberg’s hub measures the importance of a given node connecting the other important nodes. 36
Defining the human-P. vivax protein associations
To define the initial associations between human and P. vivax proteins, we used the information obtained from a sequence similarity search. When 2 protein sequences shared significant similarity with the BlastP expectation value (E-value) less than 1e−05, they were inferred to be homologous. This means that they did not arise independently, but rather shared a common ancestor. 41 Therefore, we could define an association between 2 sequences when they share more similarity than that would be expected by chance. However, when no statistically significant match was found between the 2 protein sequences, we could not ensure that no homologs were present. Thus, the machine-learning method may be able to reveal hidden homologs. The P. vivax protein sequences were retrieved from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database42,43 using the Rcpi package 44 and then searched against all human protein sequences from the NCBI database. We defined that 2 protein sequences were homologous when BlastP (https://blast.ncbi.nlm.nih.gov) gave rise an E-value less than 1e−05. Then the pair of these 2 proteins was labeled to be associated.
In addition, the relationship between network topologies and functions has been revealed in several studies with the assumption that for each function, the wiring patterns of the proteins are similar. 45 Different standard network topologies can be used to understand the information contained in the wiring of a protein in the PPI.45,46 Therefore, we integrated initial associations from the protein sequence similarity search and the similarities from network topological features and fed them into machine-learning algorithms to predict new associations using both types of similarity information. It is worth noting that our method is a homology-based method that relies on sequence similarity, similar to previous studies.29-34 Protein associations were predicted based on the initial associations from sequence similarity. Moreover, homology-based methods have been used to infer functionally interacting proteins in previous studies.29-34
Features of topological differences for machine learning
Based on the 5 network topology features, we established a vector , that is a similarity profile, representing a relationship between the topological values of a human protein and a P. vivax protein , as follows
where , i = 1, 2,. . ., m and j = 1, 2,. . ., n. m and n are the number of human and P. vivax proteins, respectively. k is the index for each topological feature, ranging from 1 to 5. represents the kth centrality value of a human protein and represents the kth centrality value of a P. vivax protein, . Therefore, denotes the topological similarity between the kth centrality values of human protein i and P. vivax protein j. A low value of indicates a high similarity between the topological features k of these 2 different types of proteins.
Training and validating of the association classifiers and calculating association scores
We investigated all possible pairs of proteins to identify human-parasite protein associations. To this end, we employed machine-learning techniques to classify defined and undefined associations. Four classification algorithms, namely naïve Bayes, neural network, random forest, and support vector machine algorithms, were employed. Each of these classifiers is a well-known algorithm for recognizing and creating classifiers in different ways. The naïve Bayes’ approach uses the statistics and likelihoods to make a final decision. A neural network calculates a set of optimal weights for a weighted network structure to separate different classes based on the features. Random forest creates complex and hierarchical rules along the features to provide a predicted class. The support vector machine builds a hyperplane to identify an optimal classifier with maximum margin. With the different calculation methods to search for the best solution for the classifier, all 4 classifiers were applied to search for the best classifier. Different parameters of each algorithm were optimized to determine the optimal models of each algorithm.
For the naïve bayes classification, we tuned 3 hyperparameters. The first parameter was to allow to use a kernel density estimation or a Gaussian density estimation. The second parameter was used to adjust the bandwidth of the kernel density when using kernel density estimation. Using this parameter, we optimized it from 0 to 5. The third parameter was the parameter for the Laplace smoother, which we tuned from 0 to 5.
For neural networks, we optimized the number of units in the hidden layers (H) and weight decay to avoid overfitting (d) by employing a grid search with H = 1, 2, 3,. . ., 10 and d = 0.5, 0.1, 1e−2, 1e−3, 1e−4, 1e−5, 1e−6, and 1e−7. The maximum iterations were set to 1000.
For the random forest algorithm, we varied the number of variables randomly sampled at each split time with a value of 2n for n ∈ {0, 1, 2, 3, 4, 5}.
For the support vector machine, we used a radial basis kernel, and optimized the cost of false classification (C) and kernel width (γ) by employing a grid search with C = {0.75, 1.0, 1.25} and γ = {0.01, 0.015, 0.2}.
Ten 10-fold cross-validations were performed to evaluate the performance of the classifiers. At each time, the undefined association set was randomly selected with an equal size to the defined set. A total of 80% of these data were used to optimize the parameters using the cross-validation technique. At each time of the cross-validation, the defined and undefined associations were randomly split into 10 equal sizes. Nine parts were concatenated and used to train and optimize the parameters. Testing was performed with the remaining part and the performance was measured by comparing the predictions and the true class labels. This experiment was repeated with a randomly undefined set 10 times. Several cutoffs on the probabilities of positive class predictions were calculated, yielding an ROC curve, which is a plot of the true-positive rate (TPR) against the false-positive rate (FPR) at the different cutoffs. Using the ROC curve, a broader view of the performance over various cutoffs could be measured by calculating the area under the curve (AUC). An AUC of 1 indicated the best performance of the classifier in which it can recognize and classify the samples, whereas an AUC of 0.5 indicated that the performace could achieve the same as random prediction by chance.
Subsequently, the AUCs of the aforementioned 4 classification algorithms were compared. The algorithm with the highest AUC was used as the prediction model. Ten classifiers from the final model were employed as the ensemble classifiers. Each classifier provided the probabilities of positive prediction for a human-parasite protein pair. The voting score (S) was calculated from the average probabilities of the 10 classifiers. Therefore, the score was computed as follows
where is the probability of a positive prediction derived from the output of the Mth machine. The score was applied to all defined and undefined associations in this study.
Ranking score calculation for each human protein
Using machine-learning algorithms to perform the classifications, we obtained a promising list of human-parasite protein associations. It would be interesting to use these associations to identify human proteins crucial for the P. vivax malaria mechanism. It is worth noting that one human protein could be associated with more than 1 P. vivax protein. To identify the impact of a human protein on the list, we applied a ranking method for all human proteins in the list. The probability of a positive prediction for a pair of human and P. vivax proteins was used to rank the protein pairs. The pair with the highest probability value was ranked first. Notably, several pairs can have the same probability value. In this case, they were assigned the same rank. The ranking score of a human protein was calculated as follows
where is the rank of a pair of a human protein and P. vivax protein , for all possible , according to the prediction probability score of the association.
Gene ontology enrichment analysis
To infer gene functions from the human candidate sets, we employed Gene Ontology (GO) enrichment analysis to determine which GO terms were overrepresented in our candidate proteins. To this end, the Cytoscape 3.7.2 47 plugin ClueGO v2.5.6 48 was used. ClueGO constructed a gene network based on GO terms by employing all differentially expressed genes. A 2-sided hypergeometric test with Benjamin-Hochberg corrections was performed to calculate the significant GO terms. Only GO terms with adjusted p-values less than 0.05 were considered.
Results
Network structures and node properties of human and P. vivax networks
In this study, we constructed 2 PPI networks of human and P. vivax from the information of the STRING database. 37 The reconstructed human PPI network consisted of 12 038 proteins and 313 359 edges, while the malaria PPI network comprised 1787 proteins and 11 477 edges. The structures of the human PPI network and malaria PPI network followed the power-law distribution (Figure 2A and andB,B, respectively), indicating that there are small numbers of high-degree nodes and large numbers of low-degree nodes in the networks. The topological network features of each protein were calculated based on node properties in the networks, namely betweenness centrality, closeness centrality, degree, eccentricity, and Kleinberg’s hub. The deviations of these features are shown as boxplots in Figure 3. Interestingly, both networks had similar average betweenness centrality, degree and eccentricity, but large differences in closeness centrality and a small difference in Kleinberg’s hub. A node with a high betweenness score was indicative of a node with overloading paths passing through it, that is, the node may act as a bridge between 2 or more communities. The boxplot of betweenness centrality scores showed that both human and parasite networks had a similar mean overload for each node in the entire network. Evidently, there were the similar mean of degrees and eccentricities for both networks.
Degree distributions of 2 networks: the degree distributions of (A) human protein-protein interaction network and (B) malaria protein-protein interaction network.
Closeness centrality provides a good measure of a given node located in the middle location, such that it can reach the other nodes in the shortest way. The human network showed lower values of closeness scores than those of the parasite network. This may be due to the fact that, in the human network, there were several proteins, and several protein interactions caused a protein complex, compared to that in the parasite network. Kleinberg’s hub represents the protein nodes that may connect to other important nodes in the network. The boxplot shows that, on average, human proteins are slightly more likely to connect with other important nodes than that are parasite proteins. Although the boxplots show the overall distributions of each node property in the entire network, they do not represent all single differences of each protein in both networks. In addition, these differences may provide a good view of how human and parasite proteins relate to each other in terms of the cooperative community in the network. Thus, the similarity profiles of these topological node properties for each pair of human and Plasmodium proteins were determined. This profile was used as a feature to train the machine-learning classifiers.
We calculated the topological similarity of each feature for each pair of human and Plasmodium proteins. All possible combinations of these 2 types of proteins resulted in 225 675 478 human-Plasmodium protein pairs. Next, the similarity features based on the node properties were calculated (see Materials and Methods) for each pair of human-Plasmodium proteins. Initially, we defined 19 939 pairs as positive association pairs based on protein sequence similarities. The remaining pairs, namely 225 655 539 pairs, were defined as an undefined set. These data sets were prepared to be fed into the established classification processes. Before the classification process, it was interesting to analyze the topology features to determine the relationship between proteins in the positive pairs. We then calculated an uncentered correlation of each node property between human and parasite proteins in the positive set, as shown in Table 1. This uncentered correlation provides the value of the relationship, ranging from 0 to 1. As expected, we found a high correlation of closeness centrality between the human and parasite proteins, with a correlation coefficient of 0.9805. In addition, a moderate correlation of eccentricity between the human and parasite proteins with a correlation coefficient of 0.6827 in the positive set was observed. A low correlation of degree and betweenness centrality between human and parasite proteins was observed, with correlation coefficients of 0.3507 and 0.1316, respectively. With Kleinberg’s hub, no correlation was observed, with correlation coefficient of 0.0556 between human and parasite proteins. The characterization of the topological features of human and parasite protein interaction networks may help to identify underlying proteins that cooperate with host cell recognition and invasion by parasite proteins.
Table 1.
Correlation coefficient values of each topological feature between human and parasite proteins in the positive set.
| Degree | Closeness | Betweenness | Eccentricity | Kleinberg’s hub |
|---|---|---|---|---|
| 0.3507 | 0.9805 | 0.1316 | 0.6827 | 0.0556 |
Performance of the classifications used to recognize human-parasite protein associations
Four classification algorithms, naïve Bayes, neural network, random forest, and support vector machine, were used to recognize human-parasite protein associations. Their performances were compared to select the best classifier for the recognition of human-parasite protein similarities, based on topological features. Ten 10-fold cross-validations were applied for each algorithm, which yielded the performance in terms of an ROC curve with an AUC, as shown in Figure 4. The random forest algorithm provided the best classifier, with an AUC of 0.85. The neural network algorithm yielded a slightly lower performance, with an AUC of 0.79. Similarly, the support vector machine achieved an AUC of 0.77. The naïve Bayes classifier yielded a slightly lower performance compared with that of the neural network and support vector machine with an AUC of 0.74. Notably, the random forest algorithm provided the best performance, with an AUC that was relatively far from that of the other algorithms. This is of great interest because the results obtained for this algorithm indicate its potential in identifying new human-parasite protein associations and, furthermore, in selection of key human proteins for the parasite.
Receiver operating characteristic (ROC) curves for the predictions of human-parasite protein associations of each machine-learning algorithm.
AUC indicates area under the curve; ROC, receiver operating characteristic.
The classifier showed a better performance than that did random selection, which may result in 50% correct predictions. Moreover, we attempted to demonstrate the reliability of the relationship between sequence similarity and network topologies by performing several random experiments. These experiments could be performed by randomly shuffling class labels and retraining the random forest classifiers. Ten 10-fold cross-validations were performed in the same procedures. An AUC of 0.5 was obtained for these random experiments. This was also a good indication that the network topologies of protein nodes in the PPI networks could be used to infer the relationship between human and parasite proteins in terms of sequence similarity, reflecting the homologs and similar cooperation in the network community.
Based on the best performance and the results of the random forest classifiers, we defined a voting score for a pair of human and parasite proteins. Ten probability values of the positive prediction for a pair of human and parasite proteins were obtained. The average of these probability values was calculated and defined as a voting score for a pair of human and parasite proteins (see Materials and Methods). This score was used to define the stringency of predicting human-parasite protein associations. Initially, we identified 12 038 human proteins in the human PPI network and 1787 parasite proteins in the parasite PPI network. This resulted in a total of 225 675 478 human-parasite protein pairs. A total of 19 939 pairs were initially defined as positive association pairs based on protein sequence similarities. After performing the random forest classification, the average voting score was calculated for each pair. It is worth noting that these scores indicated associations based on the network topological profiles of the human-parasite protein pairs using machine learning. It was also interesting to combine these scores with the other association scores from other aspects such as the heterogeneous network study. 19 With the heterogeneous network model, the network propagation algorithm with a decay factor of 0.1 was performed on the network to prioritize human-parasite protein associations. 19 A total of 21 511 906 overlap pairs from both machine-learning and network propagation techniques with scores greater than 0 were obtained and used for the further analysis and selection of key human proteins. Of these pairs, 831 had the highest voting scores of the predictions according to our machine-learning analysis (Supplementary Table S1).
Identifying promising key human proteins from predicted associations
All human proteins among the 21 511 906 pairs were ranked to calculate their ranking scores under the assumption that human proteins in association with high ranking scores may be important for parasite mechanisms. The final ranking score for each human protein was obtained by the production of the ranking score (see section “Ranking score calculation for each human protein”) calculated from the ranked pairs obtained using the machine-learning method and the ranking score calculated from the ranked pairs using the network propagation methods. The histogram of the logarithmic transformation of the final ranking scores of all 12 038 human proteins is shown in Figure 5. Notably, most of the ranking scores were less than 0.0001, while the top best-ranking score was 1 (the logarithm of 1 is 0). Using this top-ranking score, we obtained 411 human proteins. These human proteins were defined as the first list of promising target proteins in human hosts. A complete list of these 411 human proteins is provided in Supplementary Table S2. The bar plot representing the number of highest-score associations for these 411 proteins is shown in Figure 6. Note that only proteins found in more than 2 association pairs are presented in the figure. Overall, we identified Ras-related proteins, kinesin family members, and proteasome 20 S subunit alpha and beta in the list.
Histogram showing the frequency of ranking scores in logarithm scale for human proteins in the predicted human-parasite associations.
Clusters of human protein candidates associated to malaria
As mentioned in section “Identifying promising key human proteins from predicted associations,” we integrated the association scores from our machine-learning techniques and the heterogeneous network model. First, the association scores of candidate human-parasite protein pairs from the heterogeneous network method were ranked to calculate their ranking scores for each protein in the same manner as in our study (see Materials and Methods). Next, we combined the ranking scores of these 2 methods as the attributes to cluster the human proteins using hierarchical clustering. The aim was to group human proteins with similar levels of importance in both aspects. Figure 7 shows the hierarchical clustering of these proteins. By selecting the cut height of the dendrogram tree as 8, we obtained 7 groups of proteins consisting of 2 groups of Ras-related proteins, a single group of histone H2B proteins, kinesin family members, ubiquitin specific peptidase 17 like family members, zinc finger proteins, and a remaining group of mixed types of proteins. Figure S1 shows the high-resolution circular dendrogram of the clustering analysis. The complete list of these proteins in each cluster is provided in Supplementary Table S3. Ras proteins are members of a superfamily of small GTPases that are involved in many processes of cell growth control. Ubiquitin-specific peptidase 17 like family members regulate different cellular processes, such as cell proliferation, cell migration, progression through the cell cycle, apoptosis, and cellular response to viral infection.49-51
Functional characteristics of annotated human proteins
Interpreting the functions of these 411 annotated human proteins may reveal the related mechanisms of the human host and parasite. We investigated these human proteins using functional enrichment analyses. Gene ontology annotations were performed to obtain an overview of the biological processes. The analysis was performed using Cytoscape plugins, ClueGO. Gene ontology associations based on biological processes were selected using intermediate detail in the panel setting of ClueGO. This covered 3 to 8 levels of GO terminology. Based on the PPI of STRING, a second enrichment analysis was performed with a group of genes that were connected in the GO network using CluePedia (version 1.5.6). This analysis revealed 9 functional groups of GO terms, as shown in Table 2 and Figure 8. The complete list of these overrepresented GO terms in the biological process category is provided in Supplementary Table S4. Interestingly, we found the term of regulation of transcription, DNA-templated (GO:0006355), with the most significant term. In addition, Rab protein signal transduction (GO:0032482) and regulation of vesicle size (GO:0097494) were found in a high proportion of our candidate proteins. Rab proteins are a subfamily of the Ras protein family 52 and commonly possess a GTPase fold. These Rab GTPases regulate the processes of membrane trafficking, vesicle formation, and membrane fusion.52-54 Most of our candidate proteins are involved in the regulation of membrane and vesicle formation. These proteins may assist parasite transports in the host and could be potential targets for the treatment of malaria. Figure 8 presents the network of the main enriched GO terms of the 9 clusters, denoted as 9 different colors. Each cluster contained associated GO terms and was named with its principal GO term.
Table 2.
Nine functional groups based on principal gene ontology (GO) terms.
| Cluster number | GO ID | Principle GO term | Adjusted P value* | Percentage of associated proteins |
|---|---|---|---|---|
| 1 | GO:0006355 | Regulation of transcription, DNA-templated | 8.52E−112 | 7.29 |
| 2 | GO:0003700 | DNA-binding transcription factor activity | 6.50E−27 | 6.80 |
| 3 | GO:0032482 | Rab protein signal transduction | 9.87E−20 | 30.26 |
| 4 | GO:0070647 | Protein modification by small protein conjugation or removal | 9.98E−20 | 6.85 |
| 5 | GO:0006511 | Ubiquitin-dependent protein catabolic process | 8.17E−12 | 7.23 |
| 6 | GO:0090382 | Phagosome maturation | 5.41E−03 | 11.11 |
| 7 | GO:0097494 | Regulation of vesicle size | 5.67E−03 | 21.43 |
| 8 | GO:0001217 | DNA-binding transcription repressor activity | 1.43E−02 | 4.53 |
| 9 | GO:0006904 | Vesicle docking involved in exocytosis | 2.67E−02 | 8.33 |
Abbreviation: GO, gene ontology.
Protein complexes to potential protein targets
To identify sets of these 411 proteins that interact with each other and play essential roles in regulatory processes, cellular functions, and signaling cascades, we performed enrichment analysis in protein complexes. Enrichment analysis of these proteins was performed on the CORUM protein complex database (version 3.0). 55 Four protein complexes were found using Bonferroni-adjusted P values for the enrichment tests <0.05. These 4 protein complexes consisted of the 20S proteasome, 26S proteasome, PA28gamma-20S proteasome, and PA28-20S proteasome. Most of the proteins overrepresented in these protein complexes were PSMA4, PSMB2, PSMB4, PSMB5, PSMB6, and PSMB7. Only the 26S proteasome contained 1 more protein (PSMC1) in the list. Thus, these proteins may be interesting targets in future studies. Table 3 presents a list of the overrepresented protein complexes.
Table 3.
The list of protein complexes enriched in 411 promising candidate proteins.
| Protein complex | Adjusted P value | Associated proteins |
|---|---|---|
| 20S proteasome | 8.34E−03 | PSMA4, PSMB2, PSMB4, PSMB5, PSMB6, PSMB7 |
| 26S proteasome | 1.29E−02 | PSMA4, PSMB2, PSMB4, PSMB5, PSMB6, PSMB7, PSMC1 |
| PA28gamma-20S proteasome | 1.35E−02 | PSMA4, PSMB2, PSMB4, PSMB5, PSMB6, PSMB7 |
| PA28-20S proteasome | 2.10E−02 | PSMA4, PSMB2, PSMB4, PSMB5, PSMB6, PSMB7 |
Furthermore, to examine the importance of the proposed human proteins, these proteins were searched for in the Drugbank database. 56 Interestingly, Proteasome 20S Subunit Beta 2 (PSMB2) and Proteasome 20S Subunit Beta 5 (PSMB5) were identified, which are known to be drug targets, in the Drugbank database. PSMB2 and PSMB5 play several roles. They were found to be enriched in the principal GO terms of regulation of transcription, DNA-templated, protein modification by small protein conjugation or removal, and ubiquitin-dependent protein catabolic process. Interestingly, PSMB2 was found to be a drug target of carfilzomib (DB08889), while PSMB5 is a drug target of carfilzomib and bortezomib (DB00188). Carfilzomib is a synthetic proteasome inhibitor. It is an analogue of the natural product epoxomicin, which effectively kills parasites. Bortezomib is the first therapeutic proteasome inhibitor to be tested in humans, which induces cell cycle arrest and apoptosis. Bortezomib interrupts the degradation of proapoptotic proteins in cancerous cells. It is currently used for the treatment of relapsed multiple myeloma and mantle cell lymphoma. Both carfilzomib and bortezomib have been reported to be related to malaria treatment. 57 Carfilzomib has been reported to potently block P. falciparum replication at effective concentrations as well as killing asexual blood-stage P. falciparum. 58 Bortezomib exhibits antiplasmodial activities and has been examined for efficacy against P. falciparum. 59 PSMB2 and PSMB5 were found in all our resulting protein complexes (Table 3). Thus, these complexes may be a valuable starting point for further studies aiming to design and develop drugs against malaria. In addition, PSMB2 and PSMB5 were observed in mixed types of protein group of 62 proteins in our clustering results (see section “Clusters of human protein candidates associated to malaria” and Supplementary Table S3). Therefore, the remaining 60 proteins in the same cluster of these proteins may be promising therapeutic targets for P. vivax malaria. A list of these proteins is provided in Supplementary Table S5. In addition, the relationship of these 411 human proteins and P. vivax malaria was evaluated to determine orthologous proteins of P. vivax and the 411 human proteins from EggNOG database (version 5.0). 60 The results are presented in Supplementary Table S6.
Discussion
Our understanding of the invasion mechanism of P. vivax remains deficient due to the lack of a robust in vitro culture system for this parasite. In an attempt to resolve this, the host-parasite interactions were studied, including direct interactions at the protein level inside the cell. In this study, we initially reconstructed the human and parasite PPI networks, and compared their network structures. In principle, both networks follow the power distribution, and the analysis of network topologies between these 2 networks revealed a correlation of the connections within their own network between human and parasite proteins in the positive set. The high correlation of closeness centrality between these proteins indicated that most of the similar proteins between human and parasite responded to minimum paths that connect the other proteins. These proteins also formed a similar local community around them, as the high correlation was observed in terms of eccentricity. Although the degree, betweenness centrality, and Kleinberg’s hub did not show significant correlations among these proteins, the machine-learning approaches applied here may help reveal several more human and parasite protein associations in future studies.
A ranking score calculation for the human proteins was developed based on the rank of the associations according to their voting scores. A total of 411 human proteins with the best-ranking score were selected as promising target candidates. Based on the histogram shown in Figure 5, the second-best score had a gap jumping from the top best, while the rest of the scores were far away from the best one. The majority of these proteins had a ranking score of approximately 0.00001, which was very low in terms of the probability of being a reliable association. Thus, these 411 proteins were selected for further analysis together with heterogeneous network prioritization and qualified in terms of clusters, functions, and protein complexes.
The results showed that Ras-related proteins, a single group of histone H2B proteins, kinesin family members, ubiquitin-specific peptidase 17 like family members, and zinc finger proteins were the most prominent in our candidate list. These proteins are involved in several processes of cell growth control and regulation of membrane and vesicle formation. Several proteins related to proteasome 20S subunits have been previously reported as promising multistage targets for malaria therapy. 59 These proteins may be used for the invasion of parasites to the host cell and have been identified as potential drug targets in the human host.
Conclusion
In this study, we established an analysis framework that uses machine-learning approach based on a heterogeneous network structure. We used the network topology features of proteins in the human PPI network and the P. vivax PPI network and integrated protein sequence similarities to the framework to predict human-parasite protein associations. We also developed a ranking score calculation to identify promising protein targets in humans for the treatment of malaria infections. The candidate human proteins that were selected as promising targets were then qualified by clustering analysis together with the information on the existing targets from the heterogeneous network prioritization, as well as by functional and protein complex enrichment analyses. We found that proteins in the cluster of PSMB2 and PSMB5 (known drug targets), human proteins involved in the regulation of membrane and vesicle formation, and complexes such as the 20S proteasome, 26S proteasome, and PA28gamma/-20S proteasomes are potential targets for the design and development of drugs for the treatment of malaria.
In conclusion, the integration of data related to network topologies and sequence similarity provides us with an opportunity to define associations between human and P. vivax proteins. Human protein candidates extracted from these associations were used to compile a list of promising targets in humans for further validation in wet-laboratory experiments in future studies. An enhanced understanding of potential host proteins at the molecular level will provide insights to support malaria control efforts and the production of novel antimalarial drugs.
Supplemental Material
Supplemental material, sj-pdf-1-bbi-10.1177_11779322211013350 for Prediction of Human-Plasmodium vivax Protein Associations From Heterogeneous Network Structures Based on Machine-Learning Approach by Apichat Suratanee, Teerapong Buaboocha and Kitiporn Plaimas in Bioinformatics and Biology Insights
Supplemental material, sj-xls-2-bbi-10.1177_11779322211013350 for Prediction of Human-Plasmodium vivax Protein Associations From Heterogeneous Network Structures Based on Machine-Learning Approach by Apichat Suratanee, Teerapong Buaboocha and Kitiporn Plaimas in Bioinformatics and Biology Insights
Supplemental material, sj-xls-3-bbi-10.1177_11779322211013350 for Prediction of Human-Plasmodium vivax Protein Associations From Heterogeneous Network Structures Based on Machine-Learning Approach by Apichat Suratanee, Teerapong Buaboocha and Kitiporn Plaimas in Bioinformatics and Biology Insights
Acknowledgments
The authors acknowledge National e-Science Infrastructure Consortium (http://www.e-science.in.th) for providing computing resources that have contributed to the research results reported within this article.
Footnotes
Funding: The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was funded by the Office of the Higher Education Commission (OHEC) and Thailand Research Fund (TRF), grant no. MRG6180021.
Declaration of conflicting interests: The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Author Contributions: Conceptualization and writing—review and editing were performed by A.S, T.B. and K.P.; A.S. contributed in data curation, funding acquisition, and writing the original draft; formal analysis, methodology and validation was by A.S. and K.P. All authors have read and agreed to the published version of the manuscript.
Supplemental Material: Supplemental material for this article is available online.