Schizophrenia is a serious mental disease characterized by abnormalities in thinking and cognition [1], whose occurrence is widely believed to be closely related to genetics and gene expression. This chronic and disability-causing disease not only has great effects on the quality of life of patients but also imposes a heavy burden on their families and society as a whole; therefore, schizophrenia has long been a key focus of research in the medical field [2]. Searching for the associations between diseases and genes is an effective way to study complex diseases, allowing an in-depth exploration of the mechanisms and molecular basis of diseases, which is crucial to establishing accurate treatments and diagnoses.

Numerous network-based approaches have recently been proposed for disease-gene association prediction, such as protein-protein interaction (PPI), gene regulatory, gene coexpression, and metabolic interaction networks. Among these molecular networks, PPI networks are widely used as a conducive approach to discover potential disease-causing genes [3] since proteins work together to perform common biological functions. In addition, proteins associated with a common set of biological properties tend to have common topological properties in the network, such as node degree and centrality, and the pathways elucidating disease mechanisms are typically represented as strongly connected paths in the PPI network [4].

However, existing PPI networks are often incomplete and noisy. Thus, it is necessary to collect multiple types of data and build an integrated network that includes multiple, heterogeneous types of resources. This method will greatly extend the scope and ability for disease gene prediction [5]. However, scientific data do not consider the properties of biological data themselves, whereas the scientific literature represents a record of the latest scientific discoveries and important research results, containing a large amount of additional biological knowledge such as protein biological functions and sequence characteristics. Therefore, data from the scientific literature along with primary scientific data can be used as a complement to discover the implicit information and enhance the ability to predict disease-associated genes.

Many new techniques have been developed to study heterogeneous networks containing multimodal biological data. Recently, a general framework that considers heterogeneity by defining type-specific graphs was introduced [6]. The advantage of metagraph-based approaches is that they capture rich semantics, comprehensively represent different features, and effectively identify influential components of the given network, making it possible to preserve the network structure and providing flexibility to explore a diverse set of descriptors.

The problem of disease-related genes prediction tackled in our study can be regarded as a disease protein classification problem that aims to identify disease-related proteins, and the proteins are then linked to their gene products to obtain disease-related genes. The goal is to train a classifier to judge whether the protein is associated with a disease based on a training set, and we can then predict each protein’s likelihood of falling into a given category in the test set with minimal prediction errors using the learned protein features.

In this study, we attempted to integrate the PPI network from the STRING database with the keywords in the UniProt database to form a heterogeneous PPI-keyword (PPIK) network for disease protein prediction. Based on the PPIK network, we extended the metagraph methodology to predict the probability that an association between a gene and disease exists. Each protein was represented as a series of metagraphs. A previous study that utilized metagraph representations to exploit a keywords-supplemented PPI network achieved good performance on predictions for breast cancer [7]. We made some improvement on the basis of this method. In particular, we extracted basic metagraph structures from the network to represent proteins, which fully revealed the traits of proteins. We further considered the use of a more appropriate algorithm for metagraph representations that is suitable for complex networks to capture more implicit features. In addition, we used the latest machine learning models to understand the ever-growing scientific data, which were applied to schizophrenia as a more specific disease to enable better disease-gene association prediction, thereby uncovering potential therapeutic targets.

We employed metagraph representations based on the PPIK network method to improve disease-associated genes prediction. The general framework is shown in Figure 1.

A PPI network is frequently noisy and incomplete. The keywords in the UniProt database represent annotated information based on the literature or empirical evidence, covering various biological aspects of proteins [20], as summarized in Table 1. As disease-associated proteins are likely to share similar properties, these properties may reveal their function. Furthermore, proteins with the same domain as the identified disease proteins could also be associated with the disease. Therefore, we developed a method to associate these properties with different proteins together to complement existing PPI networks, which could encode both the interactive functions and biological properties of proteins. We used data in the UniProt database to directly construct the relations between proteins and keywords. Since each protein corresponds to multiple keywords in the database, we established the links between proteins and keywords based on this correspondence. On the one hand, protein-keyword associations can reinforce useful PPIs. On the other hand, proteins with no direct interactions can newly become related through keywords.

In this study, we focused only on human proteins. We obtained PPIs from the STRING database and exploited protein keywords from the UniProt database to construct the PPIK network. The network contains two types of nodes, including protein and keyword nodes. The relationships in the network indicate the interaction between two proteins and the association between the protein and keyword, which are not distinguished. A representative PPIK network is shown in Figure 2.

A metagraph is a graph structure capturing a particular topology of the network, which can obtain the location information of each node and the type-specific path pattern between a source node and a destination node in the network. Leveraging a metagraph to characterize a network will not only preserve the structural features but will also help to identify rich pathway information, which offers strong flexibility in heterogeneous networks [21]. A heterogeneous PPIK network comprises both proteins and keywords. Proteins with similar functional roles such as their disease-causing property tend to interact with other proteins and associate with certain keywords in a similar arrangement on the PPIK network. Thus, proteins with similar roles tend to have similar topological characteristics. To model topological similarities, we used a metagraph approach to define the specific structure and differentiated between different labels of nodes on the PPIK network. Each metagraph describes a particular heteronomous biological arrangement between one or more proteins and keywords. Two proteins associated with the same metagraph tend to have similar functional roles. Therefore, we could use different types of metagraphs to represent each protein, while identifying its interactions with other proteins and associations with keywords. Each instance of a metagraph represents specific evidence of the proteins’ functional properties. In particular, we only considered metagraphs of three and four nodes, which represented a good balance between efficiency and accuracy. We defined seven types of metagraphs to reveal the network feature and their common structures are shown in Figure 3. The seven metagraphs cover all instances of network motifs with three and four nodes on the PPIK network, and thus enable the thorough capture of structural characteristics of the heterogeneous network, which in turn allows for an accurate representation of protein features.

We used a series of metagraphs to construct the vector representation of each protein to characterize its multiple features. To compute the prevalence of a type-specific metagraph, we employed a type-specific metapath counts approach known as the degree-weighted path count (DWPC) [22], which represents an improvement of the path count metric, as a basic social network analysis method, by dampening each edge between a source and target node as an effective feature extraction methodology accommodating a heterogeneous network of any size. This method involves first calculating the path-degree product (PDP) of each edge. All metaedge-specific degrees along the path (D_path) are determined, where each edge composing the path contributes two degrees. Subsequently, each degree is raised to the –w power, where w≥0, which is known as the damping exponent. This parameter adjusts the intensity for each path traversing highly connected nodes to be down-weighted, and the best performance is usually achieved with w=0.4. Finally, all exponentiated degrees are multiplied to yield the PDP of each path according to the following formula:

The DWPC is then calculated as the sum of PDPs:

An example of the computation process for the DWPC metric is shown in Figure 4.

Protein representations were used to train the model through machine learning. First, disease labels for proteins were obtained from the UniProt and OMIM databases. We collected the corresponding genes for schizophrenia from the OMIM database, and then mapped the genes to their protein products according to the UniProt database to tag these proteins. Next, we adopted four machine learning models as classifiers, random forest (RF), extreme gradient boosting (XGB), light gradient boosting machine (LGBM), and logistic regression (LR), which aimed to identify the schizophrenia-associated proteins based on the extracted metagraphs. For experimental settings, considering the sparsity of the protein-disease network where the number of unconfirmed protein-disease associations is much greater than the number of confirmed associations, we randomly sampled the nonschizophrenia (negative) proteins and schizophrenia (positive) proteins with ratios r∈{1,3,5,10,20} as the data set, and 10-fold cross-validation was used in model training and testing. Specifically, to avoid the bias of data splitting, we repeated the cross-validations 10 times in different random seeds and report the mean (SD) of four performance metrics: precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC).

To demonstrate the validity and superiority of the PPIK-based metagraph representations method, we compared our proposed work to other baseline methods developed for link prediction in networks, especially in the task of gene-disease association prediction. We selected three representative indices for performance comparison. The random walk with restart (RWR) index [23] is a random-walk algorithm based on similarity. RWR considers a random walker starting from node x, who will iteratively move to a random neighbor with probability c and return to node x with probability 1–c, and this process is repeated until it converges to a steady state. The average commute time (ACT) index [24] is also a random walk–based similarity index, which assumes that two nodes are considered to be more similar if they have a smaller average commute time. ACT is denoted by the average number of steps required by a random walker starting from node x to reach node y and from node y to node x. The Katz algorithm [25] is a path-dependent index that obtains global knowledge of the network topology; it can distinguish the impact of different neighbors by assigning them different weights.

UniProt is a protein database with the most complete sequencing and extensive annotation data available internationally. The STRING database contains data on various types of interactions, including direct physical interactions between proteins, indirect functional correlations, results extracted from PubMed abstracts, and outcomes predicted using bioinformatics techniques. The original data collection in the two databases received institutional review board (IRB) approval [26,27] and the scientific data in these two databases are freely available for users under a Creative Commons BY 4.0 license [28,29]. In addition, this study involved a secondary analysis, and the publicly available data sets were prepared with the intent of making them available for the public. The data available to the public are not individually identifiable and therefore the analysis would not involve human subjects. The Ethics Review Committee of Peking Union Hospital, Chinese Academy of Medical Sciences recognizes that the analysis of deidentified, publicly available data does not constitute human subjects research and therefore does not require IRB review; only biomedical research involving humans requires the approval of an ethics committee [30]. Several studies have used these two databases [31,32]. Over dozens of years, the UniProt and STRING databases have accumulated vast amounts of protein information and ethical approval was granted for collection of the original data with informed consent from the included patients. Analyses of these data have led to improvements in our ability to understand the molecular mechanisms of diseases, thereby promoting diagnosis and treatment. Since these data are open source, they remain publicly available for anyone in the research community to use.

The performance comparison results of the four machine learning models is shown in Table 2, with each metric representing the average score of the data set with different sampling ratios. The results showed that our proposed method could achieve relatively good performance with an AUC between 0.72 and 0.76. Specifically, with a ratio of 1, the RF model outperformed other models, demonstrating good discriminatory power with precision of 0.76, recall of 0.73, F1-score of 0.72, and AUC of 0.76. Therefore, we selected RF as the algorithm of choice for further investigation.

We compared the AUC values of the three baseline methods (RWR, ACT, and Katz index) on the PPI and PPIK networks, respectively. The performance of these methods on the PPIK network was better than that on the PPI network under most circumstances, according to the AUC values, as shown in Table 3. This demonstrated that protein keywords can indeed complement and enrich the PPI network.

We further compared AUC values between the baseline methods and our proposed method on the PPIK network. Our method achieved better performance on the whole, which outperformed the Katz, ACT, and RWR methods by 0.156, 0.294, and 0.457, respectively, as shown in Table 3. Thus, it can be concluded that metagraph representation is a feasible method for network topological features.

We found that the AUC values of these baseline methods were quite low overall, which may be due to the difficulty for these methods to accurately capture the characteristics of the heterogeneous and sparse network we constructed. However, our framework can effectively capture various features in the network and demonstrated good performance.

We focused on the results from RF, as the model with the best performance, for the discovery of potential schizophrenia-causing proteins. As the goal of our study was to identify new proteins associated with schizophrenia, we excluded confirmed protein-schizophrenia associations in the database and retained the top 20 proteins identified as risk proteins for schizophrenia. We then transformed our predicted disease proteins to their producer gene IDs based on UniProt. The top 20 predicted genes and their probability scores are listed in Table 4.

In this work, we proposed an optimized model to enhance schizophrenia-related genes prediction. Our framework identified 20 potential schizophrenia risk genes: EYA3, CNTN4, HSPA8, LRRK2, AFP, CTNNB1, CYP1A1, FGFR1, PRKAR2A, HTT, SLIT1, MEPCE, CENPC, NDUFA13, PDE4D, PIEZO1, PACSIN2, GSPT1, LOXL2, and CD2AP. Given these results, it is promising to examine what our framework produced. Toward this end, we here further evaluate the top risk genes associated with schizophrenia based on publications to support our predictions and analyze these genes from four different perspectives: function, gene-set enrichment, signaling pathways, and gene character (ie, verification). The evidence for these predictions can be found in the up-to-date literature, which has proven to be reasonable.

The model adopted in our study is a metagraph representation based on the PPIK network framework. We used this model to identify potential schizophrenia-associated genes. This model attempted to integrate the PPI network with the keywords that describe the biological aspects of the proteins to ultimately form a PPIK network for disease gene prediction. The heterogeneous network contains rich information, involving not only protein interactions with one another but also their functional and structural similarities, which could better compensate for the limitations of the standard PPI network. Based on the PPIK network, we used metagraphs to solve the disease protein classification problem. This type of graph structure can thoroughly capture the topology of a heterogeneous network. By permuting all instances of tiny motifs, we can obtain all biological arrangements between proteins and keywords on the PPIK network, allowing an accurate representation of each protein using a series of metagraphs. In particular, the DWPC indicator was used to quantify metagraph representations, which could capture the prominent feature of metagraphs. Finally, we built a classifier using four advanced machine learning algorithms for disease proteins based on these metagraph representations and selected the best-performing model for disease gene prediction. The utilization of machine learning models to prioritize selected proteins according to predicted probabilities enhanced the interpretability of the results. Our proposed framework can be a practical and efficient prediction model. Of the four machine learning methods, RF achieved the best overall performance with precision of 0.76, recall of 0.73, F1-score of 0.72, and AUC of 0.76. Therefore, the framework using the RF classifier demonstrates superior predictive power.

We used the Shapley additive explanation value to analyze the importance of the metagraph features contributing to our model. The features are arranged in descending order of importance in Figure 6. We found that features containing keywords ranked on top, indicating their important role for the prediction capability of the model and their ability to capture more information and provide a better representation of each protein. Among them, larger values of the features pk, kpk, pkp, and pp indicate a higher likelihood of being a schizophrenia-associated protein, which indicated that these metagraph features reflected more characteristics describing the molecular mechanism related to the pathogenesis of schizophrenia, whereas other metagraph features seemed to have the opposite impact. The interactions between proteins form the basis of many crucial life activities, and they are of great significance in unraveling complex biological processes in living organisms and understanding the molecular mechanisms of diseases. In particular, the protein complexes that are tightly bound and have specific biological functions are closely related to diseases. The ppk, kpk, pkp, and ppp metagraphs are denoted as tightly coupled proteins with specific functions. For example, the 14-3-3 epsilon protein and the reticulon-4 receptor protein are associated with the same biological process of cerebral cortex development. Therefore, such metagraphs may be positively correlated with the incidence of diseases. By contrast, the pppp, pkpp, and kppp metagraphs each contained four nodes. An increase in the number of nodes in the pathway may result in some redundancy of features and thus weaken their meaning. Such paths typically include the binding and interaction of two protein complexes that are not closely linked. For example, the cadherin-related family member 1 protein and the meurexin-2 and neurite extension and migration factor protein complex are associated with the same biological process of replication factor C subunit 2, which therefore have a diminished impact on diseases or may even be negatively correlated with diseases. However, all of these features offer important information for the description of a network, showing various aspects of the network characteristics, and thus these features should be fully employed to represent the proteins.

In our study, we chose the classifier with the best performance for disease gene prediction. The proposed framework successfully discovered associations hidden from the network and achieved relatively favorable results. The main benefits of our framework are as follows. First, the keywords we considered in the network described rich biological properties of proteins and thus encompassed reasonable predictive power for disease genes. Second, the vector representations for proteins we constructed improved our ability to capture the topological arrangement on the PPIK network for interactions between both proteins and keywords, and each protein can provide a comprehensive representation of its characteristics from various aspects. In addition, it is noteworthy that this framework has general applicability, which can not only be applied to schizophrenia but can also be fine-tuned to achieve better performance by tailoring to other specific diseases.

Despite these benefits, our method still has some limitations. Notably, our method failed to adequately leverage pleiotropy and diverse biological features. To resolve this issue, we can propose a strategy to build a heterogeneous network covering multimodal data. Based on this comprehensive network, we can excavate the potential characteristics of the biological entities and relationships among them to explore candidate disease genes. In addition, our method was based on machine learning techniques, and effective algorithms are required to transform features into a training data set. However, these algorithms are complex and cannot fully reflect the biological information of various aspects in the network. For future work, we can employ deep learning approaches such as convolutional neural networks and autoencoders into our framework for network embedding, which can capture both the structural and nonstructural properties of networks as well as dig out implicit features, thus leading to better predictive performance.

Predicting potential associations between diseases and genes is crucial to the diagnosis and treatment of many diseases. In this work, we proposed an optimized method to mine and predict schizophrenia-associated risk genes. We incorporated PPI and biological keywords of proteins to construct a PPIK network. We then extracted features from the PPIK network and proposed metagraph representations for proteins. Further, we trained and optimized our model using four machine learning algorithms, including RF, XGB, LGBM, and LR, for disease protein prediction. Our method achieved relatively good performance, outperforming the RWR, ACT, and Katz baseline methods. We applied RF, as the machine learning model with better comprehensive performance, to make the prediction. In particular, we mapped these proteins to their gene IDs and obtained the top 20 novel potential schizophrenia-associated genes. Finally, we performed feature analysis and searched for evidence in the literature to explain the potential association between specific genes and diseases. The results indicate that application of our approach was quite reasonable and reliable in the discovery of schizophrenia-associated risk genes. Overall, our method provides a means to enrich protein interactions with more detailed information, which can encourage the deeper mining of relations between proteins. This method is likely to contribute to research on schizophrenia pathology and can offer some guidance for the diagnosis and treatment of schizophrenia. In addition, this approach can help to save time and costs, thereby facilitating basic research. After more testing and optimization in independent samples, these results are expected to be applied in future clinical practice and clinical trials.

The data sets generated and analyzed in this study are available from the corresponding author on reasonable request.