BEROLECMI: a novel prediction method to infer circRNA-miRNA interaction from the role definition of molecular attributes and biological networks

Circular RNA (CircRNA)–microRNA (miRNA) interaction (CMI) is an important model for the regulation of biological processes by non-coding RNA (ncRNA), which provides a new perspective for the study of human complex diseases. However, the existing CMI prediction models mainly rely on the nearest neighbor structure in the biological network, ignoring the molecular network topology, so it is difficult to improve the prediction performance. In this paper, we proposed a new CMI prediction method, BEROLECMI, which uses molecular sequence attributes, molecular self-similarity, and biological network topology to define the specific role feature representation for molecules to infer the new CMI. BEROLECMI effectively makes up for the lack of network topology in the CMI prediction model and achieves the highest prediction performance in three commonly used data sets. In the case study, 14 of the 15 pairs of unknown CMIs were correctly predicted.

CircRNA (circular RNA) is a kind of non-coding RNA with a closed-loop structure that exists in eukaryotic cells [1]. CircRNA was first discovered in the 1980s [2] and is considered to be a noise or by-product in the process of transcription. However, with the development of high-throughput sequencing technology, researchers have re-focused on circRNA and found that they have a variety of biological functions. By screening and analyzing the differentially expressed circRNA, researchers revealed the expression patterns of circRNA in different tissues, developmental stages, and disease states [3]. CircRNA affects cell function by regulating miRNA activity or interacting with RNA-binding proteins. Therefore, it plays a role as a potential biomarker [4].

MicroRNA (miRNA) is a kind of non-coding RNA, which can regulate gene expression by binding to the 3' untranslated region (3'UTR) of the target gene [5]. In the CMI, circRNA binds to miRNA through the "miRNA response element (MRE)" with complementary sequences to form a circRNA-miRNA complex [6, 7]. This binding occurs mainly through two main mechanisms: the sponge effect and the competitive binding mechanism. In the sponge effect, circRNA absorbs multiple miRNA molecules as the "sponge" of miRNA, thus reducing the binding between miRNA and target mRNA, and affecting the regulation of target genes by miRNA. In the competitive binding mechanism, circRNA and miRNA compete directly with the 3'UTR binding sites of the same target gene to hinder the effect of miRNA on the target gene. CMI plays an important role in the regulation of gene expression, cell proliferation, and tumorigenesis, which provides new hope for the diagnosis, treatment, and prognosis of diseases [8, 9].

Although circRNA has been proven to play an important role in biology, its specific function is not completely clear, so there is an urgent need for a comprehensive study of circRNA. At present, a large number of circRNA have been identified, which brings new opportunities and challenges to the related data mining work. Limited by manpower, materials, and resources, it is impractical to match a large number of CMI data aimlessly. Therefore, the use of computing technology to provide a high probability pre-selection range for wet experiments has become the main means of CMI discovery. CMI inference methods based on computing technology are mainly divided into two kinds: one is based on sequence matching, such as miRanda [10], and TargetScan [11], and the other is the prediction model based on known association. Based on the principle of complementary binding sites, the method based on sequence matching can achieve large-scale CMI prediction, but it will produce too many false positive samples, and only relying on a single feature will cause the known molecular association and biological structural features to be ignored; the use of advanced prediction model can effectively make up for the shortcomings of sequence matching methods, but it is often affected by data scale, network expansion, attribute collection and so on.

Currently, models that use computational methods to predict associations between biomedical entities are constantly being proposed, such as drug–drug interaction prediction [12], drug–target interaction prediction [13, 14], LncRNA-disease prediction [15,16,17], lncRNA–protein interaction [18, 19], circRNA-disease prediction [20,21,22], disease-associated Piwi-interacting RNAs prediction [23], cell–cell communication inference [24,25,26] and phage-host interaction prediction [27]. However, there are still few studies on prediction models for circRNA–miRNA interactions. Most of the existing CMI prediction models use the nearest-neighbor relationship in the biological network for modeling. For example, Wang et al. [28] proposed the KGDCMI method, which uses the HOPE method to embed the node association structure in the CMI network to predict the unknown CMI; Guo et al. [29] used the SDNE method to embed the similarity between nodes and the similarity of neighbor sets in the CMI network to predict the new CMI. Yu et al. [30] and He et al. [31] use graph convolution neural network to aggregate the feature prediction of nearest neighbor nodes in high-dimensional space CMI; Qian et al. [32] proposed CMASG, by extracting linear and nonlinear features from the CMI network to predict unknown CMI. Wang et al. used the signed graph convolutional neural network to aggregate friend and foe relationships in circRNA-miRNA-cancer networks to predict unknown CMI [33]. These methods are used to model and extract the nearest neighbor relationship in the CMI network to realize the effective prediction of CMI. However, as a kind of community relationship in the network, the application of the nearest neighbor relationship in the biological network has some limitations: 1. CMI biological network is sparse, most nodes have little or no nearest neighbor relationship; 2. The nearest neighbor relationship will be bound to the node representation, so it is difficult to achieve network expansion; 3. Limited by the capture mode of neighboring nodes, the similarity (structural similarity) between nodes with similar network topology features but not adjacent nodes is ignored. This makes it difficult to achieve a breakthrough in the accuracy of CMI prediction. As a kind of network structure, the local topological structure information in the network has been applied to the CMI prediction model. Wang et al. added local topological structure similarity to the model as a supplement to obtain reliable prediction performance [34]; Wang et al. used the wavelet diffusion mode to extract the topological structure feature of nodes in the CMI network and obtain the highest prediction performance [35]. This means that the topological structure feature is effective in model feature extraction, but these methods only use the local topological structure in the network as a supplement to the features and do not clearly define the topological structure, which limits the value of local topological structure features.

In this paper, we propose a CMI prediction method, BEROLECMI, which defines role attributes for each molecule through molecular attribute features, molecular self-similarity networks, and molecular network features for advanced prediction tasks. Specifically, BEROLECMI first uses the pre-trained Bidirectional Encoder Representations from the Transformers model for DNA language in genome (DNABERT) [36] to extract attribute features from RNA sequence, then constructs RNA self-similarity networks through Gaussian kernel function and sigmoid kernel function respectively, and the high-level representation is learned by sparse autoencoder (SAE) [37]. Next, the proposed model uses the structured embedding method to extract the role similarity of each molecule in the CMI network. Finally, these features are organically integrated as molecular exclusive role features and sent to the classifier for training and prediction tasks. In the performance verification, the prediction performance of the BEROLECMI in common data sets exceeds that of all known models. In the case study, the proposed method accurately predicts 14 of 15 pairs of CMIs. The flow chart of BEROLECMI is shown in Fig. 1.

The flow chart of BEROLECMI

Full size image

Molecular attribute feature construction

In this part, the BEROLECMI method combines the sequence structure of molecules and self-similarity network to construct molecular attribute descriptors. Specifically, we first cut the sequence of RNA molecules to form effective sequence fragments, and then use the pre-training model, DNABERT, to learn the potential features in RNA sequences. Next, we use the Gaussian kernel function and sigmoid kernel function to construct molecular self-similarity networks and use SAE to learn the advanced representation of molecular self-similarity. Finally, the RNA sequence feature and molecule self-similarity feature are fully connected, and the exclusive attribute feature is assigned to each molecule.

RNA sequence feature extraction

The Bidirectional Encoder Representations from Transformers (BERT) model has achieved great success in the field of natural language processing (NLP), which provides a new pattern for the use of large-scale preprocessing models. However, due to the unique biological characteristics, and physical and chemical structure of DNA sequences, the traditional NLP algorithm is difficult to effectively apply to the modeling of DNA data. Therefore, through the improvement and adjustment of BERT, Ji et al.[36] developed a DNABERT model suitable for DNA sequences. In this part, we introduce the DNABERT model to learn the sequence features of RNA. Specifically, BEROLECMI first cuts the RNA sequence according to the sequence aggregate of 6-mer, which is determined by the training corpus of the DNABERT model; then, the obtained RNA sequence aggregate is used as the corpus, and the pre-trained DNABERT model is used to capture the local and global representation of the DNA sequence. The feature extraction process of the RNA sequence is shown in Fig. 2. In addition, detailed information about the DNABERT pre-training model is referred to in the research of Ji et al. [36].

The feature extraction process of the RNA sequence

Full size image

Molecular self-similarity feature construction

RNA molecular self-similarity features are based on the functional similarity hypothesis, that is, molecules with similar binding targets may have the same functions. BEROLECMI introduces the Gaussian kernel function and sigmoid kernel function to construct RNA self-similarity in the CMI network. For the dataset CMI-9905 stored in the adjacency matrix M, when circRNA a interacted with miRNA b, M_ab is 1, otherwise 0. The circRNA Gaussian kernel representation in the matrix can be calculated as:

where U_a and U_b represent circRNA a and circRNA b respectively, K_circRNA(U_a, U_b) denotes the GIP kernel similarity between circRNA a and b, λ is a parameter controlling the bandwidth of the GIP kernel function, calculated as:

Similarly, the miRNA Gaussian kernel representation can be calculated as:

The circRNA sigmoid kernel representation is defined as:

where η = 1/N, N is the dimension of the input data.

Similarly, the miRNA sigmoid kernel representation is defined as:

Molecular self-similarity integration

BEROLECMI organically integrates Gaussian kernel similarity and sigmoid kernel similarity of molecules to obtain a highly representative self-similarity matrix. For circRNA, the similarity matrix can be calculated as:

For miRNA, the similarity matrix can be calculated as:

The Gaussian kernel similarity, sigmoid kernel similarity, and self-similarity matrix of circRNA and miRNA are shown in Fig. 3.

The self-similarity of molecular

Full size image

Molecular self-similarity feature enhancement

BEROLECMI performs dimensionality reduction and feature enhancement on molecular self-similarity networks by introducing SAE. The SAE has the same encoding and decoding structure as the ordinary autoencoder, and the input layer maps the input data to the hidden layer L for encoding:

where X(L₀) is the input data, W is the parameter of the hidden layer.

The difference is that the SAE adds sparsity constraints during the training process. By adding sparsity constraints to the neurons in the hidden layer, the hidden layer keeps a low activation value to obtain highly representative key features. The sparse penalty term T can be calculated as:

where T represents the sum of the degree to which the penalty item P deviates from F, and N represents the number of neurons in the hidden layer. In this experiment, the KL divergence (Kullback–Leibler) is used to calculate as follows:

where P is the sparse parameter of KL, the closer F is to P, the smaller the value of KL.

By adding sparsity constraints in the encoding layer, SAE can learn a low-dimensional robust representation of the original features, thereby improving model performance.

Network embedding based on role definition

To effectively extract molecular structured embeddings in biological networks, BEROLECMI introduces the Role2vec algorithm [38] combined with attribute random walks to capture molecular structure similarities in networks. For an undirected graph G < N, E > composed of N nodes and E edges constructed based on the CMI biological network, the Role2vec algorithm defines different roles for nodes according to the network topology, such as motifs, graphlets, etc., as shown in Fig. 4. The structure type is flexibly selected via parameters.

Different roles for nodes according to the network topology

Full size image

The detailed description of the Role2vec algorithm is shown in Algorithm 1. Role2vec uses graph G, sub-graph structure M, embedded dimension d, node walk r, walk length l, and context window size z as the algorithm input. In steps 1–3, the subgraph structure in graph G is first extracted and transformed into a specific representation, and then the subgraph structure is mapped to x by the function f. In step 4, the transfer probability p is calculated. In steps 6–10, the nodes were reordered in the reconstructed graph G^’, and the attribute walk to extract the feature representation a of node n. Then, add a to the set A. Finally, the role embedding representation of each node is obtained by using the stochastic gradient descent algorithm.

Role2vec

Full size image

Different from the traditional random walk, the walk method with the structure type as the node attribute only focuses on the local topology of the node, so it can effectively define roles for different nodes. This means that the nodes only rely on structural features for embedding, which has high scalability and can be extended to nodes in distant or even different graphs.

Evaluation criteria

In this study, we introduce five-fold cross-validation (five-fold CV) to evaluate the performance of the proposed method. five-fold CV divides the CMI data into five subsets at random, each time four subsets are used as the training set, one subset is used as the test set, and five experiments are performed until the predicted score of each subset. We comprehensively evaluate the predictive performance of the proposed model by combining multiple evaluation criteria including Accuracy. (Acc.), Precision. (Prec.), Recall and F1-score. The evaluation criteria can be represented as:

Among them, TP and TN respectively represent the number of positive samples and negative samples predicted correctly by the model; FP and FN respectively represent the number of negative samples and positive samples predicted incorrectly by the model. In addition, we introduce the receiver operating characteristic curve (ROC), and precision-recall curve (PR).

In this work, we evaluate model performance through the five-fold CV based on three commonly used datasets in the field of CMI prediction. The CMI-9905 dataset was compiled by Wang et al. [28], including 9905 interactions between 2346 circRNA and 962 miRNA. The data set contains CMI with high confidence, which is used as the benchmark data set in this study. The CMI-9589 dataset comes from the circBank database [39], and we select 9589 interactions of CMI between 2115 circRNA and 821 miRNA with high confidence as training data. The CMI-753 dataset is collected from the circR2Cancer database [40]. Through strict screening and processing, we have obtained 753 interactions of CMI between 515 circRNA and 469 miRNA in the latest version of the data, all of which are supported by experiments. In this experiment, we use this data for a case study.

In addition, we construct negative samples to balance the data set based on the uniqueness principle of sequence complementarity. Since miRNA has the response components (MRE), endogenous RNAs with a common MRE regulate each other’s expression by competitively binding to miRNA. This theory is called the competitive endogenous RNA hypothesis [41]. If a circRNA is determined to contain an MRE, it may be a potential target of a miRNA, and vice versa. In this study, negative samples are defined as interacting pairs that do not share common MREs, and we adopt specific negative sample construction methods for different data sets. Specifically, for the CMI-9905 dataset and CMI-9589 dataset, we construct all possible interactions between circRNA and miRNA, then delete the CMI with confidence score (score > 0) in the circBank database, and finally randomly select the same number of CMI as negative samples to participate in model training; For the CMI-753 data set based on real cases, since known interaction pairs are reported by experiments or papers, we select interaction pairs that have not been reported in existing studies as negative samples. Using this method can effectively avoid the potential CMI as a negative sample and ensure the reliability of the model performance.

Performance evaluation

In this section, we use the CMI-9905 as the benchmark dataset for performance evaluation. The data of the model in the five-fold CV is objectively recorded in Table 1.

The data in Table 1 shows that in the five-fold CV based on the benchmark data set, the average values of the six evaluation criteria of BEROLECMI are 0.8395, 0.8427, 0.8396, 0.8392, 0.9104, and 0.9086, respectively, which means that the proposed model can efficiently complete the prediction task of CMI. The ROC and PR curves of the BEROLECMI are shown in Fig. 5.

The ROC curve (A) and PR curve (B) of the BEROLECMI

Full size image

Performance on different datasets

To reflect the generalization ability of BEROLECMI in CMI prediction, we perform prediction tasks based on all commonly used datasets in the field of CMI prediction (CMI-9589, CMI-753). According to our statistics, more than 80% of CMI prediction models use the dataset adopted in this work as benchmark data. The experimental results based on commonly used datasets are shown in Table 2.

The data in Table 2 shows that in all common data sets in the field of CMI prediction, the AUROC of BEROLECMI based on the CMI-9589 data set exceeds 90%, and the AUROC based on the CMI-753 data set exceeds 75%, which means that the proposed model can effectively complete the CMI prediction task in commonly used data sets, and is expected to become a reliable candidate tool for CMI prediction.

The validity of the model feature extraction

In this section, we verify the effectiveness of feature extraction for each part of the BEROLECMI through independent experiments. Specifically, we divide BEROLECMI into three modules: sequence feature extraction (BE-A), self-similarity feature extraction (BE-B), and structured embedding (BE–C), and then use the three modules separately for feature extraction and perform prediction tasks to evaluate the proposed model for each module feature extraction effectiveness. The experimental results are objectively recorded in Table 3. To facilitate comparison, we use histograms to visualize the data in Table 3, as shown in Fig. 6.

Performance comparison of different modules

Full size image

The data in Table 3 shows that all the feature modules of the BEROLECMI can effectively complete the CMI prediction, which shows the effectiveness of the feature extraction strategy of the proposed method; among all three modules, the sequence feature extraction module has the lowest prediction results, which shows that the sequence feature is useful complements to model features; self-similarity features and structured embeddings achieve high predictive results, which means that combining functional similarity assumptions and role-defined structural embeddings can effectively improve the predictive performance of the model. Through the organic integration of the three feature extraction modules, we achieved the highest model prediction performance, verifying the effectiveness of the model construction.

Optimal classification strategy

In this study, we conduct prediction tasks based on different classifiers to determine the best classification strategy for the proposed method. In prediction tasks based on CMI-9905 datasets, we use the lightGBM [42], Random forest (RF) [43], Logistic Regression (LR) [44], Support Vector Machine (SVM) [45], Linear Regression (LinR) [46] for CMI prediction tasks, and the best classification strategy was selected by comparing the performance of the proposed methods. The prediction results are shown in Fig. 7.

Comparison of prediction results of different classifiers (A is the comparison of AUC results, B is the comparison of AUPR results)

Full size image

The data in Fig. 7 shows that the model using the lightGBM classifier achieves the highest predictive performance. LightGBM (Light Gradient Boosting Machine) [34] is an ensemble learning classifier based on Gradient Boosting Decision Tree (GBDT), which performs well in prediction tasks.LightGBM adopts the ensemble learning method to build a powerful prediction model by iteratively training multiple weak prediction models. It repeatedly optimizes the loss function, and each iteration builds a new decision tree on the residual of the previous model and then combines multiple decision trees to generate the final prediction result. LightGBM has the advantages of high efficiency, low memory consumption, accuracy, and support for large-scale data sets. Through the comparison of various classifiers, we finally choose LightGBM as the final classification strategy of the model.

Compared with the existing models

To evaluate the advantages of BEROLECMI in the CMI prediction, we compared the proposed model with other models in the CMI prediction field based on three commonly used data sets.

Lan et al. proposed the NECMA model, which combines circRNA-miRNA association, circRNA Gaussian kernel similarity and miRNA Gaussian kernel similarity to construct a heterogeneous network, then uses the NetMF algorithm based on matrix decomposition to extract hidden features in the heterogeneous network, and finally uses weighted neighborhood regularized logistic matrix decomposition and inner product obtain the circRNA-miRNA association probability [47]; Qian et al. proposed the CMIVGSD model, using the singular value decomposition algorithm and variational autoencoder to extract linear and nonlinear features from the circRNA-miRNA interaction network to predict unknown circRNA-miRNA interactions [32]; Wang et al. proposed the KGDCMI model, which combines RNA sequence feature and CMI network behavior feature to predict unknown CMI; Yu et al. proposed the first comprehensive prediction model of circRNA, SGCNCMI, which uses the graph neural network based on the contributing mechanism aggregates multi-modal information of molecules in biological networks and can achieve multiple predictions of circRNA-miRNA interactions, circRNA-gene interactions, and circRNA-cancer associations [30]; Guo et al. proposed the WSCD model, combined with the word2vec algorithm in natural language processing to process RNA sequences, used the SDNE algorithm to extract behavior features in the CMI network, and finally used a deep neural network to predict CMI [29]; He et al. proposed the GCNCMI model, using graph convolutional neural networks to aggregate node information to predict potential circRNA-miRNA interaction; Wang et al. proposed the JSNDCMI model, which for the first time combined denoising methods and local topological structure information in the CMI network for molecular feature extraction to predict unknown CMI [34]; Yao et al. proposed the IIMCCCMA model, which combined matrix factorization and improved inductive matrix completion algorithms predict unknown CMI [48]. Wang et al. proposed BioDGW-CMI, which combines BERT and wavelet diffusion to extract sequence and association network structure information of RNA molecules to predict potential CMI [35]. These models have achieved exciting results in CMI prediction, and we compare the BEROLECMI with these models to reflect the superior performance of the proposed model.

The comparison data are recorded in Table 4. It is worth noting that all the comparison data in this study use the same data and verification methods as the comparison models, and the number of comparison models exceeds 70% of all models in the field of CMI prediction.

Data based on the CMI-753 dataset comes from the work of Yao et al. [48].

The data in Table 4 shows that BEROLECMI surpasses all known models in the performance comparison of all three datasets.

In addition, we conduct a paired t-test based on the CMI-9905 data set and known advanced models in the CMI field to evaluate the statistical difference between the proposed model and the state-of-the-art (SOTA) model. The experimental results are recorded in Table 5.

The data in Table 5 shows that in the verification of the proposed method with the SOTA model, the P values were less than 0.05 confidence level, which means that there is a significant difference between the proposed model and the comparison model, and it has better statistical validity predict performance. There is no doubt that the proposed method is currently the most competitive in the field of CMI prediction.

Case study

To verify the practicability of BEROLECMI, we conducted a case study based on the CMI-753 dataset. The data in this dataset are all manually collected from existing literature and research, and all have experimental support.

In the case study, we perform a prediction task based on 15 pairs of CMIs to simulate the prediction performance of the proposed model in unknown CMIs. Specifically, we remove the interacting pairs for case studies in the CMI-753 data and then use the known 738 pairs of CMIs for model training to predict unknown CMIs. The prediction results are recorded in Table 6.

The data in Table 6 shows that among the 15 pairs of CMI used for prediction, 14 pairs were successfully predicted, which means that BEROLECMI can effectively predict CMI in real cases, and it is expected to be a powerful tool to provide pre-selection for wet experiments.

The circRNA-miRNA-mediated model provides new hope for the diagnosis and treatment of many complex diseases, but the biological properties of the mediated model are not completely clear, so there is an urgent need to speed up the discovery of CMI.

Influenced by the neighbor relationships in community networks, existing CMI prediction models typically rely on the nearest neighbor relationships in biological networks as the main modeling method. For example, KGDCMI and WSCD employ High-Order Proximity preserved Embedding (HOPE) and Structural Deep Network Embedding (SDNE) respectively, to capture high-order neighbor information and network structural features in heterogeneous CMI networks. Others like SGCNCMI, GCNCMI, and KS-CMI utilize graph convolutional neural networks to aggregate features of central nodes and neighbor nodes in the graph. In essence, these models use the nearest neighbor relationship set of the network or the nodes in the graph for feature extraction and aggregation. Although they can effectively predict CMI, only paying attention to a single structural feature may lead to difficult performance improvement. In addition, some models try to use other types of features in the CMI network for modeling, such as BioDGW-CMI uses the method based on wavelet diffusion to obtain the node network structure; JSNDCMI uses the multi-structure feature extraction framework to extract the topology features and functional similarity features of the nodes in the network. The prediction performance of these models is better than that of other models by adding a variety of molecular network structure features to feature extraction, which proves the advantage of multi-structure feature extraction in CMI prediction. However, although the existing methods have made progress, the existing methods still lack a targeted structure definition in the network. In this work, we propose a CMI prediction method BEROLECMI to simulate the molecular topology in the network. In this method, nine kinds of graphlet between each node and its adjacent four nodes in the CMI network are defined to extract topological features, then the attributes, self-similarity, and topological features of molecules are modeled to define unique role features for each molecule, thus inferring unknown CMIs.

BEROLECMI achieved the highest prediction performance in all commonly used datasets in the field of CMI prediction. Among them, in CMI-9905, the prediction results were 0.78% higher than the second-highest model; in the sparser CMI-753, the prediction results were higher than the second-highest model. 0.86%, which is much higher than other models by more than 10%. This means that topological structure features are an effective means of extracting network features for molecular association prediction, especially in the context of sparse relationships. In the case study, the proposed model accurately predicted 14 out of 15 pairs of CMIs. Excellent experimental results show that this method can effectively improve the performance of CMI prediction.

Although our method achieved promising results, there are still certain limitations that need to be addressed. In terms of feature extraction, although topological structures show higher advantages in sparse networks compared with nearest-neighbor structures, modeling of a single structure type shows limited capabilities in prediction tasks involving multi-type data. Therefore, the targeted use of multi-structure feature extraction will be the key to further improving prediction performance. In data construction, we use an equal amount of randomly generated negative samples for model training, which may cause potential CMI to be used as false negative samples, which can lead to a decline in model performance. We hope to continue to optimize the data used for training in subsequent research to further improve prediction performance. However, it is undeniable that BEROLECMI is currently the most competitive CMI prediction method and has proven the effectiveness of the feature construction method based on role definition, which is expected to provide a reference for subsequent research.

The data and source code can be found at https://github.com/1axin/BEROLECMI.

This work was supported by the National Natural Science Foundation of China (No.62273284), and in part by the NSFC Program, under Grant 62072378, and 62002297.

Authors and Affiliations

1. School of Information Engineering, Xijing University, Xi’an, China

   Xin-Fei Wang & Chang-Qing Yu

2. School of Computer Science, Northwestern Polytechnical University, Xi’an, China

   Zhu-Hong You

3. Key Laboratory of Symbol Computation and Knowledge Engineering of Ministry of Education, College of Computer Science and Technology, Jilin University, Changchun, China

   Yan Wang & Lan Huang

4. School of Artificial Intelligence, Jilin University, Changchun, China

   Yan Wang

5. College of Agriculture and Forestry, Longdong University, Qingyang, China

   Yan Qiao

6. School of Computer Science and Technology, China University of Mining and Technology, Xuzhou, China

   Lei Wang & Zheng-Wei Li

7. Guangxi Academy of Sciences, Nanning, China

   Lei Wang

Contributions

X-FW, C-QY, Z-HY, Y-W: conceptualization, methodology, software, and data curation; L-H, Y-Q, L-W, Z-WL: conceptualization, validation, resources; X-FW wrote the main manuscript. All authors contributed to the manuscript revision and approved the submitted version.

Corresponding authors

Correspondence to Chang-Qing Yu, Zhu-Hong You or Yan Wang.

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions