Learning Representations to Predict Intermolecular Interactions on Large-Scale Heterogeneous Molecular Association Network.

Molecular components that are functionally interdependent in human cells constitute molecular association networks. Disease can be caused by disturbance of multiple molecular interactions. New biomolecular regulatory mechanisms can be revealed by discovering new biomolecular interactions. To this end, a heterogeneous molecular association network is formed by systematically integrating comprehensive associations between miRNAs, lncRNAs, circRNAs, mRNAs, proteins, drugs, microbes, and complex diseases. We propose a machine learning method for predicting intermolecular interactions, named MMI-Pred. More specifically, a network embedding model is developed to fully exploit the network behavior of biomolecules, and attribute features are also calculated. Then, these discriminative features are combined to train a random forest classifier to predict intermolecular interactions. MMI-Pred achieves an outstanding performance of 93.50% accuracy in hybrid associations prediction under 5-fold cross-validation. This work provides systematic landscape and machine learning method to model and infer complex associations between various biological components.

Introduction

A key goal of life science research is to understand the complex association between biomolecules in various functional systems of a cell, which is important for many biomedical researches, for instance, exploring the pathogenesis of cancer, analyzing genetic diseases, and developing drugs and vaccines. Various molecular components and their interactions play important roles in life activities in cells. For example, proteins are the direct bearers of many fundamental life activities (Zhang et al., 2012; You et al., 2010; Marcotte et al., 1999). Most drugs work by binding to a specific protein, altering its biochemical and/or biophysical activities, thereby having multiple effects on multiple functions (Ay et al., 2007). Emerging evidence shows that non-coding RNA (ncRNA), genes that cannot be translated into protein, also play a significant biological role in metabolism, tumorigenesis, and cellular development (Gibb et al., 2011; Yi et al., 2018; Bartel, 2004), including microRNAs, long ncRNAs, and circular RNAs. Microbes, as environment or co-evolved partner, also have critical impacts on human's health and disease (Dethlefsen et al., 2007; Ma et al., 2016; Jostins et al., 2012). These molecules and their synergistic interactions maintain the special cellular activities, operating as part of a highly interconnected molecular association network.

Owing to the rapid development of related molecular biology, computational biology, and omics research, many valuable researches on individual intermolecular associations in human were developed and a variety of valuable experimental data have been released, e.g., mRNA-protein interactions (McCarthy and Kollmus, 1995; Peritz et al., 2006), long non-coding RNA (lncRNA)-protein interactions (Yi et al., 2018), protein-protein interactions (You et al., 2017b), micro RNA (miRNA)-protein interactions (Dweep and Gretz, 2015), and miRNA-lncRNA interactions (Huang et al., 2017, 2018). Considering exogenous chemical compound or complex disease, there are drug-disease interactions (Wang et al., 2018; Dumbreck et al., 2015; Zhang et al., 2018), miRNA-disease associations (You et al., 2017a; Wang et al., 2019), drug-protein interactions (Hu et al., 2016; Li et al., 2017), protein-disease associations (Lee et al., 2012; Wang et al., 2018), and lncRNA-disease associations (Chen, 2015). Emerging research on circular RNA (circRNA) shows there are also circRNA-miRNA associations (Zhang et al., 2017), circRNA-protein interactions (Chen et al., 2017), and circRNA-disease associations (Zhao et al., 2018). The microbes and drugs also have been involved into many biological systems (Ma et al., 2016; Sun et al., 2018).

These researches focus on individual associations between two molecules, and there are several studies that have considered the association between multiple biomolecules; e.g., Davis et al. manually compiled interactions among chemical, gene, and disease from publications to construct a chemical-gene-disease network (Davis et al., 2008). Liu et al. connected the associations between miRNA, target gene, lncRNA, and disease to build a network for calculating the similarity of miRNA and disease to predict miRNA-disease associations (Liu et al., 2017). The concerns of these studies are on two or very limited intermolecular relationships. However, intermolecular interactions are widespread and interconnected.

Inspired by this systematic perspective, to address some limitations of existing studies, we propose a molecular association network (MAN)-based framework to predict molecule-molecule associations by learning behavior and attribute feature of biomolecules, named MMI-Pred. The workflow of MMI-Pred is shown in Figure 1. First, a comprehensive MAN network is generated by connecting extensive associations between miRNAs, lncRNAs, circular RNAs, mRNAs, proteins, drugs, microbes, and diseases. It contains 14,315 molecular nodes and 18 kinds of, 114,150 association entries. Then, a random walk and skip-gram algorithm-based network embedding model node2vec is adopted to learn the behavioral features of biomolecular nodes. And the attribute feature is also calculated from sequence, structure, and phenotype information of different biomolecules. Moreover, both the attribute and behavior features are combined to train a Random Forest classifier (Breiman, 2001) to predict intermolecular associations. To evaluate the performance of MMI-Pred, the predictive ability of the entire MAN is first evaluated under 5-fold cross-validation. Furthermore, MMI-Pred was applied to predict miRNAs most relevant to Breast Neoplasms and Colon Neoplasms as a case study. Experimental results demonstrate that this work brings new insights and a promising prediction method for discovering and understanding intermolecular associations.

Figure 1

The Workflow of the MMI-Pred

The molecular association network is formed by connecting multitype intermolecular associations among mRNAs, proteins, miRNAs, lncRNAs, circRNAs, drugs, microbes, and diseases. Both the handcrafted attribute features and behavior features learned by network embedding method of biomolecules are jointly fed into a random forest classifier for training to predict potential intermolecular interactions.

Results

Molecular Association Network

The extensive associations between mRNAs, proteins, miRNAs, drugs, lncRNAs, circRNAs, microbes, and diseases are interconnected and form a complex molecular association network. Considering that same biomolecule may have different naming in different databases, we used the same naming convention to unify the naming of the eight molecules (nodes), e.g., STRING ID for protein (Szklarczyk et al., 2018), miRBase ID for miRNA (Kozomara et al., 2018), NONCODE ID for lncRNA (Fang et al., 2017), circBase ID for circRNA (Glažar et al., 2014), DrugBank ID for drug (Wishart et al., 2017), and NIH MeSH ID for microbe and disease. Then, the duplicate and completely isolated associations are removed. Finally, there are 14,315 molecule nodes and 114,150 association links in the MAN. The distribution of molecules nodes and association types is shown in Figure 2. The MAN obtained 39,674 protein-protein interactions from STRING v11 (Szklarczyk et al., 2018); 421 circRNA-disease associations (links) from Circ2Disease (Yao et al., 2018), CircRNA disease (Zhao et al., 2018), LncRNADisease 2.0 (Bao et al., 2018), and CircR2Disease (Fan et al., 2018); 1,378 circRNA-miRNA associations from SomamiR 2.0 (Bhattacharya and Cui, 2015); 3,416 mRNA-disease associations from DisGeNET (Piñero et al., 2017); 175 microbe-disease associations from HMDAD (Ma et al., 2016); 17,414 drug-disease interactions from CTD (Davis et al., 2018); 3,915 drug-mRNA associations from PharmGKB (Hewett et al., 2002); 8 drug-microbe associations from PharmacoMicrobiomics (R Rizkallah et al., 2012); 11,396 drug-protein interactions from DrugBank (Wishart et al., 2017); 874 lncRNA-disease associations from LncRNADisease (Chen et al., 2012) and lncRNASNP2 (Miao et al., 2017); 525 lncRNA-mRNA interactions from LncRNA2Target (Cheng et al., 2018); 8,634 lncRNA-miRNA interactions from lncRNASNP2 (Miao et al., 2017); 5,115 lncRNA-protein interactions from NPInter v2.0 (Yuan et al., 2013); 10,696 miRNA-disease associations from HMDD (Li et al., 2013); 3,012 mRNA-protein associations from NCBI data; 269 miRNA-drug associations from SM2miR (Liu et al., 2012); 5,186 miRNA-mRNA associations from MiRTarBase (Chou et al., 2017); and 2,042 miRNA-protein interactions from NPInter v2.0 (Yuan et al., 2013) and TransmiR v2.0 (Tong et al., 2018).

Figure 2

The Number and Type Distribution of Biomolecule Nodes and Intermolecular Associations in the Molecular Association Network

Predictive Performance Evaluation of MMI-Pred

The overall performance of MMI-Pred for predicting potential associations between arbitrary molecules in the MAN network was first evaluated under 5-fold cross-validation. In each fold validation, only the associations in the train set can be used to exploit the latent high-level representation of biomolecules nodes by network embedding model, which can avoid label leakage. As many studies have confirmed, there is a bias in measuring the performance of machine learning models using only precision or recall rates. When evaluating the classification performance of a model, the precision-recall curve and area under the precision-recall curve (AUPR) values that balance these two metrics are adopted. The overall performance of MMI-Pred is shown in Figure 3 and Table 1.

Figure 3

The Performance of MMI-Pred on Entire MAN Dataset under 5-Fold Cross-Validation

On the left is the ROC curve and AUC value, and on the right is the precision-recall curve and AUPR value.

Table 1

The 5-Fold Cross-Validation Performance of MMI-Pred on MAN Dataset

Fold	Acc. (%)	Sen. (%)	Spec. (%)	Prec. (%)	MCC (%)	AUC (%)
0	93.42	91.64	95.2	95.02	86.9	97.81
1	93.51	91.60	95.43	95.25	87.09	97.84
2	93.48	92.08	94.87	94.72	86.99	97.72
3	93.43	91.62	95.24	95.06	86.91	97.76
4	93.64	91.82	95.47	95.3	87.35	97.86
Average	93.50 ± 0.09	91.75 ± 0.20	95.24 ± 0.24	95.07 ± 0.23	87.05 ± 0.19	97.80 ± 0.06

As Figure 3 shows, in each fold cross-validation, the performance of MMI-Pred is very closed, which means the robust of our model. In whole MAN network, the model obtained a remarkable performance with high accuracy of 93.50% and high area under the curve (AUC) value of 0.9780. And the sensitivity, specificity, and precision of the model are 91.75%, 95.24%, 95.07%, respectively. The MMI-Pred receives a high AUPR value of 0.9707. In the case of class imbalance in classification tasks, accuracy is meaningless, for example, suppose there are 90 negative samples and 10 positive samples in a dataset, even if the model directly classified all samples into negative samples, the accuracy even is 90%, but this is obviously meaningless. And when the thresholds are different, the outputs are different. So, receiver operating characteristic (ROC) curve that can avoid these problems was used to measure our model's performance. The standard deviation (SD) of each performance value is 0.09%, 0.20%, 0.24%, 0.23%, 0.19%, and 0.06%, respectively, which can show the stable and robust of MMI-Pred in predicting any molecule-molecule associations in the MAN.

Evaluate the Impact of Network Behavior and Attribute Feature

Molecules in the association network are similar to people in social networks, and they have both attributes and network behavior features. Both the network behavior and attribute features are adopted as representations of biomolecules. For mRNA, miRNA, lncRNA, circRNA, and protein, their attributes are nucleic acid or amino acid sequence. The k-mer is used to transfer sequences into numerical vector. For disease and microbe, their direct attribute is hard to gain, their phenotypes are employed to calculate their semantic similarity as attribute feature. The fingerprints of drug compounds that stand for the chemical structure are used as their attribute. All nodes in the MAN network can be calculated for their network embedding based on their behavior with other nodes in the network. We tested them under the same experimental conditions to verify the performance of these features and their impact on the predicted results.

As Figure 4 and Table 2 show, the MMI-Pred model can achieve high accuracy more than 90% whether using attribute features or behavior features, which indicate that the distinguishing power of features is acceptable. In general, the performance of behavior feature is a bit better than the attribute features, whereas the best performance is obtained when using both two features. In addition, when the nodes or network behavior attributes of some new molecules are missing, the combination of these two features can enhance the robustness of the model and ensure that the prediction can be performed normally.

Figure 4

The Comparison of Network Behavior and Attribute Features Using Random Forest Classifier

On the left is the ROC curve and AUC value, and on the right is the PR curve and AUPR value.

Table 2

Comparison of Attribute and Behavior Features Using Random Forest Classifier under 5-Fold Cross-Validation

Feature	Acc. (%)	Sen. (%)	Spec. (%)	Prec. (%)	MCC (%)	AUC (%)
Attribute	90.69 ± 0.14	89.49 ± 0.19	91.89 ± 0.15	91.69 ± 0.15	81.40 ± 0.27	95.85 ± 0.12
Behavior	91.64 ± 0.18	88.44 ± 0.15	94.83 ± 0.21	94.48 ± 0.23	83.45 ± 0.36	96.87 ± 0.17
Combined	93.50 ± 0.09	91.75 ± 0.20	95.24 ± 0.24	95.07 ± 0.23	87.05 ± 0.19	97.80 ± 0.06

Compared with Widely Used Machine Learning Classifiers

To verify the impact of different machine learning models on performance, in this section, we compared the performance of the Logistic Regression (LR), AdaBoost, Naive Bayes (NB), XGBoost, and Random Forest as classifier of our framework using the attribute and behavior feature under the same experimental conditions. The Random Forest classifier and other contrast classifiers are implemented by Scikit-learn (Pedregosa et al., 2013) and use only default parameters.

As shown in Figure 5 and Table 3, the proposed method MMI-Pred that uses Random Forest classifier achieves the best performance. LR is a commonly used binary classification algorithm that directly models the possibility of classification without the assumption of data distribution in advance. AdaBoost is the most famous representative of the Boosting algorithm. It requires the base classifier to learn specific data distributions, which can be achieved by re-weighting. The NB classifier is a series of simple probability classifiers based on Bayesian theorem based on independence between hypothetical features. XGBoost is an improvement of the Gradient Boosting Decision Tree (GBDT) implementation. Random Forest is an efficient, fast, and easy-to-use decision tree-based algorithm, which was proved to be the most effective model in this task by rigorous experimental results.

Figure 5

The Performance Comparison between MMI-Pred and Four Different Comparison Models Include Naive Bayes, Adaboost, Logistic Regression, and XGBoost Classifiers

Table 3

The Performance Comparison of Different Machine Learning Classifiers

Method	Acc. (%)	Sen. (%)	Spec. (%)	Prec. (%)	MCC (%)	AUC
NB	59.64	31.35	87.94	72.2	23.38	75.57
LR	80.61	82.44	78.79	79.54	61.27	87.21
AdaBoost	80.91	82.68	79.14	79.86	61.86	88.5
XGBoost	85.67	78.66	92.68	91.48	72.05	94.44
Proposed method	93.50	91.75	95.24	95.07	87.05	97.80

Case Study: Predicting Human Disease-Associated miRNAs

To demonstrate the predictive ability of the proposed model on specific types of interactions, the MMI-Pred was executed to predict the miRNAs that are most relevant to two diseases, including Breast neoplasms and Colon neoplasms, as case studies. In the MAN, all miRNA-disease associations are from the HMDD database. When conducting case studies for individual disease, we trained the MMI-Pred predictor with a MAN network that removed those miRNA-Breast neoplasms (or Colon neoplasms) association pairs that overlapped with the dbDEMC 2.0 database (Yang et al., 2016). Then, the trained model performs prediction on testing Breast neoplasms or Colon neoplasms-miRNAs pairs. This processing can also be considered as cross-dataset validation. In the context of screening for disease-associated miRNAs, the candidate rankings are more valuable than the report of the overall false-positive, false-negative, and other indicators of the framework. Therefore, when the MMI-Pred is executed on the test samples, we rank the possible associated miRNAs based on the probability values output by the MMI-Pred. And then, the top 30 high-scored miRNAs for each disease are validated through the dbDEMC database.

Breast cancer is the most terrible killer of women's health. In 2018, about 2.1 million new cases of Breast tumor in women were diagnosed globally. And breast cancer accounts for about a quarter of the globally diagnosed cases of female cancer (Bray et al., 2018). Among the world's latest cancer incidence rates, female breast cancer also ranks second, accounting for 11.6% of the total cancer population. Studies have shown that miRNAs have the most significant expression difference between normal and cancer tissues, which can be used as tumor markers (Iorio et al., 2005). As shown in Table 4, the top 30 highest ranked breast cancer-associated miRNAs are predicted by MMI-Pred, and 25 of them were confirmed.

Table 4

The Top 30 miRNAs Relevant to Breast Cancer Predicted by MMI-Pred

miRNA	dbDEMC	miRNA	dbDEMC
hsa-mir-186-5p	Confirmed	hsa-mir-539-5p	Confirmed
hsa-mir-216a-5p	Unconfirmed	hsa-mir-330-5p	Confirmed
hsa-mir-154-5p	Confirmed	hsa-mir-543	Confirmed
hsa-mir-181d-5p	Confirmed	hsa-mir-4262	Unconfirmed
hsa-mir-449b	Confirmed	hsa-mir-384	Confirmed
hsa-mir-211-5p	Confirmed	hsa-mir-4458	Confirmed
hsa-mir-504-5p	Unconfirmed	hsa-mir-28-5p	Confirmed
hsa-mir-1271-5p	Confirmed	hsa-mir-136-5p	Confirmed
hsa-mir-300	Confirmed	hsa-mir-99b-5p	Confirmed
hsa-mir-337-5p	Confirmed	hsa-mir-518-5p	Unconfirmed
hsa-mir-637	Confirmed	hsa-mir-217	Confirmed
hsa-mir-517a-3p	Confirmed	hsa-mir-664	Confirmed
hsa-mir-671-5p	Confirmed	hsa-mir-508-5p	Confirmed
hsa-mir-525-5p	Unconfirmed	hsa-mir-431-5p	Confirmed
hsa-mir-532-5p	Confirmed	hsa-mir-483-5p	Confirmed

Colon cancer ranks fourth in overall cancer incidence, accounting for 6.1%, but ranks second in mortality, accounting for 9.2% (Bray et al., 2018). And recent research confirms that miRNAs play a role in carcinogenesis through DNA methylation and histone modifications and human colorectal tumorigenesis (Bandres et al., 2009). The predicted top 30 miRNAs with the highest score that associated with Colon Neoplasms are shown in Table 5; among them, 26 of miRNA-disease associations were confirmed.

Table 5

The Top 30 miRNAs Relevant to Colon Cancer Predicted by MMI-Pred

miRNA	dbDEMC	miRNA	dbDEMC
hsa-mir-186-5p	Confirmed	hsa-mir-16-5p	Confirmed
hsa-mir-485-5p	Confirmed	hsa-mir-497-5p	Confirmed
hsa-mir-206	Confirmed	hsa-mir-33b-5p	Confirmed
hsa-mir-19b-3p	Confirmed	hsa-mir-7-5p	Unconfirmed
hsa-mir-361-5p	Confirmed	hsa-mir-185-5p	Confirmed
hsa-mir-154-5p	Confirmed	hsa-mir-26b-5p	Confirmed
hsa-mir-9-5p	Unconfirmed	hsa-mir-34c-5p	Confirmed
hsa-mir-122-5p	Confirmed	hsa-mir-449b-5p	Confirmed
hsa-mir-590-5p	Confirmed	hsa-mir-139-5p	Confirmed
hsa-mir-340-5p	Confirmed	hsa-mir-134-5p	Unconfirmed
hsa-mir-211-5p	Confirmed	hsa-mir-153-3p	Unconfirmed
hsa-mir-149-5p	Confirmed	hsa-mir-449a-5p	Confirmed
hsa-mir-183-5p	Confirmed	hsa-mir-129-5p	Confirmed
hsa-mir-503-5p	Confirmed	hsa-mir-136-5p	Confirmed
hsa-mir-324-5p	Confirmed	hsa-mir-10a-5p	Confirmed

Discussion

In this research, we proposed a computational framework based on network representation learning to predict any associations between molecules. First, the molecular association network is constructed by integrating 18 types of associations, 14,315 nodes, 114,150 molecular associations between mRNA, lncRNA, protein, miRNA, circRNA, drug, disease, and microbe. The performance of the framework is evaluated on the entire network under 5-fold cross-validation. To demonstrate the predictive ability, we use MMI-Pred to predict miRNAs most relevant to Breast cancer and Colon cancer as case studies. Experimental results proved that the MMI-Pred can predict any potential associations between molecules. Moreover, network embedding representations obtained based on MAN network and network representation learning algorithms can serve as efficient low-rank representations of disease, microbes, and other biological components whose features are difficult to be extracted by computational algorithms. In addition, randomly sampled unknown samples without known association are used as negative samples in this work; high-quality negative samples or sampling techniques are worth studying. It is anticipated that this work can help to advance related intermolecular associations research in a long term.

Limitations of the Study

In this study, we provide a systematic and holistic perspective on intermolecular interactions and provide a machine learning method to model molecular properties and intermolecular behaviors in order to promote understanding and discover new intermolecular interactions. This work still has some limitations that deserve attention and further study. First, the interactions screened from public databases for building molecular association networks are not complete, although to our best knowledge, these databases are already of high quality and relatively comprehensive. For nodes that do not exist in the network, the network embedding feature will be not applicable. More complete data will be more conducive to comprehensive modeling of the relationship between biomolecules. Second, the MAN network is a heterogeneous information network that contains many types of molecules and many different association relationships. When characterizing the network behavior of biomolecule nodes, the network embedding algorithm does not use the heterogeneous information. The further study of network representation learning algorithms for heterogeneous information networks will be very helpful.

Methods

All methods can be found in the accompanying Transparent Methods supplemental file.

Resource Availability

Lead Contact

Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Zhu-Hong You (zhuhongyou@ms.xjb.ac.cn).

Materials Availability

This study did not generate new materials.

Data and Code Availability

The datasets/code generated during this study are available at https://github.com/haichengyi/MAN.