Predicting miRNA-disease associations based on graph attention networks and dual Laplacian regularized least squares
Gendai:GitHub - shine-lucky/MKGAT-main
https://academic.oup.com/bib/article/23/5/bbac292/6645486
Abstract
Increasing biomedical evidence shows that miRNA dysregulation is associated with complex diseases in humans. The identification of disease-related miRNAs is of great significance for the prevention, diagnosis, and treatment of diseases. To reduce the time and cost of biomedical experiments, there is a strong incentive to develop efficient computational methods to infer potential miRNA-disease associations. Although many computational methods have been proposed to solve this problem, the accuracy of predictions needs to be further improved. Inthis study, we propose a computational framework MKGAT to predict possible associations between miRNAs and diseases using double Laplacian regularized least squares via graph attention networks (GATs) . We use GATs to learn each layer of miRNA and disease embeddings from initial input features of known miRNA-disease associations, intra-miRNA similarities, and intra-disease similarities. Then, we compute the kernel matrices of miRNAs and diseases based on the Gaussian interaction spectrum (GIP) and the learned embeddings. The kernel matrix and initial similarity of each layer are further fused using the attention mechanism. Bi-Laplacian regularized least squares was finally applied with fused miRNA and disease kernels for novel miRNA-disease association predictions. Compared with the 6 latest methods through 5-fold cross-validation, our method MKGAT obtained the highest AUROC value of 0.9627 and AUPR value of 0.7372. We used MKGAT to predict related miRNAs in three cancers and found that the top 50 predictions for the three diseases were confirmed by existing databases. The excellent performance suggests that MKGAT will be a useful computational tool to reveal disease-related miRNAs.
Keywords: miRNA-disease association, graph attention network, multi-kernel fusion, double Laplacian regularized least squares
Table of contents
2.1.1.Known human miRNA–disease associations
2.1.2.miRNA functional similarity
2.1.3.Disease semantic similarity
2.1.4.GIP kernel similarity for diseases and miRNAs
2.1.5.Integrated similarities for miRNAs and diseases
2.1.6.miRNA-disease bipartite network
2.2.1. GATs for feature extraction (GATs for feature extraction)
2.2.3. The predicted Double Laplas regular minimum daily method
3.2. The impact of different checks on prediction performance
3.3. Performance comparison with other methods
1. Introduction
In this study, we propose a computational framework, MKGAT, which combines GATs [45] and double Laplacian regularized least squares to predict potential miRNA-disease associations. First, input features are constructed based on known miRNA-disease associations, intra-miRNA similarities, and intra-disease similarities. GATs were used to study the embedding of miRNAs and diseases at each layer. Then, the kernel matrix of each layer of miRNA and disease embeddings is calculated based on Gaussian interaction spectrum (GIP), and the attention mechanism is used to fuse the kernel matrix of each layer with the initial similarity. Finally, novel miRNA-disease associations were predicted using Bi-Laplacian regularized least squares in the space of combined miRNA and disease kernels. Five-fold cross-validation shows that MKGAT’s area under the ROC curve (AUROC) is 0.9627 and the area under the precision recall curve (AUPR) is 0.7372, which is better than the six state-of-the-art prediction methods. Case studies of three cancers showed that all top 50 predictions were supported by established databases, further demonstrating the effectiveness of MKGAT in detecting disease-associated miRNAs.
2. Materials and methods
2.1.Benchmark datasets
2.1.1.Known human miRNA–disease associations
The dataset used in the study was downloaded from reference [26], in which Chen et al. collected 495 miRNAs, 383 diseases, and 5430 experimentally verified miRNA-disease associations from HMDD v2.0 [46]. We use 
and 
to represent the number of miRNAs and diseases, respectively, and 
to describe the adjacency matrix of miRNA-disease associations, where (= 495) represents the row number (miRNA), (= 383) represents the column number (disease). If there is a known association between miRNA m(i) and disease d(j), the value of 
at the corresponding position in the matrix is set to 1, otherwise it is 0.
2.1.2.miRNA functional similarity
Wang et al. [47] provided a method for calculating miRNA functional similarity based on the assumption that phenotypically similar diseases are more likely to be related to functionally similar miRNAs. We downloaded the functional similarities of miRNAs from their study https://www.cuilab.cn/files/images/cuilab/misim.zip. We constructed a matrix FS to describe the functional similarity between two miRNAs, where 
represents the sum of miRNA
and 
miRNA functional similarity scores between.
2.1.3.Disease semantic similarity
We use MeSH terminology to describe each disease as a directed acyclic graph (DAG). Specifically, we formulate disease di using 
, where 
represents the node set consisting of node di and its ancestor nodes, 
represents the corresponding set of edges containing direct links from ancestors to child nodes. According to reference [47], we calculate the semantic contribution of disease dt to di as follows:
where represents the semantic contribution attenuation factor, which is set to 0.5 in our study. The semantic value of a disease di can be calculated based on the contribution of the ancestral disease and the disease di itself as follows:
The semantic similarity between diseases di and dj 
is defined by the following formula:
At the same time, Chen et al. [26] proposed another method to calculate the semantic contribution of disease dt to di
Correspondingly, the semantic value of disease di and the semantic similarity between disease di and dj can be calculated by equations (5) and (6) respectively,
In this study, we calculate the final disease semantic similarity by combining the semantic similarities of the two diseases, and the disease semantic similarity between the two diseases can be calculated using Equation (7) SS (di, dj)
2.1.4.GIP kernel similarity for diseases and miRNAs
Similar to reference [48], a binary vector
was constructed to record the association between miRNA m (i) and all diseases. If there is an experimentally supported association between the two, the corresponding value of is set to 1, otherwise it is set to 0. The GIP core similarity between miRNA mi and mj can then be calculated 
as follows:
where 
represents the normalized core bandwidth and is obtained from equation (9) as follows:
where Nm represents the number of all miRNAs, 
is the original bandwidth, which is set to 1 in our study. Similarly, the GIP kernel similarity 
between two diseases di and dj can be calculated using the following two equations:
where 
represents the normalized kernel bandwidth, 
represents the number of all diseases, and 
is set to 1.
2.1.5.Integrated similarities for miRNAs and diseases
Considering that not all pairs of miRNAs have functional similarities, the integrated similarity between miRNA mi and mj 
is calculated as follows:
Similarly, the integrated similarity between diseases di and dj 
is calculated as follows:
This similarity integration strategy has been applied to miRNA-disease association inference in references [23, 26, 49].
2.1.6.miRNA-disease bipartite network
MiRNA-disease bipartite network G is defined by the adjacency matrix 
and its transpose 
2.2.Method architecture
In this section, we introduce the architecture of MKGAT for miRNA-disease association prediction. The workflow of MKGAT is shown in Figure 1.
2.2.1. GATs for feature extraction (GATs for feature extraction)
As a new neural network structure, GATs [45] were applied in our study to extract miRNA and disease features. Specifically, given the adjacency matrix G of the above bipartite network, GATs are defined as follows:
where
is the l-layer embedding of the node, l = 1,... , L, σ (•) is the nonlinear activation function (ReLU), and GAT represents a single graph Attention layer, the entire L layer structure is composed of multiple graph attention layers superimposed. The initial input is a set of node features 
, where N is the number of nodes and F is the number of features per node. This layer produces a new set of node features
, using a learnable linear mapping to transform the input features into higher values by applying a weighting matrix
to each node hierarchical characteristics. Then we calculate the attention coefficient
After normalization with the softmax activation function, the coefficients we get are
Substituting equation (16) into equation (17), the coefficient of the attention mechanism can be expressed as follows:
where α represents the attention coefficient, 
represents the parameterized weight vector, LeakyReLU represents the activation function (negative slope is 0.2), T represents the matrix transpose, | | represents the connection operation, Ni represents the neighbor of node i. After calculating the normalized attention coefficient, the final output feature of each node can be calculated as
In our study, we constructed the initial embedding of the first layer 
as follows:
2.2.2.Kernel combination
The multi-layer GAT model calculates the embedding of different layers, and the embedding of each layer represents different graph structure information. We compute the multi-kernel matrix by embedding each layer as a different feature vector respectively. Indicates that the embedding of each layer is 
, where 
is the embedding of miRNA in the l-th layer, and 
is the embedding of the disease in the l-th layer. layer embedding. We use GIP to calculate the kernel matrix embedded in each layer of miRNA and disease as follows:
where 
represents the l-th row in the l-th layer of the miRNA and disease embeddings, and 
represents the corresponding width.
Since the contributions of different embeddings at different layers are inconsistent, the kernel of the embedding calculation represents the similarity between different view nodes. For the miRNA space 
is the initial similarity matrix SM, for the disease space 
, 
is the initial similarity matrix SD. Use the attention mechanism to combine multiple kernel matrices (in two spaces) to obtain the final combined kernel
where
is the i-th core of the miRNA and disease cores, 
is the attention factor corresponding to each core, and L is the number of layers.
2.2.3. The predicted Double Laplas regular minimum daily method
50. Belkin M, Niyogi P, Sindhwani V. Manifold regularization: a
geometric framework for learning from Labeled and Unlabeled
examples. J Mach Learn Res 2006;7:2399–434.
We apply Bi-Laplacian regularized least squares [50] to the joint kernel matrix of two feature spaces to predict potential associations between miRNAs and diseases. The loss function is defined as
where
is the Frobenius norm, 
is the adjacency matrix of miRNA-disease association in the training set, 
is The trainable matrices, 
and 
are the combined kernels in the two feature spaces, and the parameters λm and λd are the coefficients of the regularization terms.
is the Laplacian regular matrix, defined as follows:
where 
is a diagonal matrix. The final miRNA-disease association predictions are as follows:
2.2.4. Parameter optimization
We use the Adam optimizer [51] to optimize the parameters in GAT and the attention factor for kernel fusion. For the parameters of double Laplacian regular least squares, we obtain the iterative function by directly calculating the partial derivatives. In order to optimize the parameter Wm, we fix the parameter Wd as a constant, and then calculate the partial derivative of the loss function with respect to Wm as follows:
3.Results
3.1. Experimental setup
3.2. The impact of different checks on prediction performance
3.3. Performance comparison with other methods
3.4. Case studies
4 Conclusion
The latest advances in life sciences indicate that miRNA plays a key role in development and physiological regulation. Therefore, miRNAs are becoming important biomarkers for disease diagnosis. Computational efforts to predict disease-associated miRNAs represent an excellent alternative to biomedical experiments. In this paper, we propose a computational framework MKGAT to discover associations between miRNAs and diseases. Comprehensive experiments including cross-validation and case studies demonstrated the effectiveness and superiority of MKGAT in revealing disease-related miRNAs.
Our framework MKGAT mainly consists of two parts. In the first part, we used an attention mechanism to extract features. Experiments show that exploiting this mechanism can produce more reliable inference information. The other part is the dual Laplacian regular least squares method for prediction. As an entity and semi-supervised method, our double Laplacian regularized least squares method fully utilizes the information from both the miRNA side and the disease side for prediction. Compared with HGANMDA [44], a method that uses hierarchical GAT for inference, there are three main differences between our model MKGAT and HGANMDA. First, MKGAT uses GIP to compute the kernel matrix. Secondly, MKGAT uses an attention mechanism to combine multiple kernel matrices. Third, MKGAT uses different strategies to make the final prediction. Compared tosupervised methods for miRNA-disease association prediction, our framework does not require negative samples for inference. As we stated in the introduction section, data on negative samples for miRNA-disease association prediction is difficult to collect, and randomly selected negative samples will output less satisfactory results.
It is worth noting that our method relies heavily on similarity measures for reasoning. As mentioned in previous research [59], integrating appropriate features for similarity calculation is a challenging task due to shortcomings in data availability in biomedical sciences. For example, a disease may not be included in the MeSH database. Furthermore, as we mentioned in the introduction section, parameter tuning and optimization in MKGAT is a tricky process. This is a common problem that needs to be solved in deep learning methods. Finally, as regulators of human disease progression, miRNAs can upregulate or downregulate gene expression. However, we did not differentiate between modes of regulation in this study. This will be a new direction for our future research.
