A multi-omics data classification method based on global-local graph transformer

Abstract

The present invention belongs to the field of intelligent computing and data mining, and discloses a multi-omics data classification method based on global-local graph transformer. First, for each type of omics data, the topological structure and semantic information of the graph are captured through the global-local graph transformer, combined with the graph convolution layer and the multi-head self-attention layer, so as to capture the global-local feature representation of different omics data; then, a comparative learning strategy is adopted to learn the similarities and differences between samples, so as to mine more discriminative and robust graph representations from the heterogeneous information of different omics data, and deeply understand the relationship between patients with complex diseases; finally, multi-omics data classification is realized through a deep fusion network. The method of the present invention fully captures the potential information at the feature level and the omics level, and has good classification performance, which is of great significance to improving the accuracy of clinical diagnosis results of patients and promoting precision medicine.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent computing and data mining, and specifically to a multi-omics data classification method based on a global-local graph Transformer, referred to as GCNFormer. Background Art

[0002] With the development of high-throughput sequencing technology, the amount and diversity of various types of omics data, including genomics, transcriptomics, proteomics, metabolomics, and epigenomics, have increased exponentially. Studies have shown that each type of omics data provides a unique perspective for revealing the complex details of biological systems. The integration of multi-omics data provides an opportunity to fully understand the complex molecular mechanisms behind health and disease. Therefore, it is crucial to propose efficient integrated analysis methods to process rich and heterogeneous multi-omics data.

[0003] Multi-omics data integration and omics feature extraction are important research topics in multi-omics data integration and analysis. The integration process requires merging heterogeneous data from different omics into a unified framework. This task requires addressing issues such as data heterogeneity, different measurement scales, and complex interactions between biological dimensions. Effective integration of multi-omics data can fully present biological systems and provide a deep understanding of biological processes and disease mechanisms. Moon et al. proposed a multi-omics integration method using an attention mechanism to improve diagnostic performance. Li et al. combined similarity networks and graph convolutional networks for omics data integration. Han et al. proposed a dynamic multimodal fusion strategy that uses an encoding network to handle the dynamics at the feature level and modality level to achieve reliable multimodal fusion. On the other hand, omics feature extraction aims to identify and capture key information and patterns in each omics data, including selecting appropriate feature extraction methods, handling noise and missing values, and ensuring that the extracted features can accurately reflect the biological characteristics of each omics data. The quality of omics data feature extraction directly affects the results and interpretation of multi-omics data integration analysis. Wang et al. proposed MOGONET, a supervised multi-omics integration method that uses graph convolutional networks for omics-specific learning. Zheng et al. combined feature confidence learning and label confidence learning to deeply explore the potential feature information of each omics data. In-depth research on multi-omics data integration and omics feature extraction is of vital importance for understanding complex biological systems and revealing disease mechanisms. However, due to the heterogeneity of data and the complex interactions between different biological layers, there are still many challenges in the integration and analysis of multi-omics data. Therefore, new multi-omics integrated analysis methods are needed to fully capture the potential information at the feature level and the omics level to further improve the classification performance. Summary of the invention

[0004] In order to overcome the defects of the above-mentioned multi-omics data classification method, the present invention provides a multi-omics data classification method based on a global-local graph Transformer.

[0005] The present invention comprises the following steps:

[0006] Step 1: Given three types of omics data, ,For each type of omics data, the corresponding sample similarity network is constructed using cosine similarity;

[0007] Step 2: A global-local graph transformer is proposed, which combines graph convolutional layers with multi-head self-attention layers to capture the topological structure and semantic information of the graph, so as to comprehensively consider the global-local feature representation of different omics data;

[0008] Step 3: Using contrastive learning strategies, learn the similarities and differences between samples, mine discriminative and robust graph representations from the heterogeneous information of different omics data, deeply understand the relationships between patients with complex diseases, and achieve accurate classification of patients with complex diseases;

[0009] Step 4: Use deep fusion network to realize multi-omics data classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0011] Figure 1 It is a flow chart of a multi-omics data classification method based on a global-local graph Transformer provided by an embodiment of the present invention;

[0012] Figure 2 It is a graph showing the experimental results of comparing the classification performance of different types of omics data on four biomedical datasets, namely ROSMAP, LGG, BRCA and KIPAN, according to an embodiment of the present invention; DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the present invention clearer, the specific implementation methods of the present invention are further described in detail. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0014] A multi-omics data classification method GCNFormer based on global-local graph Transformer, such as Figure 1 As shown, the following steps are included:

[0015] Step 1: Given three kinds of omics data, namely mRNA expression, DNA methylation and miRNA expression, for each type of omics data, the corresponding sample similarity network is constructed using cosine similarity to obtain the connection relationship between samples.

[0016] Step 2: A global-local graph transformer is proposed, which combines the graph convolution layer with the multi-head self-attention layer to capture the topological structure and semantic information of the graph, so as to comprehensively consider the global-local feature representation of different omics data. The specific steps are as follows:

[0017] Step 2.1: For each omics data, Layer Norm normalization is first used to ensure numerical stability during training. The original feature matrix After Layer Norm standardization, , is the number of samples, and then a graph convolution layer is trained to aggregate neighborhood information to learn the discriminative local information of omics data. The graph convolution layer is defined as:

[0018]

[0019] in, is the output of the graph convolution layer, i.e. the key local structure information learned. is the corresponding weight matrix, , is the initial adjacency matrix, is the identity matrix, yes The corresponding degree matrix, represents a non-linear activation function.

[0020] Step 2.2: For each omics data, train a multi-head self-attention layer, and a single attention mechanism is defined as:

[0021]

[0022] in, It is The output of the head attention mechanism, , , , , and are the corresponding parameter matrices respectively, and the multi-head self-attention layer is defined as:

[0023]

[0024] in, represents the number of heads of the multi-head self-attention mechanism, represents the corresponding weight matrix, Represents the feature representation learned by the multi-head self-attention layer, which is input into the multi-layer perceptron to obtain omics-specific feature representations under different omics data types .

[0025] Step 3: Use contrastive learning strategies to learn the similarities and differences between samples, mine discriminative and robust graph representations from the heterogeneous information of different omics data, deeply understand the relationships between patients with complex diseases, and achieve accurate classification of patients with complex diseases. The specific steps are as follows:

[0026] Step 3.1: Represent the learned specific features under different omics data types Splicing to initially obtain a consistent representation of multi-omics data, and using the self-attention layer to obtain a consensus representation of multi-omics data and the global structure relationship matrix .

[0027] Step 3.2: Introduce a contrastive learning strategy to learn the similarities and differences between samples. Specifically, calculate the cosine similarity between the specific feature representations of various omics data and the consensus feature representation to measure the similarity between samples, better understand the heterogeneity information between various omics data, and thus improve the discrimination and robustness of classification. The specific calculation formula is as follows:

[0028]

[0029] in, and express and No. and samples, Indicates Omics-specific feature representations learned by global-local graph Transformer under various omics data.

[0030] Step 3.3: Construct a contrast loss function to maximize the similarity between samples of the same type and minimize the similarity between samples of different types. The calculation method is as follows:

[0031]

[0032] in, is the contrast loss function, Represents the temperature parameter.

[0033] Step 4: Use deep fusion network to classify multi-omics data. Input into the deep fusion network to get the final classification prediction result. The deep fusion network is defined as follows:

[0034]

[0035] in, Represents a deep fusion network, which is essentially a two-layer fully connected network. Predict the probability distribution for the final sample classification. Multiple networks are jointly trained end-to-end to achieve the final accurate classification.

[0036] In order to verify the effectiveness of the present invention, the following verification and analysis are performed:

[0037] (1) Classification performance comparison experiment verification

[0038] In order to evaluate the effectiveness of GCNFormer, a multi-omics data classification method based on global-local graph Transformer proposed in the present invention, it was applied to four datasets of different biomedical classification tasks, namely, Alzheimer's disease dataset (ROSMAP), invasive breast cancer dataset (BRCA), low-grade glioma dataset (LGG) and kidney cancer dataset (KIPAN) for patient classification, and three types of omics data were used: mRNA expression data (mRNA), DNA methylation data (meth), and miRNA expression data (miRNA). Specifically, for the ROSMAP dataset, Alzheimer's disease patients (AD) and normal samples (NC) were classified into two categories; for the LGG dataset, they were graded for two categories; for the BRCA dataset, breast cancer patients were divided into 5 subtypes by PAM50; for the KIPAN dataset, they were multi-classified according to the type of kidney cancer. The dataset information is shown in Table 1.

[0039] Table 1 Dataset information used

[0040]

[0041] In order to evaluate the performance of the present invention in classification tasks, multiple evaluation indicators are used. For binary classification tasks, three indicators, ACC, F1 and AUC, are used. For multi-classification tasks, the classification performance is evaluated by ACC, F1_weighted and F1_macro. In the experiment, each data set is randomly divided into a training set consisting of 70% of its samples and a test set consisting of the remaining 30% of the samples. In order to comprehensively analyze the classification performance of the GCNFormer of the present invention, training sets and test sets are randomly generated and 10 experiments are calculated, and the average and standard deviation of the experimental results are taken. The comparison methods are: 1) k-nearest neighbor classifier (KNN); 2) random forest (RF); 3) a gradient tree boosting based classifier (XGBoost); 4) fully connected neural network classifier (NN); 5) MOGONET, a multi-omics graph convolutional network, used to explore omics-specific learning and cross-omics related learning to achieve multi-omics data classification; 6) MOMA, a multi-task attention learning algorithm based on multi-omics data; 7) MOADLN: a multi-omics integration framework based on multi-omics attention deep learning network; 8) MMdynamics, a reliable multimodal classification algorithm that can dynamically evaluate the amount of information at the feature level and modality level; 9) MLCLNet: multi-level confidence learning for multi-trust multimodal classification, combining feature confidence learning, label confidence learning and cross-modal label fusion to achieve accurate classification of multi-omics data of complex diseases; 10) GCFANet: a multi-omics data classification network through global and cross-modal feature aggregation.

[0042] Table 2 Classification results on the ROSMAP dataset

[0043]

[0044] Table 3 Classification results on the LGG dataset

[0045]

[0046] The comparative experimental results on the binary classification dataset are shown in Tables 2 and 3. Compared with other methods, GCNFormer has higher ACC, F1 and AUC values, indicating its superior performance in binary classification tasks. The comparative experimental results on the multi-classification dataset are shown in Tables 4 and 5. GCNFormer has superior classification performance than other methods in terms of ACC, F1_weighted and F1_macro.

[0047] Table 4 Classification results on the BRCA dataset

[0048]

[0049] Table 5 Classification results on the KIPAN dataset

[0050]

[0051] (2) Experimental verification of classification performance comparison under different omics data types

[0052] To verify that the integration of three types of omics data by GCNFormer can improve classification performance, the results of integrating single omics data and multiple omics data were compared. A total of 7 combinations of omics data types were compared. The experimental results are shown in Figure 2. Figure 2 As shown in the figure, when integrating three types of omics data, mRNA expression, DNA methylation, and miRNA expression, GCNFormer achieved advantages in multiple evaluation indicators. Overall, on different datasets, as the type of omics data increases, the classification performance gradually increases.

[0053] Obviously, the above verification of the present invention is only an example for clearly illustrating the present invention, and is not intended to limit the implementation methods of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all implementation methods here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A multi-omics data classification method based on global-local graph transformer, characterized by: The following steps are involved: Step 1: We propose a global-local graph transformer that combines graph convolutional layers with multi-head self-attention layers to capture the topological structure and semantic information of the graph, thereby comprehensively considering the global-local feature representation of different omics data. Step 2: Using contrastive learning strategies to learn the similarities and differences between samples, we can mine discriminative and robust graph representations from the heterogeneous information of different omics data, gain a deep understanding of the relationships between patients with complex diseases, and achieve accurate classification of patients with complex diseases. Step 3: Use the deep fusion network to classify multi-omics data, and input the consensus features into the deep fusion network to obtain the final classification prediction results; The global-local graph transformer described in step 1 combines graph convolutional layers with multi-head self-attention layers to effectively capture the topological structure and semantic information of the graph to learn global-local feature representations of different omics data. The specific steps are as follows: Step 1.1: For each omics data, first use Layer Norm normalization to ensure numerical stability during training. The original feature matrix X∈R N×N After LayerNorm normalization, H (0) , N is the number of samples, and then a graph convolution layer is trained to aggregate neighborhood information to learn the discriminative local information of omics data. The graph convolution layer is defined as: Among them, H (1) is the output of the graph convolution layer, that is, the key local structure information learned, W (0) is the corresponding weight matrix, A is the initial adjacency matrix, I is the identity matrix, yes The corresponding degree matrix, LeakyReLU represents the nonlinear activation function; Step 1.2: For each omics data, train a multi-head self-attention layer. The single attention mechanism is defined as: Among them, h i is the output of the i-th attention mechanism, and are the corresponding parameter matrices respectively, and the multi-head self-attention layer is defined as: MultiHead(Q,K,V)=Concat(h1,h2,h3,...,hheads)W°, Among them, heads represents the number of heads of the multi-head self-attention mechanism, W° represents the corresponding weight matrix, MultiHead(Q, K, V) represents the feature representation learned by the multi-head self-attention layer, which is input into the multi-layer perceptron to obtain omics-specific feature representations under different omics data types. The contrastive learning strategy described in step 2 mines discriminative and robust graph representations from the heterogeneous information of different omics data to improve classification accuracy. The specific steps are as follows: Step 2.1: Represent the learned specific features under different omics data types Splicing, to initially obtain a consistent representation of multi-omics data, and using the self-attention layer to obtain the consensus representation H of multi-omics data and the global structural relationship matrix S; Step 2.2: Introduce a contrastive learning strategy to learn the similarities and differences between samples. Specifically, calculate the cosine similarity between the specific feature representations of various omics data and the consensus feature representation to measure the similarity between samples, better understand the heterogeneity information between multiple omics data, and thus improve the discrimination and robustness of classification. The specific calculation formula is as follows: Among them, H(i,:) and Represents H and The i-th and j-th samples of represents the omics-specific feature representation learned by the global-local graph Transformer under the m-th omics data; Step 2.3: Construct a contrastive loss function to maximize the similarity between samples of the same class and minimize the similarity between samples of different classes. The calculation method is as follows: Among them, L c To compare the loss functions, K represents the temperature parameter and M represents the number of omics data types.

2. A multi-omics data classification method based on global-local graph transformer according to claim 1, characterized in that: The deep fusion network described in step 3 is defined as follows: Y=f DFN (H); Among them, f DFN (·) represents a deep fusion network, which is essentially a two-layer fully connected network, and Y is the classification prediction probability distribution of the final sample.

Images