A method for predicting brain tissue DNA methylation using peripheral tissue DNA methylation
By utilizing peripheral tissue DNA methylation data, combined with convolutional attention mechanisms and an improved ResNet-18 model, the problem of obtaining brain tissue DNA methylation data was solved, achieving accurate prediction of brain tissue DNA methylation and improving the model's generalization ability and prediction accuracy.
Patent Information
- Application Number
- CN202411762114.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-03
- Publication Date
- 2025-11-04
- Estimated Expiration
- 2044-12-03
AI Technical Summary
Because brain tissue samples are difficult to obtain and limited in quantity, the accuracy of predicting brain tissue DNA methylation using brain tissue DNA data is insufficient and cannot meet practical needs.
By acquiring DNA methylation data from peripheral tissues such as blood, oral cavity, and saliva, cross-tissue features were extracted using a convolutional attention mechanism module. The model was then trained using an improved ResNet-18 model, with hybrid pooling layers replacing global pooling layers. The model parameters were optimized using 10-fold cross-validation to predict DNA methylation in brain tissue.
This method enables accurate prediction of DNA methylation in brain tissue based on peripheral tissue data. It is simple to obtain samples, provides stable data results, and improves the accuracy of prediction and the generalization ability of the model.
Smart Images
Figure CN119724375B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of intelligent medical technology, and in particular to a method for predicting DNA methylation of brain tissue by using DNA methylation of peripheral tissue. BACKGROUND
[0002] The DNA methylation level is closely related to the pathogenesis and progression of brain diseases. It is of great clinical application value to predict the DNA methylation of brain tissue by using the DNA methylation of peripheral tissue which is easy to obtain. With the progress of sequencing technology, researchers have obtained a large amount of DNA methylation data of blood, oral cavity and saliva, three kinds of peripheral tissues of brain. Experiments have shown that the DNA methylation of the three kinds of peripheral tissues has significant correlation with the DNA methylation of brain tissue. Therefore, it is theoretically based to predict the DNA methylation of brain tissue by using multi-tissue integrated data. In view of the differences in the DNA methylation characteristics of different peripheral tissues, it is particularly important to effectively integrate multiple peripheral tissue data to extract key features in each tissue.
[0003] In summary, the change of the DNA methylation level of brain tissue helps to reveal the pathogenesis, disease progression and clinical treatment of brain diseases. However, brain tissue is difficult to obtain and limited in quantity, which is usually obtained by surgery or donation, which is a great limitation for researchers. Moreover, brain tissue samples are relatively scarce in quantity, especially in the collection of human samples, which is often affected by many factors such as age, disease state, sample source, etc. Therefore, due to the limited number of samples, the data set may not be sufficient, which affects the generalization ability of the model. Limited samples cannot fully represent the diversity of different individuals, different environments or different genetic backgrounds, which further affects the accuracy and prediction ability of the model. Using DNA data of brain tissue alone to obtain DNA methylation data of brain tissue has insufficient accuracy and cannot meet the actual use requirements. SUMMARY
[0004] The present application provides a method for predicting DNA methylation of brain tissue by using DNA methylation of peripheral tissue, to overcome the technical problem that using DNA data of brain tissue to obtain DNA methylation data of brain tissue has insufficient accuracy and cannot meet the actual use requirements.
[0005] In order to achieve the above purpose, the technical scheme of the present application is:
[0006] A method for predicting DNA methylation of brain tissue by using DNA methylation of peripheral tissue, comprising:
[0007] S1: obtaining DNA methylation data of brain peripheral tissue, the brain peripheral tissue comprising blood tissue, oral cavity tissue and saliva tissue;
[0008] S2: feature extraction is performed on the brain peripheral tissue DNA methylation data using a convolution attention mechanism module, cross-tissue feature data matrices in the peripheral tissue DNA methylation data are extracted, and the cross-tissue feature data matrices are divided into a training set and a test set;
[0009] S3: a ResNet-18 model is introduced, and a hybrid pooling layer is used to replace a global pooling layer in the ResNet-18 network model, to obtain an improved ResNet-18 model;
[0010] S4: the training set is input into the improved ResNet-18 model for training, to obtain a brain tissue DNA methylation prediction model, ten-fold cross-validation is used to optimize parameters of the brain tissue DNA methylation prediction model, to obtain an optimized brain tissue DNA methylation prediction model;
[0011] S5: the test set is input into the optimized brain tissue DNA methylation prediction model, to obtain predicted brain tissue DNA methylation data.
[0012] Further, a ResNet-18 model is introduced, and a hybrid pooling layer is used to replace a global pooling layer in the ResNet-18 network model, to obtain an improved ResNet-18 model, including:
[0013] The ResNet-18 model includes an initial convolution layer, a maximum pooling layer, eight residual modules, a global pooling layer, and a fully connected layer connected in sequence;
[0014] The global pooling layer is replaced with the hybrid pooling layer, to obtain the improved ResNet-18 model.
[0015] Further, S4 uses ten-fold cross-validation to optimize parameters of the brain tissue DNA methylation prediction model, to obtain an optimized brain tissue DNA methylation prediction model, including:
[0016] S41, the training set is evenly divided into 10 subsets, one subset is randomly selected as an optimization validation set, and the other 9 subsets are used as optimization training sets, which are input into the brain tissue DNA methylation prediction model, to obtain a first optimized brain tissue DNA methylation prediction model;
[0017] S42, a different subset is selected as an optimization validation set, and the other 9 subsets are used as optimization training sets, which are input into the first optimized brain tissue DNA methylation prediction model, to obtain a second optimized brain tissue DNA methylation prediction model;
[0018] S43, repeat the above selecting a different 1 subset as the optimization validation set, and the other 9 subsets as the optimization training set, until the 10th training, the last selected subset is taken as the optimization validation set, and the other 9 subsets are taken as the optimization training set, the ninth optimization brain tissue DNA methylation prediction model is input, and the final optimized brain tissue DNA methylation prediction model is obtained.
[0019] Further, the test set is input into the optimized brain tissue DNA methylation prediction model, and predicted brain tissue DNA methylation data is obtained, including:
[0020] S51, input the test set into the initial convolution layer in the optimized brain tissue DNA methylation prediction model and perform batch normalization processing to obtain a first feature matrix, as shown in formula (1),
[0021] X' = ReLU(BN(Conv2D(X, ω1))) (1)
[0022] Wherein, X' represents the obtained first feature matrix, X represents the cross-tissue feature data matrix extracted from the peripheral tissue DNA methylation data by the convolution attention mechanism module, ω1 represents the convolution kernel weight matrix of the initial convolution layer, BN is the batch normalization operation, Conv2D is the two-dimensional convolution operation, and ReLU is the activation function;
[0023] S52, input the first feature matrix into the maximum pooling layer to obtain a second feature matrix, as shown in formula (2),
[0024] X1 = MaxPool2D(X', ω2) (2)
[0025] Wherein, ω2 represents the convolution kernel weight matrix of the maximum pooling layer, MaxPool2D represents the two-dimensional maximum pooling operation, and X1 represents the generated second feature matrix;
[0026] S53, input X1 into the first residual module to obtain a third feature matrix, then input the third feature matrix into the second residual module to obtain a fourth feature matrix, and so on, input the feature matrix output by the last residual module into the next residual module to obtain the final predicted brain tissue DNA methylation data.
[0027] Further, input the feature matrix output by the last residual module into the next residual module to obtain the final predicted brain tissue DNA methylation data, including:
[0028] S531, input X α Into the αth residual module to obtain the feature mapping in the αth residual module, as shown in formula (3) and formula (4),
[0029] F(X α )=BN(Conv2D(ReLU(BN(COnv2d(X α ))))) (3)
[0030] Output α =F(X α )+X α (4)
[0031] In formula (3), COnv2D is a two-dimensional convolution operation, BN is a batch normalization operation, ReLU is an activation function, F(X α ) is a feature mapping in the a-th residual block, X α is an input feature matrix of any residual block; Conv2d represents a two-dimensional convolution operation in the residual module;
[0032] In formula (4), Output α is the output of the a-th residual block;
[0033] S532, input the output feature matrix of the last residual block into the mixed pooling layer and the fully connected layer to obtain the final prediction value, as shown in formulas (5) and (6),
[0034] X mix =a·Maxpool(X last )+(1-a)·Avgpool(X last ) (5)
[0035] Y=Softmax(FC(X mix )) (6)
[0036] In formula (5), MaxPool is a maximum average pooling operation, AvgPool is an average pooling operation, X last is the output feature matrix of the last residual block, a is a hyperparameter for adjusting the maximum pooling and the average pooling, and X mix is the output feature vector of the mixed pooling layer; in formula (6), Softmax is an activation function, FC is a fully connected layer, and Y represents the final brain DNA methylation prediction value.
[0037] Further, a convolution attention mechanism module is used to extract features from the brain peripheral tissue DNA data, and a cross-tissue feature data matrix in the peripheral tissue DNA methylation data is extracted, including:
[0038] S21, define the blood methylation data matrix as B i,j , the oral methylation data matrix as O i,j , and the saliva methylation data matrix as S i,jwhere i represents a sample, i = 1, 2, …, n, j represents a feature, j = 1, 2, …, m, and the peripheral tissue DNA methylation data matrix D is defined as D = [D1, D2, …, Dn] where Dn represents the DNA methylation data of the nth sample i,j where D e B, O, S
[0039] S22, performing global average pooling and global maximum pooling on the peripheral tissue DNA methylation data matrix D respectively to obtain a one-dimensional feature vector processed by global average pooling and a one-dimensional feature vector processed by global maximum pooling, as shown in formulas (7) and (8), i,j
[0040] F avg = AvgPool(D i,j ) (7)
[0041] F max = MaxPool(D i,j ) (8)
[0042] where the dimensions of F avg and F max are both n x 1 x c, that is, there is one value for each channel of the sample; F avg represents global average pooling, and F max represents global maximum pooling; n represents the total number of samples, and c represents the number of channels.
[0043] S23, inputting both one-dimensional feature vectors into shared fully connected layers W0 and W1 to obtain two weight vectors, as shown in formulas (9) and (10),
[0044] M avg = σ(W1·ReLU(W0·F avg )) (9)
[0045] M max = σ(W1·ReLU(W0·F max )) (10)
[0046] where σ represents a Sigmoid function, W0 and W1 are two weight matrices of the fully connected layer; M avg represents the weight vector of the one-dimensional feature vector processed by global average pooling, and M max represents the one-dimensional feature vector processed by global maximum pooling.
[0047] S24, summing the two weight vectors by addition to obtain a final channel weight matrix, as shown in formula (11),
[0048] M c = σ(M avg +M max ) (11)
[0049] S25, the channel weight matrix M c is applied to each channel of the peripheral tissue DNA methylation data matrix D i,j , and the channel weighted output matrix is obtained, as shown in formula (12),
[0050] Z=M c ·D i,j (12)
[0051] S26, the channel weighted output matrix is aggregated for features, and two two-dimensional feature vectors are obtained, as shown in formula (13) and (14),
[0052] F′ avg =AvgPool(Z) (13)
[0053] F′ max =MaxPool(Z) (14)
[0054] S27, the two two-dimensional feature vectors are obtained by convolution operation, and the spatial attention weight matrix is obtained, as shown in formula (15),
[0055] M s =σ(Conv([F′ avg ;F′ max ])) (15)
[0056] Wherein, Conv represents convolution operation;
[0057] S28, the spatial attention weight matrix M s is used to the channel weighted output matrix to obtain the final output matrix, as shown in formula (16),
[0058] X=M s ·Z (16)
[0059] Wherein, X represents the final output matrix, that is, the cross-tissue feature data matrix in the peripheral tissue DNA methylation data.
[0060] Beneficial effects: the application provides a method for predicting brain tissue DNA methylation by using peripheral tissue DNA methylation, which obtains brain peripheral tissue DNA methylation data, finds DNA methylation data related to brain tissue in brain peripheral tissue DNA methylation data through convolution attention mechanism module, and inputs the data into improved ResNet-18 model to obtain brain tissue DNA methylation data based on brain peripheral tissue DNA methylation data. The brain tissue DNA methylation data can be obtained without brain tissue, the sample is simple to obtain, the data result is stable and accurate. BRIEF DESCRIPTION OF DRAWINGS
[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0062] Figure 1 A flowchart of a method for predicting brain tissue DNA methylation by peripheral tissue DNA methylation provided by the present application is shown in the figure.
[0063] Figure 2 A structure diagram of an improved ResNet-18 provided by the present application is shown in the figure. DETAILED DESCRIPTION
[0064] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0065] The present embodiment provides a method for predicting brain tissue DNA methylation by peripheral tissue DNA methylation, as shown in the figure, comprising: Figure 1
[0066] S1: Obtain brain peripheral tissue DNA methylation data, the brain peripheral tissue including blood tissue, oral tissue and saliva tissue;
[0067] S2: Use a convolution attention mechanism module to extract features from the brain peripheral tissue DNA methylation data, extract a cross-tissue feature data matrix in the peripheral tissue DNA methylation data, and divide the cross-tissue feature data matrix into a training set and a test set;
[0068] S3: Introduce a ResNet-18 model, and replace a global pooling layer in the ResNet-18 network model with a hybrid pooling layer to obtain an improved ResNet-18 model;
[0069] S4: Input the training set into the improved ResNet-18 model for training to obtain a brain tissue DNA methylation prediction model, use ten-fold cross-validation to optimize the brain tissue DNA methylation prediction model parameters, and obtain an optimized brain tissue DNA methylation prediction model;
[0070] S5: inputting the test set into the optimized brain tissue DNA methylation prediction model to obtain predicted brain tissue DNA methylation data.
[0071] Specifically, brain peripheral tissue DNA methylation data is acquired, the brain peripheral tissue including blood tissue, oral tissue and saliva tissue, brain DNA methylation data can be inferred from the peripheral tissue DNA methylation data, the result obtained is accurate, and the sample is convenient to obtain; secondly, a convolution attention mechanism module is used to extract features from the brain peripheral tissue DNA methylation data, extract a cross-tissue feature data matrix in the peripheral tissue DNA methylation data, and divide the cross-tissue feature data into a training set and a test set, which can enhance the expression ability of important features, automatically select features and improve the generalization ability of the model; a ResNet-18 model is introduced, and a mixed pooling layer is used to replace a global pooling layer in the ResNet-18 network model to obtain an improved ResNet-18 model, which can extract local significant features and global information at the same time, thereby improving the expression ability of the model; the training set is input into the improved ResNet-18 model for training to obtain a brain tissue DNA methylation prediction model, and ten-fold cross-validation is used to optimize the brain tissue DNA methylation prediction model parameters to obtain an optimized brain tissue DNA methylation prediction model, through which more accurate brain DNA methylation data can be obtained.
[0072] In specific embodiments, the scheme for acquiring brain peripheral tissue DNA data, the brain peripheral tissue including blood tissue, oral tissue and saliva tissue, is as follows:
[0073] In this embodiment, the brain peripheral tissue DNA methylation data is obtained from GSE111165, and the brain peripheral tissue DNA methylation data is acquired, which can be used to infer brain DNA methylation data, the result obtained is accurate, and the sample is convenient to obtain.
[0074] In specific embodiments, the scheme for using a convolution attention mechanism module to extract features from the brain peripheral tissue DNA methylation data, extract a cross-tissue feature data matrix in the peripheral tissue DNA methylation data, and divide the cross-tissue feature data matrix into a training set and a test set is as follows:
[0075] S21, define blood methylation data matrix B i,j , oral methylation data matrix O i,j , and saliva methylation data matrix S i,jwhere i represents a sample, i = 1, 2, …, n, j represents a feature, j = 1, 2, …, m, and the peripheral tissue DNA methylation data matrix D is defined as D = [D1, D2, …, Dn] (1) i,j where D e B, O, S;
[0076] S22, obtaining the peripheral tissue DNA methylation data matrix D i,j respectively, to obtain a one-dimensional feature vector processed by global average pooling and a one-dimensional feature vector processed by global maximum pooling, as shown in formulas (17) and (18),
[0077] F avg = AvgPool(D i,j ) (17)
[0078] F max = MaxPool(D i,j ) (18)
[0079] where the dimensions of F avg and F max are n x 1 x c, that is, there is a value for each channel of the sample; F avg represents global average pooling, and F max represents global maximum pooling; D i,j has a size of n x m x c, where n represents the total number of samples, m represents the total number of features, and c represents the number of channels.
[0080] Global average pooling calculates the average value of each channel, captures the spatial information of the whole channel, and reflects the relatively stable or global features on the channel; global maximum pooling calculates the maximum value on each channel, and captures the most prominent local features in the channel. It can highlight the most representative and important feature points in the local area;
[0081] S23, inputting the two one-dimensional feature vectors into shared fully connected layers W0 and W1 to obtain two weight vectors, as shown in formulas (19) and (20),
[0082] M avg = σ(X1·ReLU(W0·F avg )) (19)
[0083] M max = σ(W1·ReLU(W0·F max )) (20)
[0084] where σ represents a Sigmoid function, W0 and W1 are two weight matrices of the fully connected layer; M avg represents the weight vector of the one-dimensional feature vector processed by global average pooling, and M maxrepresents a one-dimensional feature vector obtained by global max-pooling processing;
[0085] S24, two weight vectors are summed by addition to obtain a final channel weight matrix, as shown in formula (21),
[0086] M c =σ(M avg +M max ) (21)
[0087] S25, apply the channel weight matrix M c to each channel of the peripheral tissue DNA methylation data matrix D i,j to obtain a channel weighted output matrix, as shown in formula (22),
[0088] Z=M c ·D i,j (22)
[0089] S26, feature aggregation is performed on the channel weighted output matrix to obtain two two-dimensional feature vectors, as shown in formulas (23) and (24),
[0090] F′ avg =AvgPool(Z) (23)
[0091] F′ max =MaxPool(Z) (24)
[0092] S27, two two-dimensional feature vectors are obtained by convolution operation to obtain a spatial attention weight matrix, as shown in formula (25),
[0093] M s =σ(Conv([F′ avg ;F′ max ])) (25)
[0094] Where, Conv represents convolution operation, and the convolution kernel size is usually 7x7;
[0095] S28, the spatial attention weight matrix M s is used to the channel weighted output matrix to obtain a final output matrix, as shown in formula (26),
[0096] X=M s ·Z (26)
[0097] Where, X represents the final output matrix, i.e., the cross-tissue feature data matrix in the peripheral tissue DNA methylation data;
[0098] S29, the obtained data is divided into a training set and a test set, and the proportion is 9:1.
[0099] The convolutional attention mechanism used in the present scheme consists of two parts: a channel attention module and a spatial attention module:
[0100] The channel attention module aims to generate a channel weight matrix by selecting more important features, and each channel is weighted;
[0101] The spatial attention module aims to aggregate the weighted channel weight matrix by focusing on more important spatial positions, aggregate into a channel information, generate two two-dimensional feature vectors and then perform convolution to generate a spatial weight matrix, and each position is weighted;
[0102] The present scheme uses the convolutional attention mechanism module to extract peripheral tissue DNA data that is more relevant to brain tissue CpG site data, which has the following advantages:
[0103] 1. Enhance the expression ability of important features:
[0104] For peripheral tissue DNA methylation data, the convolutional attention mechanism can improve the performance of the model in the key area, automatically learn which features in the peripheral tissue are most important for predicting brain tissue DNA methylation. Reduce the interference of irrelevant information and improve the accuracy and robustness of prediction.
[0105] 2. Automatically select features:
[0106] When the DNA methylation data of the peripheral tissue is complex, the convolutional attention mechanism can automatically learn the features related to brain tissue methylation from the original data, saving time and improving efficiency.
[0107] 3. Combination of local information and global information:
[0108] Convolutional neural networks are good at capturing local features, while attention mechanisms can enhance the model's expression ability through global information interaction. By combining the two, not only can local features be captured when processing peripheral tissue DNA methylation data, but global context information can also be integrated, thereby improving the model's generalization ability and helping to better predict brain tissue DNA methylation levels.
[0109] In specific embodiments, a ResNet-18 model is introduced, and a hybrid pooling layer is used to replace the global pooling layer in the ResNet-18 network model to obtain an improved ResNet-18 model scheme:
[0110] The ResNet-18 model includes an initial convolutional layer, a maximum pooling layer, 8 residual modules, a global pooling layer, and a fully connected layer connected in turn;
[0111] The size of the convolution kernel of the initial convolutional layer is 7*7, and the step is 2;
[0112] The size of the maximum pooling layer convolution kernel is 3*3, the step is 2, and the padding is 1; the residual module structure is as follows:
[0113] Each residual module contains two convolution layers: the first convolution layer and the second convolution layer are 3x3 convolution kernels with a step of 1;
[0114] Residual connection: the input of each convolution layer is directly added to the output to form a residual (jump connection), which can speed up training and improve the performance of the network; the global pooling layer averages the features of each channel;
[0115] Linear layer: used to output the final classification result or regression value;
[0116] The global pooling layer is replaced with a hybrid pooling layer to obtain an improved ResNet-18 model, as shown in Figure 2 .
[0117] In this scheme, the global pooling layer in the ResNet-18 network model is replaced with a hybrid pooling layer, which can extract both local significant features and global information, thereby improving the expression ability of the model. The model mainly consists of eight residual blocks, each of which has two convolution layers and a residual connection. Data is continuously processed through convolution, batch normalization, activation function and jump connection in each residual block, gradually learning and extracting deep features.
[0118] In specific embodiments, the training set is input into the improved ResNet-18 model for training to obtain a brain tissue DNA methylation prediction model, and ten-fold cross-validation is used to optimize the model parameters to obtain an optimized brain tissue DNA methylation prediction model.
[0119] S41, the training set is evenly divided into 10 subsets, one of which is randomly selected as an optimization validation set, and the other 9 subsets are used as optimization training sets, which are input into the brain tissue DNA methylation prediction model to obtain a first optimized brain tissue DNA methylation prediction model;
[0120] S42, a different subset is selected as an optimization validation set, and the other 9 subsets are used as optimization training sets, which are input into the first optimized brain tissue DNA methylation prediction model to obtain a second optimized brain tissue DNA methylation prediction model;
[0121] S43, repeat the above selecting a different subset as the optimization validation set, and the other 9 subsets as the optimization training set, until the 10th training, the last selected subset is taken as the optimization validation set, and the other 9 subsets are taken as the optimization training set, input the ninth optimization brain tissue DNA methylation prediction model, and obtain the final optimized brain tissue DNA methylation prediction model.
[0122] In this embodiment, the learning rate, batch size, loss function and regularization parameter of the model are optimized, and the model is trained with the training set in each cross-validation iteration. The model learns different features on different training sets, and each time tries to optimize the learned weights or parameters. The model is optimized through ten-fold cross-validation, and each optimization process selects a different subset as the optimization validation set, which can obtain a more accurate optimization model and more accurate brain DNA methylation data.
[0123] In specific embodiments, the scheme for inputting the test set into the optimized brain tissue DNA methylation prediction model to obtain predicted brain tissue DNA methylation data is:
[0124] S51, input the test set into the initial convolution layer in the optimized brain tissue DNA methylation prediction model and perform batch normalization processing to obtain a first feature matrix, as shown in formula (27),
[0125] X' = ReLU(BN(Conv2D(X, ω1))) (27)
[0126] wherein X' represents the obtained first feature matrix, X represents the cross-tissue feature data matrix extracted from the peripheral tissue DNA methylation data by the convolution attention mechanism module, ω1 represents the convolution kernel weight matrix of the initial convolution layer, the convolution kernel size of which is 7x7, BN is the batch normalization operation, Conv2D is the two-dimensional convolution operation, and ReLU is the activation function;
[0127] S52, input the first feature matrix into the maximum pooling layer to obtain a second feature matrix, as shown in formula (28),
[0128] X1 = MaxPool2D(X', ω2) (28)
[0129] wherein ω2 represents the convolution kernel weight matrix, the convolution kernel size of which is 3x3 and the step is 2, MaxPool2D represents the two-dimensional maximum pooling operation, and X1 represents the generated feature matrix.
[0130] S53, input X1 into the first residual module to obtain a third feature matrix, and input the third feature matrix into the second residual module to obtain a fourth feature matrix, and so on, input the feature matrix output by the last residual module into the next residual module to obtain the final predicted brain tissue DNA methylation data:
[0131] S531, input X α into the first residual module to obtain a first feature matrix, and input the first feature matrix into the second residual module to obtain a second feature matrix, and so on, input the feature matrix output by the last residual module into the next residual module to obtain the final predicted brain tissue DNA methylation data: α is the input feature matrix of any residual block, represents the input into the αth residual block, α = 1, 2, …, 8, and the feature mapping in the αth residual block is obtained through two convolution layers and batch normalization, as shown in formula (29) and formula (30),
[0132] F(X α ) = BN(Conv2D(ReLU(BN(Conv2d(X α )))))) (29)
[0133] Output α = F(X α ) + X α (30)
[0134] In formula (29), Conv2D is a two-dimensional convolution operation, BN is a batch normalization operation, ReLU is an activation function, and F(X α ) is the feature mapping in the αth residual block; X α is the input feature matrix of any residual block; and Conv2d represents a two-dimensional convolution operation in the residual module.
[0135] In formula (30), Output α is the output of the αth residual block.
[0136] S532, input the output feature matrix of the last residual block into the mixed pooling layer and the fully connected layer to obtain the final prediction value, as shown in formula (31) and (32),
[0137] X mix = a Maxpool(X last ) + (1-a) Avgpool(X last ) (31)
[0138] Y = Softmax(FC(X mix )) (32)
[0139] In formula (31), MaxPool is a maximum average pooling operation, AvgPool is an average pooling operation, and X lastThe output feature matrix of the last residual block, a is a hyperparameter that adjusts the max pooling and average pooling, and X mix is the output feature vector of the hybrid pooling layer; in equation (32), Softmax is an activation function, FC is a fully connected layer, and Y represents the final brain DNA methylation prediction value.
[0140] The optimized brain tissue DNA methylation prediction model was evaluated for performance, and in this scheme, the mean absolute error (MAE) and mean squared error (MSE) were used for evaluation, and the results are shown in Table 1:
[0141] Table 1 Evaluation Performance Table
[0142] MAE MSE Blood tissue 0.111 0.042 Oral tissue 0.121 0.049 Saliva tissue 0.107 0.036 Blood, oral, saliva 0.037 0.003
[0143] From Table 1, it can be seen that the MAE score for predicting brain tissue DNA methylation using blood tissue is 0.111, and the MSE score is 0.042. The MAE score for predicting brain tissue DNA methylation using oral tissue is 0.121, and the MSE score is 0.049. The MAE score for predicting brain tissue DNA methylation using saliva tissue is 0.107, and the MSE score is 0.036. However, the MAE score for predicting brain tissue DNA methylation using multiple peripheral tissues (blood, oral, and saliva tissues) is 0.037, and the MSE score is 0.003.
[0144] In summary, the results obtained by predicting brain tissue DNA methylation using multiple peripheral tissues have the smallest error and are more accurate.
[0145] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for part or all of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.
Claims
1. A method of predicting brain tissue DNA methylation using peripheral tissue DNA methylation, characterized by, Comprise: S1: Obtain peripheral tissue DNA methylation data of the brain, including blood tissue, oral tissue and saliva tissue; S2: Feature extraction is performed on the peripheral tissue DNA methylation data of the brain using a convolution attention mechanism module, a cross-tissue feature data matrix is extracted from the peripheral tissue DNA methylation data, and the cross-tissue feature data matrix is divided into a training set and a test set; The specific steps are as follows: S21, define the blood methylation data matrix as , the oral methylation data matrix is , the saliva methylation data matrix is , wherein represent samples, , represent features, , define the peripheral tissue DNA methylation data matrix as , wherein ; S22, peripheral tissue DNA methylation data matrix Global average pooling and global max pooling are performed respectively to obtain a one-dimensional feature vector processed by global average pooling and a one-dimensional feature vector processed by global max pooling, as shown in formulas (1) and (2). （1） （2） wherein, and each of the dimensions of i.e. one value per channel of the sample; denotes a global average pooling, denotes a global max pooling; denotes the total number of samples, denotes the number of channels; S23, inputting both one-dimensional feature vectors into a shared fully connected layer and , obtaining two weight vectors as shown in equations (3) and (4), （3） （4） wherein, represents a Sigmoid function, and are two weight matrices of the fully connected layer; represents a weight vector of the one-dimensional feature vector obtained by global average pooling processing, represents a one-dimensional feature vector obtained by global maximum pooling processing. S24, two weight vectors are summed by addition to obtain the final channel weight matrix, as shown in formula (5), （5） S25, applying a channel weight matrix to each of the peripheral tissue DNA methylation data matrices to obtain a channel-weighted output matrix, as shown in equation (6), （6） S26, feature aggregation is performed on the output matrix after channel weighting to obtain two two-dimensional feature vectors, as shown in formulas (7) and (8), （7） （8） S27, two two-dimensional feature vectors are obtained by convolution operation to obtain a spatial attention weight matrix, as shown in formula (9), （9） wherein denotes a convolution operation; S28, the spatial attention weight matrix The final output matrix is obtained using the channel weighted output matrix, as shown in equation (10). （10） wherein, represents the final output matrix, i.e., the cross-tissue feature data matrix in the peripheral tissue DNA methylation data; S3: Introduce a ResNet-18 model, and replace the global pooling layer in the ResNet-18 network model with a hybrid pooling layer to obtain an improved ResNet-18 model; The hybrid pooling layer is a weighted hybrid pooling layer of the maximum pooling layer and the average pooling layer; S4: The training set is input into the improved ResNet-18 model for training to obtain a brain tissue DNA methylation prediction model, and the brain tissue DNA methylation prediction model parameters are optimized using ten-fold cross-validation to obtain an optimized brain tissue DNA methylation prediction model; S5: The test set is input into the optimized brain tissue DNA methylation prediction model to obtain predicted brain tissue DNA methylation data, and the specific steps are as follows: S51, input the test set into the initial convolution layer of the optimized brain tissue DNA methylation prediction model and perform batch normalization processing to obtain a first feature matrix, as shown in formula (11), （11） wherein, represents the obtained first feature matrix, represents a cross-tissue feature data matrix extracted from the peripheral tissue DNA methylation data by the convolution attention mechanism module, represents a convolution kernel weight matrix of an initial convolution layer, and BN represents a batch normalization operation, is a two-dimensional convolution operation, is an activation function; S52, input the first feature matrix into the maximum pooling layer to obtain a second feature matrix, as shown in formula (12), （12） wherein, denotes a convolution kernel weight matrix of the max pooling layer, denotes a two-dimensional max pooling operation, denotes a generated second feature matrix; S53、will input into the first residual module to obtain a third feature matrix, and then input the third feature matrix into the second residual module to obtain a fourth feature matrix, and so on, input the feature matrix output by the last residual module into the next residual module to obtain the final predicted brain tissue DNA methylation data.
2. The method of predicting brain tissue DNA methylation using peripheral tissue DNA methylation according to claim 1, wherein, Introduce a ResNet-18 model, and replace the global pooling layer in the ResNet-18 network model with a hybrid pooling layer to obtain an improved ResNet-18 model, comprising: The ResNet-18 model comprises an initial convolution layer, a maximum pooling layer, eight residual modules, a global pooling layer and a fully connected layer connected in turn; The global pooling layer is replaced with a hybrid pooling layer to obtain an improved ResNet-18 model.
3. The method of predicting brain tissue DNA methylation using peripheral tissue DNA methylation according to claim 1, wherein, S4 uses ten-fold cross-validation to optimize the brain tissue DNA methylation prediction model parameters to obtain an optimized brain tissue DNA methylation prediction model, comprising: S41, divide the training set into 10 subsets, randomly select one subset as the optimization validation set, and input the other 9 subsets as the optimization training set into the brain tissue DNA methylation prediction model to obtain a first optimized brain tissue DNA methylation prediction model; S42, select a different subset as the optimization validation set, and input the other 9 subsets as the optimization training set into the first optimized brain tissue DNA methylation prediction model to obtain a second optimized brain tissue DNA methylation prediction model; S43, repeat the above selecting different 1 subset as the optimization verification set, and the other 9 subsets as the optimization training set, until the 10th training, the last selected subset is taken as the optimization verification set, and the other 9 subsets are taken as the optimization training set, the ninth optimized brain tissue DNA methylation prediction model is input, and the final optimized brain tissue DNA methylation prediction model is obtained.
4. The method of predicting brain tissue DNA methylation using peripheral tissue DNA methylation of claim 1, wherein, The feature matrix output by the last residual module is input into the next residual module to obtain the final predicted brain tissue DNA methylation data, including: S531、the input into the first residual module, to obtain the feature mapping in the first residual module, as shown in equations (13) and (14). （13） （14） In formula (13), is a two-dimensional convolution operation, is a batch normalization operation, is an activation function, is the feature map in the th residual block, is the input feature matrix of any one residual block; denotes a two-dimensional convolution operation in the residual module; In equation (14) is the output of the th residual block; S532, input the output feature matrix of the last residual block into the mixed pooling layer and the full connection layer to obtain the final prediction value, as shown in formulas (15) and (16), （15） （16） In formula (15), is a max-pooling operation, is an average-pooling operation, is an output feature matrix of the last residual block, is a hyper-parameter for adjusting the max-pooling and average-pooling, is an output feature vector of the mixed pooling layer; in formula (16), is an activation function, is a fully connected layer, denotes obtaining a final brain DNA methylation prediction value.
Citation Information
Patent Citations
Systems and methods for determining cancer therapy via deep learning
US20250104827A1