Method and system for predicting lens opacification degree of Rncat mouse

Abstract

The invention discloses a prediction method and system for the crystalline lens opacification degree of an Rncat mouse. The invention relates to the technical field of experimental animals and comparative medicine. Three nodes n1, n2 and n3 are respectively established for the mating mode, the day age and the G489A mutation degree of the Crygs gene of the Rncat mouse. And extracting a sample feature Pi of each node ni through the multi-scale residual network ResNet. Mutually independent sample features Pi are converted into a structured graph structure G, and a prediction problem is converted into a node prediction task in a graph neural network framework. And sequentially aggregating each adjacent node ni by using an ordered clustering algorithm, and updating embedded vectors evi of different nodes ni. Through the deep learning algorithm, the complex nonlinear relationship between data can be automatically learned, so that the prediction accuracy is improved. Compared with a traditional linear regression or simple classification model, the ResNet-GNN neural network model can capture more fine interaction between features and variables, so that the prediction result is more reliable.

Description

Technical Field

[0001] The present invention relates to the field of experimental animals and comparative medical technology, and in particular to statistical analysis and machine learning technology of mutant proteins of congenital cataract mice, and in particular to a prediction method and system for the degree of lens opacity of Rncat mice. Background Art

[0002] Cataract is an ophthalmic disease caused by multiple factors, and its core pathological mechanism is closely related to abnormalities in lens proteins. In 1993, the research team of the Xinjiang Uygur Autonomous Region Experimental Animal Research Center accidentally discovered an individual with ocular lens opacity and clinical manifestations of cataracts in a study of the Kunming mouse breeding group. Through 15 consecutive generations of breeding and selection, the team successfully retained this genetic characteristic and established a congenital cataract mouse strain with a stable genetic background, namely the Rncat mouse.

[0003] Currently, research on mutant proteins in Rncat mice provides another perspective for revealing the pathogenesis of congenital cataracts. Current studies on Rncat mice have shown that shortened mutations in the amino acid sequence may destroy the original structure of the lens protein, induce protein aggregation or abnormal conformation, and ultimately lead to the formation of cataracts. By analyzing the structure and function of such mutant proteins and exploring their intrinsic connection with cataracts, it is expected to open up new paths for the prevention and treatment of cataracts. Therefore, existing studies have used a variety of technical means to explore the degree of lens opacity in Rncat mice. For example:

[0004] (1) The paper "Zhang Yan, Wang Ying, Yuan Jiangling et al. Comparison of reproductive performance and lens opacity of Rncat mice with different mating modes [J]. Laboratory Animals and Comparative Medicine, 2018, 38(1): 60-62" pointed out that the mating mode significantly affects the reproductive performance, offspring growth and lens opacity of Rncat mice, revealing the possible relationship between genetic background and disease manifestation;

[0005] (2) The paper "Yuan Jiangling, Xu Xiaohui, Zhang Yan, et al. Analysis of the weight of major organs and blood physiological and biochemical indicators of mice with congenital cataracts [J]. Chinese Journal of Comparative Medicine, 2016, 6(26): 32-36" points out that the changes in age can reflect the physiological and pathological dynamics of mutant proteins during the development of Rncat mice, providing information on the time dimension for understanding the disease process;

[0006] (3) The literature "Yan Shunsheng, Chen Xinru, An Ran, et al. Preliminary study on recombination and cloning of mutant proteins in a mouse model of congenital cataract [S]. Bulletin of Disease Prevention and Control, 2018; 33(1): 16-17" pointed out that the G489A mutation of the Crygs gene in Rncat mice caused truncation of γS crystallin at position 163 (less than 178 amino acids). This finding directly linked gene mutation to the occurrence of cataracts. At the same time, through the Crygs gene knockout experiment, the researchers found that the gene knockout did not directly cause cataracts, and the 162 amino acid mutant protein produced in Rncat mice may be the key molecular mechanism for the occurrence and development of congenital cataracts.

[0007] However, from the perspective of statistical analysis, the above existing technologies have failed to establish or reveal the nonlinear relationship between the independent variables (mating mode, age and degree of G489A mutation of Crygs gene) and the dependent variable (degree of lens opacity). Traditional manual testing and detection recording methods have limitations in predicting the development of cataracts under different conditions, which in turn limits the ability to provide a scientific basis for drug development and disease treatment strategy formulation.

[0008] To this end, the present invention proposes a method and system for predicting the degree of lens opacity in Rncat mice. Summary of the invention

[0009] In view of this, the present invention hopes to provide a prediction method and system for the degree of lens opacity of Rncat mice to solve or alleviate the technical problems existing in the prior art, namely, how to establish a nonlinear relationship between the independent variables (mating mode, age and G489A mutation degree of Crygs gene) and the dependent variable (degree of lens opacity) of Rncat mice, and predict the corresponding dependent variable based on the given independent variables, reduce the experimental cost, and further provide a reference for the development of related drugs and the formulation of disease treatment strategies; the technical solution of the present invention is achieved as follows:

[0010] First, the prediction method for the degree of lens opacity in Rncat mice:

[0011] 1. Overview:

[0012] This solution aims to solve the above technical problems. The ResNet-GNN neural network model is used to predict the degree of lens opacity in Rncat mice. Combining the advantages of residual network (ResNet) and graph neural network (GNN), nodes are established for independent variables such as mating mode, age and G489A mutation degree through the feature extraction layer, and sample features are extracted. Then, the graph neural network layer is used to convert independent sample features into structured graph structures to capture the interactions and complex relationships between independent variables. The convolutional neural network layer is used to extract features and linearly combine the graph structure to form a prediction result. Through multi-scale feature extraction, graph structure conversion and deep learning algorithm, the accuracy and robustness of the prediction are improved, providing a new and effective method for predicting the degree of lens opacity in Rncat mice.

[0013] (II) Technical solution:

[0014] The present invention chooses to combine a multi-scale residual network (ResNet) with a graph neural network (GNN) to form a ResNet-GNN neural network model to achieve the above technical objectives.

[0015] First, the independent variables X = [M, A, μ] of the target Rncat mice are collected, where M is the mating mode of the Rncat mice, which are inbreeding and closed group breeding (inbreeding is represented by [1, 0], and closed group breeding is represented by [0, 1]); A is the age of the Rncat mice; μ is the G489A mutation degree of the Crygs gene of the Rncat mice, which refers to the degree of truncation of the γS crystallin at position 163, and is an interval value of [0, 1] (the normal standard is 178 amino acids, the smaller the number of amino acids is, the greater the G489A mutation degree; the minimum is 0, and the maximum is 1). Then, the prediction method (process) of the ResNet-GNN neural network model is as follows:

[0016] 2.1 Step S1, feature extraction layer performs extraction tasks:

[0017] Create node n for the independent variable X of Rncat mice i , divided into three nodes n 1 、n 2 and n 3 Corresponding to the mating mode, age and G489A mutation degree of Crygs gene respectively; the multi-scale residual network ResNet is used to extract each node n i The sample characteristics P i .

[0018] 2.1.1 Step S100, residual learning:

[0019] The multi-scale residual network ResNet includes corresponding to node n 1 、n2 and n 3 There are three residual blocks, each of which contains two or three convolutional layers and a shortcut connection layer. The shortcut connection layer allows the input to skip some layers directly and be added to the output of these layers;

[0020] Each residual block reads the independent variable X (to be precise, x i , that is, mating mode M, age A or G489A mutation degree μ of Crygs gene) The residual block uses residual learning to output each node n i The sample characteristics P i :

[0021]

[0022] in, is the output of the lth residual block, is the input of the lth residual block (for the first residual block, That is node n i The original input variable x i ), F(·) represents the residual function, which contains two convolutional layers. represents the weight parameters of these convolutional layers.

[0023] 2.1.1.1 Step S1000, residual function F(·):

[0024] For the i-th node (i=1, 2, 3) and the l-th residual block (l represents the index of the residual block), the residual function F(·) is expressed as:

[0025]

[0026] in, and are the learnable weight parameters of the first and second convolutional layers in the lth residual block respectively; and are the bias parameters of the first and second convolutional layers in the lth residual block, and are also the parameters that need to be learned. The symbol * represents the convolution operation; σ(·) is the ReLU activation function, that is, σ(·) = max(0, x).

[0027] 2.1.1.2 Step S1001, execute the shortcut connection layer:

[0028] For the two convolutional layers of the residual function F(·), the shortcut connection layer allows the input to skip some layers directly and add the output of these layers; the specific method is:

[0029] Calculate the output of the i-th node (i=1,2,3) and the l-th residual block And when size matching or dimensionality adjustment is required (for example, when the input and output have different numbers of channels or different spatial dimensions), perform the following operations:

[0030]

[0031] in, is a linear transformation operation (convolution, batch normalization, and / or padding operation).

[0032] 2.1.2 Step S101, feature extraction:

[0033] Since there are multiple residual blocks in the multi-scale residual network ResNet, the feature extraction process of the entire multi-scale residual network ResNet is:

[0034]

[0035] Among them, H(·) represents the ResNet network function, L represents the total number of residual blocks, represents the weight parameters of all residual blocks; (·; ·) is the vector concatenation operation;

[0036] Among them, the ResNet network function H(·) is the continuous application of residual blocks, and the operation method is as follows:

[0037]

[0038] Among them, the weight parameter set of all residual blocks is represented as W i At the same time, F l (·) represents the residual function of the lth residual block, which receives the input and the corresponding weight parameters and outputs the features of the residual block. In addition, the output of each residual block is added with its input (i.e., the shortcut connection of S1001).

[0039] 2.2 Step S2, the graph neural network layer performs the conversion task:

[0040] The independent sample features P i Transformed into a structured graph structure G, that is, the prediction problem is transformed into a node prediction task within the graph neural network framework. Use the ordered clustering algorithm to aggregate each adjacent node n in turn. i , constantly updating different nodes n i The embedding vector e vi . This step is responsible for capturing the interactions and complex relationships between the independent variables.

[0041] 2.2.1 Step S200, graph structure construction:

[0042] All sample features P i The sample feature set Construct a graph structure G = (V, E), where the collection of nodes in the graph Each node n i Corresponding to a sample feature P i ; E is the edge set, which represents the connection relationship between nodes.

[0043] 2.2.2 Step S201, node embedding vector initialization:

[0044] For each node n i , initialize its embedding vector

[0045] 2.2.3 Step S202, perform ordered clustering:

[0046] The ordered clustering algorithm is Fisher's optimal partitioning method, which aggregates each adjacent node in turn. This process can be iterated, and the embedding vector of the node is updated each iteration.

[0047] Assume the current iteration round is t, for node n i , whose neighbor node set is N(i). Fisher optimal partitioning method is used to determine how to partition n i Aggregate with its neighboring nodes. The aggregation process is expressed as:

[0048]

[0049] Aggregate(·) represents an aggregation function, which receives node n i and the embedding vectors of its neighboring nodes As input, and output the updated embedding vector W agg is the weight parameter in the aggregation process.

[0050] Considering that the technical problem to be solved by this solution is essentially an LDA problem (Linear Discriminant Analysis), the goal of the aggregation function Aggregate(·) is to maximize the inter-class variance. The specific execution process is as follows:

[0051] S2020, all nodes n i Divide into two categories: C1 and C2; the goal of Fisher's optimal partitioning method is to find a partition that maximizes the variance between the two categories;

[0052] S2021, node n i Combine the embedding vectors of its neighbor nodes:

[0053]

[0054] Among them, the updated embedding vector Refers to node n i The embedding vector after the t+1th iteration, w ij is node n i and node n j The weight between them can be determined based on the similarity or distance between nodes. N(i) is the number of nodes n. i The set of neighbor nodes of is node n j Embedding vector after iteration t.

[0055] S2022, calculate the mean vector and covariance matrix for each class and use them to calculate the between-class variance, including:

[0056] (1) Calculate the mean vector:

[0057]

[0058] Among them, μ Ck Any category C k The mean vector of (k=1,2), |C k ∣ is any category C k The number of nodes in .

[0059] (2) Calculate the intra-class scatter matrix S W :

[0060]

[0061] Where T' represents the matrix transpose operation.

[0062] (3) Calculate the inter-class scatter matrix S B :

[0063]

[0064] (4) Calculate the Fisher criterion function J (maximize the ratio of between-class variance to within-class variance):

[0065]

[0066] Here, tr(·) represents the trace of the matrix (i.e., the sum of the diagonal elements of the matrix).

[0067] S2023, loop through the above steps S2021 to S2022; when step S202 is iterated for a predetermined number of times or when the Fisher criterion function J converges, proceed to step S203.

[0068] 2.2.4 Step S203, output graph structure:

[0069] Get the updated graph structure G′=(V′,E′), where is the updated collection, each node n i ′ corresponds to an embedding vector Where T represents the final iteration round; E′ is the updated edge set, which is updated according to the aggregation process.

[0070] 2.3 Step S3, the convolutional neural network layer performs the prediction task:

[0071] Different nodes n i Embedding vector after Tth iteration Transformed into a 64*64 matrix M as the input of the convolutional neural network layer. The convolutional neural network layer includes features for feature extraction to form the output result r 1 Conv layer, and used to output the result r 1 Linear combination to form the output result r 2 Linear layer; where:

[0072] Conv layer processing: For the matrix M, two convolution layers with a convolution kernel of 3*3 are applied to extract features, and then two maximum pooling layers with a kernel of 2*2 are applied to downsample, reducing the size of the feature map and retaining key information to form the output result r 1 .

[0073] Linear layer processing: based on the output result r 1 To three fully connected layers, ReLU is used as the activation function to increase the nonlinear expression ability of the model. After passing through three layers of full connection, the output result r 2 .

[0074] 2.3.1 Step S300, convert the embedded vector into a matrix:

[0075]

[0076] Reshape(·) indicates the operation of reshaping. Assume that the dimension of the embedding vector is d. If d is less than 64*64, zero padding is used; if d is greater than 64*64, selective truncation or dimensionality reduction is used.

[0077] 2.3.2 Step S301, Conv layer processing:

[0078] Two convolution layers with kernels of 3*3 are applied to the matrix M for feature extraction, and then two maximum pooling layers with kernels of 2*2 are applied for downsampling. The second convolution layer is defined similarly to the maximum pooling layer, but with different weights and biases, and the input is the output of the previous maximum pooling layer.

[0079] (1) The first convolutional layer:

[0080] M 1 =Conv(M,W 1 ,b 1 ,ks=(3,3));

[0081] Among them, W 1 and b 1 are the weights and biases of the first convolutional layer, Conv(·) represents the convolution operation, ks is the size of the convolution kernel; M 1 is the output of the first convolutional layer;

[0082] (2) The first maximum pooling layer:

[0083]

[0084] Among them, MaxPooling(·) represents the maximum pooling operation; stride = (2,2) means that the stride of the convolution or pooling operation in the horizontal and vertical directions is 2. It moves 2 units in the horizontal and vertical directions each time. A stride greater than 1 is usually used to reduce the size of the output feature map, thereby reducing the amount of calculation and parameters, and also helps to control overfitting. Represents the output of the first max pooling layer.

[0085] (3) The second convolutional layer:

[0086]

[0087] Among them, W 2 and b 2 are the weights and biases of the second convolutional layer, Conv(·) represents the convolution operation, ks is the size of the convolution kernel; M 2 is the output of the second convolutional layer;

[0088] (4) The second maximum pooling layer:

[0089] r 1 =MaxPooling(M 2 ,ks=(2,2),stride=(2,2));

[0090] Among them, MaxPooling(·) represents the maximum pooling operation.

[0091] 2.3.3 Step S302, Linear layer processing:

[0092] The output result r 1 Input to three fully connected layers, using ReLU as the activation function:

[0093] (1) The first fully connected layer:

[0094]

[0095] Among them, W f1 and b f1 are the weights and biases of the first fully connected layer, and Linear(·) represents the fully connected operation; is the output of the first fully connected layer.

[0096] Similarly, define the second and third fully connected layers.

[0097] (2) The second fully connected layer:

[0098]

[0099] Among them, W f2 and b f2 are the weights and biases of the second fully connected layer, and Linear(·) represents the fully connected operation. is the output of the second fully connected layer.

[0100] (3) The third fully connected layer:

[0101]

[0102] Among them, W f3 and b f3 are the weights and biases of the third fully connected layer, and Linear(·) represents the fully connected operation.

[0103] 2.4 Step S4, the output layer inputs the predicted dependent variable Y:

[0104] The output result r of the Linear layer 2 Apply the Sigmoid function and output the continuous value (probability or score) of the dependent variable Y as the degree of lens opacity:

[0105]

[0106] Where e is the base of the natural logarithm; if the ResNet-GNN neural network model is trained for a binary classification problem (e.g., determining whether the lens is cloudy), then Y can be interpreted as the probability of being cloudy; if the ResNet-GNN neural network model is trained for a regression problem (e.g., predicting the degree of lens opacity), then Y is the direct score output. This depends on the dataset D used by the ResNet-GNN neural network model during training:

[0107]

[0108] Among them, x i is the independent variable of the ith Rncat mouse, M i1 and M i2 is the two mating methods of the i-th Rncat mouse using one-hot encoding, namely inbreeding and closed group housing; A i is the age of the ith Rncat mouse (continuous numerical variable); μ i is the G489A mutation degree of the Crygs gene of the i-th Rncat mouse, which refers to the degree of truncation of γS crystallin at position 163, and is an interval value of [0,1] (the normal standard is 178 amino acids, and the number of amino acids is approximately less than 178, the greater the G489A mutation degree; the minimum is 0, and the maximum is 1); Y i is the degree of lens opacity (%) of the i-th Rncat mouse. i1 *A i 、Item M i2 *A i 、Item M i1 *μ i 、Item M i2 *μ i and term Ai*μ i They are all interactive feature processing items.

[0109] (III) Mechanism for solving technical problems:

[0110] 3.1 Feature extraction and representation:

[0111] Through the feature extraction layer, the model can establish nodes for the independent variables of Rncat mice (mating mode, age, and G489A mutation degree) and extract representative sample features. These features provide the data basis for the establishment of nonlinear relationships in subsequent model processing.

[0112] 3.2 Graph structure conversion and relationship capture:

[0113] The graph neural network layer transforms independent sample features into structured graph structures. This process not only retains the correlation information between features, but also captures the complex interactions between independent variables. The graph structure conversion mechanism helps the model better understand the intrinsic connections between data and provides a structural basis for the establishment of nonlinear relationships.

[0114] 3.2 Deep Learning and Nonlinear Modeling:

[0115] The convolutional neural network layer realizes nonlinear modeling of graph structure features through the combination of multiple convolutional layers and fully connected layers. The deep learning algorithm adopted in the present invention has nonlinear expression ability and can automatically learn the complex mapping relationship between data, thereby establishing a nonlinear relationship between independent variables and dependent variables.

[0116] 3.4 Forecasting and decision support:

[0117] Based on the trained model, the degree of lens opacity can be predicted for given independent variables (mating mode, age and G489A mutation degree). This prediction is not only highly accurate, but also can greatly reduce the cost and time of the experiment, providing an effective reference for the development of related drugs and the formulation of disease treatment strategies.

[0118] Second, a prediction system for the degree of lens opacity in Rncat mice:

[0119] The prediction system is used to implement the above-mentioned prediction method. The system includes a processor and a memory connected to the processor. The memory stores program instructions. When the program instructions are executed by the processor, the processor executes the prediction method for the degree of lens opacity as described above.

[0120] Compared with the prior art, the present invention has the following beneficial effects:

[0121] 1. Improve prediction accuracy: The present invention can automatically learn the complex nonlinear relationship between data through deep learning algorithms, thereby improving the accuracy of predictions. Compared with traditional linear regression or simple classification models, the ResNet-GNN neural network model can capture more subtle features and interactions between variables, making the prediction results more reliable.

[0122] Second, reduce experimental costs and time: Traditional research methods often require a large amount of experimental data to establish the relationship between independent variables and dependent variables, which is not only time-consuming and labor-intensive, but also costly. However, the present invention can achieve rapid prediction of new data by using existing data for training and learning, thereby greatly reducing experimental costs and time.

[0123] 3. Provide strong support for drug development: The present invention can provide strong support for the development of related drugs by predicting the degree of lens opacity of Rncat mice. For example, in the drug screening stage, the model of the present invention can be used to predict the effect of different drugs on the degree of lens opacity, so as to quickly screen out potential effective drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0124] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0125] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0126] Figure 2 It is a schematic diagram of the ResNet-GNN neural network model architecture of the present invention;

[0127] Figure 3 Schematic diagram of the architecture of the multi-scale residual network ResNet of the present invention;

[0128] Figure 4 It is a schematic diagram of the graph neural network layer of the present invention;

[0129] Figure 5 It is a schematic diagram of the process flow of Fisher optimal segmentation method of the present invention;

[0130] Figure 6 Schematic diagram of the convolutional neural network layer architecture of the present invention;

[0131] Figure 7 It is a schematic diagram of the relationship between independent variables of the present invention;

[0132] Figure 8 It is a schematic diagram of the training effect of the ResNet-GNN neural network model of the present invention;

[0133] Fig. 9 It is a comparative schematic diagram of the test example of the present invention. DETAILED DESCRIPTION

[0134] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below;

[0135] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0136] Embodiment 1: Figure 1 As shown, this embodiment discloses a method for predicting the degree of lens opacity in Rncat mice; the prediction method includes the following steps: Figure 2 The ResNet-GNN neural network model shown in the figure needs to be trained by the training scheme disclosed in Example 2 before deployment; after actual deployment, based on the different independent variables of the Rncat mice to be predicted (age, mating mode, and G489A mutation degree of the Crygs gene), the corresponding dependent variable (degree of lens opacity) can be quickly and conveniently obtained by performing the following steps:

[0137] First, the independent variable X = [M, A, μ] of the target Rncat mouse is collected, where M is the mating mode of the Rncat mouse, which is inbreeding and closed group housing (inbreeding is represented by [1, 0], and closed group housing is represented by [0, 1]); A is the age of the Rncat mouse; μ is the G489A mutation degree of the Crygs gene of the Rncat mouse, which refers to the degree of truncation of γS crystallin at position 163, and is an interval value of [0, 1] (the normal standard is 178 amino acids, the smaller the number of amino acids is, the greater the G489A mutation degree; the minimum is 0 and the maximum is 1).

[0138] Then, through the ResNet-GNN neural network model, perform the following steps S1 to S4:

[0139] In this embodiment, if Figure 3 As shown, regarding step S1, the feature extraction layer performs the extraction task: establish node n for the independent variable X of Rncat mice i ; Extract each node n through the multi-scale residual network ResNet i The sample characteristics P i .in:

[0140] (1) Node n 1 represents the mating mode M, which is an important factor affecting the degree of lens opacity in Rncat mice;

[0141] (2) Node n 2 represents the age of the mouse, which is also a key factor affecting the degree of lens opacity;

[0142] (3) Node n 3 μ represents the G489A mutation degree of the Crygs gene, which is the influence of genetic level on the degree of lens opacity.

[0143] The sample feature P of each node ni is extracted through the multi-scale residual network ResNet i The specific process includes the following steps S100 to S101.

[0144] Specifically, step S100, residual learning: the multi-scale residual network ResNet includes corresponding to node n 1 、n 2 and n 3 Each residual block contains two or three convolutional layers and a shortcut connection layer. The shortcut connection layer (also called skip connection) allows the input to skip some layers directly and be added to the output of these layers; it helps to alleviate the gradient vanishing problem in deep neural networks and improve the training efficiency and performance of the model.

[0145] Each residual block reads the independent variable X (to be precise, x i , that is, the mating mode M, the age A or the G489A mutation degree μ of the Crygs gene) The residual block is learned through residual learning to output each node n i The sample characteristics P i :

[0146]

[0147] in is the output of the lth residual block, is the input of the lth residual block (for the first residual block, That is node n i The original input variable x i ), F(·) represents the residual function, which contains two convolutional layers. represents the weight parameters of these convolutional layers.

[0148] It is understandable that the sample feature P iIt is obtained by deep learning and feature extraction of the independent variables of Rncat mice (mating mode, age, and G489A mutation of Crygs gene). These features can capture the initial nonlinear relationship between the independent variables and the degree of lens opacity, providing strong support for subsequent prediction tasks. i As the input of the graph structure G in the subsequent step S2, it can enhance the representation ability of the graph neural network. Through the graph structure, the relationship between nodes can be better modeled, and the potential patterns and structural information in the data can be captured, thereby improving the prediction accuracy and generalization ability of the model. In the graph structure G, the node n i Connected to other nodes through edges, forming an information transmission network. Sample feature P i As a node attribute, it can be propagated and updated in the graph neural network, allowing the model to learn richer feature representations and node relationships.

[0149] Further, in step S1000, the residual function F(·) is a key part of the multi-scale residual network ResNet. It is responsible for learning the residual representation of the input data to capture the complex relationship between the independent variable and the degree of lens opacity.

[0150] For the i-th node (i=1, 2, 3) and the l-th residual block (l represents the index of the residual block), the residual function F(·) is expressed as:

[0151]

[0152] in, and are the learnable weight parameters of the first and second convolutional layers in the lth residual block respectively; and are the bias parameters of the first and second convolutional layers in the lth residual block, and are also parameters that need to be learned. The symbol * represents the convolution operation, which is used to extract local features of the input data. σ(·) is the ReLU activation function, that is, σ(·) = max(0, x), which is used to increase the nonlinear representation ability of the model.

[0153] The residual function learns the residual representation of the input data through a combination of two convolutional layers and the ReLU activation function. This design helps the model capture subtle changes and complex patterns in the data. By learning the residual, the model can more effectively handle noise and redundant information in the input data, improving the accuracy and robustness of the prediction.

[0154] For the two convolutional layers of the residual function F(·), in deep learning, the convolutional layer usually changes the number of channels of the input data (that is, the depth of the feature map). When the number of channels of the input and output are different, direct addition will result in dimension mismatch, and thus the operation cannot be performed. Similarly, the convolutional layer may also change the spatial size of the input data (that is, the height and width of the feature map). When the spatial sizes of the input and output are different, direct addition will also result in dimension mismatch. Therefore, further, step S1001, regarding the execution of the shortcut connection layer: the shortcut connection layer allows the input to directly skip some layers and be superimposed with the outputs of these layers. It helps to alleviate the gradient vanishing problem in the above-mentioned deep neural network and improve the training efficiency and performance of the model. Among them, for the two convolutional layers of the residual function F(·), the shortcut connection layer allows the input to directly skip some layers and be added to the outputs of these layers; the specific method is:

[0155] Calculate the output of the i-th node (i=1,2,3) and the l-th residual block And when size matching or dimensionality adjustment is required (for example, when the input and output have different numbers of channels or different spatial dimensions), perform the following operations:

[0156]

[0157] in, is a linear transformation operation, including performing any one or more of the following operations:

[0158] (1) For the input feature map and any convolution kernel W in the convolution layer conv , perform the convolution operation:

[0159]

[0160] Among them, (p,q) is the position on the output feature map, and (m,n) is the position on the convolution kernel.

[0161] (2) For the input feature map Perform batch normalization to speed up the convergence of the neural network and improve the stability of the training process:

[0162]

[0163] Among them, μ batch and are the mean and variance of the input feature map over the batch, ∈ is a small constant to prevent division by zero errors, and γ and β are weights (learnable).

[0164] (3) For the input feature map Perform a zero-padding operation:

[0165]

[0166] Among them, p and q are rows and columns respectively; p pad and q pad are the number of filled rows and columns respectively.

[0167] It can be understood that by performing the above operations, the model can process input data with different numbers of channels or spatial sizes, thereby improving the flexibility and versatility of the model. This enables the model to be applied to more diverse practical scenarios, such as image classification of different resolutions, data prediction of different feature dimensions, etc. Moreover, in the residual network, each residual block can extract features at different levels. When the number of channels or spatial sizes of the input and output are different, by performing the above operations, the model can more effectively integrate feature information at different levels, thereby improving the accuracy and richness of feature extraction. This helps the model better capture the key information in the input data and improve the accuracy and robustness of the prediction. At the same time, in some cases, the computational efficiency of the model can be optimized by adjusting the number of channels or spatial size of the input data;

[0168] For example, using 1x1 convolution to adjust the number of channels can reduce the amount of computation while maintaining the performance of the model. This helps to reduce the computational cost in practical applications and improve the real-time performance and usability of the model.

[0169] Specifically, step S101, feature extraction: This step uses a multi-scale residual network ResNet for feature extraction, aiming to extract features useful for predicting the degree of lens opacity from the input independent variables (mating mode, age, and G489A mutation degree of the Crygs gene). Since there are multiple residual blocks in the multi-scale residual network ResNet, the feature extraction process of the entire multi-scale residual network ResNet is:

[0170]

[0171] Among them, H(·) represents the ResNet network function, L represents the total number of residual blocks,

[0172] Represents the weight parameters of all residual blocks.

[0173] Among them, the multi-scale residual network ResNet consists of multiple residual blocks, each of which contains a series of convolutional layers, batch normalization layers, and activation function layers to capture different levels of features in the input data. The ResNet network function H(·) represents the continuous application of these residual blocks, and the operation method is as follows:

[0174]

[0175] Among them, the weight parameter set of all residual blocks is represented as W i At the same time, F l (·) represents the residual function of the lth residual block, which receives the input and corresponding weight parameters and outputs the features of the residual block. In addition, the output of each residual block is added with its input (i.e., the shortcut connection of S1001). Furthermore, each residual block contains multiple convolutional layers to extract local features of the input data. The convolutional layer is usually followed by a batch normalization layer and an activation function layer to improve the training stability and nonlinear expression ability of the model. In addition, the output of each residual block is added with its input (i.e., the shortcut connection), which helps to alleviate the gradient vanishing problem in deep neural networks and makes the model easier to train. After feature extraction by the multi-scale residual network ResNet, each sample will obtain a feature vector P i , which contains useful information extracted from the input features for predicting the degree of lens opacity. These feature vectors will be used as input for subsequent graph neural networks GNN to further capture the association between samples.

[0176] In this embodiment, regarding step S2, as Figures 4-5 As shown, the graph neural network layer performs the conversion task: transforming independent sample features P i Transformed into a structured graph structure G, thus transforming the prediction problem into a node prediction task within the graph neural network (GNN) framework. This step aggregates each adjacent node n in turn through an ordered clustering algorithm i , constantly updating different nodes n i The embedding vector e vi This process aims to capture the interaction and complex relationship between the independent variables (mating mode, age and G489A mutation degree of Crygs gene) and provide a richer and more accurate feature representation for the subsequent prediction model. It includes the following steps S200 to S203.

[0177] Specifically, step S200, graph structure construction:

[0178] All sample features P i The sample feature set Construct a graph structure G = (V, E), where Each node n i Corresponding to a sample feature P i ; E is the edge set, which represents the connection relationship between nodes.

[0179] In a graph neural network, each edge e in the edge set E ij ∈E is connected based on the similarity of the sample features Pi of the nodes. That is, if two sample features P i and P jIf they are similar under some metric, then their corresponding nodes n i and n j There will be an edge e between ij . This helps capture the potential associations between features and reveal the interactions between independent variables.

[0180] For example, each edge e in the edge set E ij ∈E based on the sample features P of the node i The rule for connecting based on similarity is: for any two nodes n i and n j (corresponding to sample features P i and P j ), if the difference between their sample features in a specific dimension (such as cosine similarity or Euclidean distance) is less than a preset threshold TV, then there is an edge e between them. ij ∈E: eij∈E if and only if

[0181] k max ∣P ik -P jk ∣ <TV;

[0182] This rule builds a graph structure based on the similarity between node sample features. Specifically, if the difference between two sample features in a specific dimension is small, they have similar properties or belong to the same category, so they should be connected in the graph structure. It helps the graph neural network capture the potential associations and interactions between sample features, thereby improving the prediction performance of the model. Through the edge connection rules based on the similarity of sample features, the graph neural network is able to capture the relationship between samples with similar features. This helps the model better understand the distribution and potential patterns of the data, thereby improving the accuracy of the prediction. When the model is able to learn the relationship between samples with similar features, it is more likely to make accurate predictions on unseen data. It helps to improve the generalization ability of the model, making it more robust in practical applications.

[0183] Among them, P ik and P jk Represent the sample features P i and P j The value in the kth dimension, TV is a preset threshold.

[0184] Specifically, step S201, node embedding vector initialization: for each node n i , initialize its embedding vector in Represents node n i The embedding vector at layer 0 (i.e. the initial layer) is directly equal to the sample feature P iThe purpose of initializing the embedding vector is to provide a starting point for subsequent aggregation and update operations. The embedding vector is a representation of nodes in the graph neural network, which contains the feature information of the node and the relationship information between the node and other nodes. By continuously aggregating and updating the embedding vector, the graph neural network can gradually capture the complex relationships and patterns in the graph structure.

[0185] It can be understood that by converting sample features into a graph structure and processing it with a graph neural network, this scheme can effectively capture the interactions and complex relationships between independent variables. This helps to more accurately understand the factors affecting the degree of lens opacity in Rncat mice, thereby providing a richer feature representation for the prediction model. By aggregating and updating the embedding vectors of nodes, the graph neural network can gradually learn the global and local information in the graph structure. This information fusion method helps to improve the accuracy of the prediction model and makes the prediction of the dependent variable based on a given independent variable more reliable.

[0186] Specifically, step S202, such as Figure 5 As shown, ordered clustering is performed: the ordered clustering algorithm is Fisher's optimal partitioning method; by aggregating each adjacent node in turn. This process can be iterated, and each iteration updates the node's embedding vector. Let the current iteration round be t, for node n i , whose neighbor node set is N(i). Fisher optimal partitioning method is used to determine how to partition n i Aggregate with its neighboring nodes. The aggregation process is expressed as:

[0187]

[0188] Aggregate(·) represents an aggregation function, which receives node n i and its neighbor nodes’ embedding vectors as input and output the updated embedding vector W agg It is the weight parameter in the aggregation process, which plays a regulatory role in the aggregation process to ensure the rationality of aggregation.

[0189] Considering that the above problem is essentially an LDA problem (linear discriminant analysis), the goal of the aggregation function Aggregate(·) is preferably to maximize the inter-class variance. The specific execution process is as follows:

[0190] S2020, all nodes n i Divide into two categories: C1 and C2. The goal of Fisher's optimal partitioning method is to find a partition that maximizes the variance between the two categories;

[0191] S2021, node n iCombine the embedding vectors of its neighbor nodes:

[0192]

[0193] Among them, the updated embedding vector Refers to node n i The embedding vector after the t+1th iteration, w ij is node n i and node n j The weight between them can be determined based on the similarity or distance between the nodes (e.g., the smaller the difference, the higher the weight). N(i) is the number of nodes n. i The set of neighbor nodes of is node n j Embedding vector after iteration t.

[0194] S2022, calculate the mean vector and covariance matrix for each class and use them to calculate the between-class variance, including:

[0195] (1) Calculate the mean vector:

[0196] Among them, μ Ck Any category C k The mean vector of (k=1,2), |C k ∣ is any category C k The number of nodes in .

[0197] (2) Calculate the intra-class scatter matrix S W :

[0198]

[0199] Where T' represents the matrix transpose operation.

[0200] (3) Calculate the inter-class scatter matrix S B :

[0201]

[0202] (4) Calculate the Fisher criterion function J (maximize the ratio of between-class variance to within-class variance):

[0203]

[0204] Here, tr(·) represents the trace of the matrix (i.e., the sum of the diagonal elements of the matrix).

[0205] S2023, loop through the above steps S2021 to S2022; when step S202 is iterated for a predetermined number of times or when the Fisher criterion function J converges, proceed to step S203.

[0206] It is understandable that orderly clustering through the Fisher optimal partitioning method can maximize the inter-class variance, making the clustering results more reasonable and helpful for subsequent graph structure updates and node feature extraction. Iteratively updating the embedding vector of the node can gradually optimize the representation of the node and improve the model's ability to capture node features. This step provides more accurate and effective node feature information for subsequent graph structure output and prediction tasks.

[0207] Specifically, in step S203, the graph structure is output: the updated graph structure G′=(V′, E′) is obtained, where is the updated node set, each node n i ′ corresponds to an embedding vector (After the previous ordered clustering process, the embedding vector of the node has been optimized and updated), where T represents the final iteration round; E′ is the updated edge set, which is updated according to the aggregation process (during the aggregation process, the node is aggregated with its neighboring nodes, which may lead to changes in the edge set, such as the addition, deletion or adjustment of the weight of the edge).

[0208] The final output graph structure G′ contains the updated node set V′ and edge set E′. This graph structure will be used as the input of subsequent prediction tasks to capture the relationship and feature information between nodes.

[0209] Outputting the updated graph structure provides more accurate and effective input data for subsequent prediction tasks. By optimizing and updating the embedding vectors and edge sets of nodes, the model's ability to capture node features and relationships between nodes is improved. This helps improve the prediction accuracy of the ResNet-GNN neural network model for the degree of lens opacity in Rncat mice, thereby reducing experimental costs and providing a reference for the development of related drugs and the formulation of disease treatment strategies.

[0210] In this example, see Figure 6 , regarding step S3, the convolutional neural network layer performs the prediction task: different nodes n i Embedding vector after Tth iteration Transformed into a 64*64 matrix M as the input of the convolutional neural network layer. The convolutional neural network layer includes features for feature extraction to form the output result r 1 Conv layer, and used to output the result r 1 Linear combination to form the output result r 2 Linear layer; where:

[0211] Conv layer processing: For the matrix M, two convolution layers with a convolution kernel of 3*3 are applied to extract features, and then two maximum pooling layers with a kernel of 2*2 are applied to downsample, reducing the size of the feature map and retaining key information to form the output result r 1 .

[0212] Linear layer processing: based on the output result r 1 To three fully connected layers, ReLU is used as the activation function to increase the nonlinear expression ability of the model. After passing through three layers of full connection, the output result r 2 .

[0213] The embedding vector is the node n i The low-dimensional representation in the iterative process contains the feature information of the node. The embedding vector is converted into a 64*64 matrix M to meet the input requirements of the convolutional neural network. Convolutional neural networks are usually used to process image data, which is represented in matrix form.

[0214] Specifically, step S300 converts the embedded vector into a matrix:

[0215]

[0216] Reshape(·) indicates the operation of reshaping. Assume that the dimension of the embedding vector is d. If d is less than 64*64, zero padding is used; if d is greater than 64*64, selective truncation or dimensionality reduction is used.

[0217] Use the Reshape operation to fill in zeros if the dimension d of the embedding vector is less than 64x64; if d is greater than 64x64, selectively truncate or use dimensionality reduction techniques to convert the embedding vector into a matrix form suitable for convolutional neural network processing.

[0218] Specifically, step S301, Conv layer processing: apply two convolution layers with a convolution kernel of 3*3 to the matrix M for feature extraction, and then apply two maximum pooling layers with a kernel of 2*2 for downsampling. The second convolution layer is defined similarly to the maximum pooling layer, but uses different weights and biases, and the input is the output of the previous maximum pooling layer.

[0219] (1) The first convolutional layer:

[0220] M 1 =Conv(M,W 1 ,b 1 ,ks=(3,3));

[0221] Among them, W 1 and b 1are the weights and biases of the first convolutional layer, Conv(·) represents the convolution operation, and ks is the size of the convolution kernel.

[0222] (2) The first maximum pooling layer:

[0223]

[0224] Among them, MaxPooling(·) represents the maximum pooling operation; stride = (2,2) means that the stride of the convolution or pooling operation in the horizontal and vertical directions is 2. It moves 2 units in the horizontal and vertical directions each time. A stride greater than 1 is usually used to reduce the size of the output feature map, thereby reducing the amount of calculation and parameters, and also helps to control overfitting. Represents the output of the first max pooling layer.

[0225] (3) The second convolutional layer:

[0226]

[0227] Among them, W 2 and b 2 are the weights and biases of the second convolutional layer, Conv(·) represents the convolution operation, and ks is the size of the convolution kernel.

[0228] (4) The second maximum pooling layer:

[0229] r 1 =MaxPooling(M 2 ,ks=(2,2),stride=(2,2));

[0230] Among them, MaxPooling(·) represents the maximum pooling operation.

[0231] The features in the matrix are extracted and the size of the feature map is reduced through the maximum pooling layer, which reduces the amount of calculation and parameters while retaining key information.

[0232] The convolution layer uses multiple convolution kernels to perform convolution operations on the input matrix to extract features. The size of the convolution kernel (such as 3*3) determines the size of the receptive field, that is, the range of features that the convolution operation can capture. Through the convolution operation, the convolution layer can extract multiple features from the input matrix. The maximum pooling layer downsamples the output of the convolution layer to reduce the size and computation of the feature map.

[0233] Specifically, step S302, Linear layer processing: input the output result r1 into three fully connected layers, using ReLU as the activation function:

[0234] (1) The first fully connected layer:

[0235]

[0236] Among them, W f1 and b f1 are the weights and biases of the first fully connected layer, and Linear(·) represents the fully connected operation. is the output of the first fully connected layer.

[0237] Similarly, define the second and third fully connected layers.

[0238] (2) The second fully connected layer:

[0239]

[0240] Among them, W f2 and b f2 are the weights and biases of the second fully connected layer, and Linear(·) represents the fully connected operation. is the output of the second fully connected layer.

[0241] (3) The third fully connected layer:

[0242]

[0243] Among them, W f3 and b f3 are the weights and biases of the third fully connected layer, and Linear(·) represents the fully connected operation.

[0244] The extracted features are linearly combined to form the final output result. The ReLU activation function increases the nonlinear expression ability of the model. Step S3 extracts features and linearly combines the embedded vectors through the convolutional neural network layer to finally obtain the prediction result.

[0245] The fully connected layer linearly combines the features extracted by the convolutional layer to form a higher-level feature representation. Using the ReLU activation function after the fully connected layer can increase the nonlinear expression ability of the model, enabling the model to learn more complex feature representations. The last fully connected layer is usually used as the output layer to output the prediction results.

[0246] Convolutional neural networks can extract rich feature information from input data through convolution operations and maximum pooling operations. The maximum pooling operation can reduce the size and amount of computation of the feature map, thereby improving the computational efficiency of the model. Convolutional neural networks can effectively reduce the number of model parameters through mechanisms such as shared weights and local receptive fields, thereby reducing the risk of overfitting. At the same time, the use of the ReLU activation function also helps to improve the generalization ability of the model. The conversion process of embedding vectors to matrices enables convolutional neural networks to process non-image data, such as the node embedding vectors of this embodiment.

[0247] In this embodiment, regarding step S4, the output layer inputs the predicted dependent variable Y:

[0248] The output result r of the Linear layer 2 Apply the Sigmoid function and output the continuous value (probability or score) of the dependent variable Y as the degree of lens opacity:

[0249] Where e is the base of the natural logarithm. If the ResNet-GNN neural network model is trained for a binary classification problem (such as determining whether the lens is cloudy), then Y can be interpreted as the probability of turbidity. If the ResNet-GNN neural network model is trained for a regression problem (such as predicting the opacity score of the lens), then Y is the direct score output.

[0250] For a binary classification problem, a Y value close to 1 indicates that the model is more likely to predict that the lens is cloudy, while a Y value close to 0 indicates that the model is more likely to predict that the lens is clear.

[0251] For regression problems, the Y value can be directly output as a score of the degree of lens opacity. According to the specific score definition and range, the Y value can be further interpreted and analyzed.

[0252] Embodiment 2: This embodiment further discloses the training process and method of the ResNet-GNN neural network model described in Embodiment 1, including:

[0253] P1. Dataset D construction:

[0254] The different mating modes, ages, G489A mutation degrees of Crygs genes of different Rncat mice, and the corresponding lens opacity data were collected. The collected data were preprocessed, including routine data cleaning and normalization. Then the data set D was constructed:

[0255]

[0256] Among them, x i is the independent variable of the ith Rncat mouse, M i1 and M i2 is the two mating methods of the i-th Rncat mouse using one-hot encoding, namely inbreeding and closed group housing; A i is the age of the ith Rncat mouse (continuous numerical variable); μ iis the G489A mutation degree of the Crygs gene of the i-th Rncat mouse, which refers to the degree of truncation of γS crystallin at position 163, and is an interval value of [0,1] (the normal standard is 178 amino acids, and the number of amino acids is approximately less than 178, the greater the G489A mutation degree; the minimum is 0, and the maximum is 1); Y i is the degree of lens opacity (%) of the i-th Rncat mouse. i1 *A i 、Item M i2 *A i 、Item M i1 *μ i 、Item M i2 *μ i and term Ai*μ i All are interactive feature processing items;

[0257] Then, the dataset D is divided into training set D train , test set D test , validation set D val 70% of the data can be used as training set D train , 15% as validation set D val to adjust the model parameters, and the remaining 15% is used as the test set D test To evaluate the model performance.

[0258] P2, such as Figures 2 to 6 As shown, the ResNet-GNN neural network model architecture described in steps S1 to S4 in Example 1 is constructed, and the parameters to be trained are determined as follows:

[0259] (1) Convolutional layer parameters in residual block:

[0260] and The learnable weight parameters of the first and second convolutional layers in the lth residual block;

[0261] and The bias parameters of the first and second convolutional layers in the lth residual block;

[0262] (3) Convolutional neural network layer parameters:

[0263] W 1 and b 1 : weights and biases of the first convolutional layer;

[0264] W 2 and b 2 : weights and biases of the second convolutional layer;

[0265] (3) Fully connected layer parameters:

[0266] W f1 and b f1 : weights and biases of the first fully connected layer;

[0267] W f2 and b f2 : weights and biases of the second fully connected layer;

[0268] W f3 and b f3 : Weights and biases of the third fully connected layer.

[0269] P3, training algorithm:

[0270] P3.1. Initialization parameters: Randomly initialize the above trainable parameters.

[0271] P3.2, Feature extraction:

[0272] For each Rncat mouse, the independent variable x i Create node n i And use the multi-scale residual network ResNet to extract the sample features of each node.

[0273] P3.2, Graph structure construction and node embedding:

[0274] Convert sample features into structured graph structures;

[0275] Initialize the embedding vector of each node;

[0276] Fisher optimal split method is used to update the embedding vector of the node to capture the interaction and complex relationship between independent variables.

[0277] P3.3, Convolutional Neural Network Prediction:

[0278] Convert the embedding vectors of different nodes into matrices as input to the convolutional neural network layer;

[0279] Apply convolutional layers and max pooling layers for feature extraction and downsampling;

[0280] Input the output result into the fully connected layer to form the final prediction result;

[0281] P3.4, loss calculation and back propagation:

[0282] The loss function is calculated based on the prediction results and the true label. In this embodiment, the MAPE (mean absolute percentage error) loss function is selected.

[0283] Use the back propagation algorithm to calculate the gradient and update the trainable parameters. The process includes:

[0284] P3.4.1. Forward propagation: Use the activation values ​​and weights of the previous layer to calculate the input of the current layer;

[0285] P3.4.2, Activation function: The input is passed through the ReLU activation function to calculate the activation value of the current layer;

[0286] P3.4.3, loss function: such as Figure 8 As shown, calculate the value of the MAPE loss function;

[0287] P3.4.4, Back propagation: Use the chain rule to calculate the derivative of the MAPE loss function with respect to the weights;

[0288] P3.4.5, parameter update: update the weights and biases of the corresponding layers according to the derivatives and learning rates

[0289] P3.5, iterative training: Repeat steps P3.2 to P3.4 until the predetermined number of iterations is reached or the loss function converges.

[0290] P3.6, Model evaluation: Use the test set to evaluate the performance of the model, including accuracy, recall, F1 score and other indicators.

[0291] Further, this embodiment will provide the following MATLAB execution program to construct the ResNet-GNN neural network model described in Example 1; wherein the multi-scale residual network (ResNet) is constructed using MATLAB's deep learning toolbox to extract features. Including defining residual blocks, convolutional layers, and shortcut connections; at the same time, using the graph neural network toolbox (if available) to implement the GNN graph neural network layer, including building a graph structure, initializing node embedding, and performing ordered clustering, using MATLAB's deep learning toolbox to build a convolutional neural network layer for processing the output of the GNN layer and generating the final prediction results; at the same time, it also includes combining the ResNet feature extraction layer, the GNN graph, and the CNN convolutional neural network layer, and defining the training process:

[0292]

[0293]

[0294]

[0295]

[0296] The specific training procedures are as follows:

[0297]

[0298]

[0299]

[0300]

[0301]

[0302] Embodiment 3: In step S101 of embodiment 1, the quick connection layer may have a matching problem due to non-identical mapping. To avoid this phenomenon, this embodiment further provides the following technical solutions based on step S101 of embodiment 1:

[0303] It combines convolution, batch normalization, and activation functions to match the dimensions of the residual path and the shortcut connection path. This projection mechanism is: shortcut =BN(Conv(x))+b;

[0304] Among them, Conv represents the convolution operation, BN represents the batch normalization operation, b is the bias term (optional), x is the input feature of the shortcut connection path, and x shortcut It is the feature processed by projection quick connection.

[0305] For a given input feature x and residual block F(x), the output x of the shortcut connection layer is shortcut It can be adjusted according to one of the above schemes to ensure that it matches the output feature dimension of the residual block. Then, the adjusted shortcut connection feature is added to the residual feature to obtain the final output of the residual block; ultimately, the matching problem caused by the non-identical mapping of the shortcut connection layer can be effectively solved, ensuring that the multi-scale residual network ResNet can be stably trained and accurately predict the degree of lens opacity in Rncat mice.

[0306] Embodiment 4: In step S101 of embodiment 1, if each node n can be transmitted simultaneously i The sample characteristics P i and its identification information, then the expressiveness and robustness of the features can be enhanced. To this end, this embodiment will further provide the following technical solutions based on step S101 of embodiment 1:

[0307] P1, each node n i The sample characteristics P i Expand to include identification information. Identification information can be discrete (such as category labels), continuous (such as location coordinates), or mixed. For discrete information, use One-Hot Encoding or Embedding Vector to represent it; for continuous information, it can be directly added to the feature vector.

[0308] P2, feature fusion: The expanded node features (including sample features and identification information) are input into the multi-scale residual network ResNet. In each residual block, feature fusion is performed implicitly because the residual block automatically learns how to combine these two types of information to extract useful features.

[0309] Furthermore, the above solutions are summarized as follows:

[0310] Let node n i The sample feature is P i , identification information is Q i (which can be a one-hot encoded vector, an embedded vector, or a position coordinate), the expanded node feature can be expressed as:

[0311] P i ′=[P i ,Q i ];

[0312] Among them, [·,·] represents the feature concatenation operation;

[0313] Then, the expanded feature P i ′ is input into ResNet for feature extraction:

[0314] P iout =ResNet(P i ′);

[0315] Where ResNet(·) represents the continuous application of residual blocks described in the multi-scale residual network ResNet, P iout Represents the output features after ResNet feature extraction.

[0316] Embodiment 5: In step S200 of embodiment 1, by presetting the threshold TV, some node pairs with large differences can be filtered out, thereby reducing noise interference. This helps the model focus more on learning useful feature representations and improves prediction performance. However, the subjectively assigned preset threshold TV cannot adapt to dynamic environments. To this end, this embodiment further provides two dynamic assignment methods for the preset threshold TV:

[0317] (1) Solution 1: Calculate the difference between each data point and the average value of the data set. First, calculate the feature set of the sample The average value P' is:

[0318] Then, calculate each sample feature P i The deviation D i =|P i -P'|

[0319] Finally, the threshold TV is determined according to the maximum value of the deviation:

[0320] Alternatively, a deviation value corresponding to a larger quantile (such as the 95% quantile) is selected as the threshold to exclude the influence of extreme values.

[0321] (2) Solution 2: Calculate the average that reflects the degree of dispersion and the degree of concentration of the distribution of the data. First, calculate the sample feature set The mean square error S 2 :

[0322] Then, twice the root of the mean square error (that is, twice the standard deviation) is selected as the threshold to balance the degree of variation in the data and the strictness of the threshold:

[0323] This threshold ensures that approximately 95% of the data points (due to the properties of the normal distribution) fall within the threshold range.

[0324] It can be understood that, in summary of the two schemes described above, the determination of the threshold TV is based on the actual variation of the data, rather than a subjective setting, which makes the construction of the graph structure more in line with the actual situation of the data. By considering the overall variation of the data, the influence of extreme values ​​on the determination of the threshold can be reduced, thereby improving the robustness of the graph structure.

[0325] Embodiment 6: In step S2021 of embodiment 1, w ij is the weight between node ni and node nj, determined according to the similarity or distance between the nodes (e.g., the smaller the difference, the higher the weight). The weights are assigned only according to the similarity or distance between the nodes, without considering the difference in importance of different indicators (or features) in the decision problem. This may result in some indicators that actually have a greater impact on the results being assigned lower weights, while some indicators with less impact are assigned higher weights. In practical applications, the similarity or distance between nodes may be affected by multiple factors, which may be correlated. The previous scheme did not take this correlation into account, which may lead to inaccurate weight assignment.

[0326] Because in multi-attribute decision-making problems, different indicators usually have different degrees of influence on the results. If the weight is determined only based on the similarity or distance between nodes without considering the importance of the indicators, the weight allocation may not accurately reflect the actual role of each indicator in the decision-making problem. Therefore, this embodiment further improves the weight w in a data-driven manner. ij The assignment scheme.

[0327] Let node n i and node n j The similarity between them is sij , and s ij ∈[0,1], where s ij =1 means completely similar, s ij =0 indicates that they are completely dissimilar. The similarity can be calculated based on the feature vectors of the nodes as described in the first embodiment by using a similarity measurement method (cosine similarity, Euclidean distance).

[0328] First, calculate the node n i The sum of similarities of all neighbor nodes (including itself):

[0329] S i =∑ j∈N(i)∪{i} S ij ;

[0330] Then, calculate each node n j At node n i The similarity ratio in the neighbor node set (including itself):

[0331] Next, the entropy weight method is used to calculate the importance of each node feature. The entropy weight method reflects the degree of variation of the feature by calculating the information entropy of the feature, thereby determining the weight. The smaller the information entropy, the higher the importance of the feature and the greater the weight; let the information entropy of the node feature be e j :

[0332]

[0333] Among them, |N(i)∪{i}| represents the node n i The size of the neighbor node set (including itself).

[0334] Finally, the weight w is calculated based on the information entropy ij :

[0335]

[0336] Among them, the weight w ij The similarity between nodes and the importance of each node's features are comprehensively considered.

[0337] It can be understood that the entropy weight method reflects the degree of variation of each indicator by calculating the information entropy. The smaller the information entropy, the greater the degree of variation of the indicator, the more information it provides, and therefore the higher its importance in decision-making, and a larger weight should be given. In this way, the entropy weight method can assign weights according to the actual importance of the indicator, avoiding the problem of ignoring the difference in the importance of indicators when determining weights only based on node similarity or distance. The entropy weight method reflects the degree of variation of indicators by calculating information entropy, rather than directly relying on the similarity or distance between nodes. Therefore, it is relatively insensitive to small changes in data and can reduce the impact of data disturbances on weight allocation to a certain extent.

[0338] Experimental example: This example aims to verify the superiority of the ResNet-GNN model in processing complex and irregular input data by comparing the performance of the experimental group (using the ResNet-GNN neural network model as described in Examples 1 to 6), the control group a (using the traditional CNN neural network model) and the control group b (using SGNN, i.e., the neural network model of ProtTrans combined with GNN) in predicting specific biomedical problems (such as the relationship between the degree of lens opacity of Rncat mice and time).

[0339] 1. Experimental design

[0340] Experimental group: ResNet-GNN neural network model, combining multi-scale residual network ResNet and graph neural network GNN. ResNet is used for efficient feature extraction, and GNN is combined to capture the structural relationship between data.

[0341] Control group a: Traditional CNN neural network model. It is suitable for fields such as image recognition, but may not perform well for complex and irregular biomedical data. However, it is one of the classic neural network models, so it is included in the control consideration.

[0342] Control group b: SGNN (ProtTrans combined with GNN). ProtTrans was used for protein feature extraction and combined with GNN for subsequent analysis.

[0343] (II) Experimental content:

[0344] 2.1 Dataset:

[0345] according to Fig. 9 The preset actual value curve relationship (lens opacity degree-time) is shown, and the corresponding sample data, including the age, mating mode and G156A mutation degree of Rncat mice are selected as input features, and the lens opacity degree is used as the output target.

[0346] 2.2 Training process:

[0347] The experimental group, control group a, and control group b were trained separately to ensure that each model fully learned the data characteristics. The same training parameters and conditions were used to ensure the fairness of the experimental results.

[0348] 2.3 Prediction Task:

[0349] After the training is completed, a period of time is randomly selected and each model is required to make predictions. The preset actual value curve is compared with the prediction curves of the three groups of models to evaluate the prediction effect.

[0350] (III) Experimental results and analysis:

[0351] from Fig. 9 It can be clearly seen that the predicted curve of the experimental group is closest to the actual value curve, indicating that the ResNet-GNN model has significant advantages in dealing with such problems. At the same time, based on the architecture support of the ResNet-GNN model, the nonlinear distribution relationship between the independent variables (age, mating mode and G156A mutation degree of Rncat mice) and the dependent variable (degree of lens opacity) can be output as follows: Figure 7 shown.

[0352] 3.1 Analysis of control group a:

[0353] CNN models have difficulty processing complex and irregular input data. CNN is mainly designed to process regular data such as images, and is not capable of capturing nonlinear relationships and complex features in biomedical data.

[0354] 3.2 Analysis of control group b:

[0355] 3.2.1 Question 1: The feature extraction effect of ProtTrans is not as good as ResNet.

[0356] ResNet can effectively alleviate the gradient vanishing problem in deep networks through multi-scale residual connections, while capturing richer feature information; while ProtTrans is a natural language processing model that can extract protein features, but is not sophisticated enough when processing complex biomedical data.

[0357] 3.2.2 The clustering effect of the GraphSAGE algorithm is not as good as the Fisher optimal segmentation method of the experimental group:

[0358] GraphSAGE updates node embedding by aggregating the features of adjacent nodes, but ignores global information; while Fisher's optimal partitioning method can better consider inter-class differences and intra-class similarities when performing LDA problems, achieving more accurate clustering.

[0359] 3.2.3 Poor embedding representation leads to convolutional neural network execution deviation:

[0360] Since the control group b performed poorly in the feature extraction and clustering stages, its subsequent convolutional neural network will naturally produce execution deviations after receiving inaccurate input information, affecting the accuracy of the final prediction results.

[0361] (IV) Conclusion

[0362] This example verifies the superiority of the ResNet-GNN model in processing complex and irregular biomedical data by comparing the performance of the experimental group (ResNet-GNN), control group a (CNN) and control group b (SGNN) in predicting the degree of lens opacity in Rncat mice. ResNet-GNN achieves more efficient feature extraction and structural relationship capture by combining multi-scale residual networks and graph neural networks, thereby improving the accuracy of prediction.

[0363] All the above embodiments only express the implementation methods of the relevant practical applications of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

[0364] For those skilled in the art, it can be further appreciated that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0365] At the same time, those skilled in the art can understand that all or part of the processes in all the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

Claims

1. A method for predicting the degree of lens opacity in Rncat mice, characterized in that: The method includes collecting the independent variable X=[M, A, μ] of the target Rncat mouse, wherein M is the mating mode of the Rncat mouse, A is the age of the Rncat mouse, and μ is the G489A mutation degree of the Crygs gene of the Rncat mouse, and then performing the following steps: S1, establish node n for the independent variable X i , extract each node n through the multi-scale residual network ResNet i The sample characteristics P i ; S2, the sample feature P i Transformed into a structured graph structure G, the nodes n are aggregated in sequence using an ordered clustering algorithm i , to update the node n i The embedding vector e vi ; S3, the node n i Embedding vector after Tth iteration Transformed into a matrix M as an input of a convolutional neural network layer, the convolutional neural network layer including a Conv layer for feature extraction to form an output result r1, and a Linear layer for linearly combining the output result r1 to form an output result r2; S4, converting the output result r2 into a continuous value of a dependent variable Y which is the degree of lens opacity.

2. The method for predicting the degree of lens opacity according to claim 1, characterized in that: In S1, the multi-scale residual network ResNet includes a corresponding node n i A residual block comprising a plurality of convolutional layers and a shortcut connection layer; The residual block reads the corresponding independent variable x and outputs each node n through residual learning. i The sample characteristics P i : in, is the output of the lth residual block, is the input of the lth residual block; F(·) represents the residual function, which contains two convolutional layers. Represents the weight parameters of these convolutional layers; In the shortcut connection layer, the node n is calculated i and the output of the lth residual block And when size matching or dimension adjustment is required, perform the following operations: in, is a linear transformation operation.

3. The method for predicting the degree of lens opacity according to claim 2, characterized in that: In S1, the extraction method is: Among them, H(·) represents the ResNet network function, L represents the total number of residual blocks, Represents the weight parameters of all residual blocks.

4. The method for predicting the degree of lens opacity according to claim 1, characterized in that: In S2, the conversion method includes: converting all the sample features P i The sample feature set Construct a graph structure G = (V, E); The collection of nodes in the graph Each of the nodes n i Corresponding to a sample feature P i , E is the edge set; In S2, the ordered clustering algorithm is Fisher's optimal partitioning method, which aggregates each adjacent node n in sequence and iteratively. i , each iteration updates the node’s embedding vector e vi , until the embedding vector is formed at the final iteration round T 5. The method for predicting the degree of lens opacity according to claim 4, characterized in that: In S2, let the current iteration round be t, for node n i , whose neighbor node set is N(i), uses the Fisher optimal partitioning method to divide n i The aggregation is performed with its neighbor nodes by: Aggregate(·) represents an aggregation function, which receives node n i and the embedding vectors of its neighboring nodes As input, and output the updated embedding vector W agg is the weight parameter in the aggregation process; When the iteration reaches the predetermined number of times or when the Fisher criterion function converges, the updated graph structure G′=(V′,E′) is obtained, where is the updated collection, each node n i ′ corresponds to an embedding vector Where T represents the final iteration round; E′ is the updated edge set.

6. The method for predicting the degree of lens opacity according to claim 1, 2 or 4, characterized in that: In S3, the Conv layer applies two convolution layers with a convolution kernel of 3*3 to perform feature extraction on the matrix M, and then applies two maximum pooling layers with a kernel of 2*2 to perform downsampling to form the output result r1; The Linear layer is based on the output result r1 to three fully connected layers, using ReLU as the activation function to form the output result r2.

7. The method for predicting the degree of lens opacity according to claim 6, characterized in that: The execution process of the Conv layer includes: The first convolutional layer: M1 = Conv(M, W1, b1, ks = (3, 3)); Where W1 and b1 are the weights and biases of the first convolutional layer, Conv(·) represents the convolution operation, ks is the size of the convolution kernel, and M1 is the output of the first convolutional layer. The first max pooling layer: Among them, MaxPooling(·) represents the maximum pooling operation; stride = (2,2) means that the stride of the convolution or pooling operation in the horizontal and vertical directions is 2; Represents the output of the first maximum pooling layer; The second convolutional layer: Among them, W2 and b2 are the weights and biases of the second convolutional layer; M2 is the output of the second convolutional layer; The second maximum pooling layer: r1 = MaxPooling(M2, ks = (2, 2), stride = (2, 2)).

8. The method for predicting the degree of lens opacity according to claim 6, characterized in that: The execution process of the Linear layer includes: The first fully connected layer: Among them, W f1 and b f1 are the weights and biases of the first fully connected layer, and Linear(·) represents the fully connected operation; is the output of the first fully connected layer; The second fully connected layer: Among them, W f2 and b f2 are the weights and biases of the second fully connected layer, is the output of the second fully connected layer; The third fully connected layer: Among them, W f3 and b f3 are the weights and biases of the third fully connected layer.

9. The method for predicting the degree of lens opacity according to claim 1, 2 or 4, characterized in that: In S4, the conversion method is to apply the Sigmoid function to the output result r2: Wherein, e is the base of natural logarithm; the continuous value is a probability or a score.

10. A prediction system for the degree of lens opacity in Rncat mice, characterized in that: The system includes a processor and a memory connected to the processor, wherein program instructions are stored in the memory, and when the program instructions are executed by the processor, the processor executes the method for predicting the degree of lens opacity as described in any one of claims 1-9.

Images