Training methods and devices for deep learning-based analysis models of non-insulin-dependent diabetes mellitus

Abstract

This invention relates to the field of bioinformatics and discloses a training method and apparatus for a deep learning-based analysis model of non-insulin-dependent diabetes mellitus (NIDM). The method involves acquiring target proteome data, which has binary labels characterizing the correlation between the target proteome data and NIDM; standardizing the features of the target proteome data; and performing data balancing using an oversampling algorithm to ensure that the number of target proteome data corresponding to the binary labels is equal. Based on the balanced target proteome data, the deep learning-based NIDM analysis model is trained to obtain the analysis model. By using the balanced, high-quality data to train the deep neural network model, the nonlinear correlations between high-dimensional protein features are fully explored, achieving accurate identification and analysis of the correlation between proteins and NIDM.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a training method and apparatus for a deep learning-based analysis model of non-insulin-dependent diabetes mellitus. Background Technology

[0003] In the field of bioinformatics, proteomic data, due to its comprehensive ability to reflect the expression levels, post-translational modifications, and functional states of thousands of proteins in organisms, has become a crucial data source for disease mechanism research and biomarker discovery. Compared to genomic or transcriptomic data, proteins are the direct executors of cellular functions, and their expression profile changes more closely reflect the physiological and pathological processes of diseases. This is especially true in metabolic diseases such as non-insulin-dependent diabetes mellitus (NIDDM), where abnormal expression of specific proteins often directly participates in key pathological processes such as insulin signaling pathway dysregulation, chronic inflammation, and oxidative stress. Therefore, correlation analysis between protein expression information and disease phenotypes not only helps to reveal the molecular mechanisms of diseases but also provides potential protein biomarkers with high biological interpretability for disease analysis. However, despite the rich biological information contained in proteomic data, accurately and robustly identifying protein biomarkers that are truly relevant to disease specificity remains a core challenge. Summary of the Invention

[0004] The purpose of this invention is to provide at least one training method for a deep learning-based analysis model of non-insulin-dependent diabetes mellitus, which can at least solve the technical problem of how to accurately and robustly identify protein biomarkers that are truly related to the disease, and can at least achieve accurate and robust identification of protein biomarkers related to non-insulin-dependent diabetes mellitus.

[0005] To address the aforementioned technical problems, at least one embodiment of this application provides a training method for a deep learning-based analysis model of non-insulin-dependent diabetes mellitus, comprising: acquiring target proteome data, wherein the target proteome data is the proteome data of the object to be analyzed, the target proteome data having binary labels, the binary labels being used to characterize the degree of association between the target proteome data and non-insulin-dependent diabetes mellitus; performing feature standardization on the target proteome data, and using an oversampling algorithm to perform data balancing processing on the target proteome data, so that the number of target proteome data corresponding to the binary labels is equal; and training the deep learning-based analysis model of non-insulin-dependent diabetes mellitus based on the data-balanced target proteome data to obtain the trained analysis model.

[0006] This approach effectively eliminates differences in the expression dimensions of different proteins by standardizing the features of the target proteome data, thereby improving the stability and convergence speed of model training. Secondly, an oversampling algorithm is used to balance imbalanced data, making the number of training samples in different classes more consistent, significantly alleviating the model's bias towards the majority class and improving the analytical model's ability to classify and identify proteins that are related to or unrelated to non-insulin-dependent diabetes mellitus (NIDM). Based on this, the balancing high-quality data is used to train a deep learning-based NIDM analysis model, fully exploring the nonlinear correlations between high-dimensional protein features, resulting in an analytical model that can accurately identify and analyze the correlation between proteins and NIDM.

[0007] In some examples, the data balancing process of the target proteome data using an oversampling algorithm includes: obtaining the number of target proteome data corresponding to each classification label, determining the minority class label and the minority proteome data corresponding to the minority class label; for each minority proteome data, obtaining k neighboring minority proteome data corresponding to the minority proteome data, where k≥1; randomly selecting a reference proteome data from the k neighboring minority proteome data, and generating balanced proteome data corresponding to the minority class label based on the minority proteome data and its corresponding reference proteome data; the difference between the total number of balanced proteome data and the minority proteome data group and the number of target proteome data corresponding to non-minority class labels is less than or equal to a preset difference.

[0008] In some examples, generating balanced proteomic data corresponding to the minority class label based on the minority class proteomic data and its corresponding reference proteomic data includes: obtaining the minority class proteomic features and reference proteomic features corresponding to the minority class proteomic data and its corresponding reference proteomic data, respectively; performing interpolation calculation based on a finite range of random numbers based on the minority class proteomic features and the reference proteomic features to obtain balanced proteomic features; the finite range of random numbers is a random number in [0,1]; and converting the balanced proteomic features into the balanced proteomic data.

[0009] In some examples, the step of performing interpolation calculations based on a finite range of random numbers to obtain balanced proteomic features, according to the minority proteomic features and the reference proteomic features, includes: determining the deviation distance between the minority proteomic features and the reference proteomic features based on the minority proteomic features and the reference proteomic features; randomly synthesizing the deviation distance based on a finite range of random numbers to obtain a random synthesis factor; and determining the balanced proteomic features based on the random synthesis factor and the minority proteomic features.

[0010] In some examples, training a deep learning-based non-insulin-dependent diabetes analysis model based on the target proteome data after data balancing includes setting a training patience value of 5.

[0011] In some examples, this also includes: performing reliability verification on the trained analytical model.

[0012] In some examples, the reliability verification of the trained analytical model includes: calculating at least one evaluation metric of the analytical model; if all at least one evaluation metric meets the reliability verification conditions, then the analytical model is determined to meet the reliability verification criteria; and / or obtaining the training key protein biomarkers corresponding to the analytical model through the analytical model; comparing the training key protein biomarkers with reference key protein biomarkers; if the training key protein biomarkers are consistent with the reference key protein biomarkers, then the analytical model is determined to meet the reliability verification criteria.

[0013] At least one embodiment of this application also provides a training device for a deep learning-based non-insulin-dependent diabetes mellitus analysis model, comprising: an acquisition unit for acquiring target proteome data, wherein the target proteome data is proteome data of the object to be analyzed, the target proteome data having binary labels, the binary labels being used to characterize the degree of association between the target proteome data and non-insulin-dependent diabetes mellitus; a processing unit for performing feature standardization on the target proteome data and using an oversampling algorithm to perform data balancing processing on the target proteome data, so that the number of target proteome data corresponding to the binary labels is equal; and a training unit for training the deep learning-based non-insulin-dependent diabetes mellitus analysis model based on the data-balanced target proteome data, to obtain the trained analysis model.

[0014] At least one embodiment of this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the training method of the deep learning-based non-insulin-dependent diabetes analysis model described above.

[0015] At least one embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described training method for a deep learning-based analysis model of non-insulin-dependent diabetes mellitus. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0017] Figure 1 This is a flowchart illustrating a training method for a deep learning-based analysis model for non-insulin-dependent diabetes mellitus provided in one embodiment of this application. Figure 2 This is a schematic diagram illustrating the relationship between epoch and loss value, provided in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of the analysis model obtained through training, provided in another embodiment of this application; Figure 4 This is a schematic diagram of the structure of a training device for a deep learning-based non-insulin-dependent diabetes analysis model provided in another embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0019] It should be noted that the acquisition or use of data in the embodiments of this application requires the user's consent. The relevant data can only be obtained after the user's authorization, and the acquisition or use of the data complies with the provisions of relevant laws and regulations.

[0020] To facilitate understanding of the embodiments of this application, the training method of the deep learning-based non-insulin-dependent diabetes analysis model will be introduced first.

[0021] In the field of bioinformatics, proteomic data, due to its comprehensive ability to reflect the expression levels, post-translational modifications, and functional states of thousands of proteins in organisms, has become a crucial data source for disease mechanism research and biomarker discovery. Compared to genomic or transcriptomic data, proteins are the direct executors of cellular functions, and their expression profile changes more closely reflect the physiological and pathological processes of diseases. This is especially true in metabolic diseases such as non-insulin-dependent diabetes mellitus (NIDDM), where abnormal expression of specific proteins often directly participates in key pathological processes such as insulin signaling pathway dysregulation, chronic inflammation, and oxidative stress. Therefore, correlation analysis between protein expression information and disease phenotypes not only helps to reveal the molecular mechanisms of diseases but also provides potential protein biomarkers with high biological interpretability for disease analysis. However, despite the rich biological information contained in proteomic data, accurately and robustly identifying protein biomarkers that are truly relevant to disease specificity remains a core challenge.

[0022] To address the aforementioned technical challenge of accurately and robustly identifying protein biomarkers truly relevant to disease specificity, this invention proposes a training method for a deep learning-based analysis model of non-insulin-dependent diabetes mellitus. The implementation details of this deep learning-based training method for non-insulin-dependent diabetes mellitus are described below. These details are provided for ease of understanding and are not essential for implementing this solution.

[0023] Example 1: The training method for the deep learning-based non-insulin-dependent diabetes mellitus analysis model in this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. The specific process can be as follows: Figure 1 As shown, it includes: Step 110: Obtain target proteome data. The target proteome data is the proteome data of the object to be analyzed. The target proteome data has binary labels, which are used to characterize the degree of association between the target proteome data and non-insulin-dependent diabetes mellitus.

[0024] The object to be analyzed can be a human, isolated tissue, or a deceased human or animal body. Target proteome data consists of various protein features; specifically, a single target proteome dataset includes 2923 protein features. These protein features are quantifiable indicators characterizing the state or attributes of proteins within the corresponding object to be analyzed. Each target proteome dataset has a corresponding binary label, organized and stored in CSV (Comma-Separated Values) format. The binary labels are divided into relevant and irrelevant labels, which can be represented numerically. For example, a relevant label of 1 indicates that the target proteome data is related to non-insulin-dependent diabetes mellitus (NIDM), while an irrelevant label of 0 indicates that the target proteome data is not related to NIDM. A larger value for the binary label indicates a stronger correlation between the target proteome data and NIDM; conversely, a smaller value indicates a weaker correlation.

[0025] Specifically, there are multiple target proteome datasets. One target proteome dataset serves as a sample dataset for training the analysis model. When training the analysis model, multiple sample datasets are required, meaning there are multiple target proteome datasets.

[0026] Specifically, to ensure the quality of the target proteome data, for each target proteome dataset, missing values ​​are detected. If missing values ​​are found, they are imputed using the median. Specifically, it is determined whether any protein feature is missing in the target proteome dataset; if so, the median of the protein features corresponding to the missing feature across all target proteome datasets is used for imputation.

[0027] Step 120: Perform feature standardization on the target proteome data and use an oversampling algorithm to balance the target proteome data so that the number of target proteome data corresponding to the binary classification labels is equal.

[0028] Feature standardization refers to mathematically transforming the protein features in the target proteome data to unify the data scale of each protein feature, eliminate dimensional differences, and improve model performance. Feature standardization can use either Min-Max normalization or Z-score standardization, or other methods to eliminate dimensional differences. Oversampling algorithms involve expanding the minority class target proteome data to a quantity comparable to that of the non-minority class target proteome data. This is used to alleviate class imbalance and prevent the trained analytical model from having poor prediction accuracy due to excessive bias in the number of sample data.

[0029] For example, if the number of target proteome data corresponding to relevant tags is greater than the number of target proteome data corresponding to irrelevant tags, then the target proteome data corresponding to irrelevant tags is minority class target proteome data, and the target proteome data corresponding to relevant tags is non-minority class target proteome data; if the number of target proteome data corresponding to irrelevant tags is greater than the number of target proteome data corresponding to relevant tags, then the target proteome data corresponding to relevant tags is minority class target proteome data, and the target proteome data corresponding to irrelevant tags is non-minority class target proteome data.

[0030] Specifically, in step 120 above, an oversampling algorithm is used to perform data balancing on the target proteome data, including: obtaining the number of target proteome data corresponding to each classification label, determining the minority class label and the minority proteome data corresponding to the minority class label; for each minority proteome data, obtaining k neighboring minority proteome data corresponding to the minority proteome data, where k≥1; randomly selecting a reference proteome data from the k neighboring minority proteome data, and generating balanced proteome data corresponding to the minority class label based on the minority proteome data and its corresponding reference proteome data; the difference between the total number of balanced proteome data and minority proteome data and the number of target proteome data corresponding to non-minority class labels is less than or equal to a preset difference.

[0031] In this context, minority class labels refer to the classification labels corresponding to a smaller number of target proteome data points. If the number of target proteome data points corresponding to irrelevant labels is small, then the minority class label is considered irrelevant; otherwise, it is considered relevant. Minority class proteome data refers to a smaller number of target proteome data points. Reference proteome data is any one of the k nearest minority class proteome data points. Balanced proteome data refers to new target proteome data points generated by processing the minority class proteome data and its corresponding reference proteome data using an oversampling algorithm to supplement the number of minority class proteome data points. The preset difference value can be set according to actual needs and can be 0, 1, 10, 100, or other values. It should be noted that in order to ensure that the number of sample data points with relevant labels is comparable to the number of sample data points with irrelevant labels, thus alleviating the class imbalance problem, the preset difference value should be set as small as possible. The k nearest minority class proteome data points refer to the k nearest target proteome data points in the multidimensional space (2923-dimensional space) composed of multiple protein features (2923) contained in the target proteome data. Specifically, the distance between each target proteome data can be calculated based on Euclidean distance, Manhattan distance, and cosine similarity.

[0032] For example, a first number of target proteome data corresponding to relevant tags is obtained, and a second number of target proteome data corresponding to irrelevant tags is determined. If the first number is greater than the second number, then minority class tags are determined to be irrelevant tags, and non-minority class tags are determined to be relevant tags; otherwise, minority class tags are determined to be relevant tags, and non-minority class tags are determined to be irrelevant tags.

[0033] For example, using the target proteome data of 53,013 objects to be analyzed as sample data, there are a total of 53,013 target proteome data points. Of these, 90.0% of the target proteome data points have irrelevant labels, and 10.0% have relevant labels. To balance the number of target proteome data points corresponding to each label, an oversampling algorithm is used to balance the data for the relevant labels, resulting in 38,186 target proteome data points for each label after balancing. This results in a total of 76,372 sample data points (target proteome data). 80% of these 76,372 sample data points are allocated to the training set (61,097 samples), and 20% are allocated to the test set (15,275 samples).

[0034] Furthermore, the aforementioned process of generating balanced proteomic data corresponding to minority class labels based on minority class proteomic data and their corresponding reference proteomic data includes: obtaining minority class proteomic features and reference proteomic features based on minority class proteomic data and their corresponding reference proteomic data, respectively; performing interpolation calculations based on a finite range of random numbers based on minority class proteomic features and reference proteomic features to obtain balanced proteomic features; the finite range of random numbers is a random number in [0,1]; and converting the balanced proteomic features into balanced proteomic data.

[0035] Both the minority proteome features and the reference proteome features are feature vectors with a dimension of 2923, indicating that both the minority proteome features and the reference proteome features include multiple protein features, each containing 2923 protein features.

[0036] For example, balanced proteome features are obtained by interpolation based on a finite range of random numbers, using minority proteome features and reference proteome features, as shown in the following formula: = + ( - ) in, Indicates the characteristics of a balanced proteome. It represents the characteristics of a minority of proteomes; λ is a random number between [0, 1], that is, a random number within a finite range; This indicates the characteristics of the reference proteome.

[0037] Furthermore, the aforementioned calculation of interpolation based on a finite range of random numbers to obtain balanced proteome features, based on minority proteome features and reference proteome features, includes: determining the deviation distance between minority proteome features and reference proteome features based on minority proteome features and reference proteome features; randomly synthesizing the deviation distance based on a finite range of random numbers to obtain a random synthesis factor; and determining the balanced proteome features based on the random synthesis factor and minority proteome features.

[0038] For example, the deviation distance between minority proteome features and reference proteome features can be determined based on the following formula: = -

[0039] in, Indicates the deviation distance. Indicates characteristics of a minority of proteomes; This indicates the characteristics of the reference proteome.

[0040] For example, the random synthesis factor is obtained by randomly synthesizing the deviation distance based on a finite range of random numbers, as shown in the following formula: =

[0041] in, λ represents the deviation distance, and is a random number between [0, 1], that is, a random number within a finite range; This represents a randomly synthesized factor.

[0042] For example, the characteristics of the equilibrium proteome can be determined based on random synthesis factors and minority proteome features, as shown in the following formula: =

[0043] in, Indicates the characteristics of a balanced proteome. Indicates characteristics of a minority of proteomes; This represents a randomly synthesized factor.

[0044] Step 130: Based on the target proteome data after data balancing, train the non-insulin-dependent diabetes analysis model constructed based on deep learning to obtain the trained analysis model.

[0045] The deep learning-based model for analyzing non-insulin-dependent diabetes mellitus consists of three layers: an input layer, a hidden layer, and an output layer. The hidden layer comprises three fully connected layers with 128, 64, and 32 nodes respectively. Each fully connected layer performs batch normalization and ReLU activation on the input data, and the dropout rate for each layer is 30% to prevent overfitting. The output layer has two types of outputs: relevant labels and irrelevant labels, and it uses softmax activation. The model can be used to analyze proteins in the input data that are related to or irrelevant to non-insulin-dependent diabetes mellitus.

[0046] Specifically, the connection weights from the input layer to the first fully connected layer are protein weights in the analysis model. These weights represent the strength of the influence of protein features corresponding to each protein in the input target proteome data on hidden neurons. The larger the protein weight, the stronger the influence of changes in the protein features on subsequent activation. In other words, by analyzing the average absolute value of the weights in the first layer of the non-insulin-dependent diabetes analysis model built based on deep learning, the protein features corresponding to key proteins can be efficiently identified. Combined with optional enhancement methods (such as averaging through multiple training iterations or permuting importance), potential protein biomarkers can be robustly mined, improving the interpretability of the analysis model.

[0047] Among them, the key protein is one that is associated with non-insulin-dependent diabetes mellitus.

[0048] Furthermore, to comprehensively evaluate the contribution of protein features to the analysis model, this scheme provides at least one of the following protein weighting analysis methods: perturbation importance analysis based on performance perturbation, game theory-based SHAP value analysis, gradient basis method suitable for deep nonlinear networks, and comprehensive importance scoring mechanism. Specifically, perturbation importance analysis quantifies the impact of each protein feature's value on the model's predictive performance (e.g., area under the ROC curve, AUC) by randomly shuffling the value of each protein feature. If feature perturbation leads to a significant performance decrease, it indicates a high contribution to the model's prediction. SHAP value analysis, based on game theory, calculates the marginal contribution of each protein feature to the model's prediction, considering the interactions between protein features. It is used to provide global importance (average contribution) and local importance (sample-level contribution), for example, revealing the association between GDF15 and metabolic pathways. The gradient basis method quantifies the influence of each protein in the entire neural network by analyzing the gradient contribution of protein features to the model's output. Path integral methods are used to accumulate gradients from baseline values ​​to actual values, capturing nonlinear relationships and deep network effects to improve the accuracy of protein feature importance assessment. A comprehensive importance scoring mechanism is used to fuse the results of multiple methods (such as perturbation importance analysis based on performance perturbations, SHAP value analysis based on game theory, and gradient basis methods applicable to deep nonlinear networks) to generate a comprehensive importance score. By using weighted averaging or ranking fusion, the bias of individual methods is reduced to ensure the accuracy of the determined protein weights.

[0049] For example, as shown in Table 1, the ten key proteins are identified based on the analysis model, and the protein weights of the ten key proteins identified when the analysis model was trained are also shown.

[0050] Table 1. Ten Key Proteins

[0051] Furthermore, the method also includes visualizing the protein weights and training history information of the trained analysis model to display the protein weights and training history information.

[0052] Specifically, in step 130 above, the non-insulin-dependent diabetes analysis model constructed based on deep learning is trained based on the target proteome data after data balancing, including setting the training patience value to 5.

[0053] The analysis model is used for binary classification to identify proteins that are related to or not related to non-insulin-dependent diabetes. Its goal is to optimize the class boundary, so the analysis model usually converges quickly. Setting a small patience value of 5 is suitable for tasks with fast convergence, preventing overfitting to synthetic samples and reducing unnecessary training time.

[0054] Therefore, when training a deep learning-based non-insulin-dependent diabetes mellitus analysis model based on balanced target proteome data, a patience value of 5 is set for the early stopping mechanism. That is, if the validation loss does not improve within five consecutive training cycles, training is automatically terminated, and the model weights with the best validation performance are retained. This strategy effectively prevents model overfitting and improves its stability and generalization ability in non-insulin-dependent diabetes mellitus prediction tasks.

[0055] Specifically, when training the non-insulin-dependent diabetes analysis model built based on deep learning, the training batch size is set to 16 and the maximum epoch (training rounds) is set to 200.

[0056] For example, see Figure 2 When the training epoch is between 15 and 25, the training loss and validation loss of the analysis model change relatively smoothly. At this point, the analysis model has basically converged. Therefore, setting the maximum number of training epochs to 200 provides enough training epochs for the analysis model to converge.

[0057] Therefore, the batch size is set to 16 to reduce memory usage while ensuring gradient estimation stability and adapting to high-dimensional proteomics data; the maximum number of training rounds is set to 200 to provide sufficient learning opportunities for the analysis model, and dynamic termination is achieved by combining an early stopping mechanism.

[0058] Specifically, when training the non-insulin-dependent diabetes analysis model built based on deep learning, the Adam optimizer was also used with an initial learning rate of 0.001, and end-to-end optimization was performed in combination with the class-weighted cross-entropy loss function to improve the model's classification performance under imbalanced data.

[0059] Therefore, a learning rate of 0.001 can ensure convergence speed while avoiding training oscillations. The weighted cross-entropy loss function determines the weight of the target proteome data for each classification label based on the ratio of the total number of target proteome data to the number of target proteome data corresponding to each classification label. This balances the imbalance in the number of target proteome data for each classification label, preventing the trained analysis model from favoring data with a large number of samples, which would lead to lower classification accuracy.

[0060] In some cases, this also includes: performing reliability verification on the trained analytical model.

[0061] Specifically, the reliability verification of the trained analytical model includes: calculating at least one evaluation index of the analytical model; if at least one evaluation index meets the reliability verification conditions, the analytical model is determined to meet the reliability verification; and / or obtaining the training key protein biomarkers corresponding to the analytical model through the analytical model; comparing the training key protein biomarkers with the reference key protein biomarkers; if the training key protein biomarkers are consistent with the reference key protein biomarkers, the analytical model is determined to meet the reliability verification.

[0062] Specifically, the evaluation metrics can be computational accuracy, AUC (Area Under the ROC Curve), and loss value. If the computational accuracy exceeds the preset computational accuracy, the AUC exceeds the preset AUC, and the loss value is lower than the preset loss value, then the reliability verification is considered to be satisfied.

[0063] AUC reflects the classification ability of the analytical model. Accuracy reflects the classification accuracy, and loss reflects the prediction error. Lower values ​​indicate a smaller difference between the model's output probability distribution and the true labels, resulting in a better fit. Preset accuracy can be set to any value above 50%, AUC to any value above 0.5, and loss to any value below 1. Higher preset accuracy and AUC place higher demands on the analytical model and result in better model performance. Conversely, lower preset loss also places higher demands on the analytical model and results in better model performance.

[0064] For example, in a test set of 10,603 samples, the analytical model achieved an accuracy of 88.61%, an AUC of 0.87.51, and a loss value of 0.8652. These experimental data demonstrate that the analytical model exhibits high accuracy and classification ability.

[0065] For example, to gain a clearer understanding of this solution, please refer to [link / reference]. Figure 3 The system can input 53,013 target proteome datasets into the data input module. These 53,013 datasets serve as the raw data. After median imputation, standardization, and data balancing by the preprocessing module, 15,275 preprocessed target proteome datasets are obtained. These 15,275 datasets are then used as the preprocessed sample data and input into a deep learning-based non-insulin-dependent diabetes mellitus analysis model for training. Specifically, during model training, the training module employs Adam optimization, crossover loss, and early stopping to update the model parameters. The evaluation module assesses whether the updated model can be used as an analysis model. If so, the analysis model identifies key proteins and outputs the protein weights corresponding to the key proteins obtained during model training. The key proteins and their corresponding weights are then visualized.

[0066] In summary, this approach acquires target proteome data, which is the proteome data of the object to be analyzed. The target proteome data has binary labels, which characterize the association between the target proteome data and non-insulin-dependent diabetes mellitus (NIDM). The target proteome data undergoes feature standardization, and an oversampling algorithm is used to balance the data, ensuring that the number of target proteome data points corresponding to the binary labels is equal. Based on the balanced target proteome data, a deep learning-based NIDM analysis model is trained to obtain the trained analysis model. By standardizing the features of the target proteome data, the differences in expression dimensions of different proteins are effectively eliminated, improving the stability and convergence speed of model training. Secondly, an oversampling algorithm is used to balance imbalanced data, making the number of training samples in different classes more consistent, significantly alleviating the model's bias towards the majority class and improving the model's ability to classify and identify proteins related to or unrelated to non-insulin-dependent diabetes mellitus (NIDM). Based on this, a NIDM analysis model constructed using deep learning is trained on the balanced, high-quality data. This deep learning model fully explores the nonlinear correlations between high-dimensional protein features, resulting in an analytical model that can accurately identify the correlation between proteins and NIDM. Experimental results show that the model achieves an accuracy of 88.61% and an AUC of 0.8751 on an independent test set containing 10,603 samples, demonstrating stronger classification capabilities.

[0067] Furthermore, this analytical model for identifying the association between key proteins and non-insulin-dependent diabetes mellitus has high scalability and flexibility, and can be used in applications that identify the association between key proteins and other disease-specific factors.

[0068] Example 2: Another embodiment of this application relates to a training device for a deep learning-based non-insulin-dependent diabetes mellitus analysis model. The implementation details of this deep learning-based non-insulin-dependent diabetes mellitus analysis model training device are described below. The following implementation details are provided for ease of understanding and are not essential for implementing this solution. A schematic diagram of the deep learning-based non-insulin-dependent diabetes mellitus analysis model training device 40 in this embodiment can be seen as follows: Figure 4 As shown, it includes an acquisition unit 401, a processing unit 402, and a training unit 403.

[0069] Acquisition unit 401 is used to acquire target proteome data, wherein the target proteome data is the proteome data of the object to be analyzed, and the target proteome data has a binary label, wherein the binary label is used to characterize the degree of association between the target proteome data and non-insulin-dependent diabetes mellitus.

[0070] The processing unit 402 is used to perform feature standardization on the target proteome data and to perform data balancing processing on the target proteome data using an oversampling algorithm, so that the number of target proteome data corresponding to the binary classification labels is equal.

[0071] Training unit 403 is used to train a non-insulin-dependent diabetes analysis model constructed based on deep learning based on the target proteome data after data balancing, so as to obtain a trained analysis model.

[0072] In some examples, when processing unit 402 performs data balancing processing on the target proteome data using the oversampling algorithm, it specifically performs the following steps: obtaining the number of target proteome data corresponding to each classification label, determining the minority class label and the minority proteome data corresponding to the minority class label; for each minority proteome data, obtaining k neighboring minority proteome data corresponding to the minority proteome data, where k≥1; randomly selecting a reference proteome data from the k neighboring minority proteome data, and generating balanced proteome data corresponding to the minority class label based on the minority proteome data and its corresponding reference proteome data; the difference between the total number of balanced proteome data and the total number of minority proteome data and the number of target proteome data corresponding to non-minority class labels is less than or equal to a preset difference.

[0073] In some examples, when processing unit 402 generates balanced proteome data corresponding to the minority class label based on the minority class proteome data and its corresponding reference proteome data, it specifically performs the following steps: acquiring the minority class proteome features and reference proteome features corresponding to the minority class proteome data and its corresponding reference proteome data, respectively; performing interpolation calculation based on a finite range of random numbers based on the minority class proteome features and the reference proteome features to obtain balanced proteome features; the finite range of random numbers is a random number in [0,1]; and converting the balanced proteome features into the balanced proteome data.

[0074] In some examples, when processing unit 402 performs interpolation calculations based on a finite range of random numbers to obtain balanced proteome features based on the minority proteome features and the reference proteome features, it specifically performs the following steps: determining the deviation distance between the minority proteome features and the reference proteome features based on the minority proteome features and the reference proteome features; randomly synthesizing the deviation distance based on a finite range of random numbers to obtain a random synthesis factor; and determining the balanced proteome features based on the random synthesis factor and the minority proteome features.

[0075] In some examples, when training unit 403 is used to train a deep learning-based non-insulin-dependent diabetes analysis model based on the target proteome data after the data balancing process, it specifically sets the training patience value to 5.

[0076] In some examples, the training unit 403 is also used to: perform reliability verification on the trained analysis model.

[0077] In some examples, when the training unit 403 is used to perform reliability verification on the trained analysis model, it is specifically used to: calculate at least one evaluation index of the analysis model; if all at least one evaluation index meets the reliability verification conditions, then the analysis model is determined to meet the reliability verification; and / or obtain the training key protein biomarkers corresponding to the analysis model through the analysis model; compare the training key protein biomarkers with reference key protein biomarkers; if the training key protein biomarkers are consistent with the reference key protein biomarkers, then the analysis model is determined to meet the reliability verification.

[0078] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.

[0079] Example 3: Another embodiment of this application relates to an electronic device, such as... Figure 5 As shown, it includes: at least one processor 901; and a memory 902 communicatively connected to the at least one processor 901; wherein the memory 902 stores instructions executable by the at least one processor 901, the instructions being executed by the at least one processor 901 to enable the at least one processor 901 to execute the training method of the deep learning-based non-insulin-dependent diabetes analysis model in the above embodiments.

[0080] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0081] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0082] Example 4: Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0083] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A training method for a deep learning-based analysis model of non-insulin-dependent diabetes mellitus, characterized in that, include: Acquire target proteome data, which is the proteome data of the object to be analyzed. The target proteome data has binary labels, which are used to characterize the degree of association between the target proteome data and non-insulin-dependent diabetes mellitus. The target proteome data is feature-standardized, and an oversampling algorithm is used to balance the target proteome data so that the number of target proteome data corresponding to the binary classification labels is equal. Based on the target proteome data after data balancing, a non-insulin-dependent diabetes analysis model constructed based on deep learning is trained to obtain the trained analysis model.

2. The training method for the deep learning-based non-insulin-dependent diabetes mellitus analysis model according to claim 1, characterized in that, The data balancing process for the target proteome data using an oversampling algorithm includes: The number of target proteome data corresponding to each classification label is obtained respectively, and the minority class labels and the minority class proteome data corresponding to the minority class labels are determined; For each minority proteome data point, obtain the k nearest neighbor minority proteome data points corresponding to the minority proteome data point, where k≥1; A reference proteome is randomly selected from k neighboring minority proteome datasets. Balanced proteome datasets corresponding to the minority proteome label are generated based on the minority proteome dataset and its corresponding reference proteome dataset. The difference between the total number of the balanced proteome datasets and the minority proteome datasets and the number of target proteome datasets corresponding to non-minority proteome labels is less than or equal to a preset difference.

3. The training method for the deep learning-based non-insulin-dependent diabetes mellitus analysis model according to claim 2, characterized in that, The step of generating balanced proteomic data corresponding to the minority class label based on the minority class proteomic data and its corresponding reference proteomic data includes: Obtain the minority proteome features and reference proteome features corresponding to the minority proteome data and its corresponding reference proteome data, respectively. Based on the minority proteome features and the reference proteome features, interpolation calculations based on a finite range of random numbers are performed to obtain balanced proteome features; the finite range of random numbers are random numbers in [0,1]. The balanced proteome features are converted into balanced proteome data.

4. The training method for the deep learning-based non-insulin-dependent diabetes mellitus analysis model according to claim 3, characterized in that, The step of performing interpolation calculations based on a finite range of random numbers to obtain balanced proteome features, according to the minority proteome features and the reference proteome features, includes: Based on the minority proteome features and the reference proteome features, determine the deviation distance between the minority proteome features and the reference proteome features; The deviation distance is randomly synthesized based on a finite range of random numbers to obtain a random synthesis factor; The balanced proteome features are determined based on the random synthesis factor and the minority proteome features.

5. The training method for the deep learning-based non-insulin-dependent diabetes mellitus analysis model according to claim 1, characterized in that, The training of a deep learning-based non-insulin-dependent diabetes analysis model based on the target proteome data after data balancing includes: Set the training patience value to 5.

6. The training method for the deep learning-based non-insulin-dependent diabetes mellitus analysis model according to claim 1, characterized in that, Also includes: The reliability of the trained analysis model is verified.

7. The training method for the deep learning-based non-insulin-dependent diabetes mellitus analysis model according to claim 6, characterized in that, The reliability verification of the trained analysis model includes: Calculate at least one evaluation index of the analysis model; if all at least one evaluation index meets the reliability verification conditions, then the analysis model is determined to meet the reliability verification requirements; and / or The key protein biomarkers for training the analytical model are obtained through the analytical model. The training key protein biomarkers are compared with the reference key protein biomarkers. If the training key protein biomarkers are consistent with the reference key protein biomarkers, then the analytical model is determined to meet the reliability verification.

8. A training device for a deep learning-based analysis model of non-insulin-dependent diabetes mellitus, characterized in that, include: An acquisition unit is used to acquire target proteome data, wherein the target proteome data is the proteome data of the object to be analyzed, and the target proteome data has a binary label, wherein the binary label is used to characterize the degree of association between the target proteome data and non-insulin-dependent diabetes mellitus. The processing unit is used to perform feature standardization on the target proteome data and to perform data balancing on the target proteome data using an oversampling algorithm, so that the number of target proteome data corresponding to the binary classification labels is equal. The training unit is used to train a non-insulin-dependent diabetes analysis model constructed based on deep learning, based on the target proteome data after data balancing, to obtain a trained analysis model.

9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the training method for a deep learning-based non-insulin-dependent diabetes analysis model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method of the deep learning-based non-insulin-dependent diabetes analysis model according to any one of claims 1 to 7.

Images