Multi-omics data and semi-supervised metric learning-based pan cancer classification method
By using a method based on multi-omics data and semi-supervised metric learning, principal component analysis and autoencoder network optimization embedding representation were used to construct an SVM classifier, which solved the problem of insufficient accuracy of multi-omics prediction models and achieved more efficient cancer sample classification.
Patent Information
- Application Number
- CN202510703819.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-29
- Publication Date
- 2025-09-12
AI Technical Summary
Existing multi-omics prediction models have poor prediction accuracy in cancer classification due to insufficient information utilization. In addition, the current method of constructing patient similarity networks cannot dynamically adjust the similarity metric and has limited generalization ability.
A method based on multi-omics data and semi-supervised metric learning was adopted. By obtaining multiple omics data of pan-cancer samples, preprocessing and splicing them, principal component analysis was used to obtain the principal component score matrix, which was input into the autoencoder network for pre-training. Partially labeled data was used to update the encoder network parameters, metric learning was used to optimize the embedding representation, and an SVM multi-class classifier was constructed.
The model's classification performance for pan-cancer patient categories was improved, the model construction cost was reduced, the model's discriminative ability and interpretability were enhanced, and the prediction accuracy and generalization ability for unlabeled samples were improved.
Smart Images
Figure CN120632591A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bioinformatics, and in particular to a pan-cancer classification method based on multi-omics data and semi-supervised metric learning. Background Art
[0002] According to the World Health Organization, cancer has become the world's leading threat to human life and health. Early detection and accurate diagnosis of cancer can help improve patient survival rates, thereby addressing this global health issue.
[0003] Currently, classification based on gene expression data is a common method for predicting patient categories. Most research methods are based on supervised learning classification models. These methods primarily use machine learning and deep learning to establish relationships between genotypes and phenotypes, and have been widely used in the field of bioinformatics, such as MOGONET (Multi-Omics Graph Convolutional Networks) and XOmiVAE (a variational autoencoder (VAE)-based interpretable model). However, these methods often rely on large amounts of labeled data to train the models, and obtaining labels for labeled data requires a significant amount of manpower, time, and resources, which increases the cost of model construction. Therefore, a growing number of researchers are finding that using semi-supervised methods can better utilize the large amount of unlabeled data that exists in reality, capture potential associations between patients in unlabeled data, and do not require large amounts of labeled data.
[0004] In recent years, semi-supervised learning, as an emerging machine learning method, can effectively mine the potential distribution information and data structure in unlabeled data. For example, a recent study introduced a semi-supervised learning method CG3 (Contrastive GCNs with Graph Generation, contrastive graph convolutional networks with graph generation capabilities) based on GCN (Graph Convolutional Networks). By designing a semi-supervised contrast loss, it extracts the potential relationship between data features and graph topology as supplementary supervision information to improve the classification performance of cancer samples. However, although significant progress has been made in the classification of cancer using genomic information, due to the complexity of the potential association information between samples, relying solely on genomic information to predict sample categories cannot provide accurate and comprehensive biological information. However, the above method ignores the local correlation between patients and relies on pseudo-labels generated by the model itself.
[0005] Therefore, a growing number of studies are using association information between patients during training to improve cancer classification performance. Some related studies have shown that the inclusion of patient similarity networks can capture potential association information in gene expression data, thereby improving the performance of cancer sample classification. However, current methods for constructing patient similarity networks cannot dynamically adjust the similarity metric based on different tasks, have limited generalization capabilities, and are unable to automatically learn complex nonlinear relationships in the data. Therefore, they may negatively impact the ability to capture similarity associations between samples, resulting in poor classification accuracy. Summary of the Invention
[0006] In order to solve the technical problem that current multi-omics prediction models have poor prediction accuracy due to insufficient information utilization, the present invention proposes a pan-cancer classification method based on multi-omics data and semi-supervised metric learning, aiming to solve the above problem.
[0007] The pan-cancer classification method based on multi-omics data and semi-supervised metric learning includes the following steps:
[0008] S1. Obtain multiple omics data of pan-cancer samples;
[0009] S2. After preprocessing and concatenating the multi-omics data, principal component analysis is used to obtain the corresponding principal component score matrix;
[0010] S3, inputting the principal component score matrix into an autoencoder network for pre-training to perform genetic encoding on the cancer sample;
[0011] S4. updating the parameters of the autoencoder network using the partially labeled data and optimizing the embedding representation of the autoencoder network using metric learning;
[0012] S5. Input the optimized embedding representation into the constructed SVM multi-class classifier to obtain a prediction result for the unlabeled sample category.
[0013] The beneficial effects of the above technical solution are:
[0014] (1) The present invention preprocesses and concatenates the acquired multi-omics pan-cancer sample data to obtain the corresponding principal component score matrix, then pre-trains the cancer samples through an autoencoder network to perform genetic encoding, uses metric learning to optimize the embedding representation, and constructs an SVM multi-class classifier to obtain prediction results for the unlabeled sample category. This method does not require a large amount of labeled data to train the model, reducing the cost of model construction and achieving better performance in predicting the category of pan-cancer samples than other methods.
[0015] (2) The present invention extracts compact and efficient feature representations through autoencoders, fully utilizing the potential information in unlabeled samples, making the model more discriminative and improving its interpretability and usability in clinical settings;
[0016] (3) The present invention uses metric learning to design two metrics between heterogeneous center points and homogeneous data points, optimizes the similarity metric between patients, enhances the correlation between samples, and can dynamically adjust the similarity metric according to different tasks, thereby improving the generalization ability of the model and automatically learning complex nonlinear relationships in the data. It effectively improves the model's classification performance for pan-cancer patient categories with high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is a flow chart of the overall steps of the present invention;
[0018] Figure 2 This is a flowchart of the steps of the multi-omics data preprocessing, splicing and principal component analysis method of the present invention;
[0019] Figure 3 Schematic diagram of the automatic encoder network pre-training of the present invention;
[0020] Figure 4 is a flow chart of the parameter update and metric learning optimization embedding representation steps of the encoder network of the present invention;
[0021] Figure 5 Schematic diagram of the present invention using SVM to predict the category of unlabeled samples. DETAILED DESCRIPTION
[0022] To further illustrate the technical means and effects employed by the present invention to achieve its intended objectives, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementations, structures, features, and effects of the technical solutions proposed by the present invention. In the following description, references to "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.
[0023] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
[0024] Example 1, as Figure 1 As shown, the process of an embodiment of the pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to the present invention is shown. Figures 2 to 5 Explain the various steps of the present invention and their relationships; Figure 2This is a flowchart of the steps of preprocessing and concatenating multi-omics data and obtaining the corresponding principal component score matrix using the principal component analysis method; Figure 3 is a schematic diagram of pre-training using an autoencoder provided in an embodiment of the present invention; Figure 4 This is a flowchart of the steps of using partially labeled data to update the parameters of the encoder network and using metric learning to optimize the embedding representation of the autoencoder network, provided by an embodiment of the present invention. Figure 5 Schematic diagram of using SVM to predict the category of unlabeled samples in the present invention.
[0025] The pan-cancer classification method based on multi-omics data and semi-supervised metric learning includes the following steps:
[0026] Step S101: Acquire multiple omics data of pan-cancer samples.
[0027] Specifically, an open source pan-cancer dataset is used, which can obtain multi-omics data and phenotypic data of pan-cancer patients. In this embodiment, two types of omics data, RNA-seq expression data and miRNA expression data, are selected from the open source pan-cancer dataset, including samples of 33 cancer types. The miRNA expression data contains 11,020 samples, and the corresponding number of features is 1,881; the RNA-seq expression data contains 11,768 samples, and the corresponding number of features is 60,483. Assume that each sample in the pan-cancer sample dataset has m types of multi-omics data, and each omics data has n v (The value range of v is {1,...,m}) cancer samples, and the feature data contained in each omics dataset are t v ; In this embodiment, m=2, the number of samples contained in the RNA-seq expression omics and miRNA expression omics data is n1=11768, n2=11020, and the number of features contained in the RNA-seq expression omics and miRNA expression omics data is t1=60483, t2=1881.
[0028] Step S102: After preprocessing and concatenating the multi-omics data, principal component analysis is used to obtain the corresponding principal component score matrix.
[0029] Specifically, the main purpose of the principal component analysis method is to simplify the data set by reducing the data dimension, while effectively removing the noise and redundant information in the pan-cancer multi-omics data and retaining the important features in the data set. Figure 2 As shown in the figure, after preprocessing and concatenating the multi-omics data, the principal component analysis method was used to obtain the corresponding principal component score matrix.
[0030] After preprocessing and concatenating the multi-omics data, the principal component analysis method is used to obtain the corresponding principal component score matrix, which is divided into the following steps:
[0031] Step S201 : After pre-processing the multi-omics datasets of pan-cancer samples, common samples are extracted and gene feature information is spliced.
[0032] Specifically, two omics datasets, RNA-seq expression data and miRNA expression data from pan-cancer samples, were preprocessed separately. Genes with zero expression or missing values exceeding 20% were deleted across all samples, and genes targeting the Y chromosome were removed. Samples shared by both omics datasets were then extracted, ultimately identifying 10,614 cancer samples. Their gene signatures were then concatenated and merged, resulting in a feature dimension of 60,342. The resulting omics matrix contained 10,614 samples, with a feature dimension of t = 60,342.
[0033] Step S202 , performing standardization processing on the spliced multi-omics data.
[0034] Specifically, Min-Max normalization is performed on the spliced multi-omics data to make different features have the same scale, avoiding the negative impact of some features with large deviations on classification performance. The formula corresponding to Min-Max normalization is:
[0035]
[0036] Among them, X norm is the normalized data value; X is the original data value of a column in the omics dataset; X max is the maximum value of a feature in the data set; X min It is the minimum value of a feature in the dataset.
[0037] Step S203: Calculate the covariance between all feature pairs and construct a covariance matrix.
[0038] Specifically, the covariance between all feature pairs is calculated. The covariance can measure the correlation between two random variables. The formula for constructing the covariance matrix is:
[0039]
[0040] Where C represents the constructed t×t covariance matrix; n represents the number of cancer samples; X norm T represents the transpose of the normalized matrix; X norm represents the normalized n×t omics data matrix. In this example, n=10614 and t=60342.
[0041] Step S204: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and corresponding eigenvalues. The eigenvalue represents the variance of the principal component; the larger the eigenvalue, the more variance the principal component explains. The eigenvector represents the direction of the principal component and is the basis vector of the data in the new coordinate system.
[0042] Cv i =λ i v i
[0043] Where C represents the constructed t×t covariance matrix; λ i is the eigenvalue, v i is the i-th eigenvector in the principal component direction. In this example, t=60342.
[0044] Step S205: Select the number of principal components according to the variance explanation ratio of each principal component.
[0045] Specifically, selecting the number of principal components k is a key step in PCA, and k is selected based on the variance explained by each principal component. First, calculate each eigenvalue λ i The corresponding formula is:
[0046]
[0047] Where T represents each eigenvalue λ i The sum of λ i is the eigenvalue. Indicates that the λ of each dimension i sum.
[0048] Then calculate the variance explanation ratio of the i-th principal component, the corresponding formula is:
[0049]
[0050] Among them, E i represents the variance explained by the i-th principal component; λ i is the eigenvalue; T represents each eigenvalue λ i The sum of .
[0051] Then, in order to determine the number of principal components, the cumulative variance explained ratio is calculated. The corresponding calculation formula is:
[0052]
[0053] Among them, C k Indicates the cumulative variance explained; E i represents the variance explained by the i-th principal component; Indicates the proportion of variance explained by the cumulative k principal components. In this example, the number of principal components k retained is 2000, corresponding to a cumulative contribution of 0.98.
[0054] Step S206 , selecting principal components according to the number of principal components, and using the selected principal components to project the standardized data to obtain a principal component score matrix.
[0055] Specifically, according to the order of contribution, the eigenvectors with the first k largest eigenvalues are selected. These eigenvectors constitute the principal component eigenvector matrix, and the principal component score matrix is obtained by multiplying it with the standardized data matrix. The calculation formula of the principal component score matrix is as follows:
[0056] M=X norm V
[0057] Where M is the n×k principal component score matrix; X norm is an n×t standardized omics data matrix; V is a t×k principal component eigenvector matrix; n represents the number of cancer samples, k represents the number of selected principal components; t represents the characteristic dimension of the omics data. In this embodiment, n=10614, k=2000, and t=60342.
[0058] Step S103: input the principal component score matrix into an autoencoder network for pre-training to perform gene encoding on the cancer sample.
[0059] Specifically, the principal component score matrix is input into the autoencoder network and mapped to a low-dimensional embedding representation to learn the most effective gene encoding; then the decoder in the autoencoder network reconstructs the original spatial output from the low-dimensional embedding representation. Figure 3 The figure below shows a schematic diagram of the pre-training of an autoencoder. The main idea of the autoencoder is to minimize the loss between input and output. The formula for the loss function is:
[0060]
[0061] Among them, L p represents the loss between the input and output of the autoencoder; M represents the principal component score matrix of the input; represents the reconstructed original space output of the output; Represents the mean square error between the principal component score matrix and the reconstructed original space output.
[0062] Step S104: update the parameters of the encoder network using the partially labeled data, and optimize the embedding representation of the autoencoder network using metric learning.
[0063] Specifically, metric learning optimizes the distance function so that the distance between similar samples in the low-dimensional embedding representation space is closer and the distance between dissimilar samples is farther. This optimization process can more accurately reflect the intrinsic similarity of the data itself. Figure 4 The following steps are used to update the parameters of the encoder network using partial labeled data and optimize the embedding representation of the autoencoder network using metric learning:
[0064] Step S301: Calculate the center point of each cancer category in the low-dimensional embedding representation space based on the low-dimensional embedding representation.
[0065] Specifically, the formula corresponding to the center point of each cancer category in the low-dimensional embedding representation space is:
[0066]
[0067] Among them, c i represents the center point of cancer type i; N i represents the number of samples of cancer type i; f θ () represents the mapping function of the encoder; x j Input cancer sample data; For cancer type i, N i The average value of the low-dimensional embedding representation after mapping the cancer samples;
[0068] Step S302 : Optimizing the similarity representation between samples based on the metrics between the sample points and the center point of each cancer category in the low-dimensional embedding representation space.
[0069] Specifically, the low-dimensional embedding represents the metric between the data points and the center point of each cancer category in the space, and the corresponding formula is:
[0070]
[0071] Where L1 represents the Euclidean distance between the sample point and the center point for cancer type K; d() is the distance function; d(f θ (x j ),c i ) represents the mapping f of the cancer sample θ (x j ) and the cancer center point c i The squared Euclidean distance between Indicates that for type K, N i The mean distance between the sample points and the center point; Represents the mean distance between the sample points and the center point in K cancer types.
[0072] Step S303 : Strengthen the distinguishing representation between cancer categories based on the metrics between the center points of each cancer category in the low-dimensional embedding representation space.
[0073] Specifically, the low-dimensional embedding represents the metric between the center points of each cancer category in the space, and the corresponding formula is:
[0074]
[0075] Where L2 represents the Euclidean distance between two categories in all cancer categories K; Represents the center point of cancer types i1 and i2 for all types K and The minimum squared Euclidean distance between them.
[0076] Step S304: Optimize the final objective function to reversely fine-tune the weight parameters of the encoder.
[0077] Specifically, the formula corresponding to the final optimized objective function is:
[0078]
[0079] Among them, L m Represents the final optimized objective loss function; The weight parameter of the encoder is θ, and θ is adjusted to minimize the loss, that is, L1 is minimized and L2 is maximized; L1 represents the Euclidean distance between the sample point and the center point for the cancer type K; λ is a regularization constant used to balance L1 and L2; L2 represents the Euclidean distance between two categories in all cancer categories K.
[0080] Step S105 , inputting the optimized embedding representation into the constructed SVM multi-class classifier to obtain a prediction result for the unlabeled sample category.
[0081] Specifically, SVM can effectively use the gene expression data of cancer samples for classification. The core idea is to find an optimal hyperplane to separate data points of different categories and maximize the classification boundary to improve the generalization ability of the model. Figure 5 shown.
[0082] The formula corresponding to the hyperplane is:
[0083] w T x i +b=0
[0084] Where w represents the normal vector of the hyperplane; w T represents the transpose of w; x i The vector representing the cancer sample in the embedding; w T xi Represents x i The projection size in the direction of the normal vector w of the hyperplane; b is the intercept of the hyperplane; the formula for optimizing the corresponding parameters of the hyperplane is:
[0085]
[0086] And must meet
[0087] in, represents the distance between the two interval boundaries; C represents the penalty factor; ξ i represents the slack variable; y i Represents the label of the cancer sample, in the range of {-1, +1}.
[0088] When w T x i When +b≥+1, y i =+1; when w T x i When +b≤-1, y i = -1; samples above the interval boundary belong to the positive class, and samples below it belong to the negative class.
[0089] In this example, during training, the SVM classifies samples of a certain category into one category and the remaining samples into another category. An SVM classifier is trained for each cancer category. During classification, the unlabeled samples are classified into the category with the largest classification function value, and the predicted label of each unlabeled cancer sample is finally obtained.
[0090] In Example 2, in order to comprehensively and objectively measure the performance comparison between different classification models, this method and other comparative experiments all use commonly used classification evaluation indicators: Accuracy, Precision, Recall, and F1-score. Their full definitions are as follows:
[0091] Accuracy is the accuracy rate, which is used to measure the proportion of correctly predicted samples in the model to the total samples. The corresponding definition formula is:
[0092]
[0093] Among them, TP represents the number of samples correctly predicted as positive; TN represents the number of samples correctly predicted as negative; FP represents the number of samples incorrectly predicted as positive; FN represents the number of samples incorrectly predicted as negative.
[0094] Precision is the accuracy rate, which is used to measure the proportion of samples predicted to be positive that are actually positive. The corresponding definition formula is as follows:
[0095]
[0096] Among them, TP represents the number of samples correctly predicted as positive; FP represents the number of samples that are incorrectly predicted as negative.
[0097] Recall is the recall rate, which is used to measure the proportion of all true labels that are correctly predicted as positive. The corresponding definition formula is as follows:
[0098]
[0099] Among them, TP represents the number of samples correctly predicted as positive; FN represents the number of samples that incorrectly predict positive samples as negative.
[0100] F1-Score is the F1 score, which represents the harmonic mean of precision and recall. The corresponding definition formula is as follows:
[0101]
[0102] Among them, Recall is the recall rate; Precision is the precision rate.
[0103] To validate the effectiveness of our method (SemiML-Cancer), we compared the performance of SemiML-Cancer with five other multi-omics classification methods, including the supervised learning-based MOGONET and XOmiVAE, as well as the semi-supervised meth-SemiCancer, DeepSSC, and CG3 models. MOGONET is a multi-omics integration method based on graph convolutional networks (GCNs). This method first uses a graph convolutional network (GCN) to learn specific omics features and then uses a view correlation discovery network (VCDN) to explore cross-omics correlations in label space. XOmiVAE is an interpretable deep learning model for cancer diagnosis based on variational autoencoders. It reveals the contribution of each gene and latent dimension to each classification prediction, as well as the correlation between each gene and each latent dimension, and can explain supervised classification and unsupervised clustering. meth-SemiCancer is a semi-supervised cancer subtype classification framework. It first pre-trains methylation data with subtype labels to generate pseudo-subtypes of cancer data without subtype information, and finally fine-tunes using labeled and unlabeled datasets. DeepSSC is a two-phase deep semi-supervised multi-omics integration framework for subtype classification and biomarker discovery. It first uses a denoising autoencoder to extract a compact representation for each type of omics data, then connects them and puts them into a feedforward neural network for subtype classification. A neural network classifier is then used to present specificity when identifying biomarkers. CG3 is a semi-supervised method based on graph convolutional networks. By designing a semi-supervised contrast loss, it maximizes the consistency between different views of the same data or the same type of data to generate improved node representations, extracting the potential relationship between data features and graph topology as supplementary supervision information.
[0104] The comparison between the present invention and other methods is as follows:
[0105] Comparative experiments were conducted on the Pan-Cancer Dataset, with a 10% labeling ratio used in all experiments. Specifically, 10% of the samples for each cancer type in the original Pan-Cancer Multi-Omics Dataset were labeled, leaving the remaining data as unlabeled. The model predicted the type of the unlabeled samples, and compared the model's predictions with the actual values using accuracy, recall, precision, and F1 scores.
[0106] In order to reduce the randomness and deviation of the experimental results, five repeated experiments were performed with different random seeds each time, and the average result of the five experiments was taken as the final result of the experiment. The specific experimental results are shown in Table 1, which shows the comparative experimental results of the present invention and other methods.
[0107] Table 1 Comparative experimental results of the method of the present invention and other methods
[0108]
[0109] As shown in Table 1, the method of the present invention (SemiML-Cancer) achieves classification accuracy, recall, precision, and F1 score of 0.9704, 0.9376, 0.9253, and 0.9304, respectively, which are all superior to other methods. Compared with supervised methods such as MOGONET and XOmiVAE, SemiML-Cancer achieves better classification performance because it uses a semi-supervised learning method to better utilize the complex relationships potentially present in unlabeled data. SemiML-Cancer outperforms semi-supervised models such as meth-SemiCancer, DeepSSC, and CG3 because it enhances the correlation between samples and improves the embedding representation by optimizing the similarity metric, further capturing the potential correlations and patterns between data.
[0110] Example 3. To investigate whether different modules play an important role in the model, based on Example 2, the impact of important components on the performance of SemiML-Cancer was verified through ablation experiments on the pan-cancer omics dataset, and three comparative variants were designed: First, SemiML-Cancer / -pca: SemiML-Cancer does not use principal component analysis to reduce the dimension and denoise the omics data. Instead, it directly uses the original pan-cancer multi-omics data as the input of AE after splicing according to features; Second, SemiML-Cancer / -pretrain: SemiML-Cancer does not use the pre-training module. Instead, it directly uses the data after dimensionality reduction by the encoder; Third, SemiML-Cancer / -fine-tune: SemiML-Cancer does not use the parameter fine-tuning module. Instead, it directly inputs the embedded representation processed by AE into the constructed SVM multi-class classifier; SemiML-Cancer is compared with the two variants respectively. As shown in the ablation experiment results in Table 2, removing the principal component analysis module reduces the classification performance of the present invention for pan-cancer sample categories. This is because SemiML-Cancer, trained with principal component analysis, cannot capture key features, resulting in the presence of a large number of other complex factors in the data. This noise seriously affects the accuracy and reliability of the model classification. In addition, the ablation experiment results table also shows that removing the pre-training module reduces the classification performance of the present invention for pan-cancer sample categories. This is because removing the pre-training module reduces the ability to learn pan-cancer data features, which negatively affects subsequent metric learning and thus reduces the accuracy of model classification. In addition, the ablation experiment results table also shows that removing metric learning reduces the classification performance of the present invention for pan-cancer sample categories. This is because after removing the metric learning module, SemiML-Cancer cannot utilize the correlation information and potential relationships between samples when predicting pan-cancer sample categories. As a result, the model lacks the ability to comprehensively analyze samples and cancer categories, thereby reducing the accuracy of model classification.
[0111] Table 2 Comparative experimental results of different variants of the method of the present invention
[0112]
[0113] In summary, the present invention implements a method for acquiring multiple omics data for pan-cancer samples; preprocessing and concatenating the multi-omics data, and then using principal component analysis to obtain a corresponding principal component score matrix; inputting the principal component score matrix into an autoencoder network for pretraining to genetically encode cancer samples; updating the encoder network parameters using partially labeled data, and optimizing the autoencoder network's embedding representation using metric learning; and inputting the optimized embedding representation into a constructed SVM multi-class classifier to obtain prediction results for the unlabeled sample categories. The present invention utilizes metric learning to design two metrics between heterogeneous center points and homogeneous data points, optimizing the similarity metric between patients, and enhancing the correlation between samples, effectively improving the model's classification performance for pan-cancer patient categories.
[0114] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention.
Claims
1. A pan-cancer classification method based on multi-omics data and semi-supervised metric learning, characterized by: The following steps are involved: S1. Obtain multiple omics data of pan-cancer samples; S2. After preprocessing and concatenating the multi-omics data, principal component analysis is used to obtain the corresponding principal component score matrix; S3, inputting the principal component score matrix into an autoencoder network for pre-training to perform genetic encoding on the cancer sample; S4. updating the parameters of the autoencoder network using the partially labeled data and optimizing the embedding representation of the autoencoder network using metric learning; S5. Input the optimized embedding representation into the constructed SVM multi-class classifier to obtain a prediction result for the unlabeled sample category.
2. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 1, characterized in that: The step S2 comprises: After preprocessing the multi-omics dataset of the pan-cancer samples, the samples shared by each omics were extracted and their gene feature information was spliced; performing standardization processing on the spliced multi-omics data; Calculate the covariance between all feature pairs and construct the covariance matrix; Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvectors and corresponding eigenvalues; The number of principal components was selected based on the proportion of variance explained by each principal component; The principal components are selected according to the number of principal components, and the standardized data are projected using the selected principal components to obtain a principal component score matrix.
3. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 1, characterized in that: The pre-training process of the autoencoder network in step S3 includes: The principal component score matrix is input into the autoencoder network and mapped to a low-dimensional embedding representation to learn the most effective gene encoding; then the decoder in the autoencoder network reconstructs the original spatial output from the low-dimensional embedding representation; The main idea of the autoencoder network is to minimize the loss between input and output. The formula corresponding to the loss function is: Among them, L p represents the loss between the input and output of the autoencoder network; M represents the principal component score matrix of the input; represents the reconstructed original space output of the output; Represents the mean square error between the principal component score matrix and the reconstructed original space output.
4. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 3, characterized in that: The step S4 comprises: According to the low-dimensional embedding representation, calculating the center point of each cancer category in the low-dimensional embedding representation space; Optimizing the similarity representation between samples based on the metric between the sample points and the center point of each cancer category in the low-dimensional embedding representation space; strengthening the discriminative representation between cancer classes based on a metric between the center points of each cancer class in the low-dimensional embedding representation space; The final objective function is optimized to fine-tune the encoder weight parameters in reverse.
5. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 4, characterized in that: The formula corresponding to the center point of each cancer category in the low-dimensional embedding representation space is: Among them, c i represents the center point of cancer type i; N i represents the number of samples of cancer type i; f θ () represents the mapping function of the encoder; x j Input cancer sample data; For cancer type i, N i The average of the embedded vectors after mapping of cancer samples.
6. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 4, characterized in that: The low-dimensional embedding represents the metric between the sample points and the center point of each cancer category in the space, and the corresponding formula is: Where L1 represents the Euclidean distance between the sample point and the center point for cancer type K; d() is the distance function; d(f θ (x j ),c i ) represents the mapping f of the cancer sample θ (x j ) and the cancer center point c i The squared Euclidean distance between Indicates that for type K, N i The mean distance between the sample points and the center point; Represents the mean distance between the sample points and the center point in K cancer types.
7. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 4, characterized in that: The low-dimensional embedding represents the metric between the center points of each cancer class in the space, and the corresponding formula is: Where L2 represents the Euclidean distance between two categories in all cancer categories K; Represents the center point of cancer types i1 and i2 for all types K and The minimum squared Euclidean distance between them.
8. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 4, characterized in that: The formula corresponding to the final objective function of the optimization is: Among them, L m Represents the final optimized objective loss function; The weight parameter of the encoder is θ, and θ is adjusted to minimize the loss, that is, L1 is minimized and L2 is maximized; L1 represents the Euclidean distance between the sample point and the center point for the cancer type K; λ is a regularization constant used to balance L1 and L2; L2 represents the Euclidean distance between the center points of two categories in all cancer categories K.
9. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 1, characterized in that: In step S5, the optimized embedding representation is input into the constructed SVM multi-class classifier, including the position of the hyperplane obtained by calculation, and the hyperplane is used to separate different categories of cancer samples to obtain a prediction result for the category of the unlabeled sample.
10. The pan-cancer classification method based on multi-omics data and semi-supervised metric learning according to claim 9, characterized in that: The formula corresponding to the hyperplane is: w T x i +b=0 Where w represents the normal vector of the hyperplane; w T represents the transpose of w; x i The vector representing the cancer sample in the embedding; w T x i Represents x i The projection size in the direction of the normal vector w of the hyperplane; b is the intercept of the hyperplane; the formula for optimizing the corresponding parameters of the hyperplane is: And must meet in, represents the distance between the two categories; C represents the penalty factor; ξ i represents the slack variable; y i Represents the label of cancer samples, ranging from {-1, +1}; When w T x i When +b≥+1, y i = +1; when w T x i When +b≤-1, y i =-1; samples above the interval boundary belong to the positive class, and samples below it belong to the negative class; train an SVM classifier for each cancer category and obtain the final predicted label for each unlabeled cancer sample.