An averaging method and system for a multi-omics fusion-based breast cancer subtype classification model

Abstract

This invention discloses a multi-omics fusion-based averaging method and system for breast cancer subtype classification models, relating to the interdisciplinary fields of deep ensemble learning and bioinformatics. By introducing learnable modality weight parameters and a cross-validation-based model averaging strategy, it achieves efficient fusion of multi-omics data, including single nucleotide variants, copy number variants, and gene expression, enabling accurate classification of breast cancer subtypes. The method involves: first, feature extraction from the multi-omics data; then, constructing a learnable weighted multi-omics fusion model. By introducing trainable modality weight parameters, the model can adaptively learn the contribution of different omics data during the fusion process, effectively capturing high-order cross-modal interaction information. Finally, by traversing all non-empty subsets of each omics set, a candidate model set covering different omics combinations is constructed, providing diverse and differentiated base models for subsequent model averaging and ensemble.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of deep integrated learning and bioinformatics, specifically to an averaging method and system for a multi-omics fusion breast cancer subtype classification model, which is particularly suitable for medical data classification tasks. Background Technology

[0002] Breast cancer is one of the most common malignant tumors among women worldwide, and its molecular subtyping is of significant clinical importance for developing individualized treatment plans and predicting prognosis. Currently, widely used subtyping methods in clinical practice mainly include alternative subtyping based on immunohistochemistry (IHC) and PAM50 subtyping based on gene expression profiling. With the rapid development of high-throughput sequencing technology, the accumulation of multi-omics data (including single nucleotide variants, copy number variations, gene expression, etc.) provides a richer source of information for breast cancer molecular subtyping. Deep learning methods based on multi-omics data have made significant progress in the field of breast cancer molecular subtyping in recent years. Researchers have designed various statistical modeling, machine learning, and deep learning architectures based on multi-omics data for breast cancer subtype prediction, demonstrating stronger performance in subtype discrimination, prognostic assessment, and treatment response prediction. These developments provide an important technological foundation for building more accurate and robust breast cancer molecular subtyping systems.

[0003] Patent application number 202510525610.1 provides a BRCA The patent for the construction and identification method of molecular subtyping model for mutant breast cancer was first obtained. BRCA1 / 2 Transcriptome expression profiles of patients with mutated breast cancer were analyzed, and the most significantly altered genes were selected for unsupervised clustering analysis to obtain the optimal number of clusters, classifying patients into subtypes with different clinical prognoses and molecular characteristics. Then, based on the clustering results, a molecular subtyping model was constructed using the random forest algorithm, incorporating multiple key classification genes. Finally, the molecular subtyping model was validated using a validation set and external datasets to ensure its accuracy, stability, and clinical applicability. This method can... BRCA1 / 2 Transcriptome sequencing data from tumor samples of patients with mutated breast cancer can be used for rapid and accurate molecular subtyping, providing a scientific basis for the selection of personalized treatment plans.

[0004] The existing breast cancer subtype classification scheme still has the following shortcomings: (1) Failure to fully utilize high-order interaction information of multi-omics data. Most existing machine learning methods use simple feature splicing or weighted summation for multimodal fusion, which makes it difficult to effectively capture the nonlinear and high-order interaction relationships between different omics data. The deep coupling information across omics is difficult to be fully modeled in simple fusion strategies, which limits the performance ceiling of subtype classification.

[0005] (2) Limited integration strategies and model stability. Existing methods typically train a single model based on single- or multi-omics data, lacking a systematic exploration of different combinations of multi-omics data. Different combinations of omics contain complementary information with different emphases, but existing methods fail to fully utilize this diversity to construct differentiated and complementary base model sets. In the model integration stage, existing methods mostly adopt equal-weighted averaging or simple voting strategies, failing to adaptively weight each base model according to the differences in prediction confidence and generalization error on different samples, resulting in limited robustness and generalization performance of the integrated model.

[0006] (3) Existing technologies use unsupervised clustering methods to classify patients. The subtype labels obtained by clustering lack the guidance of prior clinical knowledge. Their biological significance and prognostic association need to be verified after the fact. The clustering results on different datasets may vary greatly, and their auxiliary role in precision medicine is limited. Summary of the Invention

[0007] In view of this, the present invention provides an averaging method and system for a multi-omics fusion-based breast cancer subtype classification model, which solves the problems of insufficient modality fusion, single integration strategy, and insufficient generalization performance in the existing technology for breast cancer subtype classification, and achieves more accurate and robust multi-omics molecular subtyping of breast cancer.

[0008] To achieve the above objectives, this invention provides an ensemble learning method based on a learnable weighted multi-omics fusion model, the technical solution of which includes the following steps: Step 1: Obtain raw data containing multiple omics, including gene mutation data, copy number variation data, and gene expression data; perform feature extraction and standardization on the data of each omics separately to construct a multi-omics dataset of the samples.

[0009] Step 2: Construct a learnable weighted multi-omics fusion model. Using multi-omics data features as input, each omics data is encoded by an independent encoder to obtain encoded features. The encoded features are then nonlinearly enhanced and adaptively weighted to obtain a multi-omics feature fusion representation vector. Then, classification prediction is performed based on the multi-omics feature fusion representation vector.

[0010] Step 3: By traversing all non-empty subsets of each omics set, construct a candidate model set covering different omics combinations. For each candidate omics combination, use the K-fold cross-validation model averaging method to determine the optimal ensemble weights. Apply the learned optimal weights to the subtype prediction probability vector of each candidate model on the test set. Obtain the final subtype prediction probability vector through weighted averaging. After model averaging, the final prediction result is also the category corresponding to the maximum prediction probability.

[0011] Furthermore, in step 1, the multi-omics data used are labeled data, and the PAM50 subtype classification of each sample is used as the model learning target.

[0012] Furthermore, in step 2, the constructed learnable weighted multi-omics fusion model consists of three core modules: a multi-omics encoder module, a learnable weighted fusion module, and a classification output module. The multi-omics encoder module encodes features for each type of omics data using an independent encoder, mapping the original features to a unified latent space. The learnable weighted fusion module introduces a modality-specific transformation layer to nonlinearly enhance the encoded features before adaptive weighting, resulting in a multi-omics feature fusion representation vector. The classification output module inputs the feature fusion representation vector into a classifier for classification prediction.

[0013] Furthermore, the multi-omics encoder module specifically includes: The input data contains K sets of omics data features. The model input must include at least three sets of omics data features: single nucleotide variants (SNVs), copy number variations (CNAs), and gene expression RNA. The input feature vector for the k-th omics is... , This refers to the feature dimension corresponding to the omics data; Each type of omics data is first feature-encoded using an independent encoder, mapping the raw features to a unified latent space; for the k-th omics, the encoder processing formula is:

[0014] in The omics k-coded hidden layer feature representation obtained by the encoder processing. The unified coding latent space dimension for features of each omics. The activation function; the coding layer weight matrix of the k-th omics. Bias terms Learnable parameters for multi-omics encoder layers; The parameters in the learnable weighted feature fusion module are summarized as follows: .

[0015] Furthermore, a weight fusion module can be learned, specifically: Define a learnable weight parameter vector ,in For the k-th omics, the unnormalized weight parameter is used. The Softmax function is used to normalize the weight parameters, ensuring that all weights are non-negative and sum to 1.

[0016] in For the fusion weight of the k-th omics, satisfying and ; By introducing a modality-specific transform layer, the encoded features are nonlinearly enhanced and then adaptively weighted to obtain a multi-omics feature fusion representation vector. :

[0017] in The nonlinear activation function is tanh, and the weight matrix is... Bias terms These are the learnable parameters for the weight fusion layer; The parameters in the learnable weighted feature fusion module are summarized as follows: .

[0018] Furthermore, the classification output module specifically includes: Feature fusion representation vector Input classifier for classification prediction:

[0019] Where C is the number of categories in the classification task, and the classification layer weight matrix is... Classifier bias vector For network learnable parameters, This is the classification score vector output by the classifier; the final probability vectors for each class are obtained by normalizing the classification score vector.

[0020]

[0021] The model's final prediction result is the category corresponding to the highest predicted probability, i.e.:

[0022] The parameters in the classification output module are summarized as follows: .

[0023] Furthermore, the network employs an end-to-end joint training strategy. All learnable parameters are collaboratively optimized by minimizing the following cross-entropy loss function through backpropagation gradient updates. For a training batch containing N samples, the cross-entropy loss function is defined as:

[0024] in For the nth sample, encode the true one-hot label in class c. Let be the predicted probability of the nth sample belonging to the cth class. To prevent overfitting, an L2 regularization term is introduced into the loss function:

[0025] in This represents the set of all learnable parameters in the network. This is the regularization coefficient, used to control the strength of regularization.

[0026] Furthermore, for each candidate omics combination, this invention uses the K-fold cross-validation model averaging method to determine the optimal ensemble weights, specifically as follows: For a total of M candidate models, for each candidate model m and each fold k, m=1,2,…,M, the model is trained using data excluding the k-th fold, and a prediction is made for the k-th fold; let The prediction result at the k-th fold is based on the model parameters estimated on the data excluding the k-th fold. Concatenate all the prediction results to obtain the prediction matrix for the entire training set. , of which The elements are denoted as ; Let the weight vector be satisfy and ;remember Let n be an indicator variable indicating that sample n belongs to category c. For one-hot encoding, the value is 1 if the true class of sample n is c, and 0 otherwise. Predict the probability that sample n belongs to class c for the m-th sub-model; based on the prediction probability matrix of the entire training set, the K-fold cross-validation criterion is defined as:

[0027] The optimal model average weights are obtained by minimizing the K-fold cross-validation criterion. The learned optimal weights are applied to the subtype prediction probability vectors of each candidate model on the test set. The final subtype prediction probability vector is obtained by weighted averaging. :

[0028] After averaging the models, the final prediction result is also the category corresponding to the highest predicted probability.

[0029] Another embodiment of the present invention provides an ensemble learning system based on a learnable weighted multi-omics fusion model, characterized in that it is used to execute an ensemble learning method based on a learnable weighted multi-omics fusion model as described in any one of claims 1-8, and the system includes the following modules: The multi-omics data feature extraction module is used to acquire raw data containing multiple omics, including gene mutation data, copy number variation data, and gene expression data; feature extraction and standardization are performed on the data of each omics separately to construct a multi-omics dataset of the samples; The multi-omics fusion module is used to construct a learnable weighted multi-omics fusion model. It takes multi-omics data features as input, and each omics data is encoded by an independent encoder to obtain encoded features. The encoded features are non-linearly enhanced and then adaptively weighted to obtain a multi-omics feature fusion representation vector. Then, classification prediction is performed based on the multi-omics feature fusion representation vector. The model ensemble module is used to construct a set of candidate models covering different omics combinations by traversing all non-empty subsets of each omics set. For each candidate omics combination, the optimal ensemble weights are determined by the K-fold cross-validation model averaging method. The learned optimal weights are applied to the subtype prediction probability vectors of each candidate model on the test set. The final subtype prediction probability vector is obtained by weighted averaging. After model averaging, the final prediction result is also the category corresponding to the maximum prediction probability.

[0030] Beneficial effects: 1. This invention designs a learnable weighted fusion network, a systematic candidate model construction and cross-validation weighted integration strategy, and introduces learnable modal weight parameters and a cross-validation-based model averaging strategy to achieve efficient fusion of multi-omics data such as single nucleotide variants, copy number variants and gene expression, thereby achieving accurate classification of breast cancer subtypes. This solves the problems of insufficient modal fusion, single integration strategy and insufficient generalization performance in the existing technology for breast cancer subtype classification, and achieves more accurate and robust multi-omics molecular subtyping of breast cancer.

[0031] 2. This invention designs an effective multi-omics feature fusion mechanism and constructs a small-scale deep learning network suitable for high-dimensional, small-sample data to fully capture the nonlinear, high-order interaction information between different omics data. By introducing a learnable weight fusion mechanism, this invention enables the model to adaptively learn the contribution weights of features from different modalities, effectively modeling cross-modal nonlinear coupling relationships while preserving the original information of each modality.

[0032] 3. This invention systematically constructs a diverse and complementary base model library and adaptively optimizes the ensemble weights to improve the overall generalization performance of the model. This invention constructs a candidate model library by traversing and combining features from various omics systems, fully utilizing the complementary information of different modality combinations to enhance the diversity among base models. Furthermore, it employs a weighted model averaging strategy based on K-fold cross-validation, adaptively learning the optimal ensemble weights by minimizing the cross-validation error criterion, thereby enabling the ensemble model to achieve stronger generalization ability overall.

[0033] 4. This invention employs a supervised learning paradigm based on the PAM50 standard, using clinically recognized molecular subtyping labels for breast cancer as supervisory signals. The subtype categories output by the model have clear clinical significance and prognostic value. Compared to unsupervised clustering methods, supervised learning can directly learn the mapping relationship from multi-omics features to clinically relevant subtypes, making the subtyping results more interpretable and clinically operable, facilitating subsequent prognostic assessment and the development of individualized treatment plans. Attached Figure Description

[0034] Figure 1 The present invention provides a flowchart of an average ensemble method for a multi-omics-based breast cancer subtype classification model; Figure 2 A structural diagram of a learnable weighted multi-omics fusion model; Figure 3 Build a flowchart for candidate models. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0036] This invention aims to overcome the problems existing in the prior art and provide an ensemble learning method based on a learnable weighted multi-omics fusion model. By introducing learnable modality weight parameters and a model averaging strategy based on cross-validation, it achieves efficient fusion of multi-omics data such as single nucleotide variants, copy number variants, and gene expression, enabling accurate classification of breast cancer subtypes. The complete implementation process is as follows: Figure 1 As shown.

[0037] Step 1: Feature Extraction from Multi-omics Data We acquired raw data encompassing multiple omics, including gene mutation data, copy number variation data, and gene expression data. Feature extraction and standardization were performed on the data for each omics group to construct a multi-omics dataset. The multi-omics data used were labeled, and the PAM50 subtype classification of each sample was used as the model learning objective.

[0038] Step 2: Construct a learnable weighted multi-omics fusion model The learnable weighted multi-omics fusion model designed in the technical solution consists of three core modules: a multi-omics encoder module, a learnable weighted fusion module, and a classification prediction module. The complete model structure is as follows: Figure 2 As shown.

[0039] (1) Multi-omics feature coding module Suppose the input data contains K sets of omics data features, where K=3 in this example. The model input consists of three sets of omics data features: single nucleotide variants (SNVs), copy number variations (CNAs), and gene expression (RNA). Let the input feature vector of the k-th omics be denoted as . , This refers to the feature dimension corresponding to the omics data.

[0040] Each omics dataset is first encoded using an independent encoder, mapping the raw features to a unified latent space. For the k-th omics dataset, the encoder processing formula is:

[0041] in The omics k-coded hidden layer feature representation obtained by the encoder processing. The unified coding latent space dimension for features of each omics. The activation function is denoted as . The encoding layer weight matrix for the k-th omics approach. Bias terms These are the learnable parameters for the multi-omics encoder layer.

[0042] The parameters in the learnable weighted feature fusion module are summarized as follows: .

[0043] (2) Learnable weighted feature fusion module Define a learnable weight parameter vector ,in Let be the unnormalized weight parameters corresponding to the k-th omics. The Softmax function is used to normalize the weight parameters, ensuring that each weight is non-negative and sums to 1.

[0044] in For the fusion weight of the k-th omics, satisfying and A modality-specific transform layer is introduced to nonlinearly enhance the encoded features before adaptive weighting, resulting in a multi-omics feature fusion representation vector. :

[0045] in The nonlinear activation function is tanh, and the weight matrix is... Bias terms These are the learnable parameters for the weighted fusion layer.

[0046] The parameters in the learnable weighted feature fusion module are summarized as follows: .

[0047] (3) Classification output module Feature fusion representation vector Input classifier for classification prediction:

[0048] Where C is the number of categories in the classification task, and the classification layer weight matrix is... Classifier bias vector These are the learnable parameters of the network. This is the classification score vector output by the classifier. The final probability vectors for each class are obtained by normalizing the classification score vector.

[0049]

[0050] The model's final prediction result is the category corresponding to the highest predicted probability, i.e.:

[0051] The parameters in the classification output module are summarized as follows: .

[0052] (4) Loss calculation and parameter update The network employs an end-to-end joint training strategy, where all learnable parameters are collaboratively optimized by minimizing the following cross-entropy loss function through backpropagation gradient updates. For a training batch containing N samples, the cross-entropy loss function is defined as:

[0053] in For the nth sample, encode the true one-hot label in class c. Let be the predicted probability of the nth sample belonging to the cth class. To prevent overfitting, an L2 regularization term is introduced into the loss function:

[0054] in This represents the set of all learnable parameters in the network. This is the regularization coefficient, used to control the strength of regularization.

[0055] The multi-omics fusion model structure design of this invention: by introducing trainable modality weight parameters, the model can adaptively learn the contribution of different omics data (SNV, CNA, RNA) in the fusion process, and achieve effective capture of cross-modal high-order interaction information.

[0056] Step 3: Perform model ensemble using the model averaging method This invention proposes a systematic candidate model construction method. By traversing all non-empty subsets of each omics set, a candidate model set covering different omics combinations is constructed, providing diversified and differentiated base models for subsequent model averaging.

[0057] Define the set of omics types participating in modeling as In the breast cancer subtype classification embodiments of the present invention, three sets of omics data features are used: single nucleotide variants (SNVs), copy number variants (CNAs), and gene expression (RNA). Generate omics datasets All non-empty subsets ,exist Under the various combinations of omics data included, learnable weighted multi-omics fusion models are trained separately as candidate model sets.

[0058] for In the case of generating a total of For each candidate omics combination, A learnable weighted multi-omics fusion model is used as the basic architecture to train the corresponding subtype classification model on the training set. Specifically, for For single-modal combinations, the model degenerates into a single-omics multilayer perceptron that does not require learning multi-omics fusion weights; for The multi-omics approach combines multiple omics models and uses the learnable weighted multi-omics fusion prediction network described above to construct candidate models. The specific construction of the candidate model set is as follows: Figure 3 As shown.

[0059] For each candidate omics combination, this invention uses the K-fold cross-validation model averaging method to determine the optimal ensemble weights. In the embodiment, the training set is uniformly divided into K=5 folds.

[0060] Suppose there are M candidate models. For each candidate model m (m=1, 2, ..., M) and each fold k, train the model using data excluding the k-th fold, and then make a prediction for the k-th fold. Let... The prediction result at the k-th fold is based on the model parameters estimated on the data excluding the k-th fold. By concatenating the prediction results of all folds, the prediction matrix of the entire training set is obtained. , of which The elements are denoted as .

[0061] Let the weight vector be satisfy and ;remember The indicator variable for sample n belonging to category c (one-hot encoded, 1 if the true category of sample n is c, 0 otherwise); Predict the probability that sample n belongs to class c for the m-th sub-model. Based on the prediction probability matrix of the entire training set, the K-fold cross-validation criterion is defined as:

[0062] The optimal model average weights are obtained by minimizing the K-fold cross-validation criterion. The learned optimal weights are applied to the subtype prediction probability vectors of each candidate model on the test set. The final subtype prediction probability vector is obtained by weighted averaging. :

[0063] After averaging the models, the final prediction result is also the category corresponding to the highest predicted probability.

[0064] The systematic candidate model construction strategy is to build a candidate model library covering all modality combinations by traversing all non-empty subsets of learning, making full use of the complementary information of different modality subsets and enhancing the diversity among base models.

[0065] In this invention, a holistic paradigm of "dedicated model + candidate ensemble" for multi-omics subtype classification is proposed: This invention proposes a general technical framework for multi-omics subtype classification, namely, designing a dedicated fusion model for multimodal combinations + constructing a differentiated candidate model library through multi-omics subsets + model averaging ensemble. This paradigm is not limited to specific omics data types or disease scenarios, and can flexibly adapt to any number of omics inputs, forming a complete classification solution based on multi-omics data. Its framework itself constitutes a universally applicable technological innovation.

[0066] To implement the above method, embodiments of the present invention also provide an ensemble learning system based on a learnable weighted multi-omics fusion model, the system comprising the following modules: The multi-omics data feature extraction module is used to acquire raw data containing multiple omics, including gene mutation data, copy number variation data, and gene expression data; feature extraction and standardization are performed on the data of each omics separately to construct a multi-omics dataset of the samples.

[0067] The multi-omics fusion module is used to construct a learnable weighted multi-omics fusion model. It takes multi-omics data features as input, and each omics data is encoded by an independent encoder to obtain encoded features. The encoded features are nonlinearly enhanced and then adaptively weighted to obtain a multi-omics feature fusion representation vector. Then, classification prediction is performed based on the multi-omics feature fusion representation vector.

[0068] The model ensemble module is used to construct a set of candidate models covering different omics combinations by traversing all non-empty subsets of each omics set. For each candidate omics combination, the optimal ensemble weights are determined by the K-fold cross-validation model averaging method. The learned optimal weights are applied to the subtype prediction probability vectors of each candidate model on the test set. The final subtype prediction probability vector is obtained by weighted averaging. After model averaging, the final prediction result is also the category corresponding to the maximum prediction probability.

[0069] Compared with the prior art, the present invention has the following technical advantages: 1. Enhanced Classification Accuracy Through Multi-omics Data Integration. Existing technologies rely solely on single-omics data from transcriptome expression profiles for molecular typing, failing to fully utilize multi-level biological information such as genomic mutations and copy number variations. This application designs a learnable weighted multi-omics fusion model to jointly model SNV, CNA, and RNA omics data, enabling the model to simultaneously capture multi-dimensional information such as gene sequence variations, copy number alterations, and abnormal expression levels. Experimental results show that the AUC (0.9230-0.9454) of this application outperforms various classic machine learning methods for breast cancer subtype classification and various prediction methods based on single-omics data in the breast cancer subtype classification task.

[0070] 2. A systematic model ensemble strategy leads to more robust model performance. Existing technologies use a single random forest model for classification decisions, lacking a multi-model ensemble mechanism. This application constructs seven candidate models by traversing all non-empty subsets of omics, covering all combinations from mono-omics to tri-omics, and adaptively learns the optimal ensemble weights using a weighted model averaging method based on K-fold cross-validation. Compared to equal-weighted averaging or simple voting strategies, this application's cross-validation weighted ensemble method can differentiate weights based on the differences in prediction errors of each model on different samples, effectively improving the model's robustness and demonstrating smaller index variance in multiple randomized partitioning experiments.

[0071] 3. Existing technology is only applicable to BRCA Subtyping mutant breast cancer has limited applicability. The learnable weighted multi-omics fusion model of this application features a highly modular design, enabling flexible adaptation to any number of omics data types. It is not only applicable to breast cancer subtype classification but can also be extended to other cancer types or multi-omics classification tasks, thus possessing broader clinical application value.

[0072] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An ensemble learning method based on a learnable weighted multi-omics fusion model, characterized in that, Includes the following steps: Step 1: Obtain raw data containing multiple omics, including gene mutation data, copy number variation data, and gene expression data; perform feature extraction and standardization on the data of each omics separately to construct a multi-omics dataset of the samples; Step 2: Construct a learnable weighted multi-omics fusion model. The multi-omics data features are used as input. Each omics data is encoded by an independent encoder to obtain encoded features. The encoded features are non-linearly enhanced and then adaptively weighted to obtain a multi-omics feature fusion representation vector. Then, classification prediction is performed based on the multi-omics feature fusion representation vector. Step 3: By traversing all non-empty subsets of each omics set, construct a candidate model set covering different omics combinations. For each candidate omics combination, use the K-fold cross-validation model averaging method to determine the optimal ensemble weights. Apply the learned optimal weights to the subtype prediction probability vector of each candidate model on the test set. Obtain the final subtype prediction probability vector through weighted averaging. After model averaging, the final prediction result is also the category corresponding to the maximum prediction probability.

2. The ensemble learning method based on a learnable weighted multi-omics fusion model as described in claim 1, characterized in that, In step 1, the multi-omics data used are labeled data, and the PAM50 subtype classification of each sample is used as the model learning target.

3. The ensemble learning method based on a learnable weighted multi-omics fusion model as described in claim 1, characterized in that, In step 2, the constructed learnable weight multi-omics fusion model consists of three core modules: a multi-omics encoder module, a learnable weight fusion module, and a classification output module. The multi-omics encoder module performs feature encoding for each type of omics data using an independent encoder, mapping the original features to a unified latent space. The learnable weight fusion module introduces a modality-specific transformation layer, which nonlinearly enhances the encoded features and then adaptively weights them to obtain a multi-omics feature fusion representation vector. The classification output module inputs the feature fusion representation vector into the classifier for classification prediction.

4. The ensemble learning method based on a learnable weighted multi-omics fusion model as described in claim 1, characterized in that, The multi-omics encoder module is specifically as follows: The input data contains K sets of omics data features. The model input must include at least three sets of omics data features: single nucleotide variants (SNVs), copy number variations (CNAs), and gene expression RNA. The input feature vector for the k-th omics is... , This refers to the feature dimension corresponding to the omics data; Each type of omics data is first feature-encoded using an independent encoder, mapping the raw features to a unified latent space; for the k-th omics, the encoder processing formula is: in The omics k-coded hidden layer feature representation obtained by the encoder processing. The unified coding latent space dimension for features of each omics. The activation function; the coding layer weight matrix of the k-th omics. Bias terms Learnable parameters for multi-omics encoder layers; The parameters in the learnable weighted feature fusion module are summarized as follows: .

5. The ensemble learning method based on a learnable weighted multi-omics fusion model as described in claim 4, characterized in that, The learnable weight fusion module is specifically as follows: Define a learnable weight parameter vector ,in For the k-th omics, the unnormalized weight parameter is used. The Softmax function is used to normalize the weight parameters, ensuring that all weights are non-negative and sum to 1. in For the fusion weight of the k-th omics, satisfying and ; By introducing a modality-specific transform layer, the encoded features are nonlinearly enhanced and then adaptively weighted to obtain a multi-omics feature fusion representation vector. : in The nonlinear activation function is tanh, and the weight matrix is... Bias terms These are the learnable parameters for the weight fusion layer; The parameters in the learnable weighted feature fusion module are summarized as follows: 。 6. The ensemble learning method based on a learnable weighted multi-omics fusion model as described in claim 5, characterized in that, The classification output module is specifically as follows: Feature fusion representation vector Input classifier for classification prediction: Where C is the number of categories in the classification task, and the classification layer weight matrix is... Classifier bias vector For network learnable parameters, This is the classification score vector output by the classifier; the final probability vectors for each class are obtained by normalizing the classification score vector. The model's final prediction result is the category corresponding to the highest predicted probability, i.e.: The parameters in the classification output module are summarized as follows: .

7. The ensemble learning method based on a learnable weighted multi-omics fusion model as described in claim 6, characterized in that, The network employs an end-to-end joint training strategy. All learnable parameters are collaboratively optimized by minimizing the following cross-entropy loss function through backpropagation gradient updates. For a training batch containing N samples, the cross-entropy loss function is defined as: in For the nth sample, encode the true one-hot label in class c. Let be the predicted probability of the nth sample belonging to the cth class. To prevent overfitting, an L2 regularization term is introduced into the loss function: in This represents the set of all learnable parameters in the network. This is the regularization coefficient, used to control the strength of regularization.

8. The ensemble learning method based on a learnable weighted multi-omics fusion model as described in claim 7, characterized in that, For each candidate omics combination, this invention uses the K-fold cross-validation model averaging method to determine the optimal ensemble weights, specifically: For a total of M candidate models, for each candidate model m and each fold k, m=1,2,…,M, the model is trained using data excluding the k-th fold, and a prediction is made for the k-th fold; let The prediction result at the k-th fold is based on the model parameters estimated on the data excluding the k-th fold. Concatenate all the prediction results to obtain the prediction matrix for the entire training set. , of which The elements are denoted as ; Let the weight vector be satisfy and ;remember Let n be an indicator variable indicating that sample n belongs to category c. For one-hot encoding, the value is 1 if the true class of sample n is c, and 0 otherwise. Predict the probability that sample n belongs to class c for the m-th sub-model; based on the prediction probability matrix of the entire training set, the K-fold cross-validation criterion is defined as: The optimal model average weights are obtained by minimizing the K-fold cross-validation criterion. The learned optimal weights are applied to the subtype prediction probability vectors of each candidate model on the test set. The final subtype prediction probability vector is obtained by weighted averaging. : After averaging the models, the final prediction result is also the category corresponding to the highest predicted probability.

9. An ensemble learning system based on a learnable weighted multi-omics fusion model, characterized in that, For executing an ensemble learning method based on a learnable weighted multi-omics fusion model as described in any one of claims 1-8, the system includes the following modules: The multi-omics data feature extraction module is used to acquire raw data containing multiple omics, including gene mutation data, copy number variation data, and gene expression data; feature extraction and standardization are performed on the data of each omics separately to construct a multi-omics dataset of the samples; The multi-omics fusion module is used to construct a learnable weighted multi-omics fusion model. It takes multi-omics data features as input, and each omics data is encoded by an independent encoder to obtain encoded features. The encoded features are non-linearly enhanced and then adaptively weighted to obtain a multi-omics feature fusion representation vector. Then, classification prediction is performed based on the multi-omics feature fusion representation vector. The model ensemble module is used to construct a set of candidate models covering different omics combinations by traversing all non-empty subsets of each omics set. For each candidate omics combination, the optimal ensemble weights are determined by the K-fold cross-validation model averaging method. The learned optimal weights are applied to the subtype prediction probability vectors of each candidate model on the test set. The final subtype prediction probability vector is obtained by weighted averaging. After model averaging, the final prediction result is also the category corresponding to the maximum prediction probability.

Images