A biological sample molecular subtype classification method based on a multi-modal hierarchical attention network

Abstract

The present application relates to a kind of biological sample molecular subtype classification method based on multi-modal layered attention network, comprising: obtaining sample data to be measured and pre-processing;With the molecular subtype classification model of pre-processed sample data to be measured, by single-mode local feature extraction, cross-modal multi-view correlation, priori knowledge guided dynamic gating and dense feature aggregation strategy, output molecular phenotype classification probability, the molecular phenotype classification probability is compared with threshold value to determine molecular subtype class.The present application can overcome the limitations of the existing biological omics and other multi-source heterogeneous data fusion classification methods, such as shallow feature interaction, dimension easy to explode, and ignoring the guidance of sample macroscopic prior knowledge.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and artificial intelligence, and in particular to a method for classifying molecular subtypes of biological samples based on a multimodal hierarchical attention network. Background Technology

[0002] In the field of multimodal bioinformatics data analysis, traditional computational methods often rely on mining from a single data source or simple early feature splicing, making it difficult to fully capture the complex nonlinear dependencies between high-dimensional molecular omics data (such as gene expression profiles and copy number variations) and sample phenotypes. Existing deep multimodal fusion models also have certain limitations in cross-modal interaction and feature extraction: if conventional networks attempt to explicitly model cross-modal associations through high-order tensor products, they are prone to causing an exponential explosion in feature dimensions, often leading to severe overfitting under limited biological sample sizes; while shallow interactions fail to uncover deep complementary information. Furthermore, high-dimensional and sparse omics data are usually accompanied by a large amount of background noise. Existing models often extract each modality indiscriminately, failing to effectively utilize reliable low-dimensional prior phenotypes as a global context to guide omics noise filtering, resulting in key microstate assessment signals being easily submerged by the complex background. More importantly, most deep multimodal frameworks lack dynamic adaptive gating and interpretable design, and the connection between their fusion logic and underlying biological mechanisms is ambiguous, which greatly limits their practical application value in the accurate classification of molecular phenotypes. Summary of the Invention

[0003] The purpose of this invention is to provide a molecular subtype classification method for biological samples based on a multimodal hierarchical attention network, which overcomes the limitations of existing classification methods that fuse heterogeneous data from multiple sources, such as biological omics, where feature interactions are too shallow or dimensions are prone to explosion, and the guidance of macroscopic prior knowledge of samples is ignored.

[0004] To achieve the above objectives, the present invention provides the following solution: A method for molecular subtype classification of biological samples based on a multimodal hierarchical attention network includes: Acquire the data of the sample to be tested and perform preprocessing; The preprocessed sample data is input into the molecular subtype classification model. Through single-modal local feature extraction, cross-modal multi-view association, and prior knowledge-guided dynamic gating and dense feature aggregation strategies, the molecular phenotype classification probability is output. The molecular phenotype classification probability is compared with a threshold to determine the molecular subtype category.

[0005] Optionally, the molecular subtype classification model includes: A multimodal feature input module is used to receive multimodal heterogeneous data of a target biological sample, wherein the multimodal heterogeneous data includes high-dimensional molecular omics data and prior phenotypic features; The data preprocessing and feature dimensionality reduction module is used to impute missing values ​​and align modalities in the high-dimensional molecular omics data, perform coarse and fine dimensionality reduction on the aligned data, and perform standardization on the screened data to obtain high-dimensional molecular omics modal embedding features. The standardization module is used to perform data completion and standardization on prior phenotypic features, and input the standardized features into the multilayer perceptron to obtain prior embedded features. A single-modal feature extractor is used to mine and denoise the local feature dependencies within the modality of the high-dimensional molecular omics modal embedding features, and generate single-modal enhanced features; A multi-view association layer is used to extract cross-modal nonlinear associations based on the single-modal enhancement features and the prior embedding features, and generate a global interactive feature vector. The prior perception gating module is used to adaptively adjust the importance weight vectors of each modality based on the single-modality enhancement features and the prior embedding features, and generate a gating fusion feature vector. The hybrid feature aggregation and prediction module is used to perform channel-level concatenation of the global interactive feature vector, the gated fusion feature vector, and the multimodal heterogeneous data through a dense connection strategy, inputting them into a fully connected prediction head to output the molecular phenotype classification probability.

[0006] Optionally, the single-modal feature extractor includes: Linear projection units are used to map the high-dimensional molecular omics modality embedding features to a unified dimension; Channel-level self-attention units are used to map high-dimensional omics features after unification into query vectors, key vectors, and value vectors through a weight matrix. An affinity matrix is ​​calculated based on the query vectors and key vectors, and the value vectors are weighted and aggregated using the affinity matrix to obtain weighted features. The scaling residual connection unit is used to process the weighted features and the high-dimensional molecular omics modality embedding features using scaling residual connections to generate the single-modality enhanced features.

[0007] Optionally, the single-modal enhancement feature is generated. include: ; in, It is the residual scaling factor. It is a high-dimensional molecular omics modal embedding feature. It is a weighted feature.

[0008] Optionally, the multi-view association layer includes a cross-modal attention unit and an average aggregation unit; The cross-modal attention unit is used to calculate the cross-modal affinity matrix based on the single-modal enhancement features and the prior embedding features; The average aggregation unit is used to traverse all pairwise modal interaction combinations and average the cross-modal affinity matrices of each combination to generate the global interaction feature vector.

[0009] Optionally, calculating the cross-modal affinity matrix includes: Using any one modal feature as the query vector and the remaining modal features as the key vector and value vector, calculate the cross-modal affinity matrix, wherein the modal feature is a single-modal enhancement feature or the prior embedding feature.

[0010] Optionally, the prior perception gating module includes a joint context splicing unit, a dynamic weight generation unit, and a weighted fusion unit; The joint context splicing unit is used to splice the single-modal enhancement feature and the prior embedding feature in the channel dimension to form a global joint context vector; The dynamic weight generation unit is used to output the importance weight vectors corresponding to each modality based on the global joint context vector, using a multilayer perceptron combined with an activation function. The weighted fusion unit is used to perform element-wise weighted summation of the single-modal enhancement features and prior embedding features using the importance weight vector to generate the gated fusion feature vector.

[0011] Optionally, during the training process of the molecular subtype classification model, Focal Loss and SMOTE oversampling are combined to handle class imbalance, and the AdamW optimizer and cosine annealing strategy are used for end-to-end network training.

[0012] The beneficial effects of this invention are as follows: This invention combines the advantages of single-modal local feature mining with cross-modal multi-perspective association. Through prior knowledge-guided dynamic gating and dense feature aggregation strategies, it achieves accurate classification of molecular phenotypes and micro-state assessment of target biological samples. It is suitable for multi-dimensional omics data analysis, biomarker subtype identification, and functional genomics research.

[0013] From a theoretical perspective, the prior-aware hierarchical attention architecture of this invention achieves, for the first time, hierarchical modeling of "intramodal micro-dependency – cross-modal macro-interaction" in bioinformatics multimodal data analysis. This breakthrough challenges the simplistic assumptions of traditional multimodal fusion that treat each omics as an independent view or simply stitch them together. Through the collaborative operation of the Unimodal Feature Extractor (UMFE) and the Multi-View Association Layer (MVCL), the model not only explicitly reconstructs the internal co-expression relationships of high-dimensional molecular omics data but also accurately captures the complex nonlinear collaborative mechanism between multi-dimensional omics and phenotypic features. Furthermore, the design of the prior-aware gating module (CAGM), combined with the Focal Loss function, provides a new approach for training on noisy, imbalanced biomolecular data. This mechanism effectively suppresses omics background noise by introducing macro-phenotypic as a global context guide, enabling the model to exhibit extremely strong feature focusing capabilities when processing specific molecular subtype data with significant class imbalance. This discovery provides important methodological references for automated multi-omics analysis and the assessment of complex biological states.

[0014] In terms of practical application value, the robustness of the model in cross-domain and cross-dataset tests makes it a promising framework for automated analysis of multidimensional heterogeneous biological sequencing samples. In classic METABRIC test sets, traditional baseline models (such as simple post-weighted or unimodal models) generally have upper limits in evaluation accuracy, while the model of this invention shows significant advantages in the key ranking metrics of AUC (0.898) and C-index (0.8526). Furthermore, in the TCGA-BRCA extension experiments, the model exhibits strong modality compatibility, successfully and seamlessly integrating high-resolution image (WSI) features, raising the overall evaluation metric MCC to 0.663—this advantage stems from its hierarchical dense fusion architecture's powerful generalization and parsing capabilities for highly heterogeneous data. However, in scenarios with limited biological sample size or extreme data gaps, the model still exhibits some evaluation variance, suggesting that in complex practical bioinformatics applications, it is still necessary to combine richer sequencing metadata for comprehensive validation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a diagram of the molecular subtype classification model architecture according to an embodiment of the present invention; Figure 2This is a diagram of the architecture of a single-modal feature extractor according to an embodiment of the present invention; Figure 3 This is a diagram of the multi-view association layer architecture according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] This embodiment proposes a method for classifying molecular subtypes of biological samples based on a multimodal hierarchical attention network, including: Acquire the data of the sample to be tested and perform preprocessing; The preprocessed sample data is input into the molecular subtype classification model. Through single-modal local feature extraction, cross-modal multi-view association, and dynamic gating and dense feature aggregation strategies guided by prior knowledge, the model outputs the molecular phenotypic classification probability. Specifically, in one application scenario, the judgment threshold is set to 0.5. When the molecular phenotypic classification probability output by the model is greater than or equal to 0.5 (e.g., an output value of 0.78), the sample is determined to belong to the second molecular subtype with evolutionary risk; when the classification probability output by the model is less than 0.5 (e.g., an output value of 0.24), the sample is determined to belong to the first molecular subtype with stability. Finally, based on the above judgment results, a multi-omics automated analysis report is generated and output.

[0020] This embodiment constructs a hierarchical, progressive architecture comprising a single-modal feature extractor, a multi-view correlation layer, and a gating module. This architecture effectively extracts the internal dependencies of molecular omics data and explicitly models cross-modal nonlinear associations. By introducing prior phenotypic features of the samples as dynamic gating signals, the input feature weights are adaptively recalibrated, forming a multi-level fusion representation with strong denoising capabilities and high discriminative power. Thus, without significantly increasing computational complexity, it significantly improves the accuracy, robustness, and interpretability of classifying and assessing the microscopic state of target biological samples under high-dimensional heterogeneous biological data.

[0021] Furthermore, such as Figure 1 As shown, the molecular subtype classification model includes: A multimodal feature input module is used to receive multimodal heterogeneous data of a target biological sample, wherein the multimodal heterogeneous data includes high-dimensional molecular omics data and prior phenotypic features; The data preprocessing and feature reduction module connects the input module to the subsequent single-modality feature extractor. It is used to impute missing values ​​and align modalities in the input high-dimensional molecular omics data, and to perform coarse and fine feature reduction on the high-dimensional molecular omics data based on a two-stage algorithm (mRMR+LASSO), followed by Z-score standardization.

[0022] The standardization module processes the original prior phenotypic features. First, it performs data completion for missing terms, followed by Z-score standardization to eliminate differences in physical dimensions and numerical magnitudes between different phenotypic indicators. Then, the standardized features are input into a multilayer perceptron (MLP). The MLP not only fully exploits the complex high-order nonlinear interactions between phenotypic indicators through its deep network structure to enhance feature representation, but also adaptively maps the low-dimensional phenotypic features in the original physical observation space, precisely aligning them to the same preset hidden layer dimension as the high-dimensional molecular omics data. This process transforms the original phenotypic observation data into prior embedding features with deep semantic associations, providing precise prior control signals for subsequent cross-modal interactions and gating networks.

[0023] A single-modal feature extractor is used to mine and denoise the local feature dependencies within the modality of the high-dimensional molecular omics modal embedding features, and generate single-modal enhanced features; A multi-view association layer is used to extract cross-modal nonlinear associations based on the single-modal enhancement features and the prior embedding features, and generate a global interactive feature vector. The prior perception gating module is used to adaptively adjust the importance weight vectors of each modality based on the single-modality enhancement features and the prior embedding features, and generate a gating fusion feature vector. The hybrid feature aggregation and prediction module is used to perform channel-level concatenation of the global interactive feature vector, the gated fusion feature vector, and the multimodal heterogeneous data through a dense connection strategy, inputting them into a fully connected prediction head to output the molecular phenotype classification probability.

[0024] Specifically, the multimodal feature input module is used to receive multimodal heterogeneous data of the target biological sample (including high-dimensional molecular omics data and prior phenotypic features reflecting the sample phenotype); the high-dimensional molecular omics data includes, but is not limited to, one or more molecular-level features such as gene expression profiles, copy number variations, and DNA methylation.

[0025] Specifically, the Unimodal Feature Extractor (UMFE) is used to mine and denoise the local feature dependencies within the modality of the input high-dimensional molecular omics modality embedding features, and generate a unimodal enhanced feature vector. This module specifically includes: Linear projection unit: maps high-dimensional molecular omics modality embedding features to a unified dimension; Channel-level self-attention unit: Calculates the affinity matrix between feature channels within each modality using the query vector (Q), key vector (K), and value vector (V), and uses the affinity matrix to perform weighted aggregation of the value vector; Scaling residual connection unit: The weighted aggregated features and the high-dimensional molecular omics modality embedding features are weighted and summed according to preset fixed weight coefficients to prevent gradient vanishing in deep networks and preserve low-level semantic information.

[0026] like Figure 2 As shown, the UMFE module mainly consists of three core functional layers, which independently mine the local feature dependencies and denoise within a single molecular omics modality: (1) Linear projection layer: The main function of this layer is to map the input omics features to a specific attention feature space. (2) Channel-level self-attention calculation layer: The function of this layer is to calculate the nonlinear dependencies between different feature channels of the omics data. (3) Scaling residual and normalization layer: The function of this layer is to prevent information loss and gradient vanishing in deep networks and to stabilize feature distribution. In specific implementation, the weighted features are added element-wise to the original input features introduced through the shortcut path and multiplied by a preset scaling factor (such as 0.5), and then the final enhanced single-modality features are output through layer normalization.

[0027] The specific mathematical calculation process is as follows: First, for any high-dimensional molecular omics modality embedding features output by the data preprocessing and feature dimensionality reduction module... , To ensure a unified preset hidden layer dimension, a weight matrix is ​​used to map it to a query vector. Key vector Value vector : ; Subsequently, channel-level self-attention maps are calculated by scaling the dot product, and the affinity matrix is ​​obtained by normalization using the Softmax function. : ; By reweighting the value vector using this matrix, we obtain the weighted features. Finally, the scaled residual connections are combined to generate enhanced features. : ; Among them, coefficient It is the residual scaling factor, which ensures the stable preservation of the original molecular semantic information.

[0028] Furthermore, the Multi-View Correlation Layer (MVCL) receives the single-modal enhanced feature vector and prior embedded features, and traverses all pairwise combinations of the three modal features based on a cross-attention mechanism, calculating the cross-modal correlation features for each pair, thereby constructing a fully connected trimodal interaction graph to extract complex nonlinear cross-modal correlations and generate a global interaction feature vector. This layer specifically includes: Cross-modal attention unit: Using any one single-modal augmentation feature or prior embedding feature as the query vector, and the remaining single-modal augmentation features or prior embedding features as key vectors and value vectors, calculate the cross-modal affinity matrix; Average aggregation unit: Iterates through all possible pairwise modal interaction combinations and averages the interaction features of each combination.

[0029] like Figure 3 As shown, the prior embedding features , is a high-dimensional feature representation obtained by sequentially standardizing and nonlinearly mapping the prior phenotypic features of the target biological sample with a multilayer perceptron. The MVCL module mainly consists of a feature reprojection layer and a cross-modal aggregation layer, which explicitly captures the deep collaborative interaction between different molecular omics modalities and phenotypic features: (1) Cross-attention calculation unit: The function of this unit is to calculate the cross-view correlation between pairs of modalities. (2) Average aggregation layer: The function of this layer is to integrate cross-modal information from multiple perspectives to generate a global representation.

[0030] The specific cross-modal interaction computation is as follows: This layer establishes modality pairs through a cross-attention mechanism. The interaction and correlation between them. For a given modality pair, its attention features The calculation is as follows: ; To obtain panoramic interaction information, all possible pairwise modal interaction combinations are averaged and aggregated to generate global interaction features. : ; Furthermore, the prior perception gating module includes a joint context splicing unit, a dynamic weight generation unit, and a weighted fusion unit; The joint context splicing unit is used to splice the single-modal enhancement feature and the prior embedding feature in the channel dimension to form a global joint context vector; The dynamic weight generation unit is used to output the importance weight vectors corresponding to each modality based on the global joint context vector, using a multilayer perceptron combined with an activation function. The weighted fusion unit is used to perform element-wise weighted summation of the single-modal enhancement features and prior embedding features using the importance weight vector to generate the gated fusion feature vector.

[0031] Specifically, the prior perception gating module (CAGM) receives the single-modality enhanced feature vector and prior embedded features, and adaptively adjusts the weight contributions of each molecular and phenotypic modality feature in the final fusion. This module specifically includes: Joint Context Concatenation Unit: Concatenates single-modal augmented features with prior embedded features along the channel dimension to form a global joint context vector; Dynamic weight generation unit: It adopts a multilayer perceptron combined with an activation function to learn the competitive relationship between different modalities based on the joint context vector, and outputs the importance weight vectors of each modality. Weighted fusion unit: The importance weight vector is used to perform element-wise weighted summation of each modality feature to generate a gated fusion feature vector.

[0032] First, the molecular omics-enhanced features are concatenated with the prior phenotypic features to construct a joint context vector. : ; Gated networks learn the importance weight vectors of each modality. : ; Adaptive fusion is performed on each modality using the generated weights to obtain fused features. : ; in, These are the single-modality enhanced feature vectors output after processing the gene expression profile and copy number variation by the single-modality feature extractor, respectively. These are prior embedding features.

[0033] Specifically, the hybrid feature aggregation and prediction module integrates multi-level features and outputs a final comprehensive state assessment score. Through a dense connection strategy, the gated fusion feature vector, global interactive feature vector, high-dimensional molecular omics feature data, and prior phenotypic features are concatenated at the channel level and then input into the fully connected prediction head, outputting the molecular subtype classification probability for a specific node of the target biological sample.

[0034] Finally, the model concatenates and combines the features generated in each of the above stages to form a global comprehensive evaluation feature. : ; in, These are the original omics embedding features of gene expression profiles and copy number variations before they are input into the network; The input consists of the original prior phenotypic features.

[0035] Furthermore, during the training process of the molecular subtype classification model, Focal Loss and SMOTE oversampling are combined to handle class imbalance, and the AdamW optimizer and cosine annealing strategy are used for end-to-end network training and overfitting prevention.

[0036] In this embodiment, core metrics such as AUC, C-index, and MCC are calculated on an independent test set. Performance is evaluated by extracting the biological associations of feature weights from the attention affinity matrix.

[0037] Compared with previous studies, the method in this embodiment achieves innovative breakthroughs in three core dimensions: At the architectural level, HAG-Net's hierarchical attention and hybrid aggregation design surpasses the shortcomings of traditional models (such as the simple ensemble network En-Cox) in modeling deep cross-modal dependencies, fully verifying the necessity of preserving residual pathways and constructing trimodal fully connected interactions; At the algorithmic level, the prior-aware dynamic gating mechanism significantly improves the signal strength of high-value features compared to traditional indiscriminate feature extraction strategies, and effectively avoids the dimensionality curse and model overfitting risks caused by high-order tensor products; At the interpretability level, by extracting importance coefficients from the attention affinity matrix and dynamic gating, the model can directly associate the decision weights evaluated by the model with specific high-value subsets of genomics or phenotypic features, breaking the "black box" limitation of traditional deep neural networks in multimodal fusion, and providing transparent and reliable data-driven support for subsequent molecular mechanism hypothesis verification and biometric subtype identification.

[0038] Therefore, this embodiment combines the advantages of single-modal local feature mining with cross-modal multi-perspective association. Through prior knowledge-guided dynamic gating and dense feature aggregation strategies, it achieves accurate classification of molecular phenotypes and microstate assessment of target biological samples. It is suitable for multi-dimensional omics data analysis, biomarker subtype identification, and functional genomics research.

[0039] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for molecular subtype classification of biological samples based on a multimodal hierarchical attention network, characterized in that, include: Acquire the data of the sample to be tested and perform preprocessing; The preprocessed sample data is input into the molecular subtype classification model. Through single-modal local feature extraction, cross-modal multi-view association, and prior knowledge-guided dynamic gating and dense feature aggregation strategies, the molecular phenotype classification probability is output. The molecular phenotype classification probability is compared with a threshold to determine the molecular subtype category.

2. The method for classifying molecular subtypes of biological samples according to claim 1, characterized in that, The molecular subtype classification model includes: A multimodal feature input module is used to receive multimodal heterogeneous data of a target biological sample, wherein the multimodal heterogeneous data includes high-dimensional molecular omics data and prior phenotypic features; The data preprocessing and feature dimensionality reduction module is used to impute missing values ​​and align modalities in the high-dimensional molecular omics data, perform coarse and fine dimensionality reduction on the aligned data, and perform standardization on the screened data to obtain high-dimensional molecular omics modal embedding features. The standardization module is used to perform data completion and standardization on prior phenotypic features, and input the standardized features into the multilayer perceptron to obtain prior embedded features. A single-modal feature extractor is used to mine and denoise the local feature dependencies within the modality of the high-dimensional molecular omics modal embedding features, and generate single-modal enhanced features; A multi-view association layer is used to extract cross-modal nonlinear associations based on the single-modal enhancement features and the prior embedding features, and generate a global interactive feature vector. The prior perception gating module is used to adaptively adjust the importance weight vectors of each modality based on the single-modality enhancement features and the prior embedding features, and generate a gating fusion feature vector. The hybrid feature aggregation and prediction module is used to perform channel-level concatenation of the global interactive feature vector, the gated fusion feature vector, and the multimodal heterogeneous data through a dense connection strategy, inputting them into a fully connected prediction head to output the molecular phenotype classification probability.

3. The method for classifying molecular subtypes of biological samples according to claim 2, characterized in that, The single-modal feature extractor includes: Linear projection units are used to map the high-dimensional molecular omics modality embedding features to a unified dimension; Channel-level self-attention units are used to map high-dimensional omics features after unification into query vectors, key vectors, and value vectors through a weight matrix. An affinity matrix is ​​calculated based on the query vectors and key vectors, and the value vectors are weighted and aggregated using the affinity matrix to obtain weighted features. The scaling residual connection unit is used to process the weighted features and the high-dimensional molecular omics modality embedding features using scaling residual connections to generate the single-modality enhanced features.

4. The method for classifying molecular subtypes of biological samples according to claim 3, characterized in that, Generate the single-modal enhancement feature include: ； in, It is the residual scaling factor. It is a high-dimensional molecular omics modal embedding feature. It is a weighted feature.

5. The method for classifying molecular subtypes of biological samples according to claim 2, characterized in that, The multi-view association layer includes a cross-modal attention unit and an average aggregation unit; The cross-modal attention unit is used to calculate the cross-modal affinity matrix based on the single-modal enhancement features and the prior embedding features; The average aggregation unit is used to traverse all pairwise modal interaction combinations and average the cross-modal affinity matrices of each combination to generate the global interaction feature vector.

6. The method for classifying molecular subtypes of biological samples according to claim 5, characterized in that, Calculating the cross-modal affinity matrix includes: Using any one modal feature as the query vector and the remaining modal features as the key vector and value vector, calculate the cross-modal affinity matrix, wherein the modal feature is a single-modal enhancement feature or the prior embedding feature.

7. The method for classifying molecular subtypes of biological samples according to claim 2, characterized in that, The prior perception gating module includes a joint context splicing unit, a dynamic weight generation unit, and a weighted fusion unit; The joint context splicing unit is used to splice the single-modal enhancement feature and the prior embedding feature in the channel dimension to form a global joint context vector; The dynamic weight generation unit is used to output the importance weight vectors corresponding to each modality based on the global joint context vector, using a multilayer perceptron combined with an activation function. The weighted fusion unit is used to perform element-wise weighted summation of the single-modal enhancement features and prior embedding features using the importance weight vector to generate the gated fusion feature vector.

8. The method for classifying molecular subtypes of biological samples according to claim 1, characterized in that, The molecular subtype classification model is trained by combining Focal Loss and SMOTE oversampling to handle class imbalance, and by using the AdamW optimizer and cosine annealing strategy for end-to-end network training.

Images