Cancer-related gene comprehensive scoring method and system based on multi-modal deep learning

Abstract

The invention discloses a cancer-related gene comprehensive scoring method and system based on multi-modal deep learning. Comprising the following steps: 1, obtaining scores corresponding to gene expression, copy number variation, methylation and somatic mutation processes, forming a feature vector by using the gene expression score, the copy number variation score, the methylation score and the somatic mutation score, and preprocessing the feature vector to obtain a standardized feature vector; 2, carrying out weight fusion on the standardized feature vectors, and calculating a comprehensive score; 3, constructing a multi-modal integrated model, and training the model by taking the comprehensive score as a model training target; and 4, carrying out score prediction on the genes by utilizing the trained multi-modal integrated model, outputting predicted scores, and carrying out gene list sorting according to the predicted scores. The method is suitable for mining key genes related to cancer occurrence and development from high-throughput sequencing data, and can be applied to tumor molecular mechanism research, prognostic marker screening and precise medical scheme formulation.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics and artificial intelligence, and more specifically to a comprehensive scoring method and system for cancer-associated genes based on multimodal deep learning. Background Technology

[0002] Cancer is a complex disease caused by the accumulation of genomic abnormalities. Its molecular mechanisms involve abnormal regulation of genes at multiple levels, including abnormal gene expression, copy number variations (CNVs), epigenetic modifications (such as methylation), and somatic mutations. Accurately identifying key genes associated with cancer phenotypes is a core prerequisite for understanding tumorigenesis and developing targeted drugs.

[0003] Existing gene scoring methods have the following limitations: 1. Limited Feature Dimension: Most methods rely on only a single set of mathematical data (e.g., based solely on gene expression), neglecting the synergistic effects of molecular events at different levels. For example, some genes may be upregulated through copy number amplification (CNV), and analyzing only expression data cannot distinguish the source of this regulation.

[0004] 2. Crude fusion method: Traditional multi-omics integration methods often use simple weighting (such as arithmetic average) or early fusion (direct splicing of features), which cannot capture the nonlinear relationship between features (such as the nonlinear relationship of methylation inhibiting expression).

[0005] 3. Lack of network perspective: Genes form complex networks through pathways or protein-protein interactions (PPIs). Abnormalities in a single gene may affect the phenotype through network transmission. However, existing methods often treat genes as independent individuals and ignore their contextual relationships.

[0006] 4. Insufficient model generalization ability: Traditional machine learning models (such as linear regression and random forest) have limited ability to extract features from high-dimensional omics data and are difficult to adapt to the differences in data distribution among different cancer types or sample cohorts.

[0007] In recent years, deep learning technology has shown great potential in the field of bioinformatics. Convolutional Neural Networks (CNNs) excel at capturing local feature associations, Transformers can model long-range dependencies through self-attention mechanisms, and Graph Neural Networks (GNNs) can effectively handle network structure data. However, no research has yet organically integrated these three technologies to design an end-to-end solution for multi-omics gene scoring problems.

[0008] Therefore, developing a gene scoring method that integrates multi-omics features, fuses multimodal deep learning architectures, and takes into account both local associations and global network relationships is of great significance for improving the accuracy and biological interpretability of cancer-associated gene screening. Summary of the Invention

[0009] In view of this, the present invention provides a comprehensive scoring method and system for cancer-associated genes based on multimodal deep learning. By eliminating feature dimension differences through standardized preprocessing, the system uses a multimodal model to capture local feature associations, global feature dependencies, and gene network relationships, and finally outputs a comprehensive score through feature fusion, which can efficiently screen cancer driver genes.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: a comprehensive scoring method for cancer-associated genes based on multimodal deep learning, comprising: Step 1: Obtain the scores corresponding to gene expression, copy number variation, methylation, and somatic mutation processes. Construct a feature vector from the gene expression score, copy number variation score, methylation score, and somatic mutation score, and preprocess the feature vector to obtain a standardized feature vector. Step 2: Perform weighted fusion on the standardized feature vectors and calculate the comprehensive score; Step 3: Construct a multimodal ensemble model and train the model using the comprehensive score as the training objective; Step 4: Use the trained multimodal ensemble model to predict gene scores, output predicted scores, and sort the gene list according to the scores.

[0011] Preferably, in step 1, gene expression scores are calculated based on RNA-seq data. Copy number variation score calculated based on GISTIC2.0 data Calculate methylation score based on methylation chip data Somatic mutation score calculated based on exon sequencing data ; Z-score standardization is applied to the four types of features to obtain standardized feature vectors. .

[0012] Preferably, step 2, which calculates the comprehensive score, includes: weighting... The overall score is calculated using the following formula: , By restricting the range to [0, 100] using a truncation function, the formula is as follows: .

[0013] Preferably, the multimodal ensemble model includes a CNN module, a Transformer module, a GNN module, and a feature fusion layer; The CNN module extracts local feature correlations through convolutional layers and outputs features. ; The Transformer module captures global feature dependencies based on a self-attention mechanism and outputs features. ; The GNN module models gene network relationships using graph convolutional layers based on the KNN adjacency matrix, and outputs features. ; The feature fusion layer combines three types of features. , and And use the output of the fully connected layer to predict the score. .

[0014] Preferably, the CNN module captures local dependencies between adjacent features through sliding convolutional kernels, with the following structure: First convolutional layer: Input kernel size The output channel is 16, the activation function is ReLU, and the batch normalization is performed as follows: ; Max pooling: Step size 2, Output ; Second convolutional layer: kernel size Output channels 32, activation functions ReLU, BN: ; Max pooling: Step size 2, Output ; Fully connected layer: compresses features to 32 dimensions and outputs... .

[0015] Preferably, the Transformer module models the long-distance dependency structure of cross-omics features through a self-attention mechanism as follows: Encoder layer: Employs 2-head self-attention, 16-dimensional feedforward network, and a dropout probability of 0.3.

[0016] Multi-head attention is calculated as follows:

[0017] The attention function is: ; Encoder stacking: 2 layers of encoders, output ; Sequence pooling: Means the sequence along its length and outputs the result. ; Fully connected layer: compressed to 32 dimensions, output .

[0018] Preferably, the GNN module captures local dependencies between adjacent features through sliding convolution kernels, with the following structure: First convolutional layer: Input kernel size Output channels: 16; Activation function: ReLU; Batch normalization: ; Max pooling: Step size 2, Output ; Second convolutional layer: kernel size Output channels 32, activation functions ReLU, BN: ; Max pooling: Step size 2, Output ; Fully connected layer: compresses features to 32 dimensions and outputs... .

[0019] Preferably, the feature fusion layer combines three types of features. , and And use the output of the fully connected layer to predict the score. The expression is as follows: ; in: For feature concatenation, the dimension is 32+32+32=96; For fully connected layers, ReLU activation introduces nonlinearity; Dropout is set to 0.3 to prevent overfitting; This is the output layer, which outputs the predicted score. .

[0020] Preferably, the multimodal ensemble model is trained using the Adam optimizer and the MSE loss function, and early stopping is achieved based on the validation set loss. The model is evaluated using MSE, MAE, and R2, and the output is a score based on prediction. A sorted list of genes.

[0021] Preferably, a comprehensive cancer-associated gene scoring system based on multimodal deep learning includes: Data input module: used to obtain the scores corresponding to gene expression, copy number variation, methylation, and somatic mutation processes, construct feature vectors from gene expression scores, copy number variation scores, methylation scores, and somatic mutation scores, and preprocess the feature vectors to obtain standardized feature vectors; The comprehensive score calculation module is used to perform weighted fusion on standardized feature vectors and calculate a comprehensive score. Model training module: used to build a multimodal ensemble model and train the model using the comprehensive score as the training objective; Output module: Uses the trained multimodal ensemble model to predict gene scores, outputs the predicted scores, and sorts the gene list according to the scores.

[0022] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a comprehensive scoring method and system for cancer-associated genes based on multimodal deep learning, with the following beneficial effects: (1) Deep integration of multi-omics features: Gene abnormalities are comprehensively characterized by four core omics features. Compared with single omics methods, the information is more complete and can capture the synergistic effect of molecular events at different levels. (2) Multimodal models complement each other: CNN captures local feature associations, Transformer models global dependencies, and GNN integrates gene network relationships. The fusion of the three achieves comprehensive feature extraction and significantly improves scoring accuracy. (3) Clear biological significance: The feature definition and weight allocation are based on known biological mechanisms (such as the direct impact of expression abnormalities and the regulatory relationship between copy number and expression). The model output is not only highly accurate, but also interpretable. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 The method flowchart provided by the present invention.

[0025] Figure 2 The prediction score ranking chart is provided for an embodiment of the present invention. Detailed Implementation

[0026] 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.

[0027] The following detailed description of the invention, using lung adenocarcinoma data, illustrates the method with a practical example. The specific implementation steps are as follows: First, four feature classes—gene expression, copy number variation, methylation, and somatic mutation—are selected to form a vector. Z-score standardization is used to eliminate dimensional differences, resulting in a standardized feature vector. Next, the standardized features are fused with weights of 0.4 for gene expression, 0.3 for copy number variation, 0.15 for methylation, and 0.15 for somatic mutation. The clip function is used to limit the overall score to [0, 100] as the model training target value. Then, a CNN-Transformer-GNN ensemble model is constructed, adapting it to the input, extracting local, global, and gene network dimensional features, and concatenating them. The predicted score is then output through a fully connected layer. Finally, the training and test sets are divided in an 8:2 ratio. Within the training set, a sub-training and validation set are further divided in an 8:2 ratio. The MSE loss function and an early stopping mechanism ("stop if the validation set loss does not decrease for 10 consecutive rounds") are used to train the model. Finally, cancer-associated key genes are output in order of predicted score.

[0028] like Figure 1 As shown in the figure, this invention discloses a comprehensive scoring method for cancer-associated genes based on multimodal deep learning, including: Step 1: Obtain the scores corresponding to gene expression, copy number variation, methylation, and somatic mutation processes. Construct a feature vector from the gene expression score, copy number variation score, methylation score, and somatic mutation score, and preprocess the feature vector to obtain a standardized feature vector. Step 2: Perform weighted fusion on the standardized feature vectors and calculate the comprehensive score; Step 3: Construct a multimodal ensemble model and train the model using the comprehensive score as the training objective; Step 4: Use the trained multimodal ensemble model to predict gene scores, output predicted scores, and sort the gene list according to the scores.

[0029] Specifically, in step 1, gene expression scores are calculated based on RNA-seq data. Copy number variation score calculated based on GISTIC2.0 data Calculate methylation score based on methylation chip data Somatic mutation score calculated based on exon sequencing data ; Z-score standardization is applied to the four types of features to obtain standardized feature vectors. .

[0030] In one specific embodiment of the present invention, four types of gene features with clear biological significance are selected to form a feature vector. The calculation methods for each feature are as follows: Gene expression score ( , : Calculated based on the FPKM (Fragments Per Kilobase of transcript per Million mapped reads) value from RNA sequencing (RNA-seq), reflecting the degree of abnormality in gene transcription activity. The calculation formula is: ; in, Indicates gene In the sample The amount of expression in The total number of samples, and Genes The expression mean and standard deviation across all samples. This formula, after standardization by Z-score, takes the mean of absolute values ​​to quantify the degree to which gene expression deviates from the overall level; a higher value indicates a more significant expression abnormality.

[0031] Copy number variation score ( , (This refers to data on copy number variation, reflecting the frequency and intensity of gene copy number abnormalities. The calculation formula is:) ; in, The proportion of samples in which gene copy number amplification occurred. This represents the proportion of samples with copy number deletion. Weighting is based on the fact that amplification typically has a stronger effect on gene dosage than deletion (amplification may directly lead to overexpression, while heterozygous deletion has a weaker effect), therefore, weights of 1.0 and 0.8 are assigned respectively.

[0032] Methylation score ( , ): Based on the β value (reflecting the methylation level of CpG sites, ranging from [0,1]) of an Illumina methylation chip, this quantifies methylation abnormalities in gene promoter regions. The calculation formula is as follows: ; in, This represents the number of CpG sites in the gene promoter region. For genes No. The average methylation level of each CpG site in the tumor sample This represents the average level of this site in normal control samples. The mean absolute value of methylation difference between tumor and normal samples reflects the degree of abnormal gene epigenetic regulation.

[0033] Somatic mutation score ( , ): Based on variant call format (VCF) data from exon sequencing (WES), combining mutation frequency and harmfulness. The calculation formula is: ; in, For genes Mutation frequency in the sample set (number of mutated samples / total number of samples). For genes No. The harm score of each mutation is used, and the mean score reflects the overall harm of the mutation.

[0034] Furthermore, since the dimensions and distributions of the four types of features differ significantly (e.g., expression score range [0,5], mutation score range [0,1]), Z-score standardization is needed to eliminate the influence of dimensions. The formula is: ; in, and The first The mean and standard deviation of class features across all genes. After standardization, the features... It follows a normal distribution with a mean of 0 and a standard deviation of 1, ensuring that the model learns fairly for different features.

[0035] Specifically, step 2, calculating the comprehensive score, includes: weighting... The overall score is calculated using the following formula: , By restricting the range to [0, 100] using a truncation function, the formula is as follows: .

[0036] In one specific embodiment of the present invention, the comprehensive score target value is constructed as the training model, and the "true comprehensive score" of the gene needs to be constructed based on prior biological knowledge as the target value. Considering the varying contributions of different omics features to cancer phenotypes, a differential weighting method was used to fuse standardized features: ; Weighting basis: Gene expression (0.4) is a direct manifestation of functional abnormalities and has the most significant impact on phenotype; Copy number variation (0.3) directly alters gene expression through dose-effect, with a secondary contribution; Methylation (0.15) indirectly affects expression through epigenetic regulation, while somatic mutations (0.15) need to accumulate to a certain frequency to drive carcinogenesis; both contribute relatively little.

[0037] To make the scoring more consistent with real-world application scenarios (such as a 0-100 score scale), a truncation function is used to limit the range: .

[0038] Specifically, the multimodal ensemble model includes a CNN module, a Transformer module, a GNN module, and a feature fusion layer; The CNN module extracts local feature correlations through convolutional layers and outputs features. ; The Transformer module captures global feature dependencies based on a self-attention mechanism and outputs features. ; The GNN module models gene network relationships using graph convolutional layers based on the KNN adjacency matrix, and outputs features. ; The feature fusion layer combines three types of features. , and And use the output of the fully connected layer to predict the score. .

[0039] In a specific embodiment of the present invention, the model input adaptation processing includes: CNN input: Standardized feature vector To adapt to the requirements of convolutional layers, the dimension is expanded to... (Number of channels: 1, sequence length: 4).

[0040] Transformer input: Position-encoded sequence features are added to distinguish the biological significance of different omics features. The position encoding formula is:

[0041] in This is a feature index (0 corresponds to expression, 1 to CNV, 2 to methylation, and 3 to mutation). The encoded input is... After dimensional expansion, it becomes (Sequence length 4, feature dimension 4).

[0042] GNN input: contains feature vectors And gene adjacency matrix ( (Total number of genes). The adjacency matrix is ​​constructed using KNN (K=5): Feature similarity is calculated using cosine similarity: This reflects the similarity (potential functional association) of genes in a multi-omics feature space.

[0043] Specifically, the CNN module captures local dependencies between adjacent features through sliding convolutional kernels, with the following structure: First convolutional layer: Input kernel size The output channel is 16, the activation function is ReLU, and the batch normalization is performed as follows: ; Max pooling: Step size 2, Output ; Second convolutional layer: kernel size Output channels 32, activation functions ReLU, BN: ; Max pooling: Step size 2, Output ; Fully connected layer: compresses features to 32 dimensions and outputs... .

[0044] Specifically, the Transformer module models the long-distance dependency structure of cross-omics features through a self-attention mechanism as follows: Encoder layer: Employs 2-head self-attention, 16-dimensional feedforward network, and a dropout probability of 0.3.

[0045] Multi-head attention is calculated as follows:

[0046] The attention function is: ; Encoder stacking: 2 layers of encoders, output ; Sequence pooling: Means the sequence along its length and outputs the result. ; Fully connected layer: compressed to 32 dimensions, output .

[0047] Specifically, the GNN module captures local dependencies between adjacent features through sliding convolutional kernels, and its structure is as follows: First convolutional layer: Input kernel size Output channels: 16; Activation function: ReLU; Batch normalization: ; Max pooling: Step size 2, Output ; Second convolutional layer: kernel size Output channels 32, activation functions ReLU, BN: ; Max pooling: Step size 2, Output ; Fully connected layer: compresses features to 32 dimensions and outputs... .

[0048] Specifically, the feature fusion layer combines three types of features. , and And use the output of the fully connected layer to predict the score. The expression is as follows: ; in: For feature concatenation, the dimension is 32+32+32=96; For fully connected layers, ReLU activation introduces nonlinearity; Dropout is set to 0.3 to prevent overfitting; This is the output layer, which outputs the predicted score. .

[0049] Specifically, the multimodal ensemble model is trained using the Adam optimizer and the MSE loss function, and early stopping is achieved based on the validation set loss. The model is evaluated using MSE, MAE, and R2, and the output is a score based on prediction. A sorted list of genes.

[0050] In a specific embodiment of the present invention, the training parameter configuration includes: Dataset partitioning: The dataset is divided into training and test sets in an 8:2 ratio. The training set is further divided into sub-training and validation sets in an 8:2 ratio (for early stopping).

[0051] Loss function: Mean Squared Error (MSE), which measures the difference between the predicted and actual values. ; in For batch size, , The first in each batch Predicted and true scores for each sample.

[0052] Early stopping mechanism: When the validation set loss does not decrease for 10 consecutive rounds, stop training and save the optimal model to avoid overfitting.

[0053] The trained model is used to predict gene scores, and the predicted gene scores are displayed in order of priority to obtain key genes associated with cancer.

[0054] Specifically, a comprehensive cancer-associated gene scoring system based on multimodal deep learning includes: Data input module: used to obtain the scores corresponding to gene expression, copy number variation, methylation, and somatic mutation processes, construct feature vectors from gene expression scores, copy number variation scores, methylation scores, and somatic mutation scores, and preprocess the feature vectors to obtain standardized feature vectors; The comprehensive score calculation module is used to perform weighted fusion on standardized feature vectors and calculate a comprehensive score. Model training module: used to build a multimodal ensemble model and train the model using the comprehensive score as the training objective; Output module: Uses the trained multimodal ensemble model to predict gene scores, outputs the predicted scores, and sorts the gene list according to the scores.

[0055] In another specific embodiment of the present invention, the EMT pathway lays a key foundation for cancer metastasis by causing cancer cells to lose their epithelial characteristics, acquire stromal characteristics, and have stronger migration and invasive capabilities. Genes within this pathway are closely related to cancer development. This embodiment successfully predicted the scores of 479 EMT pathway genes by dividing the data from 447 samples into training and validation sets at an 8:2 ratio and ranking them according to priority. Figure 2 The top 30 genes with the highest prediction scores out of 479 genes were visualized, such as... Figure 2 As shown, the multimodal deep learning-based cancer-associated gene comprehensive scoring system designed in this embodiment has high accuracy and provides an effective tool for screening key cancer genes.

[0056] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0057] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for comprehensive scoring of cancer-associated genes based on multi-modal deep learning, characterized in that, Comprise: Step 1: obtain the scores corresponding to gene expression, copy number variation, methylation, somatic mutation process, construct a feature vector with gene expression score, copy number variation score, methylation score, somatic mutation score, and pretreat the feature vector to obtain a standardized feature vector; Step 2: weight fusion is performed on the standardized feature vector, and a comprehensive score is calculated; Step 3: a multi-modal integrated model is constructed, and the comprehensive score is used as the model training target to train the model; Step 4: the trained multi-modal integrated model is used to score and predict genes, output the predicted score, and sort the gene list according to the score.

2. The multi-modal deep learning based cancer associated gene comprehensive scoring method according to claim 1, characterized in that, Step 1: Calculate gene expression score based on RNA-seq data Step 2: Calculate copy number variation score based on GISTIC2.0 data Step 3: Calculate methylation score based on methylation microarray data Step 4: Calculate somatic mutation score based on exome sequencing data ; The four types of features are Z-score standardized to obtain the standardized feature vector . 3.The method of claim 2, wherein, The step 2 calculates the comprehensive score by the following process: calculating the weighted score of each index according to the weight The comprehensive score is calculated by the following formula: ； The truncation function is limited in the range of [0, 100], and the formula is as follows: 。 4.The method of claim 1, wherein, The multi-modal integrated model comprises a CNN module, a Transformer module, a GNN module and a feature fusion layer; The CNN module extracts local feature correlation through a convolution layer and outputs features ; The Transformer module captures global feature dependencies based on self-attention mechanism, outputting features ; The GNN module models the gene network relationship based on a graph convolution layer of a KNN adjacency matrix, and outputs features ; A feature fusion layer fuses the three types of features , and and outputs a prediction score using a fully connected layer .

5. The multi-modal deep learning based cancer associated gene comprehensive scoring method according to claim 4, characterized in that, The CNN module captures the local dependence between adjacent features through a sliding convolution kernel, and the structure is as follows: First layer convolution: input , convolution kernel size , output channel 16, activation function ReLU, batch normalization, formula as follows: ； Max pooling: stride 2, output ; Second layer convolution: kernel size , output channel 32, activation function ReLU, BN: ； Max pooling: stride 2, output ; Fully connected layer: compress the features to 32 dimensions, output .

6. The multi-modal deep learning based cancer associated gene comprehensive scoring method according to claim 4, characterized in that, The Transformer module models the long-distance dependence structure of cross-omics features through self-attention mechanism, and the structure is as follows: The encoder layer adopts 2-head self-attention, the feedforward network dimension is 16, and the dropout probability is 0.3: The multi-head attention calculation is as follows: The attention function is as follows: ； Encoder stack: 2 layers of encoders, output ; Sequence pooling: take mean over sequence dimension, output ; Fully connected layer: compressed to 32 dimensions, output .

7. The multi-modal deep learning based cancer associated gene comprehensive scoring method according to claim 4, characterized in that, The GNN module captures the local dependence between adjacent features through a sliding convolution kernel, and the structure is as follows: First layer convolution: input , kernel size , output channels 16, activation function ReLU, batch normalization: ； Max pooling: stride 2, output ; Second layer convolution: kernel size , output channel 32, activation function ReLU, BN: ； Max pooling: stride 2, output ; Fully connected layer: compress the features to 32 dimensions, output . 8.The method of claim 4, wherein, The feature fusion layer outputs a prediction score by concatenating the three types of features , and and using a fully connected layer , expressed as follows: ； Wherein: For feature stitching, the dimension is 32+32+32=96; For the fully connected layer, the ReLU activation introduces nonlinearity. Dropout is set to 0.3 to prevent overfitting; For the output layer, output the prediction score . 9.The method of claim 4, wherein, The multi-modal integrated model is trained using Adam optimizer and MSE loss function, and early stopping is based on the validation set loss; By MSE, MAE, R 2 Evaluate model, output by predicted score Ranked gene list. 10.A multi-modal deep learning based cancer-associated gene comprehensive scoring system, characterized in that, Comprise: Data input module: used to obtain the scores corresponding to gene expression, copy number variation, methylation, somatic mutation process, construct a feature vector with gene expression score, copy number variation score, methylation score, somatic mutation score, and pretreat the feature vector to obtain a standardized feature vector; Comprehensive score calculation module: used to weight fuse the standardized feature vector and calculate the comprehensive score; Model training module: used to construct a multi-modal integrated model, and the comprehensive score is used as the model training target to train the model; Output module: the trained multi-modal integrated model is used to score and predict genes, output the predicted score, and sort the gene list according to the score.

Images