Autism auxiliary diagnosis method and system based on gene pathway map information
By constructing a multimodal fusion network and combining pathway topology adjacency matrix and graph signal propagation, key genes were screened, which solved the problems of low accuracy and insufficient robustness of pure gene models under feature sparsity, and achieved higher diagnostic accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE CHILDRENS HOSPITAL ZHEJIANG UNIV SCHOOL OF MEDICINE
- Filing Date
- 2026-05-29
- Publication Date
- 2026-06-26
Smart Images
Figure CN122290729A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of bioinformatics and machine learning, and in particular relates to a method and system for auxiliary diagnosis of autism based on gene pathway map information. Background Technology
[0002] Autistic Spectrum Disorder (ASD) is a serious neurodevelopmental disorder, and early diagnosis is crucial for intervention and treatment. With the development of high-throughput gene expression detection technologies, building machine learning or deep learning classification models based on transcriptome data from blood or brain tissue samples has become a hot topic in ASD-assisted diagnostic research.
[0003] Classification and diagnostic techniques based on pure gene expression features directly utilize the expression levels of each gene obtained from gene chips or RNA-seq sequencing as input features. Algorithms such as support vector machines, random forests, or convolutional neural networks are used to train a binary classifier to distinguish ASD patients from healthy controls. However, in practical applications, these pure gene expression models rely heavily on pure gene expression data for model training, which severely limits their diagnostic accuracy and robustness.
[0004] First, pure gene models struggle to capture complex sample classification information, especially when only a very small number of key features are extracted, resulting in a sharp drop in classification performance. Clinical diagnosis prioritizes interpretability and often expects models to rely on only a few differentially expressed genes (such as Top 1 to Top 5) for decision-making. However, existing methods often fail to reflect the complex multi-gene, multi-pathway regulatory mechanisms behind ASD by simply relying on changes in the expression values of individual genes, leading to classification accuracy far below clinical application requirements in cases of extremely sparse features.
[0005] Second, pure gene models are highly sensitive to the partitioning of training data, exhibiting unstable feature subset selection and insufficient robustness. Due to the high dimensionality (tens of thousands of genes) and small sample size (typically tens to hundreds of cases) of gene expression data, variations in training / test set partitioning or cross-validation folds can cause significant shifts in the key gene set selected by the model. This results in large performance fluctuations for classifiers based on pure gene expression on independent validation sets, making it difficult to guarantee the reproducibility of diagnostic results.
[0006] Third, pure gene models neglect the pathway hierarchy and macro-regulatory patterns of biological systems. Biological systems are complex networks where genes interact through intricate biological pathways. The pathogenesis of ASD involves the coordinated dysregulation of multiple pathways, including synaptic function, immune response, and neurodevelopment. However, current methods treat genes merely as isolated features, ignoring the macro-regulatory patterns and biological topology at the pathway level. This "seeing the trees but not the forest" approach prevents the model from utilizing prior biological knowledge to constrain and guide feature selection, thus losing crucial discriminative patterns.
[0007] In summary, there is an urgent need for an ASD-assisted diagnostic method that can integrate gene pathway map structural information and make full use of biological network topology regulation patterns. Summary of the Invention
[0008] The purpose of this invention is to address the problems of poor robustness and large fluctuations in existing pure gene models, and to provide a method and system for autism auxiliary diagnosis based on gene pathway mapping information. This invention achieves more accurate screening of key genes by introducing pathway background knowledge. Even with a significant reduction in feature dimensions, the multimodal fusion network of this invention can still effectively maintain or even improve classification performance.
[0009] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:
[0010] In a first aspect, the present invention provides a method for auxiliary diagnosis of autism based on gene pathway mapping information, comprising the following steps:
[0011] S1: Read the raw sequencing data of the test subjects collected from the clinical setting and preprocess it to obtain a standardized differentially expressed gene matrix that can be directly input into the model;
[0012] S2: Introduce pathway functional features and topological structure information into the standardized differentially expressed gene matrix to construct a pathway topological adjacency matrix; at the same time, reduce the dimensionality of the pre-constructed original pathway feature matrix, and then combine it with the pathway topological adjacency matrix to perform graph signal propagation and feature enhancement to obtain a pathway enhancement input matrix, from which the trained TabNet model extracts gene pathway features.
[0013] S3: Input the standardized differentially expressed gene matrix and gene pathway features into the multimodal fusion network, and complete the initial training of the network through multi-view feature fusion;
[0014] S4: The top K key genes that contribute the most to ASD classification are selected by the importance scores of each gene feature, and the multimodal fusion network is trained again by combining gene pathway features.
[0015] S5: Read the gene expression levels and gene pathway characteristics of the suspected individuals to be tested and input them into the multimodal fusion network that has been trained twice to obtain the auxiliary diagnostic classification results of whether the suspected individuals have autism.
[0016] Based on the above scheme, each step can be implemented in the following preferred manner.
[0017] As a preferred embodiment of the first aspect mentioned above, the specific implementation process for obtaining the standardized differentially expressed gene matrix in S1 is as follows:
[0018] S11: Read the raw sequencing data of the test subjects, which are obtained by sequencing whole blood samples of the test subjects using transcriptome sequencing technology. The test subjects are ASD subjects or healthy controls. Perform quality control and sequence alignment on the raw sequencing data, calculate the FPKM value of each gene as a standardized index of gene expression, and then perform quantitative analysis on the obtained gene expression levels. Based on the quantitative analysis results, use the DESeq2 gene-level differential expression analysis tool to screen out differentially expressed genes with statistically significant differences in gene expression levels between ASD subjects and healthy controls, and then construct an original differentially expressed gene matrix from the differentially expressed gene values.
[0019] S12: The original differentially expressed gene matrix is standardized, and then outlier detection is performed on the processed differentially expressed gene matrix using the interquartile range method. Then, for outliers in the detected differentially expressed gene matrix that exceed the preset detection threshold, the nearest neighbor interpolation method is used for smoothing, thereby obtaining the standardized differentially expressed gene matrix.
[0020] As a preferred embodiment of the first aspect mentioned above, the specific implementation process for obtaining the path enhancement input matrix in S2 is as follows:
[0021] S21: On the standardized differentially expressed gene matrix, conduct gene ontology functional enrichment analysis covering three levels: biological process, cellular component, and molecular function. Obtain the enrichment results of each level and visualize them as directed acyclic graphs. Then, convert the directed acyclic graphs of each level into the form of adjacency matrix as the pathway topology adjacency matrix.
[0022] S22: Using the gene set variation analysis algorithm, calculate the activity score of each detected object in each GO functional pathway, construct the original pathway feature matrix from all activity scores and perform dimensionality reduction to obtain the dimensionality-reduced pathway feature matrix.
[0023] S23: The reduced path feature matrix is used as the initial signal and mapped onto the path topology adjacency matrix of each level. The graph signal propagation processing is completed through multi-order residual fusion, thereby generating the path enhancement input matrix of the fused topology structure features corresponding to each level.
[0024] As a preferred option of the first aspect mentioned above, the specific processing flow in the S3 multimodal fusion network is as follows: the part of the original TabNet model except for the last fully connected layer is used as a feature extractor. The standardized differentially expressed gene matrix is input into the feature extractor to obtain the gene feature matrix. It is then concatenated with the gene pathway features corresponding to each layer according to the layer to obtain the joint feature matrix. This joint feature matrix is then sent to the classification head for binary classification prediction to obtain the classification result.
[0025] As a preferred embodiment of the first aspect, in the secondary training of the multimodal fusion network in S4, the feature extractor of the multimodal fusion network extracts the importance scores corresponding to each gene feature, sorts all gene features in descending order of importance scores, selects the Top K key genes that contribute the most to ASD classification, and retains the differentially expressed gene values of all key genes, thereby reconstructing the differentially expressed gene matrix; the reconstructed differentially expressed gene matrix replaces the standardized differentially expressed gene matrix, generates a new pathway enhancement input matrix according to S2, and then inputs the reconstructed differentially expressed gene matrix into the feature extractor to obtain a new gene feature matrix, which is then concatenated with the new gene pathway features corresponding to each level according to the level to obtain a new joint feature matrix, which is then sent to the classification head for binary classification prediction to obtain the classification result.
[0026] As a preferred embodiment of the first aspect mentioned above, in a multimodal fusion network, the classification head is a fully connected layer.
[0027] In a second aspect, the present invention provides an autism auxiliary diagnostic system based on gene pathway map information, for implementing the autism auxiliary diagnostic method based on gene pathway map information as described in any of the solutions in the first aspect above.
[0028] Thirdly, the present invention provides a computer program product, including a computer program / instructions, which, when executed by a processor, can realize the autism auxiliary diagnosis method based on gene pathway map information as described in any of the solutions of the first aspect above.
[0029] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the autism auxiliary diagnosis method based on gene pathway map information as described in any of the embodiments of the first aspect above.
[0030] Fifthly, the present invention provides a computer electronic device, which includes a memory and a processor;
[0031] The memory is used to store computer programs;
[0032] The processor is configured to, when executing the computer program, implement the autism auxiliary diagnosis method based on gene pathway mapping information as described in any of the above-described first aspects.
[0033] Compared with the prior art, the present invention has the following advantages:
[0034] Improving accuracy and generalization: The multimodal fusion network of this invention can capture micro-expression differences at the single gene level and integrate macro-regulatory patterns at the pathway level, thereby improving the accuracy and generalization of classification tasks.
[0035] Enhanced feature discriminative power and structural information: The graph signal propagation (path topology adjacency matrix multiplication) of the present invention can effectively integrate the path topology structure and enhance the discriminative power of features.
[0036] Mitigating fluctuations and improving performance under extremely sparse features: The pathway background knowledge introduced in this invention alleviates the fluctuations of pure gene models to a certain extent. Pathway background information can effectively supplement the shortcomings of single gene features, maintaining extremely high classification performance even under extremely sparse feature conditions. Attached Figure Description
[0037] Figure 1 This is a flowchart of the steps of the method of the present invention;
[0038] Figure 2 This is a diagram of the multimodal fusion network architecture of the method of the present invention;
[0039] Figure 3 This is a flowchart of the secondary training process of the present invention;
[0040] Figure 4 This is a system block diagram of the present invention;
[0041] Figure 5 This is a schematic diagram of a computer electronic device provided by the present invention. Detailed Implementation
[0042] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.
[0043] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.
[0044] like Figure 1 As shown, in a preferred embodiment of the present invention, the autism auxiliary diagnosis method based on gene pathway map information includes the following steps S1 to S5. The specific implementation process of each step will be described in detail below.
[0045] The core of this invention is to improve the accuracy and biological interpretability of diagnostic models through multimodal feature fusion and interpretable feature screening. The specific implementation process is as follows:
[0046] S1: Read the raw sequencing data of the test subjects collected from the clinical setting and preprocess it to obtain a standardized differentially expressed gene matrix that can be directly input into the model.
[0047] It should be noted that in step S1 of this invention, the core is to extract high-quality, standardized gene features from the raw sequencing data. The specific implementation process is as follows:
[0048] S11: Read the raw sequencing data of the test subjects, which are obtained by sequencing whole blood samples of the test subjects using transcriptome sequencing technology (RNA-seq). The test subjects are ASD subjects or healthy controls. Perform quality control and sequence alignment on the raw sequencing data, calculate the FPKM value (Fragments Per Kilobase of transcript sequence per Million) of each gene as a standardized indicator of gene expression, and then perform quantitative analysis on the obtained gene expression levels. Based on the quantitative analysis results, use the DESeq2 gene-level differential expression analysis tool to screen out differentially expressed genes (DEGs) with statistically significant differences in gene expression levels between ASD subjects and healthy controls, and then construct the original differentially expressed gene matrix (DEG Matrix) from the differentially expressed gene values.
[0049] Further, in step S11 of the present invention, the screening threshold for differentially expressed genes is an adjusted p-value less than 0.05 and... A value greater than or equal to 1 indicates a fold change (FC). The ASD subjects mentioned above refer to patients who have been definitively diagnosed with autism spectrum disorder (ASD), as opposed to healthy controls, i.e., healthy individuals without ASD.
[0050] S12: The original differentially expressed gene matrix is standardized to eliminate dimensional differences between different detection objects; then, the interquartile range (IQR) method is used to detect outliers in the processed differentially expressed gene matrix to identify outliers that deviate from the normal distribution in the feature space; then, for outliers in the detected differentially expressed gene matrix that exceed the preset detection threshold, the nearest neighbor interpolation method is used to smooth them, reducing the interference of high-dimensional transcriptome data noise on the stability of subsequent model training, thereby obtaining the standardized differentially expressed gene matrix.
[0051] S2: Introduce pathway functional features and topological structure information into the standardized differentially expressed gene matrix to construct a pathway topological adjacency matrix; at the same time, reduce the dimensionality of the pre-constructed original pathway feature matrix, and then combine it with the pathway topological adjacency matrix to perform graph signal propagation and feature enhancement to obtain a pathway enhancement input matrix. The trained TabNet model then extracts gene pathway features from the pathway enhancement input matrix.
[0052] It should be noted that step S2 of this invention aims to fuse GO pathway functional features with topological structure information to generate a pathway enhancement input matrix. The specific implementation process is as follows:
[0053] S21: On the standardized differentially expressed gene matrix, conduct gene ontology (GO) functional enrichment analysis covering three levels: biological process (BP), cellular component (CC), and molecular function (MF). Obtain the enrichment results of each level and visualize them as a directed acyclic graph (DAG) to clarify the hierarchical topology of GO functional pathways. Then, convert the DAG of each level into an adjacency matrix as the pathway topology adjacency matrix.
[0054] Furthermore, in step S21 of this invention, the enrichment results in gene ontology functional enrichment analysis can be displayed using a directed acyclic graph (DAG), thereby intuitively showing the hierarchical relationships between GO terms behind the enrichment results. In this DAG, each node represents a GO functional pathway, and each edge represents the association between two GO functional pathways. Subsequently, to facilitate computer processing or graph theory analysis, this invention further constructs an adjacency matrix describing pathway associations based on the gene overlap and semantic similarity between GO terms, namely the aforementioned pathway topological adjacency matrix.
[0055] S22: Using the Gene Set Variation Analysis (GSVA) algorithm, calculate the activity score of each detected object in each GO functional pathway, construct the original pathway feature matrix from all activity scores and perform dimensionality reduction to obtain the dimensionality-reduced pathway feature matrix.
[0056] Furthermore, in step S22 of this invention, the original pathway feature matrix may contain redundant information. Therefore, this invention uses the GSVA algorithm to compress the original pathway feature matrix to the most representative latent space, eliminate the redundancy of the high-dimensional pathway feature matrix, and obtain the dimensionality-reduced pathway feature matrix.
[0057] S23: The reduced-dimensionality path feature matrix As the initial signal, it is mapped onto the path topology adjacency matrix of each level. The graph signal propagation processing is completed through multi-order residual fusion, thereby generating the path enhancement input matrix of the fused topology features corresponding to each level.
[0058] Furthermore, in step S23 of this invention, three pathway enhancement input matrices are generated, which are gene pathway features of biological processes. Gene pathway characteristics of cellular components Gene pathway characteristics of molecular function The mathematical expressions are as follows:
[0059]
[0060]
[0061]
[0062] in, The hyperparameter for the maximum propagation depth is preferably set to 3 in this embodiment; For the first Weighting coefficients of the first-order extension; , , These are the pathway topological adjacency matrices corresponding to biological processes, cellular components, and molecular functions, respectively.
[0063] After obtaining the three pathway enhancement input matrices, this invention inputs the pathway enhancement input matrices into the trained TabNet model and extracts the output of the TabNet model encoder as gene pathway features. The final gene pathway features correspond to the preceding pathway enhancement input matrices and also consist of three types: gene pathway features for biological processes, gene pathway features for cellular components, and gene pathway features for molecular functions.
[0064] S3: Input the standardized differentially expressed gene matrix and gene pathway features into the multimodal fusion network, and complete the initial training of the network through multi-view feature fusion.
[0065] It should be noted that step S3 of this invention improves the TabNet model to achieve multimodal fusion of gene and pathway features, thus completing the initial training of the model.
[0066] In the aforementioned multimodal fusion network, such as Figure 2 As shown, the specific processing flow is as follows: the part of the original TabNet model except for the last fully connected layer is used as a feature extractor. The standardized differentially expressed gene matrix is input into the feature extractor to obtain the gene feature matrix. It is then concatenated with the gene pathway features corresponding to each layer according to the layer to obtain the joint feature matrix. This joint feature matrix is then sent to the classification head for binary classification prediction to obtain the classification result.
[0067] Furthermore, in the multimodal fusion network, the classification head is a fully connected layer.
[0068] Furthermore, in this invention, to balance feature sparsity selection with model fitting capability, the dimensions of the prediction layer and attention layer of the multimodal fusion network are preferably set to 8, and the number of decision steps is set to 3. The sparsity coefficient is set to 0.3, the relaxation factor is set to 0.7, and the mask type is Entmax with the following settings: The model was trained using the Adam optimizer (with an initial learning rate of...). The weight decays to The ReduceLROnPlateau scheduler (in "min" mode, with a decay factor of 0.5 and a patience of 15 epochs) is used to dynamically adjust the learning rate in order to achieve model convergence.
[0069] It should also be noted that in the aforementioned multimodal fusion network, TabularResNet and FTTransformer can replace the original TabNet model as feature extractors. Both schemes are based on the same data source and labeling rules. Furthermore, TabularResNet, FTTransformer, and TabNet are equivalent replacements of the same functional module in this invention, with consistent input and output interfaces.
[0070] Specifically, when using TabularResNet, the input is a standardized differentially expressed gene matrix, which is then concatenated with a pathway enhancement input matrix after feature extraction. QuantileTransformer normalization is performed before training. The model structure consists of an input linear mapping + stacked residual blocks + a classification head, where the residual blocks include linear layers, batch normalization, ReLU, and Dropout. Optimal parameters include a hidden feature size of 64, a stacking layer count of 2, and a random inactivation rate of 0.3. Training uses the AdamW optimizer with an initial learning rate of... The weight decays to We employ labeled smoothed cross-entropy loss with a smoothness level of 0.1, and introduce Mixup and ReduceLROnPlateau scheduling; the number of training rounds is 50, the batch size is 16, and the training / validation ratio is 6:4.
[0071] When using FTTransformer, the input is a standardized differentially expressed gene matrix, which is then concatenated with a pathway enhancement input matrix after feature extraction. The model structure includes numerical feature embedding, CLS labeling, multi-layer Transformer encoding, and a classification head (LayerNorm+ReLU+Linear). When pathway embedding is present, it is concatenated with the CLS representation before outputting logits. The model has a feature dimension of 32, a stacking layer of 6, an attention head of 8, a random inactivation rate of 0.1 on the attention weight matrix, a random inactivation rate of 0.1 in the feed-forward network, and an output layer dimension of 1. Training uses BCEWithLogitsLoss and Adam optimizers with an initial learning rate of 0.001 and ReduceLROnPlateau scheduling. The optimal training epochs are 50, the batch size is 4, and the sigmoid probability threshold of 0.5 is used to determine the class during inference.
[0072] S4: The top K key genes that contribute the most to ASD classification are selected by the importance scores of each gene feature, and the multimodal fusion network is trained again by combining gene pathway features.
[0073] It should be noted that, as Figure 3As shown, step S4 of this invention utilizes the interpretability mechanism of the TabNet model to screen key genes. After reconstructing the data, the model undergoes secondary training. Specifically, the feature extractor of the multimodal fusion network extracts the importance scores corresponding to each gene feature. All gene features are sorted in descending order of importance scores to select the Top K key genes that contribute the most to ASD classification. The differentially expressed gene values of all key genes are retained, thereby reconstructing the differentially expressed gene matrix. The reconstructed differentially expressed gene matrix replaces the standardized differentially expressed gene matrix. New gene pathway features are generated according to step S2. The reconstructed differentially expressed gene matrix is then input into the feature extractor to obtain a new gene feature matrix. This new feature matrix is then concatenated with the new gene pathway features corresponding to each level according to the level to obtain a new joint feature matrix. This joint feature matrix is then sent to the classification head for binary classification prediction to obtain the classification result.
[0074] Furthermore, in this invention, the model's secondary training continues to use the network architecture and hyperparameter settings of S3, only adjusting the initial learning rate to... To adapt to small datasets.
[0075] S5: Read the gene expression levels and gene pathway characteristics of the suspected individuals to be tested and input them into the multimodal fusion network that has been trained twice to obtain the auxiliary diagnostic classification results of whether the suspected individuals have autism.
[0076] It should be noted that in step S5 of this invention, the method for obtaining the gene expression level of the suspected object is similar to that in step S11, the difference being that the object being processed is different. Specifically, in this embodiment, the original sequencing data corresponding to the suspected object is first obtained, that is, the whole blood sample of the suspected object is sequenced using transcriptome sequencing technology. Then, the above-mentioned original sequencing data is subjected to quality control and sequence alignment, and the FPKM value of each gene is calculated as a standardized indicator of gene expression level, thereby obtaining the gene expression level of the suspected object.
[0077] It should be noted that in step S5 of this invention, in addition to obtaining the gene expression levels of the suspected object, gene pathway features also need to be obtained. The method for obtaining gene pathway features of the suspected object is similar to that in steps S1-S2, the only difference being the object being processed. Specifically, in this embodiment, the original sequencing data of the suspected object is first read and preprocessed to obtain a standardized differentially expressed gene matrix that can be directly input into the model. Then, pathway functional features and topological structure information are introduced into the standardized differentially expressed gene matrix to construct a pathway topological adjacency matrix. At the same time, the dimensionality of the pre-constructed original pathway feature matrix is reduced. Then, graph signal propagation and feature enhancement are performed in combination with the pathway topological adjacency matrix to obtain a pathway enhancement input matrix, from which the trained TabNet model extracts gene pathway features.
[0078] The present invention will now demonstrate the application effect of the autism auxiliary diagnosis method based on gene pathway map information described in S1-S5 of the above embodiments on a specific dataset through a specific example, so as to facilitate understanding of the essence of the present invention.
[0079] Example
[0080] The specific implementation process of the autism auxiliary diagnosis method based on gene pathway map information used in this embodiment is as described above and will not be repeated here.
[0081] This embodiment is based on the TabNet model and validated on gene expression datasets and biological pathway topology datasets. During model training and evaluation, all samples were randomly divided into training and test sets in a 6:4 ratio. To eliminate the randomness of a single data partition, this partitioning process was repeated three times independently using different random seeds. All performance metrics reported in this embodiment are the mean and standard deviation (mean ± std) of these multiple independent experiments, thus comprehensively and objectively evaluating the model's robustness to changes in data distribution.
[0082] To verify the contribution of pathway features and pathway topology to model performance, this embodiment sets up an ablation experiment on the whole genome feature set to compare the performance of the pure gene model (TabNet), the model that integrates pathway features (TabNet+passway), and the model that further integrates pathway graph topology (TabNet+passway+DAG).
[0083] As shown in Table 1, the classification accuracy of the pure gene model was 0.92±0.03. In contrast, after introducing pathway information, this invention employed GSVA dimensionality reduction combined with triple DAG signal propagation (L=3), improving the model's classification accuracy to 0.96±0.04 and the Area Under the Receiver Operating Characteristic Curve (AUROC) to 0.98±0.02. This demonstrates that mapping gene expression to a hierarchical functional network can capture more discriminative molecular patterns, significantly improving classification ability. Furthermore, under the same network structure, the F-score after fully integrating pathway features and DAG topology reached 0.96±0.04, while the F-score using only pure gene features was only 0.93±0.03. Simultaneously, experiments show that compared to other dimensionality reduction methods such as t-SNE, GSVA dimensionality reduction better preserves biological representativeness variance and reduces redundancy in high-dimensional data.
[0084] The results above demonstrate that pathway features can supplement the lack of information on single gene expression, while the DAG topology further enhances the model's ability to utilize hierarchical relationships and associations between pathways, thereby improving the accuracy and stability of the ASD-assisted diagnostic model.
[0085] Table 1. Ablation experiments on the recognition performance of TabNet and the method of this invention under different structural integration components.
[0086] The advantages of this invention are even more pronounced when only a very small number of key genes (Top K) are retained. As shown in Table 2, the pure gene model achieved an accuracy of 0.86±0.01 and an F-score of 0.86±0.03 when using only the Top 5 genes. Under the same Top 5 gene condition, by introducing pathway-level structural information, the accuracy of this invention improved to 0.91±0.01, and the F-score reached 0.91±0.02. Furthermore, even under the condition of an extremely sparse Top 1 gene, the accuracy of this invention still reached 0.89±0.05, significantly better than the 0.83±0.01 of the pure gene model.
[0087] Table 2. Performance comparison of different methods for ASD detection
[0088] Experimental results show that, regardless of whether all genes are used or only a very small number of key genes are used, the model that integrates pathway topology in this invention outperforms or matches the performance of the pure gene model. Especially in the case of feature sparsity, the introduction of pathway information effectively compensates for the deficiencies of single gene features, significantly improving the robustness and accuracy of the model.
[0089] It should also be noted that the autism auxiliary diagnosis method based on gene pathway mapping information in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides an autism auxiliary diagnosis system based on gene pathway mapping information, corresponding to the autism auxiliary diagnosis method based on gene pathway mapping information provided in the above embodiments, for implementing the autism auxiliary diagnosis method based on gene pathway mapping information as described in the above embodiments.
[0090] In this embodiment, as Figure 4 As shown, the above-mentioned autism auxiliary diagnostic system includes:
[0091] The data acquisition module is used to read the raw sequencing data of the test subjects collected from the clinical setting and preprocess it to obtain a standardized differentially expressed gene matrix that can be directly input into the model.
[0092] The feature enhancement module is used to introduce pathway functional features and topological structure information into the standardized differentially expressed gene matrix to construct a pathway topological adjacency matrix. At the same time, the pre-constructed original pathway feature matrix is dimensionality reduced, and then combined with the pathway topological adjacency matrix for graph signal propagation and feature enhancement to obtain a pathway enhancement input matrix, from which the trained TabNet model extracts gene pathway features.
[0093] The initial training module is used to input the standardized differentially expressed gene matrix and gene pathway features into the multimodal fusion network, and complete the initial training of the network through multi-view feature fusion.
[0094] The secondary training module is used to select the Top K key genes that contribute the most to ASD classification by the importance scores of each gene feature, and to perform secondary training on the multimodal fusion network by combining gene pathway features.
[0095] The results acquisition module is used to read the gene expression levels and gene pathway characteristics of the suspected subjects to be detected and input them into the multimodal fusion network that has been trained twice to obtain the auxiliary diagnostic classification results of whether the suspected subjects have autism.
[0096] It is understood that the autism auxiliary diagnosis method based on gene pathway mapping information described in S1-S5 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the autism auxiliary diagnosis method based on gene pathway mapping information provided in the above embodiments, which includes a computer program / instructions. When the computer program / instructions are executed by a processor, they can implement the autism auxiliary diagnosis method based on gene pathway mapping information as described in the above embodiments.
[0097] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the autism auxiliary diagnosis method based on gene pathway map information provided in the above embodiments, such as... Figure 5 As shown, it includes a memory and a processor;
[0098] The memory is used to store computer programs;
[0099] The processor is configured to implement the autism auxiliary diagnosis method based on gene pathway map information in the above embodiments when executing the computer program.
[0100] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.
[0101] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the autism auxiliary diagnosis method based on gene pathway map information provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can realize the autism auxiliary diagnosis method based on gene pathway map information in the above embodiments.
[0102] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S5.
[0103] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.
[0104] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0105] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.
[0106] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.
Claims
1. A method for auxiliary diagnosis of autism based on gene pathway mapping information, characterized in that, Includes the following steps: S1: Read the raw sequencing data of the test subjects collected from the clinical setting and preprocess it to obtain a standardized differentially expressed gene matrix that can be directly input into the model; S2: Introduce pathway functional features and topological structure information into the standardized differentially expressed gene matrix to construct a pathway topological adjacency matrix; at the same time, reduce the dimensionality of the pre-constructed original pathway feature matrix, and then combine it with the pathway topological adjacency matrix to perform graph signal propagation and feature enhancement to obtain a pathway enhancement input matrix, from which the trained TabNet model extracts gene pathway features. S3: Input the standardized differentially expressed gene matrix and gene pathway features into the multimodal fusion network, and complete the initial training of the network through multi-view feature fusion; S4: The top K key genes that contribute the most to ASD classification are selected by the importance scores of each gene feature, and the multimodal fusion network is trained again by combining gene pathway features. S5: Read the gene expression levels and gene pathway characteristics of the suspected individuals to be tested and input them into the multimodal fusion network that has been trained twice to obtain the auxiliary diagnostic classification results of whether the suspected individuals have autism.
2. The autism auxiliary diagnostic method based on gene pathway mapping information as described in claim 1, characterized in that, In S1, the specific implementation process for obtaining the standardized differentially expressed gene matrix is as follows: S11: Read the raw sequencing data of the test subjects, which are obtained by sequencing whole blood samples of the test subjects using transcriptome sequencing technology. The test subjects are ASD subjects or healthy controls. Perform quality control and sequence alignment on the raw sequencing data, calculate the FPKM value of each gene as a standardized index of gene expression, and then perform quantitative analysis on the obtained gene expression levels. Based on the quantitative analysis results, use the DESeq2 gene-level differential expression analysis tool to screen out differentially expressed genes with statistically significant differences in gene expression levels between ASD subjects and healthy controls, and then construct an original differentially expressed gene matrix from the differentially expressed gene values. S12: The original differentially expressed gene matrix is standardized, and then outlier detection is performed on the processed differentially expressed gene matrix using the interquartile range method. Then, for outliers in the detected differentially expressed gene matrix that exceed the preset detection threshold, the nearest neighbor interpolation method is used for smoothing, thereby obtaining the standardized differentially expressed gene matrix.
3. The autism auxiliary diagnostic method based on gene pathway mapping information as described in claim 2, characterized in that, In S2, the specific implementation process for obtaining the pathway enhancement input matrix is as follows: S21: On the standardized differentially expressed gene matrix, conduct gene ontology functional enrichment analysis covering three levels: biological process, cellular component, and molecular function. Obtain the enrichment results of each level and visualize them as directed acyclic graphs. Then, convert the directed acyclic graphs of each level into the form of adjacency matrix as the pathway topology adjacency matrix. S22: Using the gene set variation analysis algorithm, calculate the activity score of each detected object in each GO functional pathway, construct the original pathway feature matrix from all activity scores and perform dimensionality reduction to obtain the dimensionality-reduced pathway feature matrix. S23: The reduced path feature matrix is used as the initial signal and mapped onto the path topology adjacency matrix of each level. The graph signal propagation processing is completed through multi-order residual fusion, thereby generating the path enhancement input matrix of the fused topology structure features corresponding to each level.
4. The autism auxiliary diagnostic method based on gene pathway mapping information as described in claim 3, characterized in that, In the S3 multimodal fusion network, the specific processing flow is as follows: the part of the original TabNet model except for the last fully connected layer is used as a feature extractor. The standardized differentially expressed gene matrix is input into the feature extractor to obtain the gene feature matrix. It is then concatenated with the gene pathway features corresponding to each layer according to the layer to obtain the joint feature matrix. This joint feature matrix is then sent to the classification head for binary classification prediction to obtain the classification result.
5. The autism auxiliary diagnostic method based on gene pathway mapping information as described in claim 4, characterized in that, In the secondary training of the S4 multimodal fusion network, the feature extractor of the multimodal fusion network extracts the importance scores corresponding to each gene feature. All gene features are sorted in order of importance scores from high to low. The Top K key genes that contribute the most to ASD classification are selected, and the differential expression gene values of all key genes are retained, thereby reconstructing the differential expression gene matrix. The reconstructed differentially expressed gene matrix is used to replace the standardized differentially expressed gene matrix. New gene pathway features are generated according to S2. The reconstructed differentially expressed gene matrix is then input into the feature extractor to obtain a new gene feature matrix. This new feature matrix is then concatenated with the new gene pathway features corresponding to each level according to the level to obtain a new joint feature matrix. This joint feature matrix is then sent to the classification head for binary classification prediction to obtain the classification result.
6. The autism auxiliary diagnostic method based on gene pathway mapping information as described in claim 5, characterized in that, In a multimodal fusion network, the classification head is a fully connected layer.
7. An autism auxiliary diagnostic system based on gene pathway mapping information, characterized in that, This method is used to implement the autism auxiliary diagnosis method based on gene pathway map information as described in any one of claims 1 to 6.
8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can realize the autism auxiliary diagnosis method based on gene pathway map information as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the autism auxiliary diagnosis method based on gene pathway map information as described in any one of claims 1 to 6.
10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the autism auxiliary diagnosis method based on gene pathway map information as described in any one of claims 1 to 6.