An AI-based gene data analysis system

Abstract

The application discloses a kind of gene data analysis systems based on AI, through multi-omics data matrix;Through 1D-CNN one-dimensional convolutional neural network, local correlation pattern mining is carried out to the multi-omics data matrix, the topological structure of gene network is identified by persistent homology analysis, and dynamic weight is allocated to sequence feature and topological feature matrix using dynamic attention mechanism, and the weighted multi-scale feature vector is output;Based on the architecture of Transformer, a multi-modal fusion model is established to fuse the multi-scale feature vector, and the general genomic features of the pre-training model are fine-tuned to adapt to the disease data set, and the fusion feature vector is output;The fusion feature vector is input into MLP multilayer perceptron for disease risk prediction, and a disease risk prediction index is generated by combining SHAP algorithm, and an auxiliary decision scheme is generated according to the prediction index. The accuracy and generalization ability of disease risk classification are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to an AI-based gene data analysis system. Background Technology

[0002] Traditional genomic data analysis methods have several shortcomings. The sheer volume of genomic data makes it difficult to comprehensively reveal the complex molecular mechanisms of diseases from a single omics dataset. Disease development is the result of multiple factors, involving gene sequence variations, abnormal epigenetic regulation, and protein expression disorders; relying solely on data from one omics dataset may lead to biased analysis results and low efficiency. Traditional data analysis methods are limited in handling high-dimensional, highly complex multi-omics datasets. Traditional statistical methods struggle to handle nonlinear relationships and complex dependencies in genomic data, resulting in low accuracy in disease risk prediction. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an AI-based gene data analysis system.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the aforementioned AI-based gene data analysis system:

[0005] The gene data acquisition module is used to acquire the patient's genomic data, epigenetic data, and proteomic data, and to process the data to obtain a multi-omics data matrix.

[0006] The feature vector extraction module is used to mine local association patterns in the multi-omics data matrix through a 1D-CNN one-dimensional convolutional neural network, identify the topological structure of the gene network through continuous homology analysis, and assign dynamic weights to the sequence features and topological feature matrix using a dynamic attention mechanism to output a weighted multi-scale feature vector.

[0007] The fusion feature generation module is used to build a multimodal fusion model based on the Transformer architecture to fuse the multi-scale feature vectors, and to fine-tune and adapt the disease dataset using the general genomic features of the pre-trained model AlphaGenome, and output the fused feature vector.

[0008] The gene data analysis module is used to input the fused feature vector into the MLP multilayer perceptron for disease risk prediction, combine it with the SHAP algorithm to generate a disease risk prediction index, and generate an auxiliary decision-making scheme based on the prediction index.

[0009] Furthermore, in the aforementioned AI-based gene data analysis system, the gene data acquisition module includes the following units:

[0010] The format conversion and cleaning unit is used to acquire the patient's genomic data, epigenetic data, and proteomic data, convert the raw sequencing data into a unified format, and remove low-quality reads to obtain cleaned gene data.

[0011] k-mer vectorization units are used to perform k-mer preprocessing on DNA sequences, breaking long sequences into short fragments and generating vectorized gene features;

[0012] The multi-omics data integration unit is used to align genomic data, epigenomic data, and proteomic data to the same gene coordinates to generate a multimodal data matrix, which includes at least gene expression, methylation status, and protein abundance.

[0013] Furthermore, in the aforementioned AI-based gene data analysis system, the feature vector extraction module includes the following units:

[0014] The convolutional computation unit is used to perform a sliding scan of the input sequence using a large convolutional kernel; and to perform secondary extraction of the feature map output from the upper layer using a small convolutional kernel.

[0015] The feature generation unit is used to retain the most significant feature values ​​in a local region and the combination of methylation features most related to gene expression changes through max pooling, and outputs a sequence feature matrix that reflects local molecular associations.

[0016] Furthermore, in the aforementioned AI-based gene data analysis system, the feature vector extraction module further includes the following units:

[0017] The association scoring unit is used to establish association scores between genes based on co-variation trends in gene expression levels, evidence of physical binding between proteins, and synergistic regulatory relationships of epigenetic modifications.

[0018] Gene association units are used to connect gene pairs with scores higher than a set threshold to form a preliminary gene association network, where each node represents a gene and the thickness of the edges indicates the association strength.

[0019] The threshold adjustment unit is used to gradually lower the threshold from gene pairs with high association strength, so as to include gene pairs with weak association strength into the network.

[0020] A structure recording unit is used to track and record key topological structures that occur, including at least connected components, loop structures, and hub nodes.

[0021] The feature transformation unit is used to calculate the duration of existence of each topological structure and transform the topological features into a numerical topological feature matrix.

[0022] Furthermore, in the aforementioned AI-based gene data analysis system, the feature vector extraction module further includes the following units:

[0023] The matrix input unit is used to input the sequence feature matrix output by 1D-CNN and the topological feature matrix obtained by continuous coherence analysis into the dynamic attention mechanism;

[0024] The feature extraction unit is used to calculate the channel-dimensional attention weights for sequence features and topological features respectively. Global average pooling is used to extract global information of feature channels to obtain the initial attention weight vector.

[0025] The weight allocation unit is used to allocate spatial dimensions of the feature sequence using a self-attention mechanism, capture long-distance dependencies between features, and output a weighted multi-scale feature vector.

[0026] Furthermore, in the aforementioned AI-based gene data analysis system, the fusion feature generation module includes the following units:

[0027] The feature processing unit is used to standardize multi-scale feature vectors and add positional encoding information, and uses the Transformer's attention head to focus on the direct interaction between genomic and proteomic features.

[0028] The vector output unit is used to process the multi-head self-attention process and then feed forward neural network, strengthen important features through nonlinear transformation, and output a fused feature vector adapted to the disease.

[0029] Furthermore, in the aforementioned AI-based gene data analysis system, the gene data analysis module includes the following units:

[0030] The multilayer perceptron unit is used to determine the first hidden layer of the MLP multilayer perceptron to screen out key pathway features related to the disease, the second layer to focus on specific gene regulatory modules, and the output layer to convert the information into a disease risk probability index, and generate auxiliary decision-making schemes based on the predicted probability index.

[0031] Furthermore, in a method for implementing an AI-based gene data analysis system, the method includes the following steps:

[0032] We acquire patients’ genomic, epigenomic, and proteomic data, process the data, and obtain a multi-omics data matrix.

[0033] The multi-omics data matrix is ​​subjected to local association pattern mining through a 1D-CNN one-dimensional convolutional neural network. The topological structure of the gene network is identified through continuous homology analysis. Dynamic attention mechanism is used to assign dynamic weights to sequence features and topological feature matrices, and a weighted multi-scale feature vector is output.

[0034] A multimodal fusion model based on the Transformer architecture is established to fuse the multi-scale feature vectors. The general genomic features of the pre-trained model AlphaGenome are used to fine-tune and adapt to the disease dataset, and the fused feature vector is output.

[0035] The fused feature vector is input into an MLP (Multilayer Perceptron) for disease risk prediction. The disease risk prediction index is generated by combining the SHAP (Shape Algorithm) with the prediction index, and an auxiliary decision-making scheme is generated based on the prediction index.

[0036] Furthermore, in a method for implementing an AI-based gene data analysis system, the method includes the following steps:

[0037] The sequence feature matrix output by 1D-CNN and the topological feature matrix obtained from continuous cohomology analysis are input into the dynamic attention mechanism;

[0038] Attention weights are calculated at the channel dimension for both sequence features and topological features. Global average pooling is used to extract global information of the feature channels to obtain the initial attention weight vector.

[0039] The self-attention mechanism is used to assign spatial dimension weights to the feature sequence, capture the long-distance dependencies between features, and output a weighted multi-scale feature vector.

[0040] Furthermore, in a method for implementing an AI-based gene data analysis system, the method includes the following steps:

[0041] After standardizing the multi-scale feature vectors, positional encoding information is added, and the attention head of the Transformer is used to focus on the direct interaction between genomic and proteomic features.

[0042] After undergoing multi-head self-attention processing, the data enters a feedforward neural network. Important features are enhanced through nonlinear transformations, and a fusion feature vector adapted to the disease is output.

[0043] Its beneficial effects are as follows: 1. It can accurately extract core information related to disease risk from fused features. Combined with optimization methods such as cross-validation and regularization, it effectively improves the accuracy and generalization ability of disease risk classification, providing reliable risk judgment results for clinical practice and helping to identify high-risk groups for diseases at an early stage. 2. By integrating multi-source data from the genome, epigenome, and proteome, the model can capture the complex relationships between gene expression, methylation status, and protein interactions, significantly improving the accuracy of disease risk prediction. 3. It overcomes the limitations of traditional single-modality modeling. Simultaneously, by combining the SHAP algorithm and knowledge distillation technology, the model achieves lightweight deployment while ensuring high accuracy, meeting the real-time analysis needs of edge devices (portable DNA sequencers). Attached Figure Description

[0044] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0045] Figure 1 This is a schematic diagram of the first embodiment of an AI-based gene data analysis system according to the present invention;

[0046] Figure 2 This is a schematic diagram of a second embodiment of an AI-based gene data analysis system according to the present invention;

[0047] Figure 3 This is a schematic diagram of a third embodiment of an AI-based gene data analysis system according to the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0050] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, an AI-based gene data analysis system includes the following modules:

[0051] The gene data acquisition module is used to acquire the patient's genomic data, epigenetic data, and proteomic data, and to process the data to obtain a multi-omics data matrix.

[0052] Specifically, this embodiment also includes a format conversion and cleaning unit, which is used to acquire the patient's genomic data, epigenetic data and proteomic data, convert the raw sequencing data into a unified format, and remove low-quality reads to obtain cleaned gene data;

[0053] k-mer vectorization units are used to perform k-mer preprocessing on DNA sequences, breaking long sequences into short fragments and generating vectorized gene features;

[0054] The multi-omics data integration unit is used to align genomic data, epigenomic data, and proteomic data to the same gene coordinates to generate a multimodal data matrix, which includes at least gene expression, methylation status, and protein abundance.

[0055] Specifically,

[0056] (I) Multi-omics data collection

[0057] Genomic data: Acquired through technologies such as whole-genome sequencing (WGS), whole-exome sequencing (WES), or targeted gene sequencing, covering variant information such as single nucleotide polymorphisms (SNPs), insertions / deletions (InDels), and copy number variations (CNVs). Data originates from gene sequencing experiments of clinical samples and must comply with ethical guidelines and data privacy protection requirements.

[0058] Epigenome data includes DNA methylation data (obtained via Illumina Infinium microarray or methylation sequencing), histone modification data (ChIP-seq data), and chromatin accessibility data (ATAC-seq data).

[0059] Proteomic data: The expression levels and modification states (phosphorylation, glycosylation, etc.) of proteins in patient samples were determined using mass spectrometry (LC-MS / MS).

[0060] (ii) Data preprocessing

[0061] Data cleaning

[0062] For genomic data, quality control tools (FastQC) are used to assess quality and remove low-quality sequencing reads (reads with more than 50% of bases having a Phred quality value below 20). Sequencing data are aligned to a reference genome using alignment tools (BWA) to remove repetitive sequences (using Picard tools) and PCR duplications.

[0063] In epigenomic data, DNA methylation data need to undergo background correction and probe filtering (removing cross-reactive probes and polymorphic probes); ChIP-seq data need to undergo peak calling (using the MACS2 tool) to filter low-confidence peaks.

[0064] The proteomic data were matched and quantified using peptide identification software (MaxQuant), proteins with low identification confidence were removed, and missing values ​​were imputed using a Gaussian mixture model-based method.

[0065] Standardized processing

[0066] SNP genotypes in genomic data were converted to 0 / 1 / 2 codes (representing homozygous reference allele, heterozygote, and homozygous alternative allele, respectively), and CNV data were converted to relative copy number values.

[0067] The epigenetic group data were standardized using the quantile normalization method to ensure a consistent data distribution across different samples.

[0068] The proteomic data were centered by median or z-score normalization to eliminate systematic errors between different samples.

[0069] Data integration

[0070] The preprocessed genomic, epigenomic, and proteomic data were aligned according to sample IDs to construct a multi-omics data matrix with consistent dimensions. Rows represented samples, and columns represented different molecular features (genes, methylation sites, proteins, etc.).

[0071] The feature vector extraction module is used to mine local association patterns in the multi-omics data matrix through a 1D-CNN one-dimensional convolutional neural network, identify the topological structure of the gene network through continuous homology analysis, and assign dynamic weights to the sequence features and topological feature matrix using a dynamic attention mechanism to output a weighted multi-scale feature vector.

[0072] Specifically, this embodiment also includes a convolution calculation unit, which is used to perform sliding scanning of the input sequence using a large convolution kernel; and to perform secondary extraction of the feature map output by the upper layer using a small convolution kernel;

[0073] The feature generation unit is used to retain the most significant feature values ​​in a local region and the combination of methylation features most related to gene expression changes through max pooling, and outputs a sequence feature matrix that reflects local molecular associations.

[0074] The association scoring unit is used to establish association scores between genes based on co-variation trends in gene expression levels, evidence of physical binding between proteins, and synergistic regulatory relationships of epigenetic modifications.

[0075] Gene association units are used to connect gene pairs with scores higher than a set threshold to form a preliminary gene association network, where each node represents a gene and the thickness of the edges indicates the association strength.

[0076] The threshold adjustment unit is used to gradually lower the threshold from gene pairs with high association strength, so as to include gene pairs with weak association strength into the network.

[0077] A structure recording unit is used to track and record key topological structures that occur, including at least connected components, loop structures, and hub nodes.

[0078] The feature transformation unit is used to calculate the duration of existence of each topological structure and transform the topological features into a numerical topological feature matrix.

[0079] The matrix input unit is used to input the sequence feature matrix output by 1D-CNN and the topological feature matrix obtained by continuous coherence analysis into the dynamic attention mechanism;

[0080] The feature extraction unit is used to calculate the channel-dimensional attention weights for sequence features and topological features respectively. Global average pooling is used to extract global information of feature channels to obtain the initial attention weight vector.

[0081] The weight allocation unit is used to allocate spatial dimensions of the feature sequence using a self-attention mechanism, capture long-distance dependencies between features, and output a weighted multi-scale feature vector.

[0082] Specifically,

[0083] (a) Data input processing

[0084] The preprocessed multi-omics data matrix is ​​arranged in biological order of molecular features (sorted by gene position on chromosomes) to ensure that adjacent features are biologically correlated. Genes arranged consecutively on a chromosome and their corresponding methylation and protein expression data are then concatenated to form sequence data suitable for one-dimensional convolution processing.

[0085] (ii) Feature capture of convolutional layers

[0086] The first convolutional layer uses a large convolutional kernel (covering a window of 5 consecutive features) to perform a sliding scan of the input sequence. This process is similar to local observation on a gene sequence, capturing the cooperative change patterns between 5 adjacent molecular features.

[0087] Subsequent convolutional layers: Using smaller convolutional kernels (3 consecutive feature windows), the feature maps output from the upper layer are extracted a second time.

[0088] (III) Pooling and Feature Selection

[0089] After each convolutional operation, max pooling is used to retain the most significant feature values ​​in the local region, reducing data dimensionality while highlighting key patterns. In regions containing multiple methylation sites, the combination of methylation features most relevant to gene expression changes is retained, ultimately outputting a sequence feature matrix reflecting local molecular associations.

[0090] II. Continuous Homologization Analysis for Identifying Gene Network Topology

[0091] (I) Construction of gene association networks

[0092] The interaction strength between all genes is calculated based on multi-omics data, such as co-variation trends in gene expression levels, evidence of physical binding between proteins, and synergistic regulatory relationships of epigenetic modifications, to establish an association score between genes. Gene pairs with scores higher than a set threshold are connected to form a preliminary gene association network, where each node represents a gene and the thickness of the edges indicates the association strength.

[0093] (ii) Dynamic network topology tracing

[0094] Gradually expand the network: Starting with the gene pairs with the strongest association, gradually lower the threshold to include more gene pairs with weaker associations, forming a series of nested networks. First, retain the top 10% of strongly associated gene pairs, then successively include the top 20% and 30% of associated pairs, and observe the dynamic changes in the network structure.

[0095] Topology feature identification: Tracking and recording key topological structures that emerge during network expansion.

[0096] Connected components: Functional modules formed by strongly associated genes, such as gene groups involved in the same metabolic pathway.

[0097] Loop structure: Closed loop associations formed by three or more genes, reflecting feedback regulatory mechanisms between genes, such as the mutual balancing loop between oncogenes and tumor suppressor genes.

[0098] Hub nodes: Core genes that are associated with multiple genes, usually key regulators of signaling pathways.

[0099] (III) Topological Feature Quantization

[0100] Through continuous homology analysis, the existence duration of each topological structure (i.e., the number of network expansion stages from appearance to disappearance) is calculated. Structures with longer existence durations indicate more stable biological significance within the gene network. These topological features are then transformed into a numerical topological feature matrix, where matrix elements represent the stability and importance of different topological structures.

[0101] III. Weighted Dynamic Attention Mechanism

[0102] (a) Preliminary assessment of the importance of features

[0103] An attention scoring module is constructed by inputting the sequence feature matrix output by the 1D-CNN and the topological feature matrix obtained from continuous cohomology analysis. Through a dynamic attention mechanism, local features directly associated with disease-related genes and core topological structures receive higher initial weights.

[0104] (ii) Dynamic adjustments during the training process

[0105] During model training, the weights are optimized in real time based on the prediction error: when a certain feature (a certain topological structure) causes the prediction result to deviate significantly from the actual condition, its weight is reduced; conversely, if a certain feature continuously improves the prediction accuracy, its weight is increased.

[0106] (III) Multi-scale feature vector generation

[0107] Dynamically weighted sequence features and topological features are integrated to form a multi-scale feature vector. This vector includes both fine-grained local associations between genes (the direct impact of methylation at specific sites on gene expression) and the overall regulatory patterns of the global network (the synergistic effects of signaling pathway modules), achieving feature fusion from the micro to the macro level and providing a comprehensive feature foundation for subsequent multimodal fusion.

[0108] The fusion feature generation module is used to build a multimodal fusion model based on the Transformer architecture to fuse the multi-scale feature vectors, and to fine-tune and adapt the disease dataset using the general genomic features of the pre-trained model AlphaGenome, and output the fused feature vector.

[0109] Specifically, this embodiment also includes a feature processing unit, which is used to standardize the multi-scale feature vectors and add position encoding information, and use the Transformer's attention head to focus on the direct interaction between genomic and proteomic features.

[0110] The vector output unit is used to process the multi-head self-attention process and then feed forward neural network, strengthen important features through nonlinear transformation, and output a fused feature vector adapted to the disease.

[0111] Specifically,

[0112] (I) Multimodal fusion process of Transformer architecture

[0113] Input feature preprocessing

[0114] The multi-scale feature vectors output from step 2 are standardized to ensure that the numerical ranges of different feature types are consistent. Then, positional encoding information is added. This encoding reflects the relative positional relationships of features within the sequence, such as which features belong to local associations in genomic data and which belong to topological structural features in epigenomic data, helping the model understand the sequential relationships between features.

[0115] Multi-head self-attention mechanism operation

[0116] The core of the Transformer is its multi-head self-attention mechanism, which simultaneously captures the relationships between features from multiple perspectives. One attention head focuses on the direct interaction between genomic and proteomic features, while another focuses on the regulatory relationship between epigenetic features and genomic features. Through this parallel processing, the model can comprehensively uncover the complex dependencies between different modalities, such as how a mutation in a gene affects the expression of downstream proteins through epigenetic modifications.

[0117] Feature fusion and enhancement

[0118] After multi-head self-attention processing, the features enter a feedforward neural network, which strengthens important features and suppresses irrelevant information through nonlinear transformations. This process is repeated across multiple encoder layers, with each layer deepening feature fusion based on the previous layer, ultimately yielding preliminary fused feature representations that integrate global correlation information from multi-omics data.

[0119] (II) Fine-tuning and adaptation of the pre-trained model AlphaGenome

[0120] Model initialization and adaptation preparation

[0121] Load the parameters of the pre-trained model AlphaGenome, which has learned basic gene sequence patterns, common regulatory modes, and other general features from massive amounts of general genomic data. For specific disease datasets, first convert the data format to match the model input requirements, and then divide it into training, validation, and test sets to prepare for fine-tuning.

[0122] Layered fine-tuning strategy

[0123] A hierarchical fine-tuning approach is adopted, freezing the parameters of the lower-level networks and adjusting only the parameters of the upper-level networks. This preserves the knowledge the model has learned on general data while allowing the upper-level networks to focus on learning the unique characteristics of specific diseases, such as the gene mutation combinations specific to a particular cancer.

[0124] Fine-tuning, optimization and adaptation

[0125] During fine-tuning, the model is trained using labeled data for a specific disease (patient's disease status, clinical phenotype). By continuously adjusting the upper-level parameters, the fusion features output by the model better reflect the molecular characteristics of the disease. Simultaneously, the training strategy is adjusted based on the performance on the validation set. For example, when the validation set accuracy no longer improves, the learning rate is reduced or training is stopped early to avoid overfitting. Ultimately, the model outputs a fusion feature vector adapted to the specific disease.

[0126] The gene data analysis module is used to input the fused feature vector into the MLP multilayer perceptron for disease risk prediction, combine it with the SHAP algorithm to generate a disease risk prediction index, and generate an auxiliary decision-making scheme based on the prediction index.

[0127] Specifically, this embodiment also includes a multilayer perceptron unit, which is used to determine the first hidden layer of the MLP multilayer perceptron to screen out key pathway features related to the disease, the second layer to focus on specific gene regulation modules, and the output layer to convert the information into a disease risk probability index and generate an auxiliary decision-making scheme based on the predicted probability index.

[0128] Specifically,

[0129] (I) Disease Risk Prediction Using Multilayer Perceptron (MLP)

[0130] Network Structure Design: The MLP (Multilayer Perceptron) consists of an input layer, hidden layers, and an output layer. The number of neurons in the input layer is consistent with the dimension of the fused feature vector, and it is used to receive the fused feature vector. There are three hidden layers: the first layer contains 256 neurons and uses the ReLU activation function to enhance the model's non-linear fitting ability; the second layer contains 128 neurons and also uses the ReLU activation function; the third layer contains 64 neurons and uses the LeakyReLU activation function to avoid neuron death. The output layer contains one neuron and uses the Sigmoid activation function, outputting a disease risk prediction probability value ranging from 0 to 1.

[0131] Model Training: The dataset was divided into training and validation sets in an 8:2 ratio, and cross-validation was used for model training. The binary cross-entropy loss function was selected, and the Adam optimizer was used. The initial learning rate was set to 0.001, which was adaptively adjusted during training iterations. The number of training iterations was set to 200. Training was stopped when the validation set loss did not decrease for 10 consecutive iterations to prevent overfitting. During training, model performance was evaluated using metrics such as accuracy, precision, recall, and F1-score.

[0132] Risk prediction: The trained MLP model is applied to the new fused feature vector to output the corresponding disease risk prediction probability value.

[0133] (ii) Generating disease risk prediction indices using the SHAP algorithm

[0134] SHAP value calculation: The SHAP algorithm is used to interpret the MLP model and calculate the SHAP value of each feature in the fused feature vector for the disease risk prediction result. This value reflects the degree of contribution of each feature to the prediction result. A positive value indicates that the feature increases the disease risk, and a negative value indicates that the disease risk decreases.

[0135] Feature weight determination: The SHAP values ​​of each feature are normalized to obtain the relative importance weight of each feature, and the total weight is 1.

[0136] Generate prediction index: The disease risk prediction probability value output by MLP is weighted and summed with the relative importance weight of each feature to obtain the disease risk prediction index. The formula is: Disease Risk Prediction Index = Σ (Disease Risk Prediction Probability Value × Relative Importance Weight of Features). The index ranges from 0 to 100. The higher the value, the higher the disease risk.

[0137] A decision support scheme is generated based on the predicted index.

[0138] Index range division: The disease risk prediction index is divided into three ranges: 0-30 is the low-risk range, 31-70 is the medium-risk range, and 71-100 is the high-risk range.

[0139] Recommendations for decision-making:

[0140] When the predictive index is in the low-risk range, patients are advised to maintain their current healthy lifestyle, have a routine physical examination once a year, and focus on monitoring the basic indicators related to the disease.

[0141] When the predictive index is in the medium-risk range, it is recommended that patients undergo a specific examination every six months, including further testing of relevant genes and examination of clinical signs. At the same time, they should adjust their lifestyle habits based on the examination results, such as a reasonable diet and increased exercise.

[0142] When the predictive index is in the high-risk range, it is recommended that patients undergo a comprehensive clinical diagnosis immediately, including imaging examinations and pathological biopsies, so that a professional doctor can develop an individualized treatment plan and conduct regular follow-up monitoring.

[0143] Its beneficial effects are as follows: 1. It can accurately extract core information related to disease risk from fused features. Combined with optimization methods such as cross-validation and regularization, it effectively improves the accuracy and generalization ability of disease risk classification, providing reliable risk judgment results for clinical practice and helping to identify high-risk groups for diseases at an early stage. 2. By integrating multi-source data from the genome, epigenome, and proteome, the model can capture the complex relationships between gene expression, methylation status, and protein interactions, significantly improving the accuracy of disease risk prediction. 3. It overcomes the limitations of traditional single-modality modeling. Simultaneously, by combining the SHAP algorithm and knowledge distillation technology, the model achieves lightweight deployment while ensuring high accuracy, meeting the real-time analysis needs of edge devices (portable DNA sequencers).

[0144] Please see Figure 2 In an AI-based gene data analysis system, the feature vector extraction module includes the following sub-modules:

[0145] The association scoring unit is used to establish association scores between genes based on co-variation trends in gene expression levels, evidence of physical binding between proteins, and synergistic regulatory relationships of epigenetic modifications.

[0146] Gene association units are used to connect gene pairs with scores higher than a set threshold to form a preliminary gene association network, where each node represents a gene and the thickness of the edges indicates the association strength.

[0147] The threshold adjustment unit is used to gradually lower the threshold from gene pairs with high association strength, so as to include gene pairs with weak association strength into the network.

[0148] A structure recording unit is used to track and record key topological structures that occur, including at least connected components, loop structures, and hub nodes.

[0149] The feature transformation unit is used to calculate the duration of existence of each topological structure and transform the topological features into a numerical topological feature matrix.

[0150] Please see Figure 3 In an AI-based gene data analysis system, the feature vector extraction module also includes the following sub-modules:

[0151] The matrix input unit is used to input the sequence feature matrix output by 1D-CNN and the topological feature matrix obtained by continuous coherence analysis into the dynamic attention mechanism;

[0152] The feature extraction unit is used to calculate the channel-dimensional attention weights for sequence features and topological features respectively. Global average pooling is used to extract global information of feature channels to obtain the initial attention weight vector.

[0153] The weight allocation unit is used to allocate spatial dimensions of the feature sequence using a self-attention mechanism, capture long-distance dependencies between features, and output a weighted multi-scale feature vector.

[0154] Specifically,

[0155] I. Definition and Objectives of Multimodal Fusion Model

[0156] The core objective of multimodal fusion is to enhance a model's ability to understand complex biological information by integrating data from different modalities (genomics, epigenetics, proteomics, etc.). In genomic data analysis, multimodal fusion aims to:

[0157] Enhanced complementarity: Different omics data (gene sequences and protein interaction networks) provide different perspectives, and their fusion can compensate for the limitations of a single modality.

[0158] Redundancy elimination: Remove redundant information and improve model efficiency through feature selection or dimensionality reduction techniques (PCA).

[0159] Cross-modal association modeling: mining potential associations between the genome and the epigenome (the effect of DNA methylation on gene expression).

[0160] II. Specific Implementation Methods of Multimodal Fusion

[0161] According to the classification in the knowledge base, multimodal fusion can be divided into three categories: early fusion, intermediate fusion, and late fusion. In gene data analysis, multiple fusion strategies are usually combined to achieve the best results.

[0162] 1. Early fusion (feature-level fusion)

[0163] Principle: In the data preprocessing stage, raw data or low-level features of different modalities are directly spliced ​​or weighted and fused.

[0164] Technical Implementation:

[0165] Feature concatenation: Concatenates feature vectors from the genome (k-mer vector), epigenome (methylation status), and proteome (protein abundance) into a feature matrix of uniform dimensions.

[0166] Linear Projection: Features from different modalities are mapped to the same latent space through a fully connected layer, and then weighted summation is performed.

[0167] Tensor Fusion Network (TFN): Models nonlinear relationships between modalities using high-order tensor interactions (coupling between gene mutations and protein interactions).

[0168] Application scenario: Preliminary integration of genome and epigenome features.

[0169] 2. Intermediate-term fusion (cross-attention mechanism)

[0170] Principle: Cross-Attention mechanism is introduced in the middle layer of the model to dynamically model the dependency relationship between different modalities.

[0171] Technical Implementation:

[0172] Co-Attention: Genomic features are used as queries, and epigenetic features are used as keys / values. The correlation between the two is calculated through attention weights.

[0173] Transformer architecture: A Transformer encoder that shares feature inputs from all modalities and extracts cross-modal features through a self-attention mechanism.

[0174] ModalityAdapter: Inserts a specific modality adapter module (Q-Former in BLIP-2) into the Transformer to control the fusion process.

[0175] Application scenario: Joint modeling of cross-omics data.

[0176] 3. Late-stage fusion (decision-level fusion)

[0177] Principle: After each modality is modeled independently, the final decision result is weighted or voted on.

[0178] Technical Implementation:

[0179] Ensemble learning: weighted averaging of classification results from the genome, epigenome, and proteome (Bayesian rule fusion).

[0180] ModalityGate: Dynamically assigns weights to each modality through a neural network (similar to an attention mechanism).

[0181] Mixture of Experts (MoE): Each modality acts as an expert, and the final output is aggregated.

[0182] Application scenario: Final decision fusion of multi-omics data.

[0183] 4. Hybrid Fusion

[0184] Principle: Combining the advantages of early, mid, and late-stage fusion, multimodal information is integrated in stages.

[0185] Technical Implementation:

[0186] Layered fusion: early fusion processes low-level features, mid-stage fusion models interaction, and late-stage fusion makes decisions.

[0187] BottleneckFusion: Introduces a fusion module into the bottleneck layer of the model (the intermediate layer of the Transformer).

[0188] Advantages: Highly flexible, balancing efficiency and performance.

[0189] Application scenario: Complex gene data analysis.

[0190] III. Specific Applications of Multimodal Fusion Models

[0191] In genomic data analysis, multimodal fusion models achieve cross-omics joint modeling through the following steps:

[0192] 1. Data preprocessing and standardization

[0193] Inputs: Genomic data (k-mer vectors), epigenome data (methylation status), and proteome data (protein abundance).

[0194] Processing: Format conversion and cleaning (FASTQ / BAM to VCF). k-mer vectorization (k-mer preprocessing of [4]). Multi-omics alignment (genomic and proteomic coordinate alignment).

[0195] 2. Feature Extraction and Modality Coding

[0196] Genomic features: 1D-CNN extracts local sequence features (regulatory elements in promoter regions).

[0197] Epigenome features: Topological features (circular structure of DNA methylation clusters) were extracted by continuous homology analysis.

[0198] Proteomics features: Graph Neural Networks (GNNs) model protein-protein interaction networks.

[0199] 3. Multimodal fusion strategy

[0200] Early fusion: The genome k-mer vector, epigenetic methylation status, and proteome abundance are spliced ​​together into a multi-scale feature vector.

[0201] Intermediate fusion: Dynamically model the dependencies between the genome and the proteome through the cross-attention mechanism of Transformer.

[0202] Late-stage fusion: The classification results (disease risk prediction) of each modality are weighted and averaged to generate the final decision.

[0203] 4. Enhanced interpretability

[0204] SHAP algorithm: Combines SHAP (Shapley Additive Explanations) to generate visual explanations that reveal the association between key genes and phenotypes.

[0205] Feature importance ranking: Key gene regions are located by attention weights or gradient backpropagation (Grad-CAM).

[0206] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based gene data analysis system, characterized in that, The gene data analysis system includes the following modules: The gene data acquisition module is used to acquire the patient's genomic data, epigenetic data, and proteomic data, and to process the data to obtain a multi-omics data matrix. The feature vector extraction module is used to mine local association patterns in the multi-omics data matrix through a 1D-CNN one-dimensional convolutional neural network, identify the topological structure of the gene network through continuous homology analysis, and assign dynamic weights to the sequence features and topological feature matrix using a dynamic attention mechanism to output a weighted multi-scale feature vector. The fusion feature generation module is used to build a multimodal fusion model based on the Transformer architecture to fuse the multi-scale feature vectors. It fine-tunes the disease dataset using the general genomic features of the pre-trained model AlphaGenome and outputs the fused feature vector. The fine-tuning and adaptation of the pre-trained model AlphaGenome includes: Model initialization and adaptation preparation; Load the parameters of the pre-trained model AlphaGenome. For a specific disease dataset, first convert the data format to match the model input requirements, and then divide it into training, validation and test sets to prepare for fine-tuning. Layered fine-tuning strategy: A layered fine-tuning approach is adopted, freezing the parameters of the bottom layer network of the model and adjusting only the parameters of the upper layer network; Fine-tuning and adaptation: During the fine-tuning process, the model is trained using labeled data for a specific disease. By continuously adjusting the upper-layer parameters, the fusion features output by the model can better reflect the molecular characteristics of the disease. At the same time, the training strategy is adjusted based on the performance on the validation set, and finally, a fusion feature vector adapted to the specific disease is output. The gene data analysis module is used to input the fused feature vector into the MLP multilayer perceptron for disease risk prediction, combine the SHAP algorithm to generate a disease risk prediction index, and generate an auxiliary decision-making scheme based on the prediction index. The gene data acquisition module includes the following units: The format conversion and cleaning unit is used to acquire the patient's genomic data, epigenetic data, and proteomic data, convert the raw sequencing data into a unified format, and remove low-quality reads to obtain cleaned gene data. k-mer vectorization units are used to perform k-mer preprocessing on DNA sequences, breaking long sequences into short fragments and generating vectorized gene features; The multi-omics data integration unit is used to align genomic data, epigenetic data, and proteomic data to the same gene coordinates to generate a multimodal data matrix, which includes at least gene expression, methylation status, and protein abundance. The feature vector extraction module includes the following units: The convolutional computation unit is used to perform a sliding scan of the input sequence using a large convolutional kernel; and to perform secondary extraction of the feature map output from the upper layer using a small convolutional kernel. The feature generation unit is used to retain the most significant feature values ​​in the local region and the combination of methylation features most related to gene expression changes through max pooling, and outputs a sequence feature matrix that reflects local molecular associations; The feature vector extraction module also includes the following units: The association scoring unit is used to establish association scores between genes based on co-variation trends in gene expression levels, evidence of physical binding between proteins, and synergistic regulatory relationships of epigenetic modifications. Gene association units are used to connect gene pairs with scores higher than a set threshold to form a preliminary gene association network, where each node represents a gene and the thickness of the edges indicates the association strength. The threshold adjustment unit is used to gradually lower the threshold from gene pairs with high association strength, so as to include gene pairs with weak association strength into the network. A structure recording unit is used to track and record key topological structures that occur, including at least connected components, loop structures, and hub nodes. The feature transformation unit is used to calculate the existence duration of each topological structure and transform the topological features into a numerical topological feature matrix. The feature vector extraction module also includes the following units: The matrix input unit is used to input the sequence feature matrix output by 1D-CNN and the topological feature matrix obtained by continuous coherence analysis into the dynamic attention mechanism; The feature extraction unit is used to calculate the channel-dimensional attention weights for sequence features and topological features respectively. Global average pooling is used to extract global information of feature channels to obtain the initial attention weight vector. The weight allocation unit is used to allocate spatial dimensions of the feature sequence using a self-attention mechanism, capture long-distance dependencies between features, and output a weighted multi-scale feature vector. The fusion feature generation module includes the following units: The feature processing unit is used to standardize multi-scale feature vectors and add positional encoding information, and uses the Transformer's attention head to focus on the direct interaction between genomic and proteomic features. The vector output unit is used to process the multi-head self-attention process and then enter the feedforward neural network. It strengthens important features through nonlinear transformation and outputs a fusion feature vector that is adapted to the disease. The gene data analysis module includes the following units: The multilayer perceptron (MLP) unit is used to determine the first hidden layer of the MLP to screen out key pathway features related to the disease, the second layer to focus on specific gene regulatory modules, and the output layer to convert information into a disease risk probability index. Based on the predicted probability index, an auxiliary decision-making scheme is generated. The MLP consists of an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is consistent with the dimension of the fused feature vector, and it is used to receive the fused feature vector. There are three hidden layers: the first layer contains 256 neurons and uses the ReLU activation function to enhance the nonlinear fitting ability of the model; the second layer contains 128 neurons and also uses the ReLU activation function; the third layer contains 64 neurons and uses the LeakyReLU activation function to avoid the problem of neuron death. The output layer contains 1 neuron and uses the Sigmoid activation function to output a disease risk prediction probability value in the range of 0-1.

Images