Gene data analysis system based on AI

Abstract

The invention discloses an AI-based gene data analysis system. The system comprises a plurality of omics data matrixes; local association pattern mining is performed on the multi-omics data matrix through a 1D-CNN one-dimensional convolutional neural network, a topological structure of a gene network is identified through continuous coherence analysis, dynamic weights are allocated to sequence features and a topological feature matrix by using a dynamic attention mechanism, and weighted multi-scale feature vectors are output; establishing a multi-modal fusion model based on a Transform architecture to fuse the multi-scale feature vectors, performing fine adjustment on the adaptive disease data set by using the general genome feature of a pre-training model, and outputting a fused feature vector; and inputting the fusion feature vector into an MLP multilayer perceptron for disease risk prediction, generating a disease risk prediction index in combination with an SHAP algorithm, and generating an auxiliary decision scheme according to the prediction index. And the accuracy and generalization ability of disease risk classification are effectively improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data analysis, in particular to an AI-based gene data analysis system. BACKGROUND

[0002] Traditional gene data analysis modes have many defects. The amount of genomic data is huge, and single omics data is difficult to comprehensively reveal the complex molecular mechanism of diseases. The occurrence of diseases is the result of the joint action of multiple factors, involving gene sequence variation, abnormal epigenetic regulation, protein expression disorder and other aspects. Relying on only one omics data may lead to one-sided analysis results and low analysis efficiency. Traditional data analysis methods have limitations in processing high-dimensional and high-complexity multi-omics data. Traditional statistical methods are difficult to handle the nonlinear relationship and complex dependence in gene data, resulting in low accuracy of disease risk prediction. SUMMARY

[0003] The application aims to solve the above problems, and designs an AI-based gene data analysis system.

[0004] To achieve the above purpose, the technical scheme of the application is that, further, in the AI-based gene data analysis system, the gene data analysis system method comprises the following steps: A gene data acquisition module is configured to acquire genomic data, epigenomic data and proteomic data of a patient, process the data, and obtain a multi-omics data matrix; A feature vector extraction module is configured to mine local correlation patterns of the multi-omics data matrix through a 1D-CNN one-dimensional convolutional neural network, identify the topological structure of a gene network through persistent homology analysis, and assign dynamic weights to sequence feature and topological feature matrices using a dynamic attention mechanism, and output weighted multi-scale feature vectors; A fusion feature generation module is configured to establish a multi-modal fusion model based on a Transformer architecture to fuse the multi-scale feature vectors, adapt a disease data set using a universal genomic feature fine-tuning of a pre-training model AlphaGenome, and output a fusion feature vector; A gene data analysis module is configured to input the fusion feature vector into an MLP multi-layer perceptron to perform disease risk prediction, generate a disease risk prediction index in combination with a SHAP algorithm, and generate an auxiliary decision scheme according to the prediction index.

[0005] Further, in the AI-based gene data analysis system, the gene data acquisition module comprises the following units: A format conversion and cleaning unit is configured to acquire genomic data, epigenomic data and proteomic data of a patient, convert original sequencing data into a unified format, and remove low-quality reads to obtain cleaned gene data; a k-mer vectorization unit for k-mer preprocessing of DNA sequences, splitting long sequences into short fragments, and generating vectorized gene features; a multi-omics data integration unit for aligning genomic data, epigenomic data, and proteomic data to the same gene coordinates, and generating a multi-modal data matrix including at least gene expression, methylation status, and protein abundance.

[0006] Further, in the AI-based gene data analysis system described above, the feature vector extraction module includes the following units: a convolution calculation unit for sliding scanning of an input sequence using a large convolution kernel, and secondary extraction of feature maps output by an upper layer using a small convolution kernel; a feature generation unit for combining the most significant feature values in a local region and the methylation features most relevant to gene expression changes by maximum pooling, and outputting a sequence feature matrix reflecting local molecular correlations.

[0007] Further, in the AI-based gene data analysis system described above, the feature vector extraction module further includes the following units: a correlation score unit for establishing correlation scores between genes based on co-variation trends of gene expression levels, physical binding evidence between proteins, and synergistic regulatory relationships of epigenetic modifications; a gene correlation unit for connecting gene pairs with scores higher than a set threshold to form a preliminary gene correlation network, wherein each node represents a gene, and the thickness of an edge represents correlation strength; a threshold adjustment unit for gradually reducing the threshold from gene pairs with high correlation strength, and including gene pairs with weak correlation in the network; a structure recording unit for tracking and recording key topological structures appearing, including at least connected components, loop structures, and hub nodes; a feature transformation unit for calculating the duration of existence of each topological structure, and transforming topological features into a numerical topological feature matrix.

[0008] Further, in the AI-based gene data analysis system described above, the feature vector extraction module further includes the following units: a matrix input unit for inputting the sequence feature matrix output by the 1D-CNN and the topological feature matrix obtained by the persistent homology analysis into a dynamic attention mechanism; an extracted feature unit for respectively calculating attention weight in the channel dimension for sequence features and topological features, and extracting global information of a feature channel by global average pooling to obtain an initial attention weight vector; The weight distribution unit is configured to distribute weights in the spatial dimension of the feature sequence by using a self-attention mechanism, capture long-distance dependency between features, and output a weighted multi-scale feature vector.

[0009] Further, in the AI-based gene data analysis system, the fusion feature generation module comprises the following units: The feature processing unit is configured to add position encoding information to the multi-scale feature vector after standardization processing, and use the attention head of the Transformer to focus on the direct interaction between the genome and the proteome features. The vector output unit is configured to input the multi-head self-attention processed vector into a feedforward neural network, strengthen important features through nonlinear transformation, and output a fusion feature vector suitable for the disease.

[0010] Further, in the AI-based gene data analysis system, the gene data analysis module comprises the following units: The multi-layer perception unit is configured to determine the first layer of the MLP multi-layer perception unit to filter out key pathway features related to the disease, the second layer to focus on specific gene regulation modules, and the output layer to convert information into a disease risk probability index, and generate an auxiliary decision scheme according to the prediction probability index.

[0011] Further, in the method for implementing the AI-based gene data analysis system, the method comprises the following steps: Obtain the genome data, epigenome data and proteome data of the patient, process the data, and obtain a multi-omics data matrix; Perform local correlation pattern mining on the multi-omics data matrix by using a 1D-CNN one-dimensional convolutional neural network, identify the topological structure of the gene network by using persistent homology analysis, and assign dynamic weights to the sequence feature and topological feature matrix by using a dynamic attention mechanism, and output a weighted multi-scale feature vector; Establish a multi-modal fusion model based on the Transformer architecture to fuse the multi-scale feature vector, fine-tune the general genomic features of the pre-trained model AlphaGenome to adapt to the disease data set, and output a fusion feature vector; Input the fusion feature vector into the MLP multi-layer perception unit for disease risk prediction, generate a disease risk prediction index by combining the SHAP algorithm, and generate an auxiliary decision scheme according to the prediction index.

[0012] Further, in the method for implementing the AI-based gene data analysis system, the method comprises the following steps: Input the sequence feature matrix output by the 1D-CNN and the topological feature matrix obtained by the persistent homology analysis into the dynamic attention mechanism; The attention weight calculation of the channel dimension is performed on the sequence features and the topological features respectively, global information of the feature channel is extracted through global average pooling, and an initial attention weight vector is obtained; The self-attention mechanism is used to allocate weight in the spatial dimension of the feature sequence, capture long-distance dependence between features, and output a weighted multi-scale feature vector.

[0013] Further, in the method for implementing an AI-based gene data analysis system, the method comprises the following steps: After the multi-scale feature vector is standardized, position encoding information is added, and the attention head of the Transformer focuses on the direct interaction between the genome and the proteome features; After the multi-head self-attention processing, the important features are strengthened through a non-linear transformation in the feedforward neural network, and a fusion feature vector suitable for the disease is output.

[0014] The beneficial effects are as follows: 1. The core information related to the disease risk can be accurately extracted from the fusion features, and the accuracy and generalization ability of the disease risk classification are effectively improved by combining the optimization methods such as cross-validation and regularization, which provides reliable risk judgment results for the clinic and helps to identify high-risk groups of diseases at an early stage. 2. By integrating multi-source data of genome, epigenome and proteome, the model can capture the complex relationship between gene expression, methylation state and protein interaction, and significantly improve the accuracy of disease risk prediction. 3. The limitations of traditional single-modal modeling are solved. At the same time, the model realizes lightweight deployment by combining the SHAP algorithm and the knowledge distillation technology, while ensuring high precision, which meets the real-time analysis requirements of edge devices (portable DNA sequencers). BRIEF DESCRIPTION OF DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not considered a limitation of the present application.

[0016] Figure 1 The first embodiment of the AI-based gene data analysis system in the embodiments of the present application is shown in the schematic diagram. Figure 2 The second embodiment of the AI-based gene data analysis system in the embodiments of the present application is shown in the schematic diagram. Figure 3 The third embodiment of the AI-based gene data analysis system in the embodiments of the present application is shown in the schematic diagram. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0018] It can be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "a", and "the" used herein can also include the plural forms. It should be further understood that the use of the terms "comprise", "comprises", "comprised" and "comprising" in the specification of the present application means that there are features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0019] The present application will be described in detail below in combination with the drawings, Figure 1 As shown, an AI-based gene data analysis system includes the following modules: A gene data acquisition module is used to acquire genomic data, epigenomic data and proteomic data of a patient, process the data, and obtain a plurality of omics data matrices; Specifically, the present embodiment further includes a format conversion and cleaning unit, which is used to acquire genomic data, epigenomic data and proteomic data of a patient, convert raw sequencing data into a unified format, and remove low-quality reads to obtain cleaned gene data; A k-mer vectorization unit is used to k-mer preprocess DNA sequences, split long sequences into short fragments, and generate vectorized gene features; A multi-omics data integration unit is used to align genomic data, epigenomic data and proteomic data to the same gene coordinates, and generate a multi-modal data matrix, which at least includes gene expression, methylation state and protein abundance.

[0020] Specifically, (I) Multi-omics data acquisition Genomic data: obtained by whole genome sequencing (WGS), whole exome sequencing (WES) or targeted gene sequencing technology, covering single nucleotide polymorphism (SNP), insertion and deletion (InDel), copy number variation (CNV) and other variation information. The data is derived from gene sequencing experiments of clinical samples, and needs to meet the requirements of ethical norms and data privacy protection.

[0021] Epigenomic data: including DNA methylation data (obtained by Illumina Infinium chip or methylation sequencing), histone modification data (ChIP-seq data), chromatin accessibility data (ATAC-seq data), etc.

[0022] Proteome data: The expression level and modification state (phosphorylation, glycosylation, etc.) of proteins in patient samples were determined by mass spectrometry analysis (LC-MS / MS).

[0023] (II) Data preprocessing Data cleaning For genomic data, quality control tools (FastQC) were used for quality assessment, and low-quality sequencing reads (reads with a Phred quality score of less than 20 accounting for more than 50% of the bases) were removed; sequencing data was aligned to the reference genome by alignment tools (BWA), and repetitive sequences (using Picard tools) and PCR duplicates were removed.

[0024] In the epigenome data, DNA methylation data needs to be background corrected and probe filtered (remove cross-reactive probes and polymorphic probes); ChIP-seq data needs to be peak called (using MACS2 tools) and low-confidence peaks filtered.

[0025] Proteome data was processed by peptide identification software (MaxQuant) for peptide matching and quantification, and proteins with low identification confidence were removed. Imputation method based on Gaussian mixture model was used for missing values.

[0026] Standardization SNP genotypes in genomic data were converted to 0 / 1 / 2 encoding (representing homozygous reference alleles, heterozygotes, and homozygous alternative alleles, respectively), and CNV data were converted to copy number relative values.

[0027] Epigenome data was standardized using the quantile normalization method to make the data distribution consistent across different samples.

[0028] Proteome data was median-centered or z-score standardized to eliminate systematic errors between different samples.

[0029] Data integration The preprocessed genomic, epigenomic, and proteomic data were aligned according to the sample ID to construct a multi-omics data matrix with consistent dimensions, with rows representing samples and columns representing different molecular features (genes, methylation sites, proteins, etc.).

[0030] The feature vector extraction module is used to mine local correlation patterns in the multi-omics data matrix through 1D-CNN one-dimensional convolutional neural network, identify the topological structure of the gene network through persistent homology analysis, and assign dynamic weights to the sequence features and topological feature matrix using a dynamic attention mechanism, outputting a weighted multi-scale feature vector. Specifically, the embodiment further includes a convolution calculation unit, configured to perform sliding scanning on the input sequence by using a large convolution kernel; and perform secondary extraction on the feature map output by the upper layer by using a small convolution kernel. The feature generation unit is configured to combine the most significant feature values in a local region and the methylation features most relevant to the gene expression changes by maximum pooling, and output a sequence feature matrix reflecting the local molecular correlation.

[0031] The correlation scoring unit is configured to establish a correlation score between genes based on the co-change trend of gene expression, physical binding evidence between proteins, and synergistic regulation relationship of epigenetic modification. The gene correlation unit is configured to connect gene pairs with a score higher than a set threshold to form a preliminary gene correlation network, wherein each node represents a gene, and the thickness of the edge represents the correlation strength. The threshold adjustment unit is configured to gradually reduce the threshold from the gene pairs with high correlation strength, and include gene pairs with weak correlation into the network. The structure recording unit is configured to track and record the key topological structures appearing, including at least a connected component, a loop structure and a hub node. The feature conversion unit is configured to calculate the duration of the existence of each topological structure, and convert the topological features into a numerical topological feature matrix.

[0032] The matrix input unit is configured to input the sequence feature matrix output by the 1D-CNN and the topological feature matrix obtained by the persistent homology analysis into the dynamic attention mechanism. The feature extraction unit is configured to respectively calculate the attention weight of the channel dimension of the sequence feature and the topological feature, and extract the global information of the feature channel by global average pooling to obtain an initial attention weight vector. The weight distribution unit is configured to use the self-attention mechanism to distribute the weight of the spatial dimension of the feature sequence, capture the long-distance dependence relationship between the features, and output a weighted multi-scale feature vector.

[0033] Specifically, (I) Data input processing The preprocessed multi-omics data matrix is arranged in the biological order of molecular features (sorted according to the position of the gene on the chromosome), so as to ensure that adjacent features have correlation in biological sense. The genes arranged continuously on a chromosome and the corresponding methylation and protein expression data are spliced in turn to form sequence data suitable for one-dimensional convolution processing.

[0034] (II) Convolution layer feature capture First layer convolution: a larger convolution kernel (covering a window of 5 continuous features) is used to perform sliding scanning on the input sequence. This process is like local observation on the gene sequence, capturing the synergistic change pattern between the adjacent 5 molecular features.

[0035] Subsequent convolutional layers: using smaller convolutional kernels (3 consecutive feature windows), the feature maps output by the upper layer are extracted again.

[0036] (Three) Pooling and feature selection After each convolutional operation, the most significant feature values in the local region are retained through maximum pooling, reducing the data dimension while highlighting the key patterns. In the region containing multiple methylation sites, the most relevant methylation feature combinations to gene expression changes are retained, and the final output is a sequence feature matrix reflecting the local molecular associations.

[0037] II. Persistent homology analysis identifies gene network topology (I) Construction of gene correlation network Based on multi-omics data, the interaction strength between all genes is calculated, such as through the co-variation trend of gene expression, physical binding evidence between proteins, and synergistic regulation relationship of epigenetic modification, etc. The correlation score between genes is established. Connect gene pairs with scores higher than the set threshold to form the initial gene correlation network, where each node represents a gene, and the thickness of the edge represents the correlation strength.

[0038] (II) Dynamic network topology tracking Step-by-step network expansion: starting from the gene pairs with the highest correlation strength, gradually lower the threshold to include more weakly correlated gene pairs into the network, forming a series of nested networks. First, retain the top 10% of strongly correlated gene pairs, then gradually include the top 20%, 30% of correlated pairs, and observe the dynamic changes in network structure.

[0039] Topology feature identification: during network expansion, track and record key topological structures that appear: Connected components: functional modules formed by strongly correlated genes, such as gene groups involved in the same metabolic pathway.

[0040] Loop structure: a closed loop correlation formed by three or more genes, reflecting the feedback regulation mechanism between genes, such as the mutual balancing loop between oncogenes and tumor suppressor genes.

[0041] Hub node: a core gene that is associated with multiple genes, usually a key regulator of signaling pathways.

[0042] (III) Quantification of topological features Through persistent homology analysis, the existence duration of each topological structure (i.e. the number of network expansion stages from appearance to disappearance) is calculated. The longer the existence duration, the more stable the structure is in the gene network, indicating its biological significance. These topological features are converted into a numerical topological feature matrix, where matrix elements represent the stability and importance of different topological structures.

[0043] III. Dynamic attention mechanism weighting (I) Preliminary assessment of feature importance An attention score module is constructed, and the sequence feature matrix output by the 1D-CNN and the topological feature matrix obtained by the persistent homology analysis are input into the module. The module uses a dynamic attention mechanism, and local features and core topological structures that are directly related to disease-related genes will obtain higher initial weights.

[0044] (II) Dynamic adjustment during the training process During model training, the weights are optimized in real time according to the prediction error: when a certain type of feature (a certain topological structure) causes a large deviation between the predicted result and the actual condition, its weight is reduced; on the contrary, if a certain type of feature continuously improves the prediction accuracy, its weight is increased.

[0045] (III) Multi-scale feature vector generation The dynamically weighted sequence features and topological features are integrated to form a multi-scale feature vector. This vector contains both local fine-grained associations between genes (direct influence of methylation at specific sites on gene expression) and global network-wide overall regulation patterns (module cooperation of signal pathways), achieving feature fusion from micro to macro and providing a comprehensive feature basis for subsequent multi-modal fusion.

[0046] The fusion feature generation module is used to establish a multi-modal fusion model based on the Transformer architecture to fuse the multi-scale feature vector, use the pre-trained model AlphaGenome to fine-tune the general genomic features to adapt to the disease data set, and output a fusion feature vector. Specifically, the feature processing unit is used to standardize the multi-scale feature vector and add position encoding information, and the attention head of the Transformer is used to focus on the direct interaction between the genomic and proteomic features. The vector output unit is used to input the multi-head self-attention processed features into a feedforward neural network, strengthen important features through nonlinear transformation, and output a fusion feature vector that adapts to the disease.

[0047] Specifically, (I) Multi-modal fusion process of Transformer architecture Input feature preprocessing The multi-scale feature vector output in step 2 is standardized to ensure that the numerical ranges of different types of features are consistent. Then, position encoding information is added. This encoding can reflect the relative position relationship of features in the sequence, such as which features belong to local associations of genomic data and which belong to topological structure features of epigenomic data, helping the model understand the sequential association between features.

[0048] Multi-head self-attention mechanism operation The core of the Transformer is the multi-head self-attention mechanism, which can capture the correlation between features from multiple angles at the same time. Part of the attention heads focus on the direct interaction between genomic and proteomic features, and the other part focuses on the regulatory relationship between epigenomic features and genomic features. Through this parallel processing, the model can comprehensively mine the complex dependencies between different modal features, such as how a mutation in a certain gene affects the expression of downstream proteins through epigenetic modification.

[0049] Feature fusion and strengthening After multi-head self-attention processing, the features enter the feedforward neural network, which strengthens important features through nonlinear transformation and suppresses irrelevant information. This process is repeated for multiple encoder layers, each layer deepening feature fusion based on the previous layer, and ultimately obtaining preliminary fused feature representations that have integrated global correlation information from multiple omics data.

[0050] (II) Fine-tuning and adaptation of pre-trained model AlphaGenome Model initialization and adaptation preparation Load the parameters of the pre-trained model AlphaGenome, which has learned basic genetic sequence rules, common regulatory patterns, and other general features from massive general genomic data. For a specific disease dataset, first convert the data format to match the model input requirements, and then divide the training set, validation set, and test set to prepare for fine-tuning.

[0051] Layered fine-tuning strategy Use a layered fine-tuning approach, freezing the bottom layer network parameters of the model and only adjusting the upper layer network parameters. This way, you can retain the knowledge learned by the model on general data while allowing the upper layer network to focus on learning unique features of a specific disease, such as a specific combination of gene mutations characteristic of a certain type of cancer.

[0052] Fine-tuning optimization and adaptation During fine-tuning, use the labeled data of a specific disease (patient disease status, clinical phenotype) to train the model, and continuously adjust the upper layer parameters to make the model output fusion features better reflect the molecular characteristics of the disease. At the same time, adjust the training strategy according to the performance of the validation set, such as reducing the learning rate or stopping training early when the validation set accuracy no longer improves to avoid overfitting, and finally output the fusion feature vector adapted to the specific disease.

[0053] Genetic data analysis module for inputting the fusion feature vector into the MLP multi-layer perceptron for disease risk prediction, combining the SHAP algorithm to generate a disease risk prediction index, and generating an auxiliary decision-making scheme based on the prediction index.

[0054] Specifically, the embodiment also includes a multi-layer perception unit for determining that the first layer of the MLP multi-layer perception filters out key pathway features related to the disease, the second layer focuses on specific gene regulatory modules, and the output layer converts the information into a disease risk probability index to generate an auxiliary decision-making scheme according to the predicted probability index.

[0055] Specifically, (I) MLP multi-layer perception disease risk prediction Network structure design: The MLP multi-layer perception includes an input layer, a hidden layer, and an output layer. The number of input layer neurons is consistent with the dimension of the fusion feature vector, and is used to receive the fusion feature vector; 3 layers of hidden layers are set, the first layer contains 256 neurons, uses a ReLU activation function to enhance the nonlinear fitting ability of the model, the second layer contains 128 neurons, also uses a ReLU activation function, and the third layer contains 64 neurons, uses a LeakyReLU activation function to avoid neuron death problem; the output layer contains one neuron, uses a Sigmoid activation function, and outputs a disease risk prediction probability value in the range of 0-1. Model training: The data set is divided into a training set and a validation set according to an 8:2 ratio, and the model is trained by using a cross-validation method. The loss function is a binary cross-entropy loss function, the optimizer is an Adam optimizer, the initial learning rate is set to 0.001, and is adjusted adaptively with training iterations. The number of training iterations is set to 200, and when the validation set loss does not decrease for 10 consecutive iterations, the training is stopped to prevent overfitting. During the training process, the model performance is evaluated by accuracy, precision, recall, and F1-score. Risk prediction: The trained MLP model is applied to a new fusion feature vector to output the corresponding disease risk prediction probability value. (II) Generating a disease risk prediction index combined with SHAP algorithm SHAP value calculation: The SHAP algorithm is used to interpret the MLP model, and the SHAP value of each feature in the fusion feature vector is calculated, which reflects the 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 feature reduces the disease risk. Feature weight determination: The SHAP values of the features are normalized to obtain the relative importance weights of each feature, and the sum of the weights is 1. Generating a prediction index: The disease risk prediction probability value output by the MLP is weighted and summed with the relative importance weights of each feature to obtain a disease risk prediction index. The formula is: Disease risk prediction index = Σ (disease risk prediction probability value x feature relative importance weight), and the index range is 0-100. The higher the value, the higher the disease risk. According to the prediction index, an auxiliary decision-making scheme is generated Index range division: divide the disease risk prediction index into three ranges, 0-30 for low risk range, 31-70 for medium risk range, and 71-100 for high risk range. Make decision recommendations: When the prediction index is in the low risk range, it is recommended that the patient maintain the current healthy lifestyle, have a regular physical examination once a year, and focus on monitoring the basic indicators related to the disease. When the prediction index is in the medium risk range, it is recommended that the patient have a special examination every half year, including further detection of related genes and clinical signs, and adjust lifestyle habits such as reasonable diet and increased exercise according to the examination results. When the prediction index is in the high risk range, it is recommended that the patient immediately undergo comprehensive clinical diagnosis, including imaging examination and pathological biopsy, and develop a personalized treatment plan with regular follow-up monitoring by a professional doctor.

[0056] The beneficial effects are: 1. The model can accurately extract core information related to disease risk from fused features, and effectively improve the accuracy and generalization ability of disease risk classification by combining cross-validation and regularization optimization methods, providing reliable risk judgment results for clinical practice and helping to identify high-risk groups early. 2. By integrating multi-source data from genome, epigenome and proteome, the model can capture the complex relationships between gene expression, methylation status and protein interaction, significantly improving the accuracy of disease risk prediction. 3. It solves the limitations of traditional single-modal modeling. At the same time, combining SHAP algorithm and knowledge distillation technology, the model realizes lightweight deployment while ensuring high accuracy, meeting the real-time analysis needs of edge devices (portable DNA sequencers).

[0057] Please refer to Figure 2 In an AI-based genetic data analysis system, the feature vector extraction module includes the following sub-modules: Correlation scoring unit, for establishing correlation scores between genes through co-variance trends of gene expression, physical binding evidence between proteins, and synergistic regulatory relationships of epigenetic modifications; Gene correlation unit, for connecting gene pairs with scores higher than a set threshold to form a preliminary gene correlation network, where each node represents a gene and the thickness of the edge represents the correlation strength; Threshold adjustment unit, for gradually reducing the threshold from gene pairs with high correlation strength to include gene pairs with weak correlation into the network; Structure recording unit, for tracking and recording key topological structures, including at least connected components, loop structures and hub nodes; Feature transformation unit, for calculating the duration of each topological structure and transforming topological features into a numerical topological feature matrix.

[0058] Referring to Figure 3 In an AI-based gene data analysis system, the feature vector extraction module further includes the following sub-modules: A matrix input unit for inputting the sequence feature matrix output by the 1D-CNN and the topological feature matrix obtained by the continuous homology analysis into the dynamic attention mechanism; An extraction feature unit for calculating the attention weight of the channel dimension of the sequence feature and the topological feature respectively, extracting the global information of the feature channel through global average pooling, and obtaining an initial attention weight vector; A weight distribution unit for using the self-attention mechanism to distribute the weight of the spatial dimension of the feature sequence, capturing the long-distance dependence relationship between the features, and outputting a weighted multi-scale feature vector.

[0059] Specifically, I. Definition and goals of multi-modal fusion model The core goal of Multimodal Fusion is to improve the model's understanding of complex biological information by integrating data from different modalities (genomics, epigenomics, proteomics, etc.). In gene data analysis, multi-modal fusion aims to: Complementary enhancement: Different omics data (gene sequences and protein interaction networks) provide different perspectives, and fusion can make up for the limitations of a single modality.

[0060] Redundancy elimination: Through feature selection or dimensionality reduction techniques (PCA), redundant information is removed to improve model efficiency.

[0061] Cross-modal association modeling: Explore the potential association between genomics and epigenomics (the influence of DNA methylation on gene expression).

[0062] II. Specific implementation methods of multi-modal fusion According to the classification in the knowledge base, multi-modal 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.

[0063] 1. Early fusion (feature-level fusion) Principle: In the data preprocessing stage, the original data or low-level features of different modalities are directly concatenated or weightedly fused.

[0064] Technical implementation: Feature Concatenation: Concatenate genomic (k-mer vector), epigenomic (methylation status), and proteomic (protein abundance) feature vectors into a unified dimensional feature matrix.

[0065] Linear Projection: Map features from different modalities to the same latent space through a fully connected layer, then perform weighted summation.

[0066] Tensor Fusion Network (TFN): Model the non-linear relationship between modalities (coupling of gene mutations and protein interactions) using high-order tensor interactions.

[0067] Application Scenario: Preliminary feature integration of genomics and epigenomics.

[0068] 2. Mid-level fusion (Cross-Attention mechanism) Principle: Introduce a cross-attention mechanism (Cross-Attention) in the middle layer of the model to dynamically model the dependency between different modalities.

[0069] Technical Implementation: Bidirectional Cross-Attention (Co-Attention): Genomic features as Query, epigenomic features as Key / Value, calculate their relevance through attention weights.

[0070] Transformer Architecture: Input features from all modalities into a shared Transformer encoder, extract cross-modal features through self-attention mechanisms.

[0071] ModalityAdapter: Insert a specific modality Adapter module (Q-Former in BLIP-2) into the Transformer to control the fusion process.

[0072] Application Scenario: Joint modeling of cross-omics data.

[0073] 3. Late fusion (decision-level fusion) Principle: After independent modeling of each modality, perform weighted or voting on the final decision result.

[0074] Technical Implementation: Ensemble Learning: Weighted average of classification results for genomics, epigenomics, and proteomics (Bayesian rule fusion).

[0075] Dynamic gating mechanism (ModalityGate): Dynamically allocate weights for each modality through a neural network (similar to attention mechanism).

[0076] Mixture of Experts (MoE): Each modality is treated as an expert, and the aggregated output is the final result.

[0077] Application scenario: Final decision fusion of multi-omics data.

[0078] 4. Hybrid Fusion Principle: Combine the advantages of early, middle, and late fusion, and integrate multi-modal information in stages.

[0079] Technical implementation: Hierarchical fusion: Early fusion handles low-level features, middle fusion models interactions, and late fusion makes decisions.

[0080] Bottleneck Fusion: Introduce a fusion module at the model bottleneck layer (middle layer of Transformer).

[0081] Advantages: Strong flexibility, balancing efficiency and performance.

[0082] Application scenario: Complex genetic data analysis.

[0083] Three, specific application of multi-modal fusion model In genetic data analysis, multi-modal fusion model realizes cross-omics joint modeling through the following steps: 1. Data preprocessing and standardization Input: Genomic data (k-mer vector), epigenomic data (methylation state), proteomic data (protein abundance).

[0084] Processing: Format conversion and cleaning (FASTQ / BAM to VCF). k-mer vectorization (k-mer preprocessing of [4]). Multi-omics alignment (coordinate alignment of genome and proteome).

[0085] 2. Feature extraction and modality encoding Genomic features: 1D-CNN extracts local sequence features (regulatory elements in promoter regions).

[0086] Epigenomic features: Topological feature extraction using persistent homology analysis (circular structure of DNA methylation clusters).

[0087] Proteomic features: Graph neural network (GNN) models protein interaction networks.

[0088] 3. Multi-modal fusion strategy Early fusion: Concatenate genomic k-mer vector, epigenomic methylation state, and proteomic abundance into a multi-scale feature vector.

[0089] Mid-fusion: Dynamically model the dependency between genomics and proteomics via the cross-attention mechanism of Transformer.

[0090] Late fusion: Weighted average of classification results (disease risk prediction) of each modality to generate the final decision.

[0091] 4. Enhanced interpretability SHAP algorithm: Generate visual explanations with SHAP (Shapley Additive Explanations) to reveal the association between key genes and phenotypes.

[0092] Feature importance ranking: Locate key gene regions through attention weights or gradient backpropagation (Grad-CAM).

[0093] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application 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, 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. 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.

2. The AI-based gene data analysis system as described in claim 1, characterized in that, 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, 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.

3. The AI-based gene data analysis system as described in claim 1, characterized in that, 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 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.

4. The AI-based gene data analysis system as described in claim 1, characterized in that, 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 duration of existence of each topological structure and transform the topological features into a numerical topological feature matrix.

5. The AI-based gene data analysis system as described in claim 1, characterized in that, 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.

6. The AI-based gene data analysis system as described in claim 1, characterized in that, 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 feed forward neural network, strengthen important features through nonlinear transformation, and output a fused feature vector adapted to the disease.

7. The AI-based gene data analysis system as described in claim 1, characterized in that, The gene data analysis module includes the following units: 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.

8. A method for implementing an AI-based gene data analysis system as described in claim 1, characterized in that, The method includes the following steps: We acquire patients’ genomic, epigenomic, and proteomic data, process the data, and obtain a multi-omics data matrix. 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. 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. 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.

9. A method for implementing an AI-based gene data analysis system as described in claim 8, characterized in that, The method includes the following steps: 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; 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. 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.

10. A method for implementing an AI-based gene data analysis system as described in claim 8, characterized in that, The method includes the following steps: 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. 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.

Images