Osteoporosis diagnosis method and system based on multi-omics joint machine learning

Abstract

The application discloses an osteoporosis diagnosis method and system based on multi-omics joint machine learning, relates to the field of artificial intelligence medicine, and is particularly designed for the needs of pilots. The method obtains multi-omics data of osteopenia and osteoporosis states, screens out a set of differentially expressed biomarkers, and performs gene function enrichment analysis to identify target biomarkers with diagnostic significance. Further, an osteoporosis diagnosis model is constructed based on these biomarkers, and the model is used to generate diagnosis information for target users. This method combines the physiological characteristics of pilots with multi-omics data to achieve personalized and accurate osteoporosis risk prediction and diagnosis, and is suitable for health management of high-intensity professional groups such as pilots.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence medicine, in particular to an osteoporosis diagnosis method and system based on multi-omics joint machine learning. BACKGROUND

[0002] With the aggravation of population aging, the incidence of osteoporosis is rapidly rising. For special occupational groups, especially high-intensity professional groups such as pilots, the health management has brought severe challenges. Osteoporosis can significantly increase the risk of bone fracture, affecting the physical health and professional safety of pilots, so early screening and accurate prediction of osteoporosis are of great significance to protect the professional health of pilots. At present, the diagnosis of osteoporosis usually relies on bone density detection methods such as dual-energy X-ray absorptiometry (DXA) and quantitative CT (QCT), but these methods require professional equipment and personnel, and it is difficult to realize high-frequency and wide-range screening. Therefore, the osteoporosis risk prediction method for special groups such as pilots needs to be explored in order to realize early detection and risk control.

[0003] For example, the Chinese patent with the granted announcement number CN115775628A discloses an osteoporosis risk prediction method based on a relationship network. The method collects physiological indicators and questionnaire data of the subjects, constructs a risk prediction model based on a relationship network, and uses network features combined with Logistic regression, SVM, random forest and other algorithms to predict the risk of osteoporosis.

[0004] The above method has the problems proposed in the background, and the effect of the method for some special groups is not good, and the network construction and analysis rely on fixed network feature construction methods, which is difficult to adapt to the special physiological characteristics and professional needs of the pilot group. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application discloses an osteoporosis diagnosis method and system based on multi-omics joint machine learning.

[0006] In a first aspect, the present application provides an osteoporosis diagnosis method based on multi-omics joint machine learning, comprising:

[0007] obtaining multi-omics data of a user to be diagnosed;

[0008] For the multi-omics data, using a preset osteoporosis diagnosis model, the bone state is diagnosed, and a risk score of osteopenia is generated; the risk score of osteopenia is used for osteoporosis diagnosis;

[0009] The osteoporosis diagnostic model is constructed based on representative differentially expressed biomarkers, which are obtained by performing gene function enrichment analysis on a set of differentially expressed biomarkers, and the set of differentially expressed biomarkers is obtained by screening from a data set of multi-omics data including bone loss and osteoporosis status.

[0010] As an optional embodiment, the multi-omics data includes: multi-omics data of a pilot group; the method further includes:

[0011] Based on a public database, a gene interaction network is constructed; wherein the nodes in the gene interaction network represent genes, and the edges represent the interactions between genes; the interactions include activation, inhibition, and physical binding;

[0012] Physiological data of a group of pilots are obtained, and the physiological data of the group of pilots are used as additional features of nodes and / or edges.

[0013] As an optional embodiment, the enrichment analysis includes:

[0014] The preset multimodal attention enhancement network was used to perform gene function enrichment analysis on the differentially expressed biomarker set to determine representative differentially expressed biomarkers in the construction of osteoporosis diagnostic models.

[0015] As an optional implementation, the multimodal attention enhancement network includes: a graph neural network layer, an attention enhancement layer, and a multimodal fusion layer;

[0016] The graph neural network layer is used to model the graph structure of the gene interaction network and extract the feature representation of each gene node through graph convolution operation;

[0017] The attention enhancement layer is used to dynamically adjust the edge weights between gene nodes based on the pilot's physiological data;

[0018] The multimodal fusion layer is used to fuse the feature representations from the multi-omics data with the pilot's physiological data to generate a comprehensive multimodal feature vector.

[0019] As an optional implementation, the dynamically adjusting edge weights between gene nodes based on the pilot's physiological data includes:

[0020] Using a multilayer perceptron, the standardized physiological data of the pilot is converted into a low-dimensional embedding vector; wherein the embedding vector has the same dimension as the gene node features in the gene interaction network;

[0021] Based on the preset additive attention mechanism, the attention weights between gene nodes are calculated;

[0022] Adjusting edge weights between gene nodes in the gene interaction network based on the attention weights;

[0023] Using the adjusted edge weights, we screened out gene node pairs with an edge weight increase of more than 20% to identify key genes associated with osteopenia;

[0024] Biological pathway enrichment analysis was performed on the key genes related to osteopenia to determine the biological pathways in which they were located.

[0025] As an optional implementation, the calculation of attention weights between gene nodes based on a preset additive attention mechanism includes:

[0026] The pilot’s physiological feature embedding vector is input into a fully connected layer to generate a query vector;

[0027] Input the feature vectors of the gene nodes into independent fully connected layers to generate key vectors and value vectors;

[0028] Generate an attention score by calculating the similarity between the query vector and each key vector;

[0029] Apply the softmax function to all attention scores to obtain normalized attention weights;

[0030] Multiply the normalized attention weight by the corresponding value vector to generate the weighted gene feature representation.

[0031] As an optional implementation, generating an attention score by calculating the similarity between the query vector and each key vector includes:

[0032] After linearly transforming the query vector and the key vector, the similarity score is calculated using a hyperbolic tangent activation function and a trainable weight matrix.

[0033] As an optional implementation manner, adjusting the edge weights between gene nodes in the gene interaction network based on the attention weights includes:

[0034] For gene-node pairs with attention weights exceeding 0.8, their edge weights are increased based on the attention weights, and the increase in edge weights is greater than 20% of the original weights.

[0035] For gene-node pairs with attention weights lower than 0.2, their edge weights are reduced based on the attention weights, and the reduction in edge weights is less than 10% of the original weights.

[0036] Secondly, this application also provides an osteoporosis diagnosis system based on multi-omics combined with machine learning, including:

[0037] An acquisition unit, used to obtain multi-omics data of the user to be diagnosed;

[0038] a diagnostic unit, configured to diagnose bone status based on the multi-omics data using a preset osteoporosis diagnostic model and generate an osteopenia risk score; the osteopenia risk score is used for osteoporosis diagnosis;

[0039] The osteoporosis diagnostic model is constructed based on representative differentially expressed biomarkers, which are obtained by performing gene function enrichment analysis on a set of differentially expressed biomarkers, and the set of differentially expressed biomarkers is obtained by screening from a data set of multi-omics data including bone loss and osteoporosis status.

[0040] Compared with existing technologies, this invention offers the following advantages: Based on a combined analysis of pilots' physiological data and multi-omics data, this invention constructs a multimodal attention-enhanced network, effectively integrating individual physiological and genetic characteristics. Compared with traditional diagnostic methods based on single-omics data or conventional classification models, this invention dynamically adjusts the weights of gene nodes, enabling the model to more sensitively capture an individual's osteoporosis risk, significantly improving the personalization and accuracy of diagnosis.

[0041] This paper innovatively constructs a multimodal attention-enhanced network architecture consisting of a graph neural network, an attention-enhancing layer, and a multimodal fusion layer. Compared to traditional neural networks, this architecture can establish complex feature interactions based on multi-layered data and effectively focus on the impact of different modal information on diagnostic results, thereby improving the robustness and flexibility of the diagnostic model under complex data conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flowchart of the osteoporosis diagnosis method based on multi-omics combined with machine learning provided in an embodiment of the present disclosure;

[0043] Figure 2 A flow chart of a method for calculating attention weights between gene nodes provided in an embodiment of the present disclosure;

[0044] Figure 3 Schematic diagram of an osteoporosis diagnostic system based on multi-omics combined with machine learning provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0046] See Figure 1 , Figure 1 This is a flow chart of a method for diagnosing osteoporosis based on multi-omics combined with machine learning provided in an embodiment of the present disclosure. The method includes steps S101 to S102, wherein:

[0047] S101: Acquire multi-omics data of the user to be diagnosed;

[0048] S102: Based on the multi-omics data, using a preset osteoporosis diagnostic model, diagnose bone status and generate an osteopenia risk score; the osteopenia risk score is used for osteoporosis diagnosis;

[0049] The osteoporosis diagnostic model is constructed based on representative differentially expressed biomarkers, which are obtained by performing gene function enrichment analysis on a set of differentially expressed biomarkers, and the set of differentially expressed biomarkers is obtained by screening from a data set of multi-omics data including bone loss and osteoporosis status.

[0050] It should be emphasized that the diagnosis in the present disclosure only involves data processing and does not involve the modification of living bodies. For example, the osteoporosis diagnostic model is used to diagnose bone status and output a bone reduction risk score. The specific processing process is to perform model calculation processing based on multi-omics data to generate a bone reduction risk score, and this score can be used for subsequent auxiliary osteoporosis diagnosis. It can be seen that the method steps of this application do not include any content about disease treatment itself.

[0051] This paper provides an osteoporosis diagnosis method based on multi-omics combined with machine learning. By comprehensively analyzing the multi-omics data of the user being diagnosed and using a pre-defined osteoporosis diagnostic model, the method aims to diagnose the user's bone status and generate an osteopenia risk score. This risk score is used to assess the user's risk of osteoporosis, providing a scientific basis for early prevention and intervention.

[0052] In a specific implementation, first, multi-omics data of the user to be diagnosed is obtained. The multi-omics data includes genomic data, transcriptomic data, proteomic data, and metabolomic data. These data can be obtained through modern biotechnology methods such as high-throughput sequencing technology and mass spectrometry analysis.

[0053] For example, genomic data can obtain the user's DNA sequence information through whole genome sequencing (WGS); transcriptomic data can obtain the user's gene expression level through RNA sequencing (RNA-Seq); proteomic data can use liquid chromatography-mass spectrometry (LC-MS / MS) to detect protein expression in serum or tissue samples; metabolomic data can analyze the concentration and type of metabolites in the user's body through nuclear magnetic resonance (NMR) or mass spectrometry technology.

[0054] After acquiring the multi-omics data, the data is pre-processed and normalized. For example, the raw data can be first subjected to quality control to remove low-quality reads and possible sequencing errors. Then, the data is normalized to eliminate systematic errors and batch effects among different samples.

[0055] For example, for RNA-Seq data, Fragments Per Kilobase of transcript per Million mapped reads (FPKM) or Transcripts Per Million (TPM) method can be used for normalization. For proteomics and metabolomics data, internal standard method or total ion current normalization method can be used. The purpose of data pre-processing is to ensure the accuracy and reliability of subsequent analysis.

[0056] Next, the pre-processed multi-omics data is subjected to bone status diagnosis using a pre-set osteoporosis diagnosis model. The osteoporosis diagnosis model is constructed based on representative differentially expressed biomarkers.

[0057] In a specific implementation, first, differentially expressed biomarkers are screened from a multi-omics dataset (such as a public database or previous research data) including osteopenia and osteoporosis states. These markers are genes, proteins or metabolites that have significant expression differences between the healthy control group and the osteoporosis patient group. The screening process can use statistical methods such as t-test, ANOVA analysis, and multiple testing correction (such as Benjamini-Hochberg method) to control the false discovery rate.

[0058] The screened set of differentially expressed biomarkers is then subjected to gene function enrichment analysis to determine representative biomarkers therein. Enrichment analysis can use databases such as Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) to reveal the commonality of these markers in biological processes, cellular components and molecular functions, as well as the signaling pathways they participate in. Through enrichment analysis, biomarkers related to key processes such as bone metabolism, bone remodeling, calcium signaling, etc. can be identified, thereby providing a biological basis for model construction.

[0059] Based on the above screening and analysis results, an osteoporosis diagnostic model is constructed. This model can be trained using machine learning algorithms such as support vector machines (SVMs), random forests, or gradient boosting decision trees (GBDTs). The model's input features are the expression levels of representative differentially expressed biomarkers, and its output is a prediction of the user's bone status.

[0060] After the model is built, the user's multi-omics data is input into a pre-set osteoporosis diagnostic model to diagnose bone status and generate an osteopenia risk score. This risk score is a numerical value between 0 and 1, derived from the model's predicted output through a specific mapping relationship. A higher score indicates a greater risk of osteoporosis. Based on the risk score, users can be categorized into low-risk, medium-risk, and high-risk groups, and appropriate health advice and intervention measures can be provided.

[0061] For example, for users with a risk score lower than 0.3, it is recommended to maintain good living habits and undergo regular bone density testing; for users with a risk score between 0.3 and 0.7, it is recommended to increase calcium and vitamin D intake, strengthen physical exercise, and avoid bad habits that affect bone health; for users with a risk score higher than 0.7, it is recommended to seek medical attention in a timely manner and receive professional osteoporosis diagnosis and treatment.

[0062] This comprehensive analysis of multi-omics data and the use of machine learning techniques have enabled early and accurate diagnosis of osteoporosis, demonstrating significant clinical application. Compared to traditional bone density testing, this method can identify the risk of osteopenia earlier, providing a new technical approach for preventing osteoporosis and its associated serious consequences, such as fractures.

[0063] As an optional embodiment, the multi-omics data includes: multi-omics data of a pilot group; the method also includes: constructing a gene interaction network based on a public database; wherein the nodes in the gene interaction network represent genes, and the edges represent interactions between genes; the interactions include: activation, inhibition, and physical binding; obtaining physiological data of the pilot group, and using the physiological data of the pilot group as additional features of the nodes and / or edges.

[0064] In specific implementations, the multi-omics data includes data from a pilot population. Pilots are exposed to specialized occupational environments for extended periods, such as high-altitude flight, simulated weightlessness, and high stress levels. These factors may have unique impacts on bone metabolism. Therefore, data analysis targeting pilots can improve the accuracy and specificity of osteoporosis diagnosis.

[0065] Samples were collected from each pilot, including blood and urine samples, for extraction of genomic DNA, total RNA, proteins, and metabolites. Simultaneously, dual-energy X-ray absorptiometry (DXA) was used to measure the pilots' bone density.

[0066] Next, high-throughput sequencing and mass spectrometry were used to generate multi-omics data for the pilot population.

[0067] For example, genomic data is obtained through whole genome sequencing (WGS) to obtain the pilots' DNA sequence information and identify genetic variations associated with osteoporosis, such as single nucleotide polymorphisms (SNPs). Transcriptomics data uses RNA sequencing (RNA-Seq) technology to obtain the pilots' gene expression profiles and identify differentially expressed genes. Proteomics data uses liquid chromatography-tandem mass spectrometry (LC-MS / MS) technology to identify and quantify proteins in serum or tissue samples. Metabolomics data uses nuclear magnetic resonance spectroscopy (NMR) or mass spectrometry to detect the types and concentrations of metabolites in the pilots' bodies.

[0068] After obtaining multi-omics data, gene interaction networks were constructed based on publicly available biological databases, including the STRING database, the BioGRID database, and the KEGG database, which provide known and predicted gene and protein interaction information.

[0069] Each node in the network represents a gene, derived from a set of differentially expressed genes from the pilot population. Edges between nodes represent interactions between genes, including activation, repression, and physical binding. Activation indicates that one gene product promotes the expression or activity of another; repression indicates that one gene product inhibits the expression or activity of another; and physical binding indicates a direct physical interaction between two gene products.

[0070] In the process of constructing the gene interaction network, the interaction information of the genes of interest is extracted from the above database, and an interaction confidence threshold (eg, 0.7) is set to retain only interactions with high confidence.

[0071] In addition, the network representation can be in the form of an adjacency matrix or an adjacency list to facilitate subsequent computational analysis.

[0072] In specific implementations, physiological data includes, but is not limited to: Bone density data: bone density values ​​obtained through DXA measurement. Biochemical indicators: testing of blood biochemical indicators related to bone metabolism, such as calcium, phosphorus, vitamin D, and parathyroid hormone (PTH) concentrations. Lifestyle and occupational factors: collecting information on pilots' flight time, cumulative flight time, frequency of exposure to high-acceleration environments, as well as dietary habits, exercise frequency, smoking and drinking habits, etc.

[0073] In the specific implementation, the collected physiological data needs to be standardized, and the Z-score normalization method is used to eliminate the dimensional differences between different indicators to facilitate fusion with genetic data.

[0074] During the data integration process, the pilots' physiological data are used as node attributes of the gene interaction network or to adjust the weights of edges.

[0075] For example, bone density data can be associated with gene nodes related to bone formation and bone absorption to reflect the individual's bone health status; the weights of the edges between genes can be adjusted according to physiological data to reflect the impact of physiological state on gene interactions. For example, under high stress conditions, the interactions between certain genes may be enhanced or weakened.

[0076] For example, to achieve effective data fusion, a graph embedding algorithm (such as GraphSAGE) can be used to convert gene interaction networks and physiological data into low-dimensional vector representations.

[0077] In practice, the Python programming language can be used in conjunction with the NetworkX library to process network structures, and the Pandas and NumPy libraries can be used to process data for data preprocessing and integration. This allows for a unified representation of multi-omics and physiological data, laying the foundation for subsequent model building and analysis.

[0078] Integrating pilots' physiological data into the gene interaction network can improve the personalized nature of the diagnostic model, targeting the unique physiological and occupational characteristics of the pilot population, and enhancing the model's accuracy and reliability. The inclusion of physiological data enriches the network's information, enabling a more comprehensive reflection of gene-environment interactions. This will help identify key genes and biological pathways associated with osteoporosis, providing a scientific basis for further research and intervention.

[0079] As an optional embodiment, the enrichment analysis includes:

[0080] The preset multimodal attention enhancement network was used to perform gene function enrichment analysis on the differentially expressed biomarker set to determine representative differentially expressed biomarkers in the construction of osteoporosis diagnostic models.

[0081] In order to perform more accurate gene function enrichment analysis on the differentially expressed biomarker set, a preset multimodal attention enhancement network can be used to identify representative differentially expressed biomarkers in the construction of osteoporosis diagnostic models.

[0082] In practice, after obtaining multi-omics data from the pilot population and screening for differentially expressed biomarkers, a multimodal attention-enhanced network was introduced to conduct in-depth functional analysis of these biomarkers. This network is designed to integrate multiple data modalities and capture complex intergene relationships and functional properties, thereby improving the accuracy and biological interpretability of enrichment analysis.

[0083] First, multiple types of information are collected for each differentially expressed biomarker to construct a multimodal input dataset. This information includes: Gene expression levels: data from transcriptomics, which reflects the expression changes of genes under different states. Protein interaction information: data based on public databases (such as STRING), describing the interaction relationship between gene-encoded proteins. Functional annotation and pathway information: using databases such as Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG), the functional classification of genes and the biological pathways involved are obtained. Other omics data: such as metabolomics data, provides information on metabolite levels.

[0084] An exemplary multimodal attention enhancement network has the following architecture design:

[0085] Input layer: Receives multimodal input data, and each data modality is processed by a separate sub-network. For example, text data (functional annotations) is converted into vector representations by the embedding layer, and numerical data (gene expression levels) is processed by the fully connected layer.

[0086] Feature extraction layer: Extracts high-level feature representations for each modality’s data. For example, convolutional neural networks (CNNs) are used to process sequence data, or graph neural networks (GNNs) are used to process protein interaction networks.

[0087] Attention mechanism layer: This layer introduces an attention mechanism to dynamically weight features of different modalities. By calculating the contribution of each modality to the enrichment analysis task, the network can automatically focus on the most important information.

[0088] Fusion layer: The weighted features of each modality are fused to form a comprehensive feature representation.

[0089] Output layer: Outputs the enrichment probability or score of each biomarker in different functional categories and pathways.

[0090] During model training, existing gene function classifications and pathway participation information can be used to construct training and validation sets. Supervised learning methods are used to minimize the loss function between the predicted output and the true label. Regularization techniques and early stopping strategies can be used to prevent overfitting and improve model generalization.

[0091] After training, the differentially expressed biomarker set is fed into the trained multimodal attention-enhanced network. The model comprehensively analyzes the multimodal information of each biomarker and outputs enrichment scores for each biomarker across different functional categories and pathways. These scores can be used to identify key functions and pathways, identifying biological processes and signaling pathways related to osteoporosis, such as bone remodeling, calcium signaling, and the Wnt signaling pathway.

[0092] In addition, representative biomarkers can be screened and biomarkers with high enrichment in key functions and pathways can be selected as core features for diagnostic model construction.

[0093] By using a multimodal attention-enhanced network, enrichment analysis not only considers gene expression levels but also integrates protein interactions, functional annotations, and pathway information, capturing the complex biological relationships between genes. The introduction of the attention mechanism enables the model to automatically identify and emphasize the most diagnostically valuable features, improving the accuracy and reliability of enrichment analysis.

[0094] As an optional implementation, the multimodal attention enhancement network includes: a graph neural network layer, an attention enhancement layer, and a multimodal fusion layer;

[0095] The graph neural network layer is used to model the graph structure of the gene interaction network and extract the feature representation of each gene node through graph convolution operation;

[0096] The attention enhancement layer is used to dynamically adjust the edge weights between gene nodes based on the pilot's physiological data;

[0097] The multimodal fusion layer is used to fuse the feature representations from the multi-omics data with the pilot's physiological data to generate a comprehensive multimodal feature vector.

[0098] Among them, the network aims to deeply explore the set of differentially expressed biomarkers, combine them with the pilots' physiological data, and build a more accurate and personalized osteoporosis diagnostic model.

[0099] In practice, the graph neural network layer is used to model the graph structure of the constructed gene interaction network. Nodes in the gene interaction network represent genes, and edges represent interactions between genes (such as activation, inhibition, and physical binding). In this layer, a graph convolutional neural network (GCN) is used to process the network.

[0100] Through graph convolution operations, the model can effectively aggregate neighborhood information for each gene node and extract rich feature representations. These feature representations not only contain the properties of the gene itself, but also incorporate the influence of its neighboring genes, reflecting the gene's position and functional role in the network topology.

[0101] In practice, the attention enhancement layer dynamically adjusts the edge weights between gene nodes based on the pilot's physiological data. This data includes bone density, blood biochemical indicators (such as calcium, phosphorus, and vitamin D levels), and stress levels. After normalization, this data serves as one of the model's inputs.

[0102] The attention enhancement layer, by introducing an attention mechanism, dynamically adjusts the strength of interactions between gene nodes based on the pilot's individual physiological characteristics. For example, this layer calculates an attention weight between each pair of gene nodes. The weight reflects the importance of that gene pair in diagnosing osteoporosis under the current physiological state. This allows the model to focus more on gene interactions that are highly relevant to the pilot's physiological state, improving diagnostic accuracy and personalization.

[0103] In practice, the multimodal fusion layer is used to fuse gene signatures from multi-omics data with the pilot's physiological data to generate a comprehensive multimodal feature vector. In this layer, a fusion strategy is employed to effectively combine data from different modalities.

[0104] An exemplary approach is feature concatenation, which directly concatenates genetic feature representations and physiological data vectors to form a high-dimensional feature vector. Another approach is weighted fusion, which assigns different weights to each modality based on its importance and then adds them together to create a composite feature. Through multimodal fusion, the model can simultaneously leverage both molecular information at the genetic level and physiological information at the individual level, providing a more comprehensive assessment of osteoporosis risk.

[0105] In this way, through the design of the multimodal attention-enhanced network, the model fully utilizes the structural information of the gene interaction network, the characteristics of differentially expressed biomarkers, and the pilots' physiological data to achieve accurate diagnosis of osteoporosis. This approach not only improves the model's predictive performance but also enhances its adaptability to individual differences.

[0106] As an optional implementation, the dynamically adjusting edge weights between gene nodes based on the pilot's physiological data includes:

[0107] Using a multilayer perceptron, the standardized physiological data of the pilot is converted into a low-dimensional embedding vector; wherein the embedding vector has the same dimension as the gene node features in the gene interaction network;

[0108] Based on the preset additive attention mechanism, the attention weights between gene nodes are calculated;

[0109] Adjusting edge weights between gene nodes in the gene interaction network based on the attention weights;

[0110] Using the adjusted edge weights, we screened out gene node pairs with an edge weight increase of more than 20% to identify key genes associated with osteopenia;

[0111] Biological pathway enrichment analysis was performed on the key genes related to osteopenia to determine the biological pathways in which they were located.

[0112] See Figure 2 , Figure 2 Flowchart of the method for calculating attention weights between gene nodes provided in an embodiment of the present disclosure. As an optional implementation, the method for calculating attention weights between gene nodes based on a preset additive attention mechanism includes steps S201 to S205:

[0113] S201: Input the pilot’s physiological feature embedding vector into a fully connected layer to generate a query vector;

[0114] S202: Input the feature vectors of the gene nodes into independent fully connected layers to generate key vectors and value vectors;

[0115] S203: Generate an attention score by calculating the similarity between the query vector and each key vector;

[0116] S204: Apply the softmax function to all attention scores to obtain normalized attention weights;

[0117] S205: Multiply the normalized attention weight by the corresponding value vector to generate a weighted gene feature representation.

[0118] Furthermore, the present invention dynamically adjusts the edge weights between gene nodes based on the pilot's physiological data to further improve the accuracy and personalization of the osteoporosis diagnosis model.

[0119] In practice, pilots' physiological data must be normalized to eliminate dimensional differences between different indicators. This data includes bone density, blood biochemical indicators (such as calcium, phosphorus, and vitamin D levels), and stress levels. After normalization, this data is converted into a low-dimensional embedding vector using a multi-layer perceptron (MLP). The dimensionality of this embedding vector matches that of the gene node features in the gene interaction network, ensuring effective integration in subsequent calculations.

[0120] In the specific implementation, the attention weights between gene nodes are calculated based on the preset additive attention mechanism. Specifically, the pilot's physiological feature embedding vector is input into a fully connected layer to generate a query vector (Query); at the same time, the feature vectors of the gene nodes are input into independent fully connected layers to generate corresponding key vectors (Key) and value vectors (Value). Then, the similarity between the query vector and each key vector is calculated to generate an attention score. The similarity is calculated by linearly transforming the query vector and the key vector, applying an appropriate activation function (such as ReLU or tanh), and then multiplying them with a trainable weight matrix to obtain the final similarity score. The softmax function is applied to all attention scores to obtain normalized attention weights.

[0121] In practice, the obtained attention weights are used to dynamically adjust the edge weights between gene nodes in the gene interaction network. For gene node pairs with attention weights exceeding 0.8, they are considered to have a stronger correlation with the current pilot's physiological state, and their edge weights are increased by more than 20% of the original weight. For gene node pairs with attention weights below 0.2, their correlation is considered to be weaker, and their edge weights are appropriately reduced by less than 10% of the original weight. This dynamic adjustment allows the model to focus more on gene interactions that are closely related to the pilot's physiological state.

[0122] Using the adjusted edge weights, we further screened for gene node pairs whose edge weights increased by more than 20%. These gene node pairs exhibited stronger interactions under the current physiological state and played a key role in the development and progression of osteopenia. After identifying these key osteopenia-related genes, we conducted biological pathway enrichment analysis.

[0123] Use public databases (such as KEGG and Reactome) to determine the biological pathways in which these genes are located, such as the Wnt signaling pathway and RANK / RANKL pathway that are closely related to bone metabolism.

[0124] This information helps us gain a deeper understanding of the molecular mechanisms of osteoporosis in pilots and provides a scientific basis for personalized prevention and treatment plans. By combining pilots' physiological data with the gene interaction network and dynamically adjusting edge weights, the model can more accurately reflect individual differences and improve the accuracy and reliability of osteoporosis diagnosis.

[0125] As an optional implementation, generating an attention score by calculating the similarity between the query vector and each key vector includes:

[0126] After linearly transforming the query vector and the key vector, the similarity score is calculated using a hyperbolic tangent activation function and a trainable weight matrix.

[0127] In a specific implementation, in order to effectively capture the complex nonlinear relationship between the pilot's physiological feature embedding vector (query vector) and the gene node feature vector (key vector), the present invention uses a trainable weight matrix to perform a linear transformation on these vectors.

[0128] Specifically, the pilot's physiological feature embedding vector is normalized and then fed into a fully connected layer. This fully connected layer, which contains a trainable weight matrix and bias term, generates a new query vector by performing a linear transformation on the input vector. This step aims to map the original physiological features into a new feature space, enabling the model to better learn and represent them.

[0129] At the same time, the feature vectors of each gene node are fed into separate fully connected layers, where they undergo similar linear transformations to generate new key vectors. These fully connected layers also contain trainable weight matrices and bias terms. By linearly transforming the key vectors, the model adjusts the representation of gene node features so that they are in the same feature space as the query vector, facilitating subsequent similarity calculations.

[0130] After the linear transformation, the hyperbolic tangent activation function (tanh) is applied to the new query and key vectors. This activation function introduces nonlinearity, allowing the model to capture more complex nonlinear relationships between the query and key vectors. Since the output range of the tanh function is (-1, 1), this helps stabilize the model training process and prevent problems such as vanishing or exploding gradients.

[0131] Next, the activated query vector is concatenated with each activated key vector to form a longer vector. This concatenated vector contains comprehensive information about the pilot's physiological characteristics and gene node features, more comprehensively reflecting the correlation between the two. This concatenated vector is then input into another fully connected layer, which also contains a trainable weight matrix and bias term. Through linear transformation and activation of the concatenated vector, the model outputs a similarity score indicating the degree of similarity between the query vector and the key vector.

[0132] With the attention weights, the value vectors of each gene node (usually a linear transformation of the key vector) are weighted and summed to generate a comprehensive weighted gene feature representation. This weighted feature representation incorporates genetic information that is highly correlated with the pilot's physiological characteristics, providing key input for subsequent osteoporosis risk assessment.

[0133] During model training, a training set containing pilot physiological and genetic data is used to update the model's trainable weight matrices using backpropagation and an optimization algorithm (such as the Adam optimizer). Specifically, the model first calculates the predicted output through forward propagation, compares it with the actual osteopenia risk score or diagnostic result, and calculates a loss function (such as mean squared error). The model then calculates the gradient of the loss function with respect to each weight matrix through backpropagation. Based on this gradient information, the optimization algorithm adjusts the weight matrix parameters to gradually reduce the loss function.

[0134] During training, the trainable weight matrix gradually learned how to effectively map the pilot's physiological characteristics and gene node features into the corresponding feature space, as well as how to calculate the similarity score between the two. Through continuous adjustment and optimization, the model can more accurately identify key gene nodes related to osteoporosis, improving the accuracy and reliability of diagnosis.

[0135] It should be noted that to prevent model overfitting, regularization techniques (such as L2 regularization) and dropout methods may be used during training. In this way, using a trainable weight matrix and tanh activation function, the model has strong learning ability and adaptability, can handle complex biological data and individual differences of pilots, and provides technical support for personalized medicine.

[0136] Based on the same inventive concept, an osteoporosis diagnosis system based on multi-omics combined with machine learning is also provided in the embodiments of the present application. Since the principle of solving the problem by the osteoporosis diagnosis system based on multi-omics combined with machine learning in the embodiments of the present application is similar to the above-mentioned osteoporosis diagnosis method based on multi-omics combined with machine learning in the embodiments of the present disclosure, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be repeated.

[0137] Reference Figure 3 As shown, Figure 3 This is a schematic diagram of an osteoporosis diagnosis system based on multi-omics combined with machine learning provided by an embodiment of the present disclosure, the system comprising:

[0138] The acquisition unit 10 is used to obtain multi-omics data of the user to be diagnosed;

[0139] The diagnostic unit 20 is configured to diagnose bone status based on the multi-omics data using a preset osteoporosis diagnostic model and generate an osteopenia risk score; the osteopenia risk score is used for osteoporosis diagnosis;

[0140] The osteoporosis diagnostic model is constructed based on representative differentially expressed biomarkers, which are obtained by performing gene function enrichment analysis on a set of differentially expressed biomarkers, and the set of differentially expressed biomarkers is obtained by screening from a data set of multi-omics data including bone loss and osteoporosis status.

[0141] For descriptions of the processing flow of each unit in the system and the interaction flow between the units, reference may be made to the relevant descriptions in the above method embodiments, which will not be described in detail here.

[0142] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0143] It should be understood that determining B based on A does not mean determining B only based on A. B can also be determined based on A and / or other information.

[0144] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0145] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0146] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present application. The preferred embodiments do not describe all details in detail, nor do they limit the present application to specific embodiments. Obviously, many modifications and variations can be made based on the contents of this specification. This specification selects and describes these embodiments in detail to better explain the principles and practical applications of the present application, so that those skilled in the art can better understand and utilize the present application. The present application is limited only by the claims and their full scope and equivalents.

Claims

1. An osteoporosis diagnosis method based on multi-omics combined with machine learning, characterized by: include: Obtain multi-omics data of the user to be diagnosed; Based on the multi-omics data, a preset osteoporosis diagnostic model is used to diagnose bone status and generate an osteopenia risk score; the osteopenia risk score is used for osteoporosis diagnosis; The osteoporosis diagnostic model is constructed based on representative differentially expressed biomarkers, which are obtained by performing gene function enrichment analysis on a set of differentially expressed biomarkers, which are obtained by screening from a dataset of multi-omics data including osteopenia and osteoporosis status; The multi-omics data includes: multi-omics data of a pilot group; the method further includes: Based on a public database, a gene interaction network is constructed; wherein the nodes in the gene interaction network represent genes, and the edges represent the interactions between genes; the interactions include activation, inhibition, and physical binding; Acquire physiological data of a group of pilots, and use the physiological data of the group of pilots as additional features of nodes and / or edges; The enrichment analysis includes: Using a preset multimodal attention-enhanced network, gene function enrichment analysis is performed on the differentially expressed biomarker set to determine representative differentially expressed biomarkers in the construction of an osteoporosis diagnostic model; The multimodal attention enhancement network includes: a graph neural network layer, an attention enhancement layer, and a multimodal fusion layer; The graph neural network layer is used to model the graph structure of the gene interaction network and extract the feature representation of each gene node through graph convolution operation; The attention enhancement layer is used to dynamically adjust the edge weights between gene nodes based on the pilot's physiological data; The multimodal fusion layer is used to fuse the feature representations from the multi-omics data with the pilot's physiological data to generate a comprehensive multimodal feature vector; The dynamic adjustment of edge weights between gene nodes based on the pilot's physiological data includes: Using a multilayer perceptron, the standardized physiological data of the pilot is converted into a low-dimensional embedding vector; wherein the embedding vector has the same dimension as the gene node features in the gene interaction network; Based on the preset additive attention mechanism, the attention weights between gene nodes are calculated; Adjusting edge weights between gene nodes in the gene interaction network based on the attention weights; Using the adjusted edge weights, we screened out gene node pairs with an edge weight increase of more than 20% to identify key genes associated with osteopenia; Biological pathway enrichment analysis was performed on the key genes related to osteopenia to determine the biological pathways in which they were located.

2. The method according to claim 1, characterized in that The calculation of the attention weights between gene nodes based on the preset additive attention mechanism includes: The pilot’s physiological feature embedding vector is input into a fully connected layer to generate a query vector; Input the feature vectors of the gene nodes into independent fully connected layers to generate key vectors and value vectors; Generate an attention score by calculating the similarity between the query vector and each key vector; Apply the softmax function to all attention scores to obtain normalized attention weights; Multiply the normalized attention weight by the corresponding value vector to generate the weighted gene feature representation.

3. The method according to claim 2, characterized in that The process of generating an attention score by calculating the similarity between the query vector and each key vector includes: After linearly transforming the query vector and the key vector, the similarity score is calculated using a hyperbolic tangent activation function and a trainable weight matrix.

4. The method according to claim 3, characterized in that The adjusting the edge weights between gene nodes in the gene interaction network based on the attention weights includes: For gene-node pairs with attention weights exceeding 0.8, their edge weights are increased based on the attention weights, and the increase in edge weights is greater than 20% of the original weights. For gene-node pairs with attention weights lower than 0.2, their edge weights are reduced based on the attention weights, and the reduction in edge weights is less than 10% of the original weights.

5. An osteoporosis diagnosis system based on multi-omics combined with machine learning, characterized by: include: An acquisition unit, used to obtain multi-omics data of the user to be diagnosed; a diagnostic unit, configured to diagnose bone status based on the multi-omics data using a preset osteoporosis diagnostic model and generate an osteopenia risk score; the osteopenia risk score is used for osteoporosis diagnosis; The osteoporosis diagnostic model is constructed based on representative differentially expressed biomarkers, which are obtained by performing gene function enrichment analysis on a set of differentially expressed biomarkers, which are obtained by screening from a dataset of multi-omics data including osteopenia and osteoporosis status; The multi-omics data includes: multi-omics data of a pilot group; Based on a public database, a gene interaction network is constructed; wherein the nodes in the gene interaction network represent genes, and the edges represent the interactions between genes; the interactions include activation, inhibition, and physical binding; Acquire physiological data of a group of pilots, and use the physiological data of the group of pilots as additional features of nodes and / or edges; The enrichment analysis includes: Using a preset multimodal attention-enhanced network, gene function enrichment analysis is performed on the differentially expressed biomarker set to determine representative differentially expressed biomarkers in the construction of an osteoporosis diagnostic model; The multimodal attention enhancement network includes: a graph neural network layer, an attention enhancement layer, and a multimodal fusion layer; The graph neural network layer is used to model the graph structure of the gene interaction network and extract the feature representation of each gene node through graph convolution operation; The attention enhancement layer is used to dynamically adjust the edge weights between gene nodes based on the pilot's physiological data; The multimodal fusion layer is used to fuse the feature representations from the multi-omics data with the pilot's physiological data to generate a comprehensive multimodal feature vector; The dynamic adjustment of edge weights between gene nodes based on the pilot's physiological data includes: Using a multilayer perceptron, the standardized physiological data of the pilot is converted into a low-dimensional embedding vector; wherein the embedding vector has the same dimension as the gene node features in the gene interaction network; Based on the preset additive attention mechanism, the attention weights between gene nodes are calculated; Adjusting edge weights between gene nodes in the gene interaction network based on the attention weights; Using the adjusted edge weights, we screened out gene node pairs with an edge weight increase of more than 20% to identify key genes associated with osteopenia; Biological pathway enrichment analysis was performed on the key genes related to osteopenia to determine the biological pathways in which they were located.

Images