Parkinson's disease evolution prediction modeling method based on high-order gene communication tensor

Abstract

The invention belongs to the field of disease evolution trajectory prediction in precision medicine, and discloses a Parkinson's disease evolution prediction modeling method based on high-order gene communication tensor. According to the method, firstly, based on an intercellular communication mechanism, a high-order gene module communication tensor is constructed as spatial topological data, so that an information interaction mode between gene modules is described; secondly, a self-adaptive Transform module is designed to extract gene expression features and high-order space topology features at the same time; and finally, integrating multi-view interaction information between the high-order gene module communication tensor and the whole blood transcriptome data through a multi-view collaborative fusion mechanism to realize effective fusion and utilization of features. Experimental results show that compared with a traditional genetic map, the high-order space representation provided by the invention has finer resolution capability, a plurality of evaluation indexes are remarkably improved, and a new technical means is provided for accurate staging of Parkinson's disease.

Description

Technical Field

[0001] This invention belongs to the field of disease evolution trajectory prediction in precision medicine, and in particular relates to a Parkinson's disease evolution prediction modeling method based on high-order gene communication tensors. Background Technology

[0002] The pathogenesis of neurodegenerative diseases is a key research focus in the field of precision medicine. These diseases possess complex pathological features, involving the dynamic synergistic effects of multiple factors, including genetic susceptibility, age-related degenerative changes, and environmental exposure. Notably, these pathogenic factors leave unique molecular imprints at the transcriptional regulatory level, making it possible to assess disease progression stages through high-throughput gene expression profiling. In recent years, breakthroughs in single-cell RNA sequencing (scRNA-seq) technology have brought new opportunities for the precise diagnosis of neurodegenerative diseases such as Parkinson's disease (PD) and Alzheimer's disease (AD). This technology can resolve cellular heterogeneity characteristics in the central nervous system at single-cell resolution and reveal abnormal activation or inhibition states of key regulatory networks.

[0003] In-depth studies based on human whole blood RNA sequencing data have successfully identified multiple transcriptomic features closely related to disease development stages. These features include, but are not limited to, differentially expressed genes (DEGs), aberrantly activated signaling pathways, alternative splicing variants, and abnormal regulation of non-coding RNAs. These findings not only significantly deepen our understanding of the molecular mechanisms of neurodegenerative diseases, but more importantly, these detectable transcriptomic features in peripheral blood are becoming highly valuable clinical biomarkers, providing novel technological approaches for early warning of brain diseases, molecular subtype identification, and monitoring of treatment responses.

[0004] In data analysis and pattern recognition, Graph Neural Networks (GNNs) have demonstrated significant advantages in high-throughput transcriptome data analysis in recent years due to their excellent data mining capabilities. Although various deep learning methods have been developed to identify anomalous patterns in gene expression data, directly applying traditional deep neural networks to RNA sequencing data often fails to achieve ideal results. This limitation stems primarily from two factors: firstly, RNA sequencing data is characterized by high dimensionality and high noise; secondly, single-perspective analysis methods struggle to comprehensively capture the multi-level and dynamically changing features of molecular systems.

[0005] Recent research advances have shown that extracting spatial pattern features from single-cell transcriptome data can not only significantly improve the predictive accuracy of computational models but also effectively enhance their biological interpretability. In particular, systems biology analysis methods based on disease-specific gene networks offer a novel research perspective for revealing the molecular mechanisms of disease development and identifying potential drug targets. Taking Parkinson's disease as an example, by constructing disease-specific geneco-expression networks (PD-GCNs), researchers have been able to preliminarily reveal the structural dysregulation features of its gene regulatory network. However, relying solely on such coarse-grained network information is still insufficient to establish high-precision diagnostic models with clinical applicability.

[0006] To more accurately capture the molecular dynamics of early-stage Parkinson's disease, it is essential to delve deeper into the fine-grained information within single-cell RNA sequencing data, thereby constructing more comprehensive and predictive computational models. Notably, PD-GCN networks naturally contain topological information at different scales; systematically exploring the information flow patterns within the network can help understand the structural dysregulation mechanisms of gene networks at high resolution.

[0007] This invention innovatively draws on the analytical paradigm of intercellular communication, conceptualizing the dynamic communication patterns between gene modules as information flow and functional hierarchy features in disease-specific gene maps. Based on this idea, a four-dimensional higher-order gene module communication tensor, also known as a four-dimensional gene module communication tensor (4D-GMC), was invented. This tensor can effectively capture the higher-order spatial pattern features in gene regulatory networks.

[0008] This invention innovatively applies high-order 4D-GMC tensors to the structural dysregulation analysis of Parkinson's disease-specific gene co-expression networks; then, an adaptive Transformer architecture is used to extract the temporal dynamic features of gene expression and the spatial topological features of the 4D-GMC tensor; finally, a multi-perspective collaborative fusion mechanism is designed to integrate the multi-perspective interaction information between the 4D-GMC tensor and gene expression levels through attention weighting, achieving optimal fusion and collaborative utilization of features at different scales and from different perspectives. Summary of the Invention

[0009] To address the problems existing in the background technology, this invention proposes a Parkinson's disease evolution prediction modeling method based on high-order gene communication tensors. By conducting in-depth research on whole blood transcriptomics data of Parkinson's disease, and synergistically fusing data from two perspectives for training, a new framework is provided for the evolutionary modeling of Parkinson's disease.

[0010] To solve the above problems, the present invention adopts the following technical solution: A Parkinson's disease evolution prediction modeling method based on high-order gene communication tensors includes the following steps: S1: Assume that the whole blood transcriptome data at each time step is X, containing m patients and n genes; use the multi-scale embedded gene co-expression network analysis algorithm (MEGENA) to perform co-expression network analysis on the whole blood transcriptome data X at each time step to obtain a Parkinson's-specific gene map G containing the correlation between genes; then use a method based on intercellular communication patterns to construct a higher-order gene communication tensor, namely the four-dimensional communication tensor 4D-GMC; S2: A multi-scale adaptive Transformer model was used to extract whole blood transcriptome data. Temporal characteristics and higher-order gene communication tensors Spatial characteristics; S3: Employs a multi-perspective collaborative fusion scheme to integrate multi-perspective interactive information between high-order gene communication tensors and whole blood transcriptome data, thereby achieving effective feature fusion and utilization; S4: Use fully connected layers as classifiers for Parkinson's disease evolution modeling, predicting motor and non-motor dysfunction scores for the next time step based on data from the previous time step.

[0011] Furthermore, the whole blood transcriptome data in step S1 specifically includes, but is not limited to: S11: Blood transcriptome (sncRNA) data from the Parkinson's Disease Progression Biomarker Initiative (PPMI) project; S12: Parkinson's disease data and blood transcriptome data from the Biomarker Project (PDBP).

[0012] Furthermore, the basic steps of the multi-scale embedded gene co-expression network analysis algorithm in step S1 include: S13: Calculate the correlation coefficient matrix between genes; S14: Fast Plane Filtering Network Construction with Parallelization, Early Termination and Prior Quality Control - FPFNC, which is part of the MEGENA algorithm in this invention; S15: Use multi-scale cluster analysis to divide the gene modules into k modules (M1, M2, ..., M...). K Its formula is: ; in denoted as gene module compactness score, used to divide gene modules, where k represents the number of gene modules; This represents the number of edges within the i-th gene module. This represents the total number of edges. This represents the resolution parameter.

[0013] Furthermore, the method for calculating the correlation coefficient of gene pairs in step S13 includes, but is not limited to, methods for calculating the correlation coefficient of gene pairs. The first method is to use the Pearson correlation coefficient method to calculate the correlation between genes. The specific formula is as follows: ; Assume there are a total of m patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and Let represent the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. and This represents the average expression level of the i-th and j-th genes; The second method is to use Spearman's rank correlation coefficient to calculate the correlation between genes. The specific formula is as follows: ; Assume there are a total of m patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and These represent the rankings of the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. The third method is to use Kendall's Tau correlation coefficient to calculate the correlation between genes. The specific formula is as follows: ; Assume there are a total of m patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and Let represent the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. and Let represent the expression values ​​of the i-th and j-th genes in the l-th patient, respectively, and C represent the number of identical pairs, i.e. The number of inconsistent pairs, D, is the number of inconsistent pairs. Quantity, This represents the number of times the expression value of the i-th gene is the same. This represents the number of times the expression value of the j-th gene is the same.

[0014] Furthermore, the basic steps for constructing the higher-order gene communication tensor in step S1 include: S16: Construct a two-dimensional tensor for the k gene modules (M1, M2, ..., MK) partitioned in S15. Specifically, for gene a from gene module A, if gene b is connected to it in gene module B, it indicates that there is information exchange between module A and module B, and the communication quantity increases by 1. For gene a, the final two-dimensional communication tensor is represented as follows: The formula is as follows: ; Where i and j represent the indices of the gene modules, respectively; S17: After calculating the two-dimensional communication tensor, for the i-th patient, use all genes to generate the three-dimensional communication tensor. The structure is as follows: G represents the total number of genes; the specific construction formula is as follows: ; Where g represents the g-th gene, and G represents the total number of genes. This represents a two-dimensional communication tensor constructed using the g-th gene; S18: Finally, the three-dimensional communication tensors of all samples are combined to form the final four-dimensional communication tensor 4D-GMC. The specific construction formula is as follows: ; Where n represents the nth patient, and N represents the total number of patients. This represents the three-dimensional communication tensor constructed using the nth patient.

[0015] Furthermore, the basic steps in step S2 for extracting spatiotemporal features using a multi-scale adaptive Transformer model include: S21: To process whole blood transcriptome data X, this invention designs a linear layer to perform linear projection on X, thereby generating feature vectors of the whole blood transcriptome data. The formula is as follows: ; Here, Liner represents a linear layer in a neural network model; S22: In order to process whole blood transcriptome data This invention uses a combination of convolutional neural networks and the KAN model to extract feature tensors, and the final result is... ; S23: Multi-scale adaptive Transformer was used to extract features at the whole blood transcriptome scale and the higher-order communication tensor scale, respectively. The two scale features obtained were as follows: and Its formula is: ; Wherein, the query vector Q, the key vector K, and the value vector V are respectively and These are obtained through linear layers, where T represents the transpose operation and softmax represents the activation function. Indicates a parameter.

[0016] Furthermore, the basic steps in step S22 using convolutional and KAN models include: S221: Four-dimensional gene module communication tensor The data is first fed into a convolutional layer with 2000 input channels and 1024 output channels; then, an activation function is applied to the output. Its formula is: ; Where ReLU represents the activation function and Conv represents the convolutional layer; S222: Subsequently, an average pooling layer is used to... The dimension is reduced to Then, the obtained tensor is flattened into two dimensions, as shown in the following formula: ; Here, `view` represents the flattening operation, and `AvgPool` represents the average pooling layer operation. This represents the feature representation after extraction by the average pooling layer. S223: Finally, a KAN layer is used to extract non-linear relationships in high-dimensional data, and its formula is as follows: ; Wherein, KAN represents the KAN layer. express The result after feature extraction using the KAN model.

[0017] Furthermore, the basic steps of multi-view collaborative fusion in step S3 include: S31: The core of multi-view collaborative fusion is the cross-attention mechanism, and its formula is: ; ; ; in, They represent from Exported queries ( ),key( ) and value ( );akin, They represent from Exported queries ( ),key( ) and value ( ); and These are the feature matrices generated after the first and second cross-attention processes, respectively. The final result obtained after two cross-attention processes is... ; S32: Subsequently, and Feature fusion is performed to obtain an enhancement matrix containing more information. The formula is as follows: ； Here, Concat represents the feature concatenation operation.

[0018] Furthermore, the basic steps for disease stage identification using fully connected layers in step S4 include: S41: Using a fully connected layer with a softmax activation function The formula for classification prediction is as follows: ; The softmax function converts the output of the fully connected layer into class probabilities, and FC represents a fully connected layer. S42: The cross-entropy loss function is used to extract characteristics and commonalities across perspectives. Its mathematical formula is: ; Where L represents the loss function, which quantifies the difference between the predicted probability distribution output by the model and the true label distribution, reflecting the model's prediction accuracy. It is a hyperparameter. for and The cross-entropy loss function between them for The cross-entropy loss function, for The cross-entropy loss function.

[0019] Furthermore, in step S4, four metrics—accuracy (ACC), F1-score, recall, and precision—are used to evaluate the model's accuracy. Accuracy refers to the proportion of correctly predicted samples out of the total number of samples, providing a direct assessment of the model's overall performance. Recall, also known as the full detection rate, represents the ratio of correctly predicted positive instances to the total number of actual positive instances, evaluating the model's ability to discover all relevant instances. The F1-score is the harmonic mean of precision and recall, comprehensively considering both the model's accuracy and completeness. Precision represents the proportion of correctly predicted positive samples out of the total number of predicted positive samples, focusing on the accuracy of the model's positive predictions. Its formula is as follows: ; ; ; ; TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative. The higher these indicators are, the more accurate the prediction results are.

[0020] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention uses a multi-scale embedded gene co-expression network analysis method to generate a silver co-expression network. Then, based on intercellular communication pattern analysis, a high-order gene communication tensor is constructed to capture high-order spatial patterns in the gene network, i.e., information interaction patterns between gene modules. The constructed communication tensor contains more information and has higher resolution compared to traditional gene co-expression networks.

[0021] (2) This invention first designs an adaptive Transformer module, which introduces a learnable adaptive weight mechanism that can simultaneously extract gene expression features and high-order spatial topological features. Finally, a collaborative fusion module based on a cross-attention mechanism is constructed to achieve deep integration and complementary enhancement of gene expression features and spatial information. Attached Figure Description

[0022] Figure 1 This is a flowchart of Parkinson's disease evolution prediction based on higher-order gene communication tensors in a specific embodiment of the present invention; Figure 2 This is a Transformer framework based on multiple perspectives and scales in a specific embodiment of the present invention; Figure 3 This is a box plot comparing different baseline methods in the Motor PPMI case in an embodiment of the present invention; Figure 4This is a visualization of the communication pattern of a Motor PPMI case at the initial treatment moment in an embodiment of the present invention; Figure 5 This is a visualization of the communication patterns of a Motor PDBP case at a six-month follow-up time in an embodiment of the present invention; Figure 6 This is a visualization of the gene co-expression network in the Motor PPMI case in this embodiment of the invention; Figure 7 This is a visualization of the gene co-expression network in the Motor PDBP case in an embodiment of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention are described below to enable any person skilled in the art to understand the present invention. However, the scope of protection of the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, any equivalent substitutions or modifications made to the technical solutions and inventive concepts of the present invention should be covered within the scope of protection of the present invention.

[0024] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0025] Example 1 like Figure 1 As shown, a Parkinson's disease evolution prediction modeling method based on high-order gene communication tensors includes the following steps: S1: Assume that the whole blood transcriptome data at each time step is X, containing m patients and n genes; use the multi-scale embedded gene co-expression network analysis algorithm (MEGENA) to perform co-expression network analysis on the whole blood transcriptome data X at each time step to obtain a Parkinson's-specific gene map G containing the correlation between genes, and then use a method based on intercellular communication patterns to construct a higher-order gene communication tensor (4D-GMC). S2: A multi-scale adaptive Transformer model was used to extract whole blood transcriptome data. Temporal characteristics and higher-order gene communication tensors Spatial characteristics; S3: Employs a multi-perspective collaborative fusion scheme to integrate multi-perspective interactive information between high-order gene communication tensors and whole blood transcriptome data, thereby achieving effective feature fusion and utilization; S4: Use fully connected layers as classifiers for Parkinson's disease evolution modeling, predicting motor and non-motor dysfunction scores for the next time step based on data from the previous time step.

[0026] The whole blood transcriptome data in step S1 includes, but is not limited to: S11: Blood transcriptome (sncRNA) data from the Parkinson's Disease Progression Biomarker Initiative (PPMI) project; S12: Parkinson's disease data and blood transcriptome data from the Biomarker Project (PDBP).

[0027] The basic steps of the multi-scale embedded gene co-expression network analysis algorithm in step S1 include: S13: Calculate the correlation coefficient matrix between genes; S14: Fast Plane Filtering Network Construction with Parallelization, Early Termination and Prior Quality Control - FPFNC, this step is part of the MEGENA algorithm; S15: Use multi-scale cluster analysis to divide the gene modules into k modules (M1, M2, ..., M...). K Its formula is: ; in This is represented by the gene module compactness score, used to divide gene modules, where k represents the number of gene modules. This represents the number of edges within the i-th gene module. This represents the total number of edges. This represents the resolution parameter.

[0028] The methods for calculating the correlation coefficient of gene pairs in step S13 include, but are not limited to, those for calculating the correlation coefficient of gene pairs. S131: The first method is to use the Pearson correlation coefficient method to calculate the correlation between genes. The specific formula is as follows: ; Assume there are a total of m patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and Let represent the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. and This represents the average expression level of the i-th and j-th genes; S132: The second method is to use Spearman's rank correlation coefficient to calculate the correlation between genes. The specific formula is as follows: ; Assume there are a total of m patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and These represent the rankings of the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. S133: The third method is to use Kendall's tau correlation coefficient to calculate the correlation between genes. The specific formula is as follows: ; Assume there are a total of m patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and Let represent the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. and Let represent the expression values ​​of the i-th and j-th genes in the l-th patient, respectively, and C represent the number of identical pairs, i.e. The number of inconsistent pairs, D, is the number of inconsistent pairs. Quantity, This represents the number of times the expression value of the i-th gene is the same. This represents the number of times the expression value of the j-th gene is the same.

[0029] Furthermore, the basic steps for constructing the higher-order gene communication tensor in step S1 include: S16: Construct a two-dimensional tensor for the k gene modules (M1, M2, ..., MK) partitioned in S15. Specifically, for gene a from gene module A, if gene b is connected to it in gene module B, it indicates that there is information exchange between module A and module B, and the communication quantity increases by 1. For gene a, the final two-dimensional communication tensor is represented as follows: The formula is as follows: ; Where i and j represent the indices of the gene modules, respectively; S17: After calculating the two-dimensional communication tensor, for the i-th patient, use all genes to generate the three-dimensional communication tensor. The structure is as follows: G represents the total number of genes; the specific construction formula is as follows: ; Where g represents the g-th gene, and G represents the total number of genes. This represents a two-dimensional communication tensor constructed using the g-th gene; S18: Finally, the three-dimensional communication tensors of all samples are combined to form the final four-dimensional communication tensor (4D-GMC). The specific construction formula is as follows: ; Where n represents the nth patient, and N represents the total number of patients. This represents the three-dimensional communication tensor constructed using the nth patient.

[0030] Furthermore, the basic steps in step S2 for extracting spatiotemporal features using a multi-scale adaptive Transformer model include: S21: To process whole blood transcriptome data X, this invention designs a linear layer to perform linear projection on X, thereby generating feature vectors of the whole blood transcriptome data. The formula is as follows: ; Here, Liner represents a linear layer in a neural network model; S22: In order to process whole blood transcriptome data This invention uses a combination of convolutional neural networks and the KAN model to extract feature tensors, and the final result is... ; S23: Multi-scale adaptive Transformer was used to extract features at the whole blood transcriptome scale and the higher-order communication tensor scale, respectively. The two scale features obtained were as follows: and Its formula is: ; Wherein, the query vector Q, the key vector K, and the value vector V are respectively and These are obtained through linear layers, where T represents the transpose operation and softmax represents the activation function. Indicates a parameter.

[0031] The basic steps in step S22 using convolutional and KAN models include: S221: Four-dimensional gene module communication tensor The data is first fed into a convolutional layer with 2000 input channels and 1024 output channels. Then, an activation function is applied to the output. Its formula is: ; Where ReLU represents the activation function and Conv represents the convolutional layer; S222: Subsequently, an average pooling layer is used to... The dimension is reduced to Then, the obtained tensor is flattened into two dimensions, as shown in the following formula: ; Here, `view` represents the flattening operation, and `AvgPool` represents the average pooling layer operation. This represents the feature representation after extraction by the average pooling layer. S223: Finally, a KAN layer is used to extract non-linear relationships in high-dimensional data, and its formula is as follows: ; Wherein, KAN represents the KAN layer. express The result after feature extraction using the KAN model.

[0032] Furthermore, the basic steps of multi-view collaborative fusion in step S3 include: S31: The core of multi-view collaborative fusion is the cross-attention mechanism, and its formula is: ; ; ; in, They represent from Exported queries ( ),key( ) and value ( );akin, Other means from Exported queries ( ),key( ) and value ( ); and These are the feature matrices generated after the first and second cross-attention processes, respectively; the final result obtained after two cross-attention processes is... ; S32: Subsequently, and Feature fusion is performed to obtain an enhancement matrix containing more information. The formula is as follows: ; Here, Concat represents the feature concatenation operation.

[0033] The basic steps for disease stage identification using fully connected layers in step S4 include: S41: Using a fully connected layer with a softmax activation function The formula for classification prediction is as follows: ; The softmax function converts the output of the fully connected layer into class probabilities, and FC represents a fully connected layer. S42: The cross-entropy loss function is used to extract characteristics and commonalities across perspectives. Its mathematical formula is: ; Where L represents the loss function, which quantifies the difference between the predicted probability distribution output by the model and the true label distribution, reflecting the model's prediction accuracy. It is a hyperparameter. for and The cross-entropy loss function between them for The cross-entropy loss function, for The cross-entropy loss function.

[0034] In step S4, four metrics—accuracy (ACC), F1-score, recall, and precision—are used to evaluate the model's accuracy. Accuracy refers to the proportion of correctly predicted samples out of the total number of samples, providing a direct assessment of the model's overall performance. Recall, also known as the full detection rate, represents the ratio of correctly predicted positive instances to the actual total number of positive instances, evaluating the model's ability to discover all relevant instances. The F1-score is the harmonic mean of precision and recall, comprehensively considering both the model's accuracy and completeness. Precision represents the proportion of correctly predicted positive instances out of the total number of correctly predicted positive instances, focusing on the accuracy of the model's positive predictions. Its formula is as follows: ; ; ; ; TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative. The higher these indicators are, the more accurate the prediction results.

[0035] Example 2 The hardware environment for this embodiment was: a laptop computer with a Xeon(R) Gold 6430 CPU and 24.0GB of RAM; software environment: Python 3.6, R3.3.3; operating platform: Win10. The communication patterns of the Motor PPMI case at the initial treatment moment are visualized as follows: Figure 4 Communication patterns of the Motor PDBP case at a six-month follow-up time are visualized as follows: Figure 5 As shown, the gene co-expression network visualization of the Motor PPMI case is as follows: Figure 6 As shown, the gene co-expression network visualization of the Motor PDBP case is as follows: Figure 7 As shown.

[0036] This embodiment tested the effectiveness of the proposed method on a whole blood transcriptome dataset, including PPMI and PDBP datasets. Analysis was conducted from two aspects: motor function assessment (Motor PPMI and Motor PDBP) and non-motor function assessment (Non-motor PPMI and Non-motor PDBP). For motor function impairment assessment, the Hoehn and Yahr scale labels were used. This rating scale, based on the severity of motor symptoms and functional impairment, is a commonly used disease staging system for assessing disease severity and progression, as detailed below: Stage 0: Asymptomatic; Phase 1: Only one side of the body is affected, but balance is not affected; Phase 2: Only both sides of the body are affected, but balance is not affected; Stage 3: Balance is affected, mild to moderate disease; but the patient can live independently; Stage 4: Severe immobility. However, the patient can walk and stand independently; Stage 5: Without assistance from others, the patient can only lie in bed or use a wheelchair.

[0037] For non-motor assessment, the cognitive impairment label from UPDRS-I was used, which mainly includes cognitive function (such as memory, attention, executive function, etc.), emotional state (depression, anxiety), and autonomic symptoms (constipation, urinary disorders), etc.; the label is divided into five categories, as follows: Stage 0: No cognitive impairment; Stage 1: Mild amnesia, no impact on daily life; Stage 2: Moderate memory or executive function impairment, mildly affecting daily life (such as forgetting appointments, difficulty making decisions); Stage 3: Severe cognitive impairment, requiring assistance to complete complex tasks (such as financial management and medication). Stage 4: Dementia (unable to live independently, requires full-time care).

[0038] In all four cases, the top 2000 hypervariable genes were used as features, and the top 10 most important gene modules were selected to construct 4D-GMCs. For data partitioning, the training set used gene expression data from the patient's baseline (BL) period, i.e., the initial stage of treatment, while the test set used time-series data from the sixth month of follow-up (Month 6, M6). The constructed 4D-GMC information is shown in Table 1.

[0039] Table 1. Information on higher-order gene communication tensors .

[0040] This embodiment proposes a Parkinson's disease evolution prediction method based on high-order gene communication tensors, the flowchart of which is shown below. Figure 1 As shown, the implementation process of this method is illustrated using the Motor PPMI dataset of whole blood transcriptome data as an example, specifically including the following steps: S1: First, the original PPMI dataset is preprocessed to separate the whole blood transcriptome data from multiple disease stages of the patients, obtaining whole blood transcriptome data X at two disease time points for the same batch of patients (4425 patients). Then, the scanpy package is used to select the top 2000 highly variable genes as features. Finally, the multi-scale embedded gene co-expression network analysis algorithm (MEGENA) is used to perform co-expression network analysis on the whole blood transcriptome data at the two time points to obtain the relevant adjacency matrix representation. Size is 2000 2000, its calculation formula is: ; Where m represents the number of patients, with a size of 4425; the correlation between gene i and gene j is... n represents the number of genes, with a size of 2000; and These represent the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. and This represents the average expression level of the i-th and j-th genes; in addition, the gene co-expression network contains information on several k gene modules, and its calculation formula is: ; in This is represented as the gene module compactness score, used to divide gene modules; k represents the number of gene modules. This represents the number of edges within the i-th gene module; This represents the total number of edges; This represents the resolution parameter.

[0041] S2: Based on the intercellular communication pattern analysis method, a high-order gene communication tensor was constructed. , Figure 1 As shown in the process flow, the specific technical solution is as follows: S21: Based on the gene importance scores obtained by the MEGENA algorithm, select the top ten gene modules to obtain a two-dimensional interaction matrix between gene modules, with a size of 10*10. If gene a in module A is connected to gene b in module B by an edge, it indicates that module A and module B have information interaction, and the value of the corresponding position in its interaction matrix is ​​increased by one. The calculation formula is as follows: ; In the above formula, i and j represent the indices of the gene module. This represents the interaction matrix between the generated gene modules; S22: Generate an interaction matrix between gene modules for each gene, thus forming a three-dimensional matrix of size 2000. 10 The calculation formula is as follows: ; In the above formula, g represents the gene index. This represents the interaction matrix of the g-th gene. This represents the generated three-dimensional matrix; S23: The three-dimensional matrix generated in step S22 is for a single patient. Therefore, following the previous steps, a three-dimensional matrix is ​​generated for each patient, resulting in a four-dimensional matrix, namely the higher-order gene communication tensor, with a size of 4425. 2000 10 The calculation formula is as follows: ; In the above formula, n represents the patient's index. A three-dimensional matrix representing the nth patient. This represents the generated higher-order gene communication tensor.

[0042] S3: Whole blood transcriptome data and higher-order gene communication tensors As two types of scale information, this invention employs a multi-scale adaptive Transformer model to extract feature information from both, such as... Figure 2 As shown; its specific technical solution is as follows: S31: For whole blood transcriptome data X, linear projection is performed using a linear layer to obtain its feature vector. The calculation formula is: ; Here, Liner represents a linear layer in a neural network model; S32: For whole blood transcriptome data First, we use a convolutional neural network for dimensionality reduction. The calculation formula is as follows: ; Where ReLU represents the activation function, Conv represents the convolutional layer with 2000 input channels and 1024 output channels, and then an average pooling layer is used to... Flattened into two dimensions, the calculation formula is: ; Here, `view` represents the flattening operation, and `AvgPool` represents the average pooling layer operation. This represents the feature representation after extraction by the average pooling layer. Finally, a KAN layer is used to extract... The nonlinear relationship present in the equation is calculated using the following formula: ; Wherein, KAN represents the KAN layer. express The result after feature extraction using the KAN model; S33: Multi-scale adaptive Transformer was used to extract features at the whole blood transcriptome scale and the higher-order communication tensor scale, respectively. The two scale features obtained were as follows: and The calculation formula is: ; Wherein, the query vector Q, the key vector K, and the value vector V are respectively and These are obtained through linear layers, where T represents the transpose operation and softmax represents the activation function. Indicates a parameter.

[0043] S4: Employs a multi-perspective collaborative fusion scheme to integrate multi-perspective interactive information between high-order gene communication tensors and whole blood transcriptome data. The core principle is the cross-attention mechanism. Its specific technical solution is as follows: S41: First, multiple cross-attention layers are used to extract feature information within their respective viewpoints. The calculation formula is as follows: ; ; ; in, They represent from Exported queries ( ),key( ) and value ( ).akin, Other means from Exported queries ( ),key( ) and value ( ). and These are the feature matrices generated after the first and second cross-attention processes, respectively, where T represents the transpose of the matrix vector. The final result after two cross-attention processes is... ; S42: Subsequently, and Feature concatenation is performed to obtain an enhancement matrix containing more information. The calculation formula is: ; Wherein, Concat represents the feature concatenation operation; S5: Use a fully connected layer with the softmax activation function to... The formula for classification prediction is as follows: ; in The softmax function, representing the final prediction result of the model, transforms the output of the fully connected layer into class probabilities. FC stands for Fully Connected Layer. The model uses the cross-entropy loss function to extract characteristics and commonalities across viewpoints, and its calculation formula is as follows: ; Where L represents the loss function of the neural network model, which reflects the model's prediction accuracy by quantifying the difference between the predicted probability distribution output by the model and the true label distribution; the smaller the difference, the lower the loss value. It is a hyperparameter. for and The cross-entropy loss function between them For gene embedding The cross-entropy loss function, For higher-order communication tensors The cross-entropy loss function.

[0044] Finally, the actual motor / non-motor function assessment labels y and the model-predicted labels are compared. The model's accuracy was evaluated using four metrics: accuracy (ACC), F1-score, recall, and precision. Accuracy refers to the proportion of correctly predicted samples out of the total number of samples, providing a direct assessment of the model's overall performance. Recall, also known as the full detection rate, represents the ratio of correctly predicted positive instances to the actual number of positive instances, evaluating the model's ability to find all relevant instances. The F1-score is the harmonic mean of precision and recall, comprehensively considering both accuracy and completeness. Precision represents the proportion of correctly predicted positive samples out of the total number of predicted positive samples, focusing on the accuracy of the model's positive predictions. Its calculation formula is as follows: ; ; ; ; TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative. The higher these indicators are, the more accurate the prediction results.

[0045] Example 3 The main contribution of this invention is the construction of a high-order gene communication tensor and the use of an adaptive Transformer model to extract feature information at two scales. Finally, a collaborative fusion scheme is constructed to fuse intra-view characteristics and inter-view commonalities, extracting information from past time periods to predict future disease states. This invention utilizes the M2GMC model (see...). Figure 1 The M2GMC model is an M2GMC method based on a spatiotemporal joint perspective. This method integrates a hybrid perspective of temporal representation patterns and high-order communication tensors (the communication tensor corresponds to the spatial perspective). To verify the effectiveness of this model, six baseline methods were selected as comparison models: (1) CNN: Convolutional Neural Networks, as a typical deep learning architecture, are mainly used in the field of visual data processing. This network automatically extracts the spatial features of images through multi-layer convolution operations, compresses the feature dimensions by combining downsampling operations, and finally uses fully connected layers to realize the mapping of high-level semantic information and task prediction.

[0046] (2) Transformer: Its core architecture is based on multi-head self-attention mechanism and feedforward neural network. Through the stacking of multi-layer encoders and decoders, it realizes context modeling and sequence generation of input data.

[0047] (3) ACmix: This model reveals a strong underlying relationship between Self-Attention and convolution, providing a new perspective for understanding the connection between the two modules and providing inspiration for designing new learning paradigms.

[0048] (4) MAET: This model is a novel multimodal adaptive emotion Transformer that can adaptively handle single-modal and multimodal inputs.

[0049] (5) Meta-Transformer: This model consists of three main components: a unified data tokenizer, a modality-shared encoder, and a task-specific header for downstream tasks. Meta-Transformer is the first framework capable of performing unified learning across 12 modalities and using unpaired data.

[0050] (6) Perceiver: This model uses a single Transformer-based architecture to manipulate inconsistent permutations of different modalities.

[0051] The experimental results of this invention in the cases of Motor PPMI, Motor PDBP, Non-motor PPMI, and Non-motor PDBP are shown in Tables 2-5: Table 2. Motion assessment examples from the PPMI dataset. .

[0052] Table 3 Motion evaluation cases from the PDBP dataset .

[0053] Table 4 Non-motion assessment cases from the PPMI dataset .

[0054] Table 5 Non-motion evaluation cases from the PDBP dataset .

[0055] The M2GMC model proposed in this implementation achieved good results in all four case studies. First, single-view models (including CNN, Transformer, and ACmix) were used to independently test the higher-order communication tensor and whole-blood transcriptome data, respectively. Then, a dual-view model (including MAET, Meta-Transformer, and Perceiver) was used for joint analysis and comparison of the higher-order communication tensor and whole-blood transcriptome data. Specific experimental results are shown in Tables 2-5 and... Figure 3 As shown in the figure, the experimental results indicate that in the single-view model, the prediction performance of the model using only higher-order communication tensors is slightly lower than that of the model using only whole blood transcriptome data, which demonstrates the effectiveness of higher-order communication tensors. The experimental results show that in the single-view model, the prediction performance of the model using only higher-order communication tensors is slightly lower than that of the model using only whole blood transcriptome data, which preliminarily verifies the effectiveness of higher-order communication tensors. However, in dual-view learning, the model performance is significantly better than the single-view model using only higher-order communication tensors or gene expression patterns, indicating a complementarity between the two data types. Multi-view joint analysis can effectively enhance feature representation capabilities, thereby improving the model's prediction performance. For example, on the Motor PDBP dataset, the accuracy of M2GMC reaches 0.732, which is 7.9% higher than the MAET method and significantly better than other methods.

[0056] While deep learning-based candidate disease modeling methods (such as ACmix and MAET) have shown some competitiveness, they still fall short in effectively capturing complex patterns in datasets. M2GMC's superior performance is attributed to its advanced design, which integrates relationships between viewpoints and view angles into a unified framework, thereby enhancing its ability to more comprehensively model the underlying data structure. Evaluation metrics obtained by M2GMC compared to candidate methods demonstrate its effectiveness in various applications, highlighting the positive role of inter-module communication patterns in disease modeling.

[0057] To further explore the roles of higher-order communication tensors, multi-scale adaptive modules, and collaborative fusion modules, the M2GMC model underwent ablation analysis on four whole blood transcriptome datasets. The experimental design employed a rigorous control method, performing ablation on each module individually and in combination to evaluate the contribution of each module to the model. Specific experimental results are shown in Tables 6 and 7.

[0058] Table 6 Ablation experiments on the motor function assessment dataset .

[0059] Table 7 Ablation experiments on the non-motor function assessment dataset .

[0060] Experimental data show that after integrating the communication tensor, the model's performance indicators in three different application scenarios are significantly improved. This result fully demonstrates the effectiveness of high-order communication tensors in improving model performance.

[0061] Furthermore, the experiments analyzed in detail the impact of the multi-scale adaptive module and the multi-view collaborative module on model performance. The comparative results show that both modules positively enhance the model.

[0062] To further verify the irreplaceable nature of the four-dimensional communication tensor, a comparative experiment was designed: a three-dimensional communication tensor (patient × sending module × receiving module) was used to replace the original four-dimensional communication tensor (patient × gene × sending module × receiving module). The experimental results are shown in Table 8. Table 8 Performance Comparison of Four-Dimensional Communication Tensor and Three-Dimensional Communication Tensor .

[0063] Table 8 shows that in all cases, the four-dimensional communication tensor outperforms the three-dimensional communication tensor in all four evaluation metrics, which confirms the independent contribution of gene-specific features to model performance.

[0064] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A Parkinson's evolution prediction modeling method based on higher-order gene communication tensors, characterized in that, Includes the following steps: S1: The whole blood transcriptome data at each time point is X, containing m patients and n genes; The MEGENA multi-scale embedded gene co-expression network analysis algorithm was used to perform co-expression network analysis on the whole blood transcriptome data X at each time step to obtain a Parkinson's-specific gene map G containing the correlation between genes; then, a high-order gene communication tensor, namely the four-dimensional communication tensor 4D-GMC, was constructed using a method based on intercellular communication patterns. S2: A multi-scale adaptive Transformer model was used to extract whole blood transcriptome data. Temporal characteristics and higher-order gene communication tensors Spatial characteristics; S3: Employs a multi-perspective collaborative fusion scheme to integrate multi-perspective interactive information between high-order gene communication tensors and whole blood transcriptome data, thereby achieving effective feature fusion and utilization; S4: Use fully connected layers as classifiers for Parkinson's disease evolution modeling, predicting motor and non-motor dysfunction scores for the next time step based on data from the previous time step.

2. The Parkinson's evolution prediction modeling method based on higher-order gene communication tensors according to claim 1, characterized in that, The types of whole blood transcriptome data in step S1 include, but are not limited to: Blood transcriptome sncRNA data and Parkinson's disease data from the Parkinson's Disease Progression Biomarkers Initiative (PPMI) and blood transcriptome data from the Parkinson's Disease Biomarkers Initiative (PDBP).

3. The Parkinson's evolution prediction modeling method based on higher-order gene communication tensors according to claim 1, characterized in that, The steps of the multi-scale embedded gene co-expression network analysis algorithm in step S1 include: S11: Calculate the correlation coefficient matrix between genes; S12: Construction of Fast Plane Filtering Networks with Parallelization, Early Termination and Prior Quality Control - FPFNC; S13: Use multi-scale cluster analysis to divide the gene modules into k modules (M1, M2, ..., M...). K Its formula is: ； in denoted as gene module compactness score, used to divide gene modules, where k represents the number of gene modules; This represents the number of edges within the i-th gene module. This represents the total number of edges. This represents the resolution parameter.

4. The Parkinson's evolution prediction modeling method based on higher-order gene communication tensors according to claim 3, characterized in that, The method for calculating the correlation coefficient matrix between genes in step S11 can be any of the following: The first method is to use the Pearson correlation coefficient method to calculate the correlation between genes. The formula is as follows: ； m represents the number of patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and Let represent the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. and This represents the average expression level of the i-th and j-th genes; The second method is to use Spearman's rank correlation coefficient to calculate the correlation between genes. The specific formula is as follows: ； m represents the number of patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and These represent the rankings of the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. The third method is to use Kendall's Tau correlation coefficient to calculate the correlation between genes. The specific formula is as follows: ； m represents the number of patients, and the correlation between gene i and gene j is... Where A has an n*n dimension, and n is the number of genes. and Let represent the expression values ​​of the i-th and j-th genes in the k-th patient, respectively. and Let represent the expression values ​​of the i-th and j-th genes in the l-th patient, respectively, and C represent the number of identical pairs, i.e. The number of inconsistent pairs, D, is the number of inconsistent pairs. Quantity, This represents the number of times the expression value of the i-th gene is the same. This represents the number of times the expression value of the j-th gene is the same.

5. The Parkinson's evolution prediction modeling method based on higher-order gene communication tensors according to claim 1, characterized in that, The basic steps for constructing the higher-order gene communication tensor in step S1 include: S14: Construct a two-dimensional tensor for the k gene modules M1, M2, ..., MK partitioned in S13; for gene a from gene module A, if gene b is connected to it in gene module B, it indicates that there is information exchange between module A and module B, and the communication quantity increases by 1; for gene a, the final two-dimensional communication tensor is represented as follows: The formula is as follows: ； Where i and j represent the indices of the gene modules, respectively; S15: After calculating the two-dimensional communication tensor, for the i-th patient, use all genes to generate the three-dimensional communication tensor. The structure is as follows: G represents the total number of genes; the construction formula is as follows: ； Where g represents the g-th gene, and G represents the total number of genes. This represents a two-dimensional gene communication tensor constructed using the g-th gene; S16: Finally, the three-dimensional communication tensors of all samples are combined to form the final four-dimensional communication tensor 4D-GMC. The formula is constructed as follows: ； Where n represents the nth patient, and N represents the total number of patients. This represents the three-dimensional communication tensor constructed using the nth patient.

6. The Parkinson's evolution prediction modeling method based on higher-order gene communication tensors according to claim 1, characterized in that, The basic steps in step S2 using the multi-scale adaptive Transformer model include: S21: To process whole blood transcriptome data X, a linear layer is used to perform a linear projection on X, thereby generating a feature vector of the whole blood transcriptome data. The formula is as follows: ； Here, Liner represents a linear layer in a neural network model; S22: In order to process whole blood transcriptome data The feature tensor is extracted using a combination of convolutional neural networks and the KAN model, and the final result is... ; S23: Multi-scale adaptive Transformer was used to extract features at the whole blood transcriptome scale and the higher-order communication tensor scale, respectively. The two scale features obtained were as follows: and Its formula is: ； Wherein, the query vector Q, the key vector K, and the value vector V are respectively and These are obtained through linear layers; T represents the transpose operation, and softmax represents the activation function. Indicates a parameter.

7. The Parkinson's evolution prediction modeling method based on higher-order gene communication tensors according to claim 6, characterized in that, The basic steps for extracting feature tensors using a combination of convolutional neural networks and KAN models in step S22 include: S221: Four-dimensional gene module communication tensor The data is first fed into a convolutional layer with 2000 input channels and 1024 output channels; then, an activation function is applied to the output. Its formula is: ； Where ReLU represents the activation function and Conv represents the convolutional layer; S222: Subsequently, an average pooling layer is used to... The dimension is reduced to Then, the obtained tensor is flattened into two dimensions, as shown in the following formula: ； Here, `view` represents the flattening operation, and `AvgPool` represents the average pooling layer operation. This represents the feature representation after extraction by the average pooling layer. S223: The KAN layer is used to extract non-linear relationships in high-dimensional data. The formula is as follows: ； Wherein, KAN represents the KAN layer. express The result after feature extraction using the KAN model.

8. The Parkinson's evolution prediction modeling method based on higher-order gene communication tensors according to claim 1, characterized in that, Step S3, which involves feature fusion using a multi-view collaborative approach, includes: S31: Multiple cross-attention layers are used to extract features from the two perspectives separately. The formula is as follows: ； ； ； in, They represent from Exported query ,key Sum ;akin, Other means from Exported query ,key Sum ; and These are the feature matrices generated after the first and second cross-attention processes, respectively, where T represents the transpose of the matrix vector; the final result obtained after two cross-attention processes is... ; S32: Subsequently, and Feature fusion is performed to obtain an enhancement matrix containing more information. The formula is as follows: ； Here, Concat represents the feature concatenation operation.

9. The Parkinson's evolution prediction modeling method based on higher-order gene communication tensors according to claim 1, characterized in that, The basic steps in step S4 for disease stage identification using fully connected layers include: S41: Using a fully connected layer with a softmax activation function The formula for classification prediction is as follows: ； The softmax function transforms the output of the fully connected layer into class probabilities; FC stands for Fully Connected Layer. This indicates the model's final prediction result; S42: The cross-entropy loss function is used to extract characteristics and commonalities across perspectives. Its mathematical formula is: ； Where L represents the loss function, which quantifies the difference between the predicted probability distribution output by the model and the true label distribution, reflecting the model's prediction accuracy. It is a hyperparameter. for and The cross-entropy loss function between them for The cross-entropy loss function, for The cross-entropy loss function; Finally, the actual motor / non-motor function assessment labels y and the model-predicted labels are compared. The accuracy of the model was evaluated using four metrics: accuracy (ACC), F1-score, recall, and precision. The F1-score is the harmonic mean of precision and recall. Precision represents the proportion of samples correctly predicted as positive out of the total number of correctly predicted positive samples, and the formula is as follows: ； ； ； ； TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative.

Images