Alzheimer's disease genetic correlation network detection method

Abstract

The invention discloses an Alzheimer's disease genetic correlation network detection method which comprises four evaluation strategies. The method comprises a hippocampus subregion assisted network importance evaluation strategy, an overlapped voxel analysis, function enrichment and pathway analysis combined strategy, a genetic interaction network driven gene ontology function annotation and pathway analysis strategy, and a network and AD correlation and hippocampus subregion comparative analysis strategy. According to the method, the effectiveness of multi-modal fusion analysis is improved, AD biomarkers are mined, and an excellent prediction result is obtained.

Description

Technical Field

[0001] The present invention relates to the field of neurogenetics, and in particular to a method for detecting a genetically related network of Alzheimer's disease. Background Art

[0002] Imaging genomics studies have collected a large amount of human brain functional imaging data and whole-genome genetic data. Genome-wide association studies (GWAS) conducted by Christopher et al. rely on these datasets to analyze the association between single nucleotide polymorphisms (SNPs) and pathological phenotypes. Voxel-based GWAS (vGWAS) subdivides phenotypes into voxels to analyze associations. Functional genome-wide association studies (FGWAS) developed by Huang et al. can effectively identify sparse signals in space. Among these methods, vGWAS makes it possible to capture subtle signals and brain structures that are easily overlooked by traditional phenotyping methods.

[0003] The role of the hippocampus in memory, learning, and spatial navigation has been demonstrated. GWAS have identified several SNPs associated with the hippocampus. The hippocampus is a combination of subregions with different functions. For example, the CA1 region (CA1) and the submandibular belt are associated with Alzheimer's disease (AD). The dentate gyrus (DG) and CA3 are also associated with AD, but not as closely as the CA1 and submandibular belt.

[0004] Based on the differences in structure and function of hippocampal subregions and the support of gene expression studies, different subregions have different genetic determinants. Therefore, it makes sense to use the hippocampus as a research object.

[0005] In previous studies, linkage disequilibrium score regression (LDSC) was used as an effective method to calculate the heritability and genetic correlation of multiple phenotypes. For example, Pairo-Castineira et al. found a genetic association that was highly correlated with inflammatory organ damage in COVID-19. This provides a biological theoretical basis for existing drug treatments. Tang et al. confirmed the assumed causal relationship between body mass index (BMI) and rheumatoid arthritis (RA) risk and proposed a feasible RA prevention strategy. Stansosdottir et al. studied the correlation between hypertension-polygenic risk score and preeclampsia and confirmed the effect of hypertension on preeclampsia. In addition to these studies on the disease, M.Dennis et al. also used hippocampal subregions as phenotypes and calculated the genetic correlation of subregion pairs. Therefore, using hippocampal voxels as phenotypes to study the interpretable links between voxels can better understand the impact of voxel sets on the hippocampus and disease. Summary of the invention

[0006] In view of the problems existing in the prior art, the present invention provides a method for detecting an Alzheimer's disease genetic correlation network, which improves the effectiveness of multimodal fusion analysis, mines AD biomarkers and obtains excellent prediction results.

[0007] The purpose of the present invention is achieved through the following technical solutions.

[0008] A method for detecting genetic-related networks of Alzheimer's disease includes four evaluation strategies: a network importance evaluation strategy assisted by hippocampal subregions, a strategy combining overlapping voxel analysis with functional enrichment and pathway analysis, a gene ontology functional annotation and pathway analysis strategy driven by genetic interaction networks, and a network-AD correlation and hippocampal subregion comparative analysis strategy.

[0009] The hippocampal sub-region-assisted network importance assessment strategy uses the FreeSurfer v6.0 neuroimaging analysis tool to accurately define and segment the hippocampal sub-regions, and then uses a specific strategy 1 to quantify the overlap ratio between the sub-regions and the network.

[0010] The strategy of combining overlapping voxel analysis with functional enrichment and pathway analysis is to conduct a comprehensive and in-depth evaluation of overlapping voxels between genetic related networks and sub-regions, carry out enriched functional analysis and pathway analysis, and use this as the core means to comprehensively measure the biological significance of the network.

[0011] Specific operation process: The discovered genetic correlation network is set as the phenotypic variable in the whole genome association analysis, and the diagnostic information, gender, age, education level and the top 4 results in the population stratification analysis are included in the analysis model as covariates, so as to accurately obtain the intrinsic correlation between single nucleotide polymorphisms and the network; then, SNPs with p-value < 0.05 are screened out according to strict statistical standards, and with the help of effective chi-square test and the "hg19" genome reference version, the SNPs are accurately located to the corresponding genes according to their chromosomal location information; then, the located genes are subjected to Bonferroni correction, and the genes with corrected p-value < 0.05 are applied to the subsequent strategy 2 and strategy 3 analysis processes to ensure the accuracy and reliability of the research results.

[0012] The genetic interaction network driven gene ontology function annotation and pathway analysis strategy: the genetic interaction network is used to annotate the gene ontology functions of the constructed network in detail, and based on this, in-depth pathway analysis is carried out.

[0013] Specific implementation process: The discovered genetic correlation network is used as the core variable, the disease diagnosis result is used as the phenotypic variable, and gender, age, education level and the top 4 results in the group stratification analysis are used as covariates. Strategy 4 is used to accurately calculate the correlation between the network and Alzheimer's disease, so as to deeply explore the potential mechanism of action of the network in the occurrence and development of the disease.

[0014] The strategy for comparative analysis of the network's correlation with AD and with hippocampal subregions is to simultaneously calculate the correlation between the network and AD and between the hippocampal subregions and AD, and conduct a meticulous comparative analysis of these two sets of correlation data to clarify the network's unique position and mechanism of action in AD-related neurobiological processes.

[0015] Validation data set result reproducibility test strategy: Use an independent validation data set to completely repeat all the above evaluation steps to strictly test the repeatability and stability of the results of this invention and ensure that the research results have a high degree of reliability and generalization ability.

[0016] vGWAS was performed on hippocampal voxels using PLINK v1.9. The image data and genetic data obtained by joint analysis were used, and the voxels were used as phenotypes. Diagnosis, sex, age, education, and the first 4 results from the population stratification analysis were used as covariates. The R of each SNP at each voxel was calculated and obtained in the vGWAS results. 2 ;

[0017] Based on the results of the vGWAS data, LDSC was used to estimate the LD score for each chromosome with a window of 1 cm. Then, to avoid inferring that the heritability of each SNP in common variants is the same as that of rare variants, non-partitioned LD score regression was performed for each hippocampal voxel using the LD file and vGWAS results. Using the LD file, the vGWAS results were converted to LDSC format, which included the number of SNPs, effect allele 1, allele 2, SNP size, p-value, and summary statistics. Subsequently, the calculation file was applied to calculate the heritability of each voxel.

[0018] The heritability explained by S is defined as equation (1),

[0019]

[0020] Where S represents a set of SNPs, y1, y2 represent voxels, j is a SNP, β, γ are the unique modulus SNPs of y1, y2,

[0021] Since voxels with low average chi2 have low heritability and the genetic correlation between these voxels is infinitely close to zero, in order to reduce the time of calculating genetic correlation, the following steps select voxels with average chi2>1.02 (https: / / github.com / bulik / ldsc / wiki / FAQ). In order to estimate the genetic correlation between y1 and y2, the total error of measuring the relationship between y1 and y2 is defined as the genetic covariance. The genetic covariance between y1 and y2 in S is defined as equation (2),

[0022] ρ s (y1,y2):=∑ j∈S β j γ j (2)

[0023] LDSC was applied to estimate the genetic correlation between genotypes of all voxel pairs. Using the obtained heritability and genetic covariance, the genetic correlation between y1 and y2 in S was defined as Equation (3),

[0024]

[0025] Since each genetic correlation calculated by LDSC has a corresponding p-value, the resulting correlations were Bonferroni-corrected, and nodes with significant correlations (p-value < 0.05) were selected to construct a voxel-based genetic correlation network, in which nodes represent voxels and connections between pairs of nodes describe the genetic correlation of voxel pairs.

[0026] Compared with the existing technology, the advantages of the present invention are: in the frontier research field of neurogenetics, the present invention pioneered the use of voxel-based genetic correlation network construction strategy, deeply explained the genetic variation mechanism of the hippocampus and opened up a new evaluation path. This method is significantly innovative in the field of hippocampal research, and is the first to focus on the exploration of genetic correlation at the hippocampal voxel level, filling the gap in related research methods in this field.

[0027] The study of the correlation between imaging and genetic information is an important area of ​​AD research. A limitation that researchers often face is how to effectively fuse imaging information and genetic information and extract important information. In previous studies, some researchers used classical methods, such as Pearson correlation analysis, to solve the problem of multimodal fusion analysis. These methods achieve information fusion, but most current studies focus on the results of multimodal fusion, such as fusion feature construction, feature selection, or sample classification. How to propose a new research framework, improve the effectiveness of multimodal fusion analysis, mine AD biomarkers, and obtain excellent prediction results are challenges faced by multimodal fusion research between neuroimaging data and genetic data, which is of great significance for the early diagnosis and treatment of AD. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is the overall process framework diagram.

[0029] Figure 2 The X-axis represents voxels, and the Y-axis also represents voxels. Figure 2 (a) shows the genetic correlation of 5,102 pairs of voxels in the experimental dataset; Figure 2 (b) indicates that 5,102 pairs of voxels in the experimental dataset have genetic correlation (corresponding p value < 0.05); Figure 2 (c) Indicates that 10,153 pairs of voxels in the experimental dataset have genetic correlation (corresponding p value < 0.05 / 10,153). Figure 2 (d) shows the genetic correlation of 10,153 pairs of voxels in the validation dataset; Figure 2 (e) indicates that 10,153 pairs of voxels in the validation dataset have genetic correlation (corresponding p value < 0.05); Figure 2 (f) Indicates that 10,153 pairs of voxels in the validation dataset have genetic correlation (corresponding p-value < 0.05 / 10,153).

[0030] Figure 3 This is the genetic correlation network of voxels in the right hippocampus. The numbers represent voxels, and the connecting lines between nodes represent the genetic correlation between voxels. The color indicates the subregion where the node is located in the network. Figure 3 (a) is the genetic correlation network in the experimental dataset. Figure 3 (b) is the genetic correlation network in the validation dataset.

[0031] Figure 4 To use the corrected GO Terms with p-value < 1e-6 in the entire network and subnetworks. The color of the inner circle represents the current number of genes / background number of genes, and the size represents the P in -log10GO Terms. The points in the outer circle represent genes, and their positions depend on the -log10P genes. Figure 4 (a) GO Terms of the entire network in the experimental dataset; Figure 4 (b) GO Terms of network A in the experimental dataset; Figure 4 (c) GO Terms of network B in the experimental dataset. Figure 4 (d) GO Terms for the entire network in the validation dataset; Figure 4 (e) GO Terms of network A in the validation dataset; Figure 4 (f) GO Terms of network B in the validation dataset.

[0032] Figure 5 Corrected pathway analysis results with p-value < 0.05 in the entire network and subnetworks. The x-axis represents the current number of genes / background number of genes, and the y-axis represents the pathway. The color represents the pathway analysis result with -log10p-value. Figure 5 (a) Pathway analysis results with p-value < 0.05 after correction for the entire network in the experimental data set; Figure 5 (b) Pathway analysis results of network A in the experimental data set with a corrected p-value < 0.05; Figure 5 (c) Pathway analysis results of network B in the experimental data set with p-value < 0.05 after correction. Figure 5 (d) Pathway analysis results with corrected p-value < 0.05 in the entire network in the validation dataset; Figure 5 (e) Pathway analysis results with corrected p-value < 0.05 in network A in the validation dataset; Figure 5 (f) Pathway analysis results with corrected p-value < 0.05 in network B in the validation dataset.

[0033] Figure 6 The correlation between Alzheimer's disease (AD) and the entire network, subnetwork, hippocampus, and hippocampal subregions. The X-axis represents the entire network, subnetwork, hippocampus, and hippocampal subregions. The primary Y-axis represents the measurement of β value and -log10P, and the secondary Y-axis represents the measurement of R2. Figure 6 (a) represents the correlation in the experimental dataset; Figure 6 (b) shows the correlation in the validation dataset.

[0034] Figure 7 The following is a comparison chart of the four algorithm models. Figure 7 (a): Louvain algorithm community division; Figure 7 (b): LPA label propagation algorithm community division; Figure 7 (c): Community division of GN community discovery algorithm; Figure 7(d): Optimized Louvain community discovery algorithm.

[0035] Figure 8 Venn diagram for the six communities. Figure 8 (a) Genetic Venn diagram; Figure 8 (b) GO Venn diagram; Figure 8 (c)KEGGPATHWAY Venn diagram. DETAILED DESCRIPTION

[0036] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0037] In the field of neurogenetic research, this invention focuses on the constructed voxel-based genetic correlation network, aiming to comprehensively and deeply analyze its biological significance. For this purpose, four rigorous and systematic evaluation strategies are carefully designed and implemented.

[0038] (1) Hippocampal subregion-assisted network importance assessment strategy

[0039] FreeSurfer v6.0, a professional neuroimaging analysis tool, was used to accurately define and segment the subregions of the hippocampus, and then a specific strategy was used to quantify the overlap between the subregions and the network. Through in-depth exploration of the functional characteristics of subregions with complete overlap (100% overlap), the potential biological nature and significance of the network can be effectively explored and derived, laying a solid foundation for subsequent research.

[0040] (2) Strategy combining overlapping voxel analysis with functional enrichment and pathway analysis

[0041] A comprehensive and in-depth evaluation of the overlapping voxels between the genetic correlation network and the sub-regions was carried out. On this basis, the enrichment function analysis and pathway analysis were further carried out to comprehensively measure the biological significance of the network. In the specific operation, the genetic correlation network found was set as the phenotypic variable in the genome-wide association analysis (GWAS), and the diagnostic information, gender, age, education level and the top 4 results in the population stratification analysis were included in the analysis model as covariates, so as to accurately obtain the intrinsic relationship between single nucleotide polymorphisms (SNPs) and the network. Subsequently, SNPs with p-value < 0.05 were screened out according to strict statistical standards, and with the help of ECS (effective chi-square test) and the "hg19" genome reference version, the SNPs were accurately located to the corresponding genes according to their chromosomal location information. Next, the Bonferroni correction was applied to the located genes, and the genes with p-value < 0.05 after correction were applied to the subsequent strategy 2 and strategy 3 analysis processes to ensure the accuracy and reliability of the research results.

[0042] (3) Gene Ontology (GO) functional annotation and pathway analysis strategy driven by genetic interaction networks

[0043] Taking full advantage of the powerful tool of genetic interaction network, the Gene Ontology (GO) function of the constructed network was annotated in detail, and in-depth pathway analysis was carried out based on it. In the specific implementation process, the discovered genetic correlation network was used as the core variable, the disease diagnosis result was used as the phenotypic variable, and gender, age, education level and the top 4 results in the group stratification analysis were used as covariates. Strategy 4 was used to accurately calculate the correlation between the network and Alzheimer's disease (AD), so as to deeply explore the potential mechanism of the network in the occurrence and development of the disease.

[0044] (4) Correlation between the network and AD and comparative analysis strategy with hippocampal subregions

[0045] The correlations between the network and AD and between the hippocampal subregions and AD were calculated simultaneously, and the two sets of correlation data were compared and analyzed in detail (e.g. Figure 5 Through this comparative study, the unique position and mechanism of action of the network in AD-related neurobiological processes are further clarified.

[0046] (5) Validation of the data set results reproducibility test strategy

[0047] All the above evaluation steps are repeated using an independent validation data set to strictly test the repeatability and stability of the results of the present invention and ensure that the research results have a high degree of reliability and generalization ability.

[0048] Through the above series of systematic and in-depth research work, the present invention has achieved a series of valuable results. In terms of the experimental data set, the heritability of 145 voxels and the genetic correlation of 5102 voxel pairs were successfully obtained; in terms of the validation data set, the heritability of 143 voxels and the genetic correlation of 10153 voxel pairs were obtained, and the voxel-based genetic correlation strategy was adopted. The data analysis results show that there are two types of genetic correlations in voxel pairs, positive correlation and negative correlation, and the number of positive correlations is significantly greater than that of negative correlations. Among them, most positive correlations can pass the strict screening of p-value<0.05, while all negative correlations are excluded. In the more stringent multiple test correction process, most positive correlations can still exceed the significance threshold (Bonferroni correction p-value=0.05 / 10,153). This phenomenon fully shows that although the absolute value of negative correlation is not smaller than that of positive correlation in numerical value, its statistical significance p-value is generally at a low level. Based on this, negative correlations are not taken into consideration in the subsequent in-depth analysis process.

[0049] from Figure 2 It can be clearly seen in the genetic correlation map of voxel pairs in the hippocampus shown, where both the x-axis and the y-axis are used to represent voxels. Figure 2 (a) Visually shows the overall picture of genetic correlation of 5102 voxel pairs in the experimental dataset; Figure 2 (b) focuses on the genetic correlation of voxels with p-value < 0.05; Figure 2 (c) shows the distribution characteristics of genetic correlation of 5102 voxels after correction (p-value<0.05 / 5,102). Figure 2 (d)-(f) respectively show in detail the overall genetic correlation of 10153 voxel pairs in the validation dataset, the genetic correlation characteristics of voxel pairs corresponding to p-value<0.05, and the distribution trend of genetic correlation of voxel pairs after correction (p-value<0.05 / 10,153).

[0050] After completing the Bonferroni correction, we carefully selected the voxel pairs with significant correlation to construct a network, which is composed of all SNPs in the data set. The network structure features are as follows: Figure 3 As shown. Figure 3 In the figure, numbers are used to represent voxels, the connecting lines between nodes are used to represent the genetic correlation between voxels, and the color difference is used to intuitively reflect the location information of the hippocampal subregion where the nodes in the network are located. Figure 3 (a) shows the genetic correlation network structure characteristics in the experimental data set. Figure 3 (b) presents the genetic correlation network structure characteristics in the validation dataset.

[0051] Of particular note is that in both independent datasets, the genetic correlation network showed a significant feature of natural differentiation into two sub-networks. Figure 3 (a) and Figure 3 It can be clearly observed in (b) that most voxels in the network tend to be concentrated in a specific network area, and most of the voxels belong to the hippocampal ampal-tail network b. This phenomenon suggests that the voxels outside the two groups are mainly concentrated in the hippocampal region, and the genetic correlation between these voxels is relatively weak or negatively correlated, which provides important clues and research directions for further exploring the genetic structure and functional mechanism of the hippocampus.

[0052] Example

[0053] 1. Imaging and genotype data

[0054] The data used in the present invention were downloaded from the Alzheimer's Disease Neuroimaging Project (ADNI) database (adni.loni.usc.edu). The genetic and imaging data of 1033 participants were then analyzed. In order to verify the validity and repeatability of the results, the ADNI data were re-downloaded and the sample size was increased to 1515. The characteristics of these subjects are shown in Table 1.

[0055]

[0056] Table 1: Characteristics of participants. HC = healthy controls; SMC = mild memory problems; EMCI = early mild cognitive impairment; LMCI = late mild cognitive impairment; AD = Alzheimer's disease

[0057] MRI scans were aligned and normalized to a T1-weighted template image and the Montreal Neurological Institute (MNI) space using voxel-based morphometry (VBM), respectively. Gray matter density (GMD) maps were segmented, extracted, and smoothed with an 8-mm FWHM kernel. The resulting maps were then downsampled to a dimension of 61 × 73 × 61 to reduce the amount of data for subsequent analysis. Two regions of interest (left and right hippocampus) and their coordinates were defined using the automated anatomical labeling (AAL) atlas and aligned to the GMD image.

[0058] SNPs were selected using the method described in . SNPs that met all of the following criteria were then extracted: (1) SNPs on chromosomes 1-22; (2) the call rate of each SNP was ≥95%; (3) the minor allele frequency was ≥5%. After quality control, 563,980 SNPs in the experimental dataset and 5,574,300 SNPs in the validation dataset were used for the following analysis.

[0059] 2. Voxel-based genetic correlation strategy

[0060] vGWAS image data and genetic data obtained by joint analysis were performed on hippocampal voxels using PLINK v1.9. Voxels were used as phenotypes, and diagnosis, sex, age, education, and the first 4 outcomes from population stratification analysis were used as covariates. R was calculated and obtained for each SNP at each voxel in the vGWAS results. 2 .

[0061] Based on the results of the vGWAS data, LDSC was used to estimate the LD score for each chromosome with a window of 1 cm. Then, to avoid inferring that the heritability of each SNP in common variants is the same as that of rare variants, non-partitioned LD score regression was performed for each hippocampal voxel using the LD file and vGWAS results. Using the LD file, the vGWAS results were converted to LDSC format, which includes the number of SNPs, effect allele 1, allele 2, SNP size, p-value, and summary statistics (β, OR, etc.). Subsequently, the calculation file was applied to calculate the heritability of each voxel.

[0062] The heritability explained by S is defined as equation (1).

[0063]

[0064] Where S represents a set of SNPs, y1, y2 represent voxels, j is a SNP, and β, γ are the unique modulus SNPs of y1, y2.

[0065] Since voxels with low average chi2 have low heritability and the genetic correlation between these voxels is infinitely close to zero, in order to reduce the time of calculating genetic correlation, voxels with average chi2>1.02 were selected for the following steps (https: / / github.com / bulik / ldsc / wiki / FAQ). In order to estimate the genetic correlation between y1 and y2, the total error of measuring the relationship between y1 and y2 is defined as the genetic covariance. The genetic covariance between y1 and y2 in S is defined as equation (2).

[0066] ρ s (y1, y2):=∑ j∈s β j γ j (2)

[0067] LDSC was further applied to estimate the genetic correlation between the genotypes of all voxel pairs. Using the obtained heritability and genetic covariance, the genetic correlation between y1 and y2 in S was defined as Equation (3).

[0068]

[0069] Since each genetic correlation calculated by LDSC has a corresponding p-value, the resulting correlations were Bonferroni-corrected, and nodes with significant correlations (p-value < 0.05) were selected to construct a voxel-based genetic correlation network. The nodes in the network represent voxels, and the connections between pairs of nodes describe the genetic correlation of the voxel pairs.

[0070] 3. Materiality Assessment

[0071] For the constructed network, the following four strategies were applied to evaluate the biological significance of the network.

[0072] Hippocampal subfields were used to assess network importance. These subregions were defined and segmented using FreeSurfer v6.0, and strategy 1 was used to calculate the overlap of subregions and networks. The functions of subregions with 100% overlap allowed the biological significance of the network to be derived.

[0073] The overlapping voxels of genetically related networks and subfields were assessed. The overlap within each subfield is shown in Table 2.

[0074] In addition to the above strategies, enrichment functions and pathway analysis were also analyzed to evaluate the biological significance of the network. Using the discovered genetic correlation network as the phenotype of GWAS, the correlation between SNPs and the network was obtained with diagnosis, sex, age, education, and the top 4 results from population stratification analysis as covariates. Then, SNPs with p-values ​​< 0.05 were selected, and these SNPs were mapped to genes according to their chromosome and base positions using ECS ​​(effective chi-square test) and "hg19". Subsequently, the obtained genes were subjected to Bonferroni correction, and genes with corrected p-values ​​< 0.05 were applied to strategies 2 and 3.

[0075] A gene interaction network was established to interpret the gene ontology function of the network. GO term functions such as Figure 3 shown.

[0076] Pathway analysis was also performed using a gene interaction network. The results of the top ten pathway analysis for Alzheimer's disease are shown in Table 4.

[0077] Strategy 4 was performed to calculate the association between the network and AD using the discovered genetic association network as a variable, diagnosis as a phenotype, and sex, age, education, and the first 4 outcomes from the population stratification analysis as covariates.

[0078] The correlation between the network and AD was calculated. In addition, the correlation between the hippocampal subregions and AD was calculated. Then, the information of the two groups was compared ( Figure 5 ).

[0079] Finally, all the above steps were repeated using the validation dataset to test the reproducibility of our results.

[0080] 4.4. Community detection algorithm model.

[0081] The Louvain community detection algorithm can be divided into two stages. For the first stage of community partitioning, a Boolean variable is initially defined to determine whether the entire algorithm can terminate. True indicates that termination is possible, while False indicates that termination is not possible. After that, the key value of the outermost dictionary is extracted and re-iterated. In each iteration, each node and its connected nodes are traversed. The modularity gain of adding each adjacent node to the community of the node is calculated. The sum() function is used to calculate the sum of the weights between each node and the community it is connected to (named "tot"). If two nodes are the same, the weight of the node needs to be subtracted. Each modularity gain is added to a list. The modularity gains in the cid_Q dictionary are sorted, and the items function is used to generate a list of (key, value) tuples from the cid_Q data. This list is then traversed and sorted using the sorted() function. The first tuple in the descending list is taken to partition the community. The condition for partitioning is that the modularity gain is greater than 0 and the nodes are different. If this condition is met, the node is added to the community and removed from the other community, thereby updating the two termination variables.

[0082] For the second phase, when the algorithm does not need to terminate in the first phase, it continues. Its main purpose is to update the communities obtained in the first phase. Each community obtained in the first phase is regarded as a new node. The weight of the new node is the weight of the cycle formed by the community. The weights between communities are converted into weights between new nodes. First, all internal nodes of the community are traversed, the node information of each updated community is stored in vid_vertex, and the relevant information of the community is stored as a new node in vid_vertex. Nested loops are used to traverse the communities obtained in the first phase. The goal is to regard each community as a brand new node and update the weights between the nodes of this community and the nodes of other communities. If the weight between a node in this community and a node in another community is greater than 0, the weight is updated in the network graph to form a new network graph.

[0083] The experimental results of the present invention found that the modularity value of community division is not ideal. The Louvain community detection algorithm divides communities based on the modularity gain Q. By adjusting the value of the modularity gain Q, the modularity can be increased and better community division can be achieved. The adjusted modularity gain formula is as follows:

[0084]

[0085] At this point, the modularity Q is 0.6411, indicating that the community division has produced six communities. The number of voxel values ​​in each divided community is relatively similar, and the modularity Q value is much better than the communities divided by other algorithms. Therefore, the result of this community division is quite satisfactory. Next, the stability of the Louvain algorithm, the LPA label propagation algorithm, and the GN algorithm are compared. Compared with other algorithms, the GN community detection algorithm has a longer running time, so it is not ideal in terms of execution time. The LPA label propagation algorithm has lower stability because it randomly assigns communities when encountering the same number of labels, and usually requires multiple runs to obtain satisfactory results. In contrast, the Louvain algorithm and the GN algorithm performed well in terms of stability.

[0086] In summary, the use of the Louvain community detection algorithm for community segmentation generally outperforms the other two algorithms in terms of modularity, outliers, balance, and runtime.

[0087] SNPs including rs13303118, rs3131972, rs12124819, rs11240777, rs4970383, rs4475691, etc. were processed by KGG software with p-value less than 0.05. Then, the SNPs were mapped to genes using the software. The obtained genes were subjected to Bonferroni correction, and genes with p-value less than Bonferroni (0.05) after correction were considered significant.

[0088] The nine most significant genes are shown in Table 2.

[0089] Gene Conventional P value Adjusted P value CSMD1 2.72E-26 1.58E-22 CDH13 5.32E-17 3.1E-13 RBFOX1 3.27E-14 1.9E-10 PTPRD 2.06E-13 1.2E-09 WWOX 1.1E-12 6.41E-09 PRKG1 2.48E-12 1.44E-08 CTNNA2 2.56E-11 1.49E-07 ADARB2 3.36E-11 1.96E-07

[0090] Previous studies on genetic associations have mainly focused on the associations between different diseases. For example, Feng et al. used genome-wide association study (GWAS) statistics to study the genetic relationship between five cancers and Alzheimer's disease (AD) to examine the co-inheritance of AD and cancer; Guerreiro et al. applied genetic association estimates to demonstrate that Lewy body dementia is equally correlated with Parkinson's disease and AD. In a recent study, hippocampal subregions were used as phenotypes to calculate genetic associations between subregions, with AD, and with schizophrenia. In the present invention, voxels were used as phenotypes for the first time, and the heritability of 302 voxels and the genetic associations between them were calculated. The voxel-based method can not only describe the outline of the subregion, but also explore its internal composition. In addition, compared with directly using subregions as phenotypes, the combination of subregions and voxels has stronger statistical power. Compared with the previous voxel-wide genome association study (vGWAS) method, the calculation of voxel-level heritability and inter-voxel genetic associations are integrated. In this process, the results of vGWAS were used to calculate heritability, taking into account the contribution of even small signals, thus avoiding the loss of these signals. In this way, a genetic association network was obtained, which was superior to hippocampal subregions in terms of influencing AD, goodness of fit, and correlation. Given the above advantages of the proposed network, it can be further studied as a new research group in the hippocampus and can replace the hippocampus to a certain extent. Using voxels as phenotypes provides more detailed spatial resolution, making it easier to capture subtle genetic signals. However, using voxels as phenotypes also has limitations, such as requiring more computing resources and time. Voxel-based methods have detailed spatial resolution and improve the precision and specificity of genetic associations, thereby enabling the identification of more subtle relationships between genetic variants and brain structure.

[0091] The modularity of the improved Louvain algorithm is 0.6411, which is better than the modularity of the communities divided by other algorithms (0.0276, 0.0021 and 0.0251). In terms of balance, the balance number of the Louvain community detection algorithm is 6, while the balance number of the optimized Louvain community detection algorithm is 5. The balance number of the LPA label propagation algorithm is 40, and the balance number of the GN community detection algorithm is 30. Therefore, compared with the LPA label propagation algorithm and the GN community detection algorithm, the Louvain community detection algorithm shows better balance in this community network. In terms of algorithm stability, the LPA label propagation algorithm usually needs to be run multiple times to obtain satisfactory results because it randomly selects a community to assign nodes when encountering communities with the same number of labels. In contrast, the Louvain community detection algorithm and the GN community detection algorithm show satisfactory performance in terms of stability. Although the Louvain algorithm is superior to other algorithms in terms of modularity, balance and running time, it also has limitations. For example, the communities identified by the algorithm tend to be small and fail to fully cover specific hippocampal subregions, thereby reducing the relevance of each community to Alzheimer's disease.

[0092] The optimized Louvain algorithm obtained a significantly higher modularity Q value than other algorithms, indicating its superior ability to identify community structures within the hippocampal voxel network. The algorithm makes the connections within the community more dense and the connections between communities more sparse. The optimized Louvain algorithm divides the network into six communities, and the number of voxels in each community is relatively balanced, avoiding extreme or overly dispersed community divisions and improving the reliability of subsequent biological analysis. In these six communities, common genes including CSMD1, PTPRD, WWOX, PRKG1 and CDH13 were identified. Through GO and KEGG pathway analysis, AD-related gene pathways such as calcium ion binding, plastid membrane and glutamatergic synapse were found in the communities identified by the optimized Louvain algorithm. These pathways are consistent with existing AD research results and provide new theoretical insights into the pathogenesis of AD.

[0093] Previous studies have shown that the inferior parahippocampal gyrus (HATA) is closely related to the amygdala. The volume of HATA and parahippocampal regions was reduced in the multiple system atrophy mild cognitive impairment (MCI) group compared with the healthy group. This suggests that changes in HATA and parahippocampal regions affect information processing. Extending these previous findings to the discovered network, the reduction of voxels in HATA suggests that it affects information processing. These findings provide support for understanding the significance of the network. As a connecting area between the hippocampus and the amygdala, HATA plays an important role in learning and memory functions. Through statistical analysis, GO term analysis, and pathway analysis, it was found that the genetic association network was more correlated with AD than HATA and comparable to that of the hippocampus. In addition, the genetic association network was able to more accurately identify genetic variants associated with AD and provide relevant pathway analysis results. This finding suggests that although HATA is a hippocampal subregion associated with AD, other voxels in the genetic association network have made important contributions to enhancing statistical power and discovering genetic variants associated with AD. Since the hippocampal genetic association network was identified using voxel-wise heritability, this network can be used to further explore the functions and interactions of the hippocampus and their changes in different disease states.

[0094] In previous studies of hippocampal subregions, these subregions were studied as a whole or as a collection of voxels without considering the correlation between individual voxels. The construction of a voxel-based genetic association network is based on the genetic correlation between voxels. Compared with studies focusing on hippocampal subregions, the network is more tightly connected internally and has higher statistical power. In neurogenetic studies, the network excludes voxels with low contribution or low correlation, thereby identifying the top genes associated with AD. Through biological significance analysis, the top pathways were also found to be associated with AD, indicating that the network has certain biological relevance.

[0095] The differences between subnetwork A and subnetwork B indicate that the two subnetworks not only have a large deviation in position, but also have their own unique functions. Network A mainly plays a role in cell function, while network B plays a vital role in cell-to-cell signaling through synaptic function and neurons. However, although the interaction between the two subnetworks integrates some functions (GO:0030198, GO:0031012, GO:0043005, and GO:0007165), it also weakens or even offsets some functions (GO:0099061, GO:0098632, GO:0014069, GO:0005829, and GO:0045202). It is not clear how the two subnetworks interact with each other, so the interaction of differentially expressed genes in the two subnetworks deserves further study.

[0096] The GO of the entire network focuses on functions such as the basic construction of cells (occupying 5 of the top 16 GOs) and information transmission between cells (occupying 11 of the top 16 GOs). Therefore, it can be inferred that the entire network plays an important role in the signal transduction of the hippocampus. In addition, as part of GO:0071944 (cell periphery), the most significant GO:0005886 (plasma membrane, corrected p-value = 1.92E-5) is the membrane surrounding the cell and is essential for the formation of the cell. Another significant GO:0007399 (nervous system development, corrected p-value = 1.17E-11) is the parent term of more than a dozen neural-related GOs, such as GO:0021675 and GO:0007417, which are definitely related to the progression of neurological diseases. For example, through enrichment analysis and pathway overexpression analysis, the target genes of nervous system development function were enriched in the pathway (GO:0007399) and were found to be a potential pathogenesis of sporadic amyotrophic lateral sclerosis. Significant gene enrichment was found in GO:0007399 in autism spectrum disorder and schizophrenia.

[0097] For the entire network, the interactions of the significant pathways have an impact on Alzheimer's disease. For example, Hsa04151 (Bonferroni-corrected p-value = 6.38E-3) is associated with Hsa04933 (Bonferroni-corrected p-value = 1.14E-3) by participating in apoptosis. Then, combined with Hsa04020 (Bonferroni-corrected p-value = 1.18E-4), they are associated with Hsa05010 (Alzheimer's disease) and lead to neurodegeneration in AD by destroying Ca2+ in cells.

[0098] The β, R2, and p-values ​​of the entire network were superior to those of all hippocampal subregions. In addition, network A and network B were also superior to subregions in these three aspects. This once again proves that even if the discovered network is divided into two subnetworks, they are still superior to hippocampal subregions in terms of affecting AD, goodness of fit, and correlation, second only to the entire hippocampus. Therefore, given that the proposed network has the above advantages, it can be further studied as a new research group in the hippocampus and can replace the hippocampus to a certain extent.

[0099] like Figure 8 As shown, in Figure 8(a), among the shared genes CSMD1, PTPRD, WWOX, PRKG1, and CDH13 in six communities, V.A.Stepanov et al. found a strong association between the rs2616984 variant in the CSMD1 gene and Alzheimer's disease in the Russian population, with a p-value of 0.018. In addition, they identified gene combinations involved in epistatic interactions, including CSMD1 and NRIP1, NOTCH4, CSMD1 and NRIP1, and TLR4, CSMD1 and NRIP1, and their association with Alzheimer's disease. Hannah A.Taylor et al. experimentally demonstrated that the PTPRD gene played a role in the use of β-site amyloid precursor protein cleaving enzyme 1 (BACE1) in the treatment of Alzheimer's disease. Brian W.Kunkle et al. analyzed patients with late-onset Alzheimer's disease and identified and confirmed five new genome-wide loci associated with Alzheimer's disease, namely IQCK, ACE, ADAM10, ADAMTS1, and WWOX. The p-value of these risk genes was determined to be 1.32×10-7. F.-F. Liu et al. confirmed through experiments that microRNA-377 can regulate the expression level of CDH13 by promoting cell proliferation and inhibiting cell apoptosis, thereby participating in the occurrence of Alzheimer's disease, and proved the close relationship between CDH13 and Alzheimer's disease.

[0100] exist Figure 8In (b), the three common genetic pathways shared by the six communities are calcium ion binding, plasma membrane, and glutamatergic synapses. In a comprehensive study, Cristovao et al. investigated the potential role of S100 proteins in several pathological processes associated with Alzheimer's disease (AD). Neuroinflammation leads to increased levels of S100 calcium-binding proteins in the brains of AD patients. The association of S100 proteins with protein inclusions suggests that they actively participate in the formation of amyloid aggregates. In addition, S100 proteins are known to regulate aggregation processes, affect AD-related signaling pathways, and influence trace metal homeostasis in the brain. Wang et al. published an article highlighting the importance of calcium (Ca2+) imbalance in the pathogenesis of Alzheimer's disease. In Alzheimer's disease, the regulation of Ca2+ concentrations in the cell membrane, extracellular space, endoplasmic reticulum, and lysosomes is disrupted. Aβ and tau proteins interact with Ca2+ channels, leading to an increase in intracellular Ca2+ levels and disruption of Ca2+ signaling pathways. The present invention also highlights the mechanisms by which Ca2+ homeostasis is disrupted in plasma membranes and organelles such as the endoplasmic reticulum and mitochondria, leading to cell damage and apoptosis. Conway discussed current treatment options for Alzheimer's disease, which primarily target the cholinesterase and glutamate systems, but with limited efficacy observed in clinical trials. Conway suggested studying the regulation of brain glutamate and the impact of atypical pathways, as well as the potential role of branched-chain amino acids (BCAAs) in modulating neurotransmitter profiles under disease conditions and their use in Alzheimer's disease treatment.

[0101] exist Figure 8In (c), three common genetic pathways shared by the six communities were insulin resistance, circadian entrainment, and hypertrophic cardiomyopathy (HCM). Studies by Kellar et al. have shown that insulin plays a crucial role in regulating glucose metabolism in the body and affects various brain functions. It enhances synaptic activity, regulates vascular function, and affects the clearance of amyloid peptides and the phosphorylation of tau protein. Insulin imbalance may lead to neurodegeneration, and new approaches to restore insulin function in the brain are being studied as potential treatments for Alzheimer's disease and related diseases. Studies by Hoyt et al. have shown that circadian rhythm entrainment affects learning, memory, and mood through cell signaling events, gene networks, and synaptic circuits. Central clocks and peripheral oscillators play a role in controlling synaptic plasticity and memory. Disruption of circadian rhythms in cortical and limbic circuits in Alzheimer's disease is associated with early and mid-stage cognitive impairment in Alzheimer's disease; therapies aimed at strengthening cellular time networks may improve cognitive abilities in Alzheimer's patients. Lee et al investigated the risk of dementia, Alzheimer's disease, and vascular dementia in patients with hypertrophic cardiomyopathy, the most common inherited heart disease (HCM). Data collected from a large longitudinal cohort showed that the risk of dementia, especially Alzheimer's disease, in the HCM group increased by 50%. The present invention proposes early monitoring and prevention of cognitive dysfunction in HCM patients to improve their quality of life.

Claims

1. A method for detecting a genetically related network of Alzheimer's disease, characterized in that It includes four evaluation strategies, namely, a network importance evaluation strategy assisted by hippocampal sub-regions, a strategy combining overlapping voxel analysis with functional enrichment and pathway analysis, a gene ontology functional annotation and pathway analysis strategy driven by genetic interaction networks, and a network-AD correlation and comparative analysis strategy with hippocampal sub-regions.

2. A method for detecting a genetic correlation network of Alzheimer's disease according to claim 1, characterized in that The hippocampal sub-region-assisted network importance assessment strategy uses the FreeSurfer v6.0 neuroimaging analysis tool to accurately define and segment the hippocampal sub-regions, and then uses a specific strategy 1 to quantify the overlap ratio between the sub-regions and the network.

3. The method for detecting a genetic correlation network of Alzheimer's disease according to claim 1, characterized in that The strategy of combining overlapping voxel analysis with functional enrichment and pathway analysis is to conduct a comprehensive and in-depth evaluation of overlapping voxels between genetic related networks and sub-regions, carry out enriched functional analysis and pathway analysis, and use this as the core means to comprehensively measure the biological significance of the network.

4. A method for detecting a genetic correlation network of Alzheimer's disease according to claim 3, characterized in that Specific operation process: The discovered genetic correlation network is set as the phenotypic variable in the whole genome association analysis, and the diagnostic information, gender, age, education level and the top 4 results in the group stratification analysis are included in the analysis model as covariates, so as to accurately obtain the intrinsic correlation between single nucleotide polymorphisms and the network; Subsequently, SNPs with p-value < 0.05 were screened out based on strict statistical standards, and with the help of an effective chi-square test and the "hg19" genome reference version, the SNPs were accurately located on the corresponding genes based on their chromosomal location information; then, Bonferroni correction was performed on the located genes, and the genes with p-value < 0.05 after correction were applied to the subsequent Strategy 2 and Strategy 3 analysis processes to ensure the accuracy and reliability of the research results.

5. The method for detecting a genetic correlation network of Alzheimer's disease according to claim 1, characterized in that The genetic interaction network driven gene ontology function annotation and pathway analysis strategy: the genetic interaction network is used to annotate the gene ontology functions of the constructed network in detail, and based on this, in-depth pathway analysis is carried out.

6. A method for detecting a genetic correlation network of Alzheimer's disease according to claim 5, characterized in that Specific implementation process: The discovered genetic correlation network is used as the core variable, the disease diagnosis result is used as the phenotypic variable, and gender, age, education level and the top 4 results in the group stratification analysis are used as covariates. Strategy 4 is used to accurately calculate the correlation between the network and Alzheimer's disease, so as to deeply explore the potential mechanism of action of the network in the occurrence and development of the disease.

7. The method for detecting a genetic correlation network of Alzheimer's disease according to claim 1, characterized in that The strategy for comparative analysis of the network's correlation with AD and with hippocampal subregions is to simultaneously calculate the correlation between the network and AD and between the hippocampal subregions and AD, and conduct a meticulous comparative analysis of these two sets of correlation data to clarify the network's unique position and mechanism of action in AD-related neurobiological processes.

8. A method for detecting a genetic correlation network of Alzheimer's disease according to any one of claims 1 to 7, characterized in that Validation data set result reproducibility test strategy: Use an independent validation data set to completely repeat all the above evaluation steps to strictly test the repeatability and stability of the results of this invention and ensure that the research results have a high degree of reliability and generalization ability.

9. A method for detecting a genetic correlation network of Alzheimer's disease according to any one of claims 1 to 7, characterized in that vGWAS was performed on hippocampal voxels using PLINK v1.

9. The image data and genetic data obtained by joint analysis were used, and the voxels were used as phenotypes. Diagnosis, sex, age, education, and the first 4 results from the population stratification analysis were used as covariates. The R of each SNP at each voxel was calculated and obtained in the vGWAS results. 2 ; Based on the results of the vGWAS data, LDSC was used to estimate the LD score for each chromosome with a window of 1 cm. Then, to avoid inferring that the heritability of each SNP in common variants is the same as that of rare variants, non-partitioned LD score regression was performed for each hippocampal voxel using the LD file and vGWAS results. Using the LD file, the vGWAS results were converted to LDSC format, which included the number of SNPs, effect allele 1, allele 2, SNP size, p-value, and summary statistics. Subsequently, the calculation file was applied to calculate the heritability of each voxel. The heritability explained by S is defined as equation (1), Where S represents a set of SNPs, y1, y2 represent voxels, j is a SNP, β, γ are the unique modulus SNPs of y1, y2, Since voxels with low average chi2 have low heritability and the genetic correlation between these voxels is infinitely close to zero, in order to reduce the time of calculating genetic correlation, the following steps select voxels with average chi2>1.02 (https: / / github.com / bulik / ldsc / wiki / FAQ). In order to estimate the genetic correlation between y1 and y2, the total error of measuring the relationship between y1 and y2 is defined as the genetic covariance. The genetic covariance between y1 and y2 in S is defined as equation (2), r s (y1,y2):=∑ j∈S b j c j (2) LDSC was applied to estimate the genetic correlation between genotypes of all voxel pairs. Using the obtained heritability and genetic covariance, the genetic correlation between y1 and y2 in S was defined as Equation (3), Since each genetic correlation calculated by LDSC has a corresponding p-value, the resulting correlations were Bonferroni-corrected, and nodes with significant correlations (p-value < 0.05) were selected to construct a voxel-based genetic correlation network, in which nodes represent voxels and connections between pairs of nodes describe the genetic correlation of voxel pairs.

Images