Asthma risk site discovery and multi-gene risk scoring system based on deep learning and multi-character analysis

Abstract

The invention discloses an asthma risk site discovery and multi-gene risk scoring system based on deep learning and multi-character analysis. According to the system, preliminary site screening is carried out through GWAS meta analysis and a multi-character statistical method (condFDR / MTAG), a deep learning model Insight GWAS based on a Transform architecture is innovatively adopted, GWAS statistics, linkage imbalance information and multifunctional genome annotation features are integrated, and high-precision site priority sorting is achieved. A polygene risk score (PRS) constructed on the basis shows excellent prediction performance in an independent verification queue. According to the invention, five novel asthma risk SNP loci including rs11766867, rs17644896 and the like and related genes thereof are identified at the same time, and an important genetic tool and a biomarker are provided for early diagnosis, risk prediction and targeted prevention and treatment of asthma.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics and computational genomics, specifically involving a system and method that integrates multi-trait statistical genetics methods and the Transformer deep learning model to systematically discover genetic loci of asthma risk and construct high-precision polygenic risk scores (PRS). Background Technology

[0002] Asthma is a common and complex inherited respiratory disease. Although large-scale genome-wide association studies (GWAS) have identified hundreds of genetic loci associated with asthma, current genetic findings still fall far short of explaining its full heritability. This is mainly due to: first, the limited statistical power of individual GWAS studies, with many loci having small or moderate effects remaining undiscovered; second, traditional single-trait analysis methods failing to fully utilize the additional information provided by disease-related endophenotypes (such as eosinophil counts); and third, the predictive accuracy and cross-population generalization ability of existing PRS models still need improvement.

[0003] In existing technologies, multi-trait analysis methods (such as conditional error discovery rate (condFDR) and multi-trait meta-analysis (MTAG)) have improved the ability to discover shared genetic loci by integrating GWAS data of multiple related traits. For example, Weichao Yuan et al. (Yuan W et al., Investigating the shared genetic basis of inflammatory bowel disease and systemic lupus erythematosus using geneticoverlap analysis. BMC Genomics. 2024 Sep 16;25(1):868.) used condFDR and MTAG methods to analyze the shared genetic basis of inflammatory bowel disease and systemic lupus erythematosus. However, these methods still rely on linear statistical models and cannot effectively handle complex patterns in high-dimensional genomic data.

[0004] On the other hand, PRS methods such as EndoPRS (Kharitonova EV et al., EndoPRS: Incorporating endophenotype information to improve polygenic risk scores for clinical endpoints-A study in asthma. Am J Hum Genet. 2025 May 1;112(5):1199-1214.) attempt to improve polygenic risk scores for asthma by utilizing information from endophenotypes (such as eosinophil count) through a weighted LASSO model. Although this method recognizes the importance of endophenotypes, its core model (LASSO) is a linear model, which cannot capture the complex nonlinear relationships and feature interactions in genetic data, and its site discovery process is independent of deep learning.

[0005] Therefore, existing technologies lack a solution that can simultaneously integrate multi-trait statistical methods and deep learning models to systematically discover new risk sites and construct high-precision PRS. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned technical problems by providing an asthma genetic locus discovery and risk scoring system that integrates multi-trait analysis and the Transformer deep learning framework, aiming to overcome the shortcomings of existing technologies in locus discovery and risk prediction.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A system for identifying asthma risk loci and a multi-gene risk scoring system, the core of which lies in a multi-method fusion architecture that organically combines three levels of analysis:

[0009] 1. GWAS Meta-analysis and Multi-trait Analysis Layer: The system first performs a fixed-effects meta-analysis on asthma GWAS summary statistics from multiple large biobanks (such as the Global Biobank Meta-analysis Initiative (GBMI) and the Million Veterans Project (MVP)) to maximize discovery power. Subsequently, multi-trait analysis is introduced, particularly condFDR and MTAG, to jointly analyze asthma with endophenotypes such as eosinophil count, leveraging genetic pleiotropy to enhance the statistical power of locus discovery. The locus set obtained in this stage provides statistically validated, high-quality prior knowledge for subsequent deep learning.

[0010] 2. Transformer-driven site priority ranking layer (InsightGWAS): This is the core innovation and substantial contribution of this invention.

[0011] Dedicated Model Architecture: This invention provides a deep learning framework called InsightGWAS, based on a simplified and efficient Transformer encoder architecture. This architecture comprises two encoder layers, an input processing layer, and an output layer. The encoder layer consists of two main modules: a multi-head self-attention mechanism and a position-wise feedforward network. Residual connections and layer normalization are added around each module to enhance convergence and stability. The encoder layer has four attention heads, and the feedforward network has a hidden layer size of 64. Unlike traditional encoder-decoder structures, this invention omits the decoder. The encoder's output corresponding to the first label is directly fed into a fully connected layer and activated by a sigmoid function, outputting an association probability score between 0 and 1. This design is specifically optimized for binary classification tasks involving genetic variation.

[0012] The multi-head self-attention mechanism includes both a self-attention mechanism and a multi-head attention mechanism.

[0013] The self-attention mechanism dynamically calculates the dependencies between features through the self-attention formula, and the multi-head attention mechanism captures different types of feature interactions from multiple subspaces through the multi-head attention formula.

[0014] Furthermore, the formula for calculating self-attention is as follows:

[0015]

[0016] Where Q, K, and V represent the query, key, and value matrices, respectively, and d is the dimension of the key vector;

[0017] Furthermore, the formula for calculating multi-head attention is as follows:

[0018]

[0019]

[0020] in It is the corresponding linear projection matrix. It outputs a linear transformation matrix;

[0021] The self-attention mechanism and the multi-head attention mechanism enable the model to automatically learn the complex, non-linear interactions between genetic traits and identify patterns that traditional linear models cannot capture.

[0022] 3. Innovative Multidimensional Feature Engineering: The model represents each genetic variation as a high-dimensional, multidimensional feature vector. Input features go far beyond GWAS summary statistics (P-value, effect size, allele frequency) and linkage disequilibrium (LD) scores; it creatively and systematically integrates over ten categories of functional genomic annotations from public databases.

[0023] Furthermore, the genome annotation includes annotations of quantitative trait loci, epigenome and chromatin accessibility features, regulatory and pathogenic variants;

[0024] Furthermore, the quantitative trait loci include GTEx eQTLs, mQTLs, and sc-eQTLs;

[0025] Furthermore, the epigenome and chromatin accessibility features include digital genome footprinting and ENCODE annotation regions;

[0026] Furthermore, the regulatory and pathogenic variant annotations include microRNA / snoRNA binding sites, transcription factor binding sites, fragment duplications, known GWAS variants, and pathogenicity annotations.

[0027] These features are all linearly projected and positionally encoded before being input into the model to preserve genomic contextual information.

[0028] Targeted Training Strategy: To address the labeling ambiguity of polygenic traits, this invention employs an innovative "two-step labeling" training strategy. The results of a large-scale, statistically powerful GBMI asthma GWAS (121,940 cases / 1,254,131 controls) are used as training labels, while the summary statistics of a smaller, less statistically powerful independent asthma GWAS (88,486 cases / 447,859 controls) and the aforementioned multifunctional features are used as model input. This design forces the model to learn patterns from clear, strong signals and generalize to identifying new, weak associations, significantly enhancing the model's generalization ability and ability to discover new loci.

[0029] Specific Implementation Results: Analysis of the GBMI dataset using this framework identified 1,427 high-priority asthma-related loci, including 684 previously unreported novel loci, demonstrating the method's powerful discovery capabilities. Endophenotypic Enhanced PRS Construction Layer: Based on the set of high-confidence risk loci identified in the above steps and validated by multiple methods (especially those prioritized by InsightGWAS), a multigene risk score was constructed. This PRS model demonstrated higher predictive accuracy and better cross-population generalization ability in independent validation cohorts compared to PRS constructed based on traditional GWAS or existing multi-trait methods (such as MTAG and EndoPRS).

[0030] This invention provides a genetic marker for asthma genetic risk assessment, wherein the genetic marker includes at least one of the following five SNP loci: rs11766867, rs17644896, rs863171, rs2568953, and rs7518602.

[0031] Furthermore, the five SNP sites correspond to the following genes: AHR / ENSG00000237773, AHR, TNFRSF9 / UTS2, RPL31P12 / NEGR1, and PIK3CD / PIK3CD-AS2.

[0032] Furthermore, the corresponding band of gene ENSG00000237773 is located at 7p21.1; the corresponding band of gene UTS2 is located at 1p36.23; the corresponding band of gene PIK3CD-AS2 is located at 1p36.22; and the corresponding band of gene RPL31P12 is located at 1p31.1.

[0033] The present invention also provides the application of the genetic markers described herein in the early auxiliary diagnosis or genetic risk prediction of asthma.

[0034] This invention provides a kit for genetic markers related to asthma genetic risk assessment, wherein the genetic markers include at least one of the following five SNP loci: rs11766867, rs17644896, rs863171, rs2568953, and rs7518602.

[0035] Furthermore, the five SNP sites correspond to the following genes: AHR / ENSG00000237773, AHR, TNFRSF9 / UTS2, RPL31P12 / NEGR1, and PIK3CD / PIK3CD-AS2.

[0036] Furthermore, the corresponding band of gene ENSG00000237773 is located at 7p21.1; the corresponding band of gene UTS2 is located at 1p36.23; the corresponding band of gene PIK3CD-AS2 is located at 1p36.22; and the corresponding band of gene RPL31P12 is located at 1p31.1.

[0037] The present invention also provides the use of a gene for assessing genetic risk of asthma in the preparation of a pharmaceutical composition for the prevention or treatment of asthma, said gene comprising at least one of the following genes: AHR / ENSG00000237773, AHR, TNFRSF9 / UTS2, RPL31P12 / NEGR1, PIK3CD / PIK3CD-AS2.

[0038] Furthermore, the pharmaceutical composition includes an inhibitor of the expression of the gene or the activity of the gene protein product.

[0039] Compared with the prior art, the advantages of this invention are:

[0040] 1. Systematic Innovation: For the first time, GWAS meta-analysis, multi-trait statistical methods, and Transformer deep learning models are organically combined within a unified framework, forming a synergistically enhanced analysis pipeline.

[0041] 2. Enhanced detection capabilities: Transformer models can capture complex patterns that traditional linear models ignore, thereby discovering more potential risk sites, especially those with weak effects or those that operate through complex mechanisms.

[0042] 3. Higher prediction accuracy: The PRS, built based on deep learning priority sites, has demonstrated superior prediction performance in multiple independent cohorts, providing a more reliable tool for precise disease prevention and stratification.

[0043] 4. Technical universality: The system architecture and methodology described in this invention are universal and can be transferred to the study of other complex human traits or diseases. Attached Figure Description

[0044] Figure 1 The overall research design and analysis flowchart provided in the embodiments of the present invention. Figure 2 : Genome-wide association and regional map of asthma in this embodiment of the invention: Figure 2 (A) is a Manhattan plot from a GWAS meta-analysis, showing genome-wide significant loci (red) exceeding P < 5 × 10⁻⁶. -8 The threshold (red dashed line);

[0045] Figure 2 (B–F) Regional association maps of representative new loci: including rs11766867 and rs17644896 at the AHR locus (B–C), rs863171 at the THEMISF9 locus (D), rs2569853 at the NEGRS1 locus (E), and rs7518602 at the PIK3CD locus (F). Colors indicate the degree of linkage disequilibrium (r²) between the European reference panel based on 1000 genomes and major variants.

[0046] Figure 3 Overview of deep learning frameworks for enhancing GWAS signal detection.

[0047] Figure 4 Genetic associations between asthma and hematological traits in individuals of European descent estimated using LD score regression (LDSC). Figure 5Upset plots of risk site overlap identified by different analysis methods (META, condFDR, MTAG, InsightGWAS) in this embodiment of the invention. The bar chart on the left shows the total number of sites identified by each method, while the main plot depicts the intersection between the methods, highlighting shared sites and unique sites. Figure 6 Assessment of asthma PRS using previously published GWAS abstracts in the CHOP cohort: Figure 6 (A) Results for the CHOP-HM cohort, PRS constructed using summary statistics from MVP, GBMI, and their combined meta-analysis (META); Figure 6 (B) is the result of the CHOP-GSA queue, and PRS is also derived from MVP, GBMI and METAGWAS.

[0048] Figure 7 Performance comparison of PRS constructed based on sites identified by different methods in independent validation queues in this embodiment of the invention: Figure 7 (A) is the association between the PRS deciles and asthma risk in the CHOP-HM cohort. Figure 7 (B) is the association between the PRS deciles and asthma risk in the CHOP-GSA cohort.

[0049] Figure 8 Differential expression of asthma-related genes between asthma patients and controls was observed using whole blood gene expression data from GSE141661: violin plots show the normalized expression distribution, and inset box plots represent interquartile ranges. Figure 8 (A) is a differential analysis plot of TNFRSF9. Figure 8 (B) is a differential analysis chart of PRKCD. Figure 8 (C) is a graph showing the difference in AHR. P-values ​​were calculated using a two-tailed Wilcoxon test. Detailed Implementation

[0050] The technical solutions described in this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. Obviously, the embodiments described in this specification are only some feasible technical solutions of this invention. Other implementation methods obtained by those skilled in the art based on the embodiments of this invention without any creative effort should be considered to fall within the scope of protection of this invention.

[0051] Example 1: Systematic discovery of asthma risk loci based on a multi-method fusion architecture

[0052] This embodiment details how to use the system of the present invention to discover new, high-confidence risk genetic loci associated with asthma.

[0053] I. Data Preparation and Preprocessing

[0054] 1. Research Subjects

[0055] Asthma cases and European controls were recruited from the Applied Genomics Center Biobank at Children's Hospital of Philadelphia. This biobank contains pediatric blood samples with whole-genome genotyping data linked to participants' electronic medical records. Asthma status was determined using a validated phenotypic analysis algorithm (see the Asthma Study Project conducted by Lyam Vazquez and John Connolly at Children's Hospital of Philadelphia; for details, see the PheKB platform standard published in 2013: https: / / phekb.org / phenotype / 146). This algorithm integrates ICD-9 / ICD-10 diagnostic codes (493.x and J45), asthma-related medication prescription records, and medical record verification. Cases were defined as individuals aged ≥4 years with a clear asthma diagnosis and related medication history; controls were individuals aged ≥4 years without an asthma diagnosis code and no asthma-related medication history. This study protocol was approved by the Ethics Committee of Children's Hospital of Philadelphia, and all participants and their parents or guardians signed written informed consent forms.

[0056] 2. Genotyping, quality control, and filling gaps

[0057] High-density SNP chips were used for genotyping of DNA samples, including the Illumina HumanHap550 / 610 and Illumina Global Screening chips. Sample-level quality control excluded individuals with the following traits: genotyping detection rate <95%, sex inconsistency, abnormal heterozygosity, or recessive kinship detected by PLINK (PI_HAT > 0.25). Population structure analysis was performed using EIGENSTRAT, retaining only individuals whose clustering was consistent with the 1000 Genomes project European reference population (based on the first two principal components).

[0058] SNP-level quality control excluded the following variants: detection rate <95%, minor allele frequency (MAF) <1%, or Hardy-Weinberg equilibrium test p-value <1×10⁻⁶ in the control group. -6 Genotyping was performed using the TOPMed imputation server and the minimac4 algorithm, with the TOPMed freeze 5b reference panel as a baseline. Variants with MAF > 1% and imputation quality Rsq > 0.5 were retained for subsequent analysis.

[0059] 3. Obtaining GWAS aggregated statistical data

[0060] Pooled GWAS statistics for asthma cases (European ancestry) were obtained from the Global Biobank GBMI, comprising 121,940 cases and 1,254,131 controls (European ancestry). Pooled GWAS statistics for asthma cases (European ancestry) were obtained from MVP, comprising 36,823 cases and 398,278 controls (European ancestry). Pooled GWAS statistics for eosinophil counts in populations of European ancestry were obtained from public databases (sample size 474,237).

[0061] 4. Summary data on blood cell characteristics

[0062] Data were obtained from a cross-ethnic genome-wide association study encompassing 746,667 participants from five major global populations. The analysis was limited to a subset of the European population, with sample sizes for each trait as follows: lymphocyte count (524,923), monocyte count (521,594), basophil count (474,001), eosinophil count (474,237), neutrophil count (519,288), platelet count (542,827), erythrocyte count (545,203), white blood cell count (562,243), mean corpuscular hemoglobin (MCH) (486,823), mean corpuscular volume (MCV) (544,127), and hemoglobin concentration (HCC) (563,946). Detailed information for all datasets is summarized in Table 1.

[0063] Table 1 Summary statistics of asthma and blood cell traits used in this invention

[0064] Traits Document Title PMID Number of Cases Number of Controls Total Sample Size Asthma Diversity and scale: Genetic architecture of 2068 traits in the VA Million Veteran Program. 39024449 36,823 398,278 435,101 Multi-ancestry meta-analysis of asthma identifies novel associations and highlights the value of increased power and diversity 36778051 121,940 1,254,131 1,376,071 Genome-wide analysis highlights contribution of immune system pathways to the genetic architecture of asthma 32296059 88,486 447,859 536,345 Lymphocyte count Trans-ethnic and Ancestry-Specific Blood-Cell Genetics in 746,667 Individuals from 5 Global Populations. 32888493 - - 524,923 Monocyte count 32888493 - - 521,594 Basophil count 32888493 - - 474,001 Eosinophil count 32888493 - - 474,237 Neutrophil count 32888493 - - 519,288 Platelet count 32888493 - - 542,827 Red blood cell count 32888493 - - 545,203 White blood cell count 32888493 - - 562,243 Mean corpuscular hemoglobin 32888493 - - 486,823 Mean corpuscular volume 32888493 - - 544,127 Hemoglobin concentration 32888493 - - 563,946

[0065] II. GWAS Meta-analysis and Multitrait Analysis (MATG)

[0066] 1. Fixed-effects inverse variance weighted meta-analysis: Fixed-effects inverse variance weighted meta-analysis was performed on the pooled asthma GWAS statistics of GBMI and MVP using PLINK software. This step pooled a total of 158,763 cases and 1,652,409 controls, significantly improving statistical power. The analysis identified 466 independent (linkage disequilibrium (LD) r² < 0.05) studies, reaching genome-wide significance (P < 5 × 10⁻⁻⁴). 8 ) sites, of which 69 were previously unreported new sites ( Figure 2 A).

[0067] 2. Multi-trait analysis to improve discovery efficiency:

[0068] MTAG Analysis: MTAG is a generalized meta-analysis framework that leverages genetic correlations between different traits to enhance the statistical power of locus discovery. Variants were filtered before analysis to exclude uncommon SNPs, ambiguous alleles, and duplicate loci. Pairwise genetic correlations between asthma and eosinophil count were estimated using LD fractional regression (LDSC) to calibrate the variance-covariance matrix of the random effects components. Subsequently, MTAG performed a random-effects meta-analysis to generate summary statistics for SNP levels for each trait. In this analysis, asthma was designated as the major phenotype, and eosinophil count as the minor phenotype. A genome-wide significance cutoff of P < 5 × 10⁻ was applied to the major phenotype. 8 For minor phenotypes, a more lenient threshold (P < 0.05) was used to prioritize secondary signals. This analysis identified 340 independent genomic significant loci, 32 of which were novel.

[0069] condFD analysis: The condFDR method expands the standard error rate (FDR) framework by reordering test statistics for the primary phenotype (asthma) using genetic association information from secondary phenotypes (eosinophil count). This reordering process integrates association evidence from secondary traits, enabling a more refined prioritization of variants associated with the primary phenotype. By dynamically adjusting the significance threshold based on eosinophil count, condFDR improves the statistical power for detecting shared genetic loci while maintaining strict control over false findings. This analysis identified 894 asthma-related loci at a condFDR level of <0.05, including 234 novel loci.

[0070] III. Transformer-Driven Site Priority Ranking (InsightGWAS)

[0071] 1. Feature Engineering: A multidimensional feature vector was constructed for each single nucleotide polymorphism (SNP). Features include: ① Statistical features: P-value, effect size (beta), standard error, and minor allele frequency from the aforementioned smaller-scale asthma GWAS. ② Genomic context features: Linkage disequilibrium (LD) score calculated based on the 1000 Genomes European reference population. ③ Functional annotation features: Integration of eQTL and sQTL data from GTEx, chromatin accessibility (DNase-seq) and transcription factor binding site (ChIP-seq) data from ENCODE, and variant harmfulness prediction scores such as GERP++ and CADD (see Table 2 for a detailed feature list).

[0072] Table 2 Summary statistics of asthma and blood cell traits used in this invention

[0073] Feature Category Specific Feature Details Data Source and Tool Genomic Data GWAS Summary Statistics Effect size, P-value, sample size, minor allele frequency (MAF) GWAS Catalog; GBMI LD Score Sum of r2 values between target variant and nearby variants within a 1 Mb window 1000 Genomes Project (East Asian population) Quantitative Trait Locus (QTL) Data sQTL Splice Quantitative Trait Locus Data GTEx v8 eQTL (Whole Blood) Gene expression quantitative trait loci from blood tissue 1. Westra et al. 2013, Nat Genet; 2. Lloyd-Jones et al. 2017, Am J Hum Genet; 3. Lappalainen et al. 2013, Nature eQTL (GTEx) Gene expression quantitative trait loci from 49 human tissues GTEx v8 mQTL Methylation Quantitative Trait Locus Data 1. Hatton et al. 2024, Nat Commun; 2. McRae et al. 2018, Sci Rep; Wu et al. 2018, Nat Commun; 3. Hannon et al. 2018, Am J Hum Genet; 4. Hannon et al. 2017, Am J HumGenet; 5. Peng et al. 2024, Nat Genet; sc-eQTL Single-cell eQTL data (immune cells) scQTLbase (Ding et al. 2024, Nucleic Acids Res ) Epigenetics and Chromatin Accessibility Digital Genome Footprint (DGF) Nucleotide resolution map of transcription factor binding sites Vierstra et al. 2020, Nature Genomic functional element annotation Transcriptional regions, promoter markers (H3K4me3, H3K9ac), enhancer markers (H3K4me1, H3K27ac), etc. ENCODE project Functional variation annotation MicroRNA and snoRNA structures Variations affecting microRNA / snoRNA and their binding sites ANNOVAR annotations: miRBase, snoRNABase, TargetScan Transcription factor binding sites (TFBS) TFBS Comments ANNOVAR Note: tfbsConsSites,TRANSFAC Fragment repetition Variations located in fragment repetition regions ANNOVAR Note Reported GWAS variants Variations associated with mental and neurological disorders GWAS Catalog; Open Targets Database Pathogenicity annotation Conservatism (GERP++), functional impact prediction (PolyPhen-2, SIFT, CADD, etc.) GERP++; PolyPhen-2; SIFT; CADD and related tools

[0074] 2. Model training and inference, such as Figure 3 As shown: ① Model Architecture: A deep learning framework based on Transformer encoders (InsightGWAS) is adopted. This architecture includes two encoder layers, each equipped with a four-head self-attention mechanism and a position-wise feedforward network. Residual connections and layer normalization are used to ensure training stability. ② Training Strategy: To address the label ambiguity of polygenic traits, a two-step labeling method is adopted. The results of the large-scale, statistically powerful GBMI asthma GWAS (121,940 cases / 1,254,131 controls) are used as training labels, while the summary statistics and the aforementioned functional features of the smaller, statistically less powerful independent asthma GWAS (88,486 cases / 447,859 controls) are used as model inputs. This allows the model to learn from explicit signals and generalize to new associations. ③ Locus Prioritization: The SNP features of the GBMI dataset are input into the trained InsightGWAS model, and the model outputs an association probability score between 0 and 1 for each SNP. By setting probability thresholds, 1,427 high-priority asthma-related loci were identified, of which 684 were novel loci that had not been previously reported.

[0075] IV. Integration and Validation of High-Confidence Locator Sets

[0076] 1. Validation of Genetic Association Analysis: The linkage disequilibrium score regression (LDSC) method was used to assess the genetic association between asthma and leukocyte traits, and the analysis was limited to HapMap3 variants according to methodological recommendations. The study used HapMap3 single nucleotide polymorphism linkage disequilibrium (LD) scores from the 1000 Genomes Project for European populations, while excluding variants with imputed information factor (INFO) <0.9 to ensure data quality. Results are as follows: Figure 4 As shown, asthma is significantly positively correlated with eosinophil count (rg > 0, P < 0.05), and also shows varying degrees of association with basophils, monocytes, etc. This finding confirms the close link between asthma and immune regulatory pathways at the genetic level, providing a theoretical basis for subsequent condFDR and MTAG multitrait analysis using eosinophil count as a conditional trait.

[0077] 2. Cross-validation of findings from multiple methods: This invention systematically compares the findings from traditional meta-analysis (META), multi-trait analysis methods (condFDR / MTAG), and deep learning frameworks (InsightGWAS), such as... Figure 5The Upset plot clearly shows that while each method can identify a considerable number of unique sites, 37 high-confidence sites are supported by all three methods. This indicates that the multi-method fusion architecture proposed in this invention can capture real disease signals from different perspectives, demonstrating both the complementarity of each method and significantly improving the reliability of the findings through cross-validation.

[0078] 3. Biological Functional Annotation: In-depth functional annotation was performed on these 37 high-confidence loci. For example, using the GTEx database for quantitative trait (eQTL) analysis, rs863171 (similar to TNFRSF9) was found to be a significant eQTL in lung tissue (P = 1.5 × 10⁻⁻⁻⁶). 5 Differential expression analysis of representative genes (such as TNFRSF9, PIK3CD, NEGR1, and AHR) confirmed their differential expression between asthma patients and controls. Single-cell eQTL analysis of rs11766867 at the AHR locus revealed a significant regulatory effect in monocytes and CD14⁺ monocytes (P < 0.001). Fine localization using LocusZoom identified a second independent signal (rs17644896) within the AHR locus, suggesting the presence of multiple functional variations (such as...). Figure 2 BF, Table 3).

[0079] Table 3: Single-organization eQTLs of newly discovered representative loci in meta-analysis and combined FDR analysis

[0080] SNP Cytogenetic bands CHR BP A1 A2 Gene symbol p-value NES Organizational Source rs863171 1p36.23 1 7932555 C T UTS2 0.0000002 -0.27 transverse colon rs7518602 1p36.22 1 9651286 T C PIK3CD-AS2 0.0000033 -0.15 thyroid rs2568953 1p31.1 1 72288273 T C RPL31P12 0.0000000000000000055 -0.89 Cerebellum rs11766867 7p21.1 7 17152931 C G ENSG00000237773 0.000014 -0.32 Tibial nerve

[0081] Captions: SNP: Reference SNP identifier (rsID); Cytogenetic bands: Chromosomal band location; CHR: Chromosome; BP: Base pair location on the GRCh37 reference genome; Gene: Gene name; P-value: Association significance; NES: Standardized effect size; Tissue origin: Tissue type from GTEx database.

[0082] This embodiment successfully discovered a large number of new, biologically plausible asthma risk loci through systematic multi-method fusion, and significantly improved the reliability of the discovery results through multi-method cross-validation and functional annotation.

[0083] Example 2: Construction and Validation of PRS Based on Deep Learning Priority Sites

[0084] This embodiment details how to construct a high-precision PRS using the risk sites discovered in Embodiment 1, and validate it in an independent cohort.

[0085] 1. PRS Model Construction

[0086] Benchmark PRS Model: As a comparison benchmark, we constructed the following PRS model:

[0087] PRS-MVP: Based on genomic significant sites and their effect sizes derived solely from MVP GWAS.

[0088] PRS-GBMI: Based on genomic significant loci and their effect sizes derived solely from GBMI GWAS.

[0089] PRS-META: Genomic significant sites and their effect values ​​identified by the MVP+GBMI meta-analysis in Example 1.

[0090] PRS-condFDR: Integrating known European asthma loci with newly discovered loci (234 loci) and their raw effect values ​​from condFDR analysis.

[0091] PRS-MTAG: Integrating known European asthma loci with newly discovered loci (32 loci) from MTAG analysis and their MTAG-corrected effect values.

[0092] The PRS model of this invention (PRS-InsightGWAS) is based on all sites (1,427 sites) prioritized by the InsightGWAS deep learning model in Example 1 and their effect values ​​in the original GWAS. This is the core PRS model of this invention.

[0093] 2. Independent verification queue

[0094] Validation was performed using two independent European pediatric asthma cohorts from the Children's Hospital of Philadelphia (CHOP):

[0095] CHOP-HM cohort: 1,480 cases and 3,008 controls (genotyped using Illumina HumanHap550 / 610 chip).

[0096] CHOP-GSA cohort: 2,623 cases and 6,002 controls (genotyping using Illumina Global ScreeningArray).

[0097] All samples underwent rigorous genotyping quality control and imputation.

[0098] 3. PRS Calculation and Performance Evaluation

[0099] PRS calculation: The PRS values ​​described above were calculated for each individual in the validation cohort using the --score function of the PLINK software. During the calculation, alleles at each locus were weighted and summed according to their effect values.

[0100] Performance evaluation:

[0101] Explained variance (R²): In a logistic regression model, the PRS is included along with covariates (age, sex, and the top 10 principal components), and compared to a null model that only includes covariates. Nagelkerke's R² is calculated, and the R² is converted to a susceptibility scale assuming a prevalence of 8%.

[0102] Decilet risk stratification: The population is divided into 10 equal parts (deciles) according to PRS values ​​from low to high. The lowest decimal is used as a reference to calculate the asthma risk ratio (OR) and its 95% confidence interval for each of the remaining decimals.

[0103] 4. Results and Analysis

[0104] This invention first evaluated the predictive performance of PRS from three GWAS sources: the results are as follows Figure 6 As shown, in both cohorts, PRS-META explained the highest proportion of asthma susceptibility (HM cohort: R² = 2.42%; GSA cohort: R² = 1.82%), followed by PRS-GBMI (HM cohort: R² = 1.95%; GSA cohort: R² = 1.80%) and PRS-MVP (HM cohort: R² = 0.77%; GSA cohort: R² = 0.67%). Stratified analysis based on decimals showed that the risk gradient was consistent with these R² values, with PRS-META exhibiting the largest increase in ratio from the lowest to the highest decimal in both cohorts.

[0105] To evaluate the performance of the Transformer-based model (InsightGWAS) relative to existing statistical methods, this invention constructed three additional PRSs. PRS-MTAG combines reported European population loci with asthma and eosinophilic pleiotropic loci identified by the MTAG method; PRS-condFDR integrates reported loci with additional loci identified by the condFDR method; and PRS-InsightGWAS includes all loci prioritized by the InsightGWAS model. Results are as follows... Figure 7As shown, in both the CHOP-HM and CHOP-GSA cohorts, PRS-InsightGWAS achieved the highest predictive performance (HM cohort: R² = 3.34%; GSA cohort: R² = 2.08%), outperforming PRS-MTAG (HM cohort: R² = 2.05%; GSA cohort: R² = 1.90%) and PRS-condFDR (HM cohort: R² = 1.96%; GSA cohort: R² = 1.99%). Decitary analysis supports these findings, with PRS-InsightGWAS producing the most significant discrimination between high-risk and low-risk groups, indicating the potential of deep learning-based variant site prioritization strategies to enhance multigene risk prediction capabilities. This comparison strongly demonstrates that a set of sites identified by integrating multifunctional genomic features through a deep learning framework is more biologically relevant and predictive of disease risk than a set of sites solely relying on GWAS significance thresholds.

[0106] This embodiment demonstrates that the PRS constructed based on the priority ranking of genetic loci using the deep learning framework (InsightGWAS) of this invention has significantly higher prediction accuracy than PRS constructed based on traditional GWAS or existing statistical methods (such as MTAG, condFDR). This provides a more powerful genetic tool for accurate risk prediction, early screening, and targeted prevention of asthma.

[0107] Example 3: Verification of biological function

[0108] Differential expression analysis validation: This invention obtained an independent whole blood gene expression dataset (GSE141661) from the GEO database and analyzed the expression levels of several asthma-related genes (such as TNFRSF9, PRKCD, and AHR) suggested by this invention in asthma patients and healthy controls. Results are as follows: Figure 8 As shown, these genes all exhibited significant differential expression in asthma patients (P < 0.05). This finding confirms the association between the genes identified in this invention and the pathophysiological processes of asthma at an independent transcriptomic level, providing external evidence to support the functional roles of these sites and enhancing the biological credibility of the findings of this invention.

[0109] Example 4: InsightGWAS Deep Learning Model Architecture and Training Details

[0110] This embodiment further details the specific architecture, training process, and key technical details of the InsightGWAS deep learning model to fully support the feasibility and inventiveness of this invention.

[0111] InsightGWAS Framework Core Technologies: (See attached diagram) Figure 3 This diagram visually illustrates the complete workflow and model architecture of the core deep learning framework (InsightGWAS) of this invention. The left half, "Training and Inference Pipeline," clarifies our innovative "two-step labeling" training strategy: using large-scale, statistically powerful GBMI asthma GWAS results as training labels, while using smaller-scale, less statistically powerful summary statistics and multifunctional features from independent asthma GWAS as model input. The right half, "Model Architecture," clearly depicts how the Transformer-based encoder integrates GWAS statistics, linkage disequilibrium information, and rich multi-omics functional annotations, and captures complex feature interactions through a self-attention mechanism, ultimately outputting SNP-level association probabilities. This diagram systematically summarizes the complete technical solution of this invention from data preparation and model training to final inference, highlighting its core advantages in integrating multi-dimensional information and capturing complex patterns.

[0112] The following is a detailed description of the InsightGWAS deep learning model architecture, input features, training, and optimization of this invention.

[0113] I. The architecture of the InsightGWAS deep learning model is as follows:

[0114] S1 encoder layer: Each encoder block consists of two main modules: a multi-head self-attention mechanism and a position-wise feedforward network. Residual connections and layer normalization are added around each module to enhance convergence and stability.

[0115] S2 Input Processing: Before entering the encoder, the genetic input features are linearly projected to an appropriate dimension (e.g., 64-dimensional) and positional encoding is added to preserve genome order and regulatory context.

[0116] S3 Output Layer: The encoder's output corresponding to the first label is propagated to a fully connected layer, which is then passed through a Sigmoid activation function to produce a binary prediction (probability between 0 and 1) for site-disease association.

[0117] Preferably, the encoder layer in S1 has two layers, each using four attention heads, and the hidden layer size of the feedforward network is 64.

[0118] Preferably, the formula for calculating self-attention described in S1 is as follows:

[0119]

[0120] Where Q, K, and V represent the query, key, and value matrices, respectively, and d is the dimension of the key vector;

[0121] Preferably, the formula for calculating multi-head attention as described in S1 is as follows:

[0122]

[0123]

[0124] in It is the corresponding linear projection matrix. It outputs a linear transformation matrix;

[0125] Each attention head independently computes self-attention, then all outputs are concatenated and linearly transformed to obtain the final result, thereby capturing the relationships between features more comprehensively and improving the model's representational power and adaptability to the complexity of biological data.

[0126] II. Input Features (as shown in Table 2)

[0127] 1. Association Statistics and Population Genetics: Key predictors for this model include GWAS summary statistics and population genetics indicators related to migraine-associated loci. These characteristics include effect size, p-value, sample size, and minor allele frequencies. To capture linkage disequilibrium patterns, linkage disequilibrium scores (LD scores) were derived from European ancestry samples from the 1000 Genomes Project. For each variant, the LD score was calculated by summing the r² values ​​between the target variant and all neighboring variants within a 1 Mb window.

[0128] 2. Functional Annotation: Integrates more than ten types of functional genomic annotations, including:

[0129] ① Quantitative trait loci:

[0130] GTEx eQTLs: This feature set integrates eQTL data from 49 human tissues based on GTEx V8, focusing on variants located within 1 Mb of the transcription start site that show significant association.

[0131] mQTLs: Methylation QTL data are from two main sources. The first dataset includes summary statistics of mQTLs from a meta-analysis of 3,701 European ancestry samples by Hatton et al. The second dataset provides summary statistics of mQTLs from a meta-analysis of 1,980 European individuals, reported by McRae et al. and Wu et al., in SMRBESD format.

[0132] sc-eQTLs: Single-cell eQTL data specific to brain tissue (including astrocytes and neurons) are available from scQTLbase (https: / / bioinfo.szbl.ac.cn / scQTLbase / ).

[0133] ② Epigenome and chromatin accessibility features:

[0134] Digital Genome Footprint: Genome DNase I footprint data, depicting the nucleotide resolution of transcription factor occupancy sites within natural chromatin, is provided by Vierstra et al. This dataset combines over 67 billion DNase I cleavage sites from more than 240 human cell types and states, providing a comprehensive map of transcription factor recognition sites. Detailed information is available at [link to relevant documentation]. https: / / www.vierstra.org / resources / dgf Obtain.

[0135] ENCODE Annotated Regions: Variants located within ENCODE-annotated genomic regions are identified and annotated using data from the ENCODE project. These annotations include transcriptional regions, active promoter markers (H3K4Me3, H3K9Ac), active enhancer markers (H3K4Me1, H3K27Ac), active extension markers (H3K36Me3, H3K79Me2), and repression markers (H3K27Me3, H3K9Me3). These features are integrated into the InsightGWAS framework to improve variant prioritization and provide insights into genome-wide regulatory activity and chromatin status.

[0136] ③Notes on regulation and pathogenic variation:

[0137] MicroRNA and snoRNA Interference: Variations affecting microRNAs, snoRNAs, and predicted microRNA binding sites were annotated using ANNOVAR, a powerful software tool for functional annotation of genetic variations across a wide range of genomes. Variations overlapping with microRNA and snoRNA loci were identified using the wgRna tables (based on miRbase) and snoRNABase from the UCSC Genome Explorer. Furthermore, the TargetScan database was used to detect variations disrupting conserved mammalian microRNA regulatory targets in the 3' UTR region of RefSeq genes. By leveraging the capabilities of ANNOVAR, these annotations provide valuable insights into posttranscriptional regulation and the functional consequences of genetic variations.

[0138] Transcription factor binding sites: Transcription factor binding sites were annotated using ANNOVAR, integrating the tfbsConsSites database. This database provides the locations and scores of conserved TFBSs in human, mouse, and rat alignments. Scores and thresholds were obtained from the Transfac database.

[0139] Fragment duplications: Genetic variations located at fragment duplications are often alignment artifacts and should be interpreted with caution. These variations may appear as SNPs with high coverage and confidence scores, but may represent non-polymorphic sites with identical flanking sequences in the duplication region. Using ANNOVAR, we identified variations overlapping with fragment duplications by leveraging a relevant annotation database. These annotations help flag potential false variations, ensuring more accurate downstream analysis.

[0140] Previously reported GWAS variants: Annotation of variants associated with mental and neurological diseases or related traits using data from the GWAS Catalog and Open Targets. The GWAS Catalog compiles published GWAS results, enabling the identification of variants associated with relevant traits or diseases. Open Targets enhances this by integrating evidence from GWAS, functional genomics, and curated literature, providing a comprehensive view of variant-disease and variant-trait relationships. These resources provide crucial context for interpreting the importance of variants in mental and neurogenetics.

[0141] Pathogenicity Annotation: This feature integrates multiple predictive scores to assess the pathogenicity and functional impact of genetic variants. Conservation indices, such as GERP++, highlight variants in evolutionarily restricted regions under strong selection pressure. Functional impact predictions from tools such as PolyPhen-2 (HDIV and HVAR), SIFT, LRT, MutationTaster, MutationAssessor, FATHMM, MetaSVM, and MetaLR assess the potential of missense variants to disrupt protein structure, biochemical properties, or function. Combined Annotation-Dependent Exhaustion (CADD) raw scores further quantify the harmfulness of variants by integrating multiple functional annotations. By synthesizing these predictive tools, this feature enables InsightGWAS to robustly prioritize variants with potential pathogenic effects, facilitating the identification of biologically relevant loci.

[0142] III. Training and Optimization

[0143] Optimizer and loss function: The Adam optimizer is used, and the binary cross-entropy is used as the loss function.

[0144] Hyperparameters: learning rate set to 0.001, batch size to 256, training epochs to 50.

[0145] Regularization: Use Dropout (ratio 0.1) and early stopping (patience=5) to prevent overfitting.

[0146] Hardware and Software: The model was implemented using the PyTorch framework on an NVIDIA Tesla V100 GPU. The code is open-source on GitHub. https: / / github.com / ziangmeng / Asthma-InsightGWAS .

[0147] This embodiment further demonstrates the repeatability, advancement, and significant advantages of the present invention compared to traditional linear models by describing the technical details and training process of the model in detail.

[0148] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and additions without departing from the principle of the present invention, and these improvements and additions should also be considered within the scope of protection of the present invention.

Claims

1. A system for identifying asthma risk loci and a multi-gene risk scoring system, characterized in that, include: The GWAS meta-analysis and multi-trait analysis module is configured to: perform fixed-effects inverse variance weighted meta-analysis on multiple asthma GWAS datasets, and integrate GWAS summary statistics of asthma with at least one blood intraphenotypic trait using the conditional error discovery rate (condFDR) and the multi-trait meta-analysis (MTAG) method to identify the first risk locus set. The Transformer-driven site prioritization module is configured to: use the InsightGWAS deep learning model based on the Transformer encoder to predict the association probability of genetic variations by integrating GWAS statistics, linkage disequilibrium information and multifunctional genome annotation features, in order to identify the set of second risk sites. The multigene risk score (PRS) construction module is configured to construct a multigene risk score for asthma risk assessment based on the loci and / or effect values ​​in the first and / or second risk loci sets.

2. The system as described in claim 1, characterized in that, In the Transformer-driven site-priority ranking module, the architecture of the deep learning model InsightGWAS includes: The input processing layer is used to perform linear projection and positional encoding on the feature vectors of genetic variations. At least one Transformer encoder layer, which consists of two main modules: a multi-head self-attention mechanism and a position-wise feedforward network, with residual connections and layer normalization in each module; The output layer is used to process the encoder's output through a fully connected layer and a sigmoid activation function to generate association probabilities between 0 and 1.

3. The system as described in claim 1 or 2, characterized in that, The multifunctional genome annotation features include at least three of the following categories: quantitative trait loci, epigenome and chromatin accessibility features, and regulatory and pathogenic variant annotations.

4. The system as described in claim 1, characterized in that, The Transformer-driven site priority ranking module employs a two-step labeling training strategy. It uses the association results of the first large-scale, high-statistical-power GBMI asthma GWAS dataset as training labels, and uses the summary statistics of the second independent GWAS dataset and the multifunctional genome annotation features as model inputs for training.

5. A genetic marker for asthma genetic risk assessment, said genetic marker comprising at least one of the following five SNP loci: rs11766867, rs17644896, rs863171, rs2568953, rs7518602.

6. The genetic marker as described in claim 5, wherein the five SNP sites correspond to the following genes: AHR / ENSG00000237773, AHR, TNFRSF9 / UTS2, RPL31P12 / NEGR1, and PIK3CD / PIK3CD-AS2.

7. The genetic marker as described in claim 6, wherein the gene ENSG00000237773 corresponds to a band located at 7p21.1; the gene UTS2 corresponds to a band located at 1p36.23; the gene PIK3CD-AS2 corresponds to a band located at 1p36.22; and the gene RPL31P12 corresponds to a band located at 1p31.

1.

8. The use of a genetic marker as described in any one of claims 5-7 in the early auxiliary diagnosis or genetic risk prediction of asthma.

9. A kit for assessing genetic risk associated with asthma, the genetic marker comprising at least one of the following five SNP loci: rs11766867, rs17644896, rs863171, rs2568953, rs7518602; The five SNP sites correspond to the following genes: AHR / ENSG00000237773, AHR, TNFRSF9 / UTS2, RPL31P12 / NEGR1, and PIK3CD / PIK3CD-AS2. The corresponding band of gene ENSG00000237773 is located at 7p21.1; the corresponding band of gene UTS2 is located at 1p36.23; the corresponding band of gene PIK3CD-AS2 is located at 1p36.22; and the corresponding band of gene RPL31P12 is located at 1p31.

1.

10. The use of a gene for assessing genetic risk of asthma in the preparation of a pharmaceutical composition for the prevention or treatment of asthma, said gene comprising at least one of the following genes: AHR / ENSG00000237773, AHR, TNFRSF9 / UTS2, RPL31P12 / NEGR1, PIK3CD / PIK3CD-AS2; The pharmaceutical composition comprises an inhibitor of the expression of the gene or the activity of the gene protein product.

Images