A Multi-Ethnic PRS Dynamic Calibration Method and System Based on Transfer Learning

Abstract

The present invention provides a method and system for dynamically calibrating multi-ethnic PRS based on transfer learning, which relates to the technical fields of bioinformatics and medical diagnosis. The method includes: obtaining genomic data of a source ethnic group and a target ethnic group, and generating a high-quality data set through standardized preprocessing; training an initial PRS model using the source data to generate SNP weights; dynamically aligning the feature distributions of the two ethnic groups through an adversarial domain adaptation algorithm, and adjusting the model parameters to adapt to the target ethnic group; screening SNPs with allele frequencies meeting the requirements in the target ethnic group, calculating a weight correction coefficient, and fusing the specific SNP effect values thereof to generate a dynamic PRS formula. The present invention breaks through the ethnic heterogeneity limitation of traditional PRS models, improves the cross-ethnic prediction accuracy and robustness through transfer learning and a dual-path scoring mechanism, supports incremental learning to achieve lightweight multi-ethnic expansion, optimizes the data utilization efficiency, and provides an efficient tool for accurate risk assessment of complex diseases.

Description

Technical Field

[0001] The present invention relates to the fields of bioinformatics and medical diagnosis technology, and in particular to a multi-cluster PRS dynamic calibration method and system based on transfer learning. Background Art

[0002] Polygenic risk scores (PRS) quantify individual disease risk by integrating the genetic effects of multiple loci in the genome and have been widely used in the early diagnosis and risk assessment of complex diseases. Existing PRS models are typically trained based on genetic data from a single ethnic group, such as European or East Asian populations. However, significant genetic heterogeneity exists between different ethnic groups (such as differences in allele frequencies and linkage disequilibrium patterns), resulting in a significant decrease in the performance of PRS models trained on a single ethnic group when applied across ethnic groups. For example, the literature shows that the AUC value of a PRS trained on a European population drops from 0.924 to 0.788 in an East Asian population, indicating insufficient cross-ethnic generalization ability. Traditional diagnostic methods (such as HLA-B27 and MRI) also have limitations such as low sensitivity, high cost, or reliance on subjective interpretation.

[0003] The existing technology has the following defects:

[0004] (1) Ethnic group-specific limitations: The PRS model lacks cross-ethnic adaptability and needs to be retrained for different ethnic groups, resulting in a waste of data and computing resources;

[0005] (2) Static model defects: most existing PRS calibration methods use static weight adjustment and cannot dynamically respond to the dynamic changes in genetic characteristics among ethnic groups;

[0006] (3) Data utilization is insufficient, and the transfer learning mechanism of cross-ethnic genetic data has not been fully integrated, making it difficult to explore the synergistic effects of shared genetic characteristics and ethnic-specific loci.

[0007] To address the above shortcomings, a method that can dynamically calibrate multi-ethnic PRS is urgently needed. Summary of the Invention

[0008] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide a multi-ethnic PRS dynamic calibration method and system based on transfer learning, which integrates multi-ethnic genetic data through transfer learning technology to improve the prediction performance of PRS in the target population while reducing model development costs and data requirements.

[0009] To achieve the above object, the present invention provides the following solutions:

[0010] A multi-cluster PRS dynamic calibration method based on transfer learning, comprising:

[0011] Obtain the genomic dataset of the source population and the genomic dataset of the target population, and preprocess each of the genomic datasets to obtain the first preprocessed data corresponding to the source population and the second preprocessed data corresponding to the target population;

[0012] Input the first preprocessed data into an initial polygenic risk score model to generate a weight vector for SNP loci;

[0013] Adopt an adversarial domain adaptation algorithm to align the feature distributions of the first preprocessed data and the second preprocessed data through a discriminator network, so as to dynamically adjust the model parameters of the initial polygenic risk score model and obtain a cross-population adapted polygenic risk score model;

[0014] Screen the allele frequency vectors in the second preprocessed data and calculate the weight correction coefficient for the specific SNP loci of the target population;

[0015] Based on the weight correction coefficient, fuse the weights of shared SNPs and target population-specific SNPs to generate a dynamically calibrated polygenic risk score formula, and calculate the polygenic risk score value according to the polygenic risk score formula.

[0016] Preferably, it further includes:

[0017] Divide the second preprocessed data into a training set and a validation set, and evaluate the AUC, PPV and NPV of the calibrated polygenic risk score value through ten-fold cross-validation.

[0018] Preferably, it further includes:

[0019] When a new population dataset is input, update the cross-population adapted polygenic risk score model based on the incremental learning algorithm.

[0020] Preferably, the genomic data includes genotypes of single nucleotide polymorphism loci, allele frequencies and phenotypic labels.

[0021] Preferably, preprocessing each of the genomic datasets to obtain the first preprocessed data corresponding to the source population and the second preprocessed data corresponding to the target population includes:

[0022] Perform data quality control on each of the genomic datasets to obtain quality control data;

[0023] Perform frequency normalization and standardization on the quality control data to obtain standardized data;

[0024] Perform linkage disequilibrium pruning on the standardized data to obtain balanced data;

[0025] Align the SNP locus coordinates of the source population and the target population in the balanced data to the same reference genome version, and fill them with population-specific allele frequencies to obtain the first preprocessed data and the second preprocessed data.

[0026] Preferably, the calculation formula for the weight correction coefficient of the specific SNP locus of the target population is:

[0027]

[0028] Wherein, is the original weight of the th SNP in the source population, and the original weight comes from the initial polygenic risk score model; is the allele frequency of the th SNP in the target population; is the allele frequency of the th SNP in the source population; is the frequency difference penalty coefficient, which is used to suppress the weight contribution of SNPs with too large frequency differences, The default value of is 0.5; is the weight correction coefficient.

[0029] Preferably, the determination criteria for the allele frequency vector in the second preprocessed data include:

[0030] In the second preprocessed data, the allele frequency > 1%;

[0031] In the first preprocessed data, the genetic association p-value with the disease phenotype of the target population < 0.05.

[0032] Preferably, the polygenic risk score formula is:

[0033]

[0034] Wherein, is the polygenic risk score value, m is the total number of shared SNPs, is the genotype coding of the th shared SNP, is the total number of target population-specific SNPs, is the th effect value of the target population-specific SNP, which is calculated by independent association analysis (such as GWAS) of the target population, is the th genotype coding of the target population-specific SNP.

[0035] A multi-population PRS dynamic calibration system based on transfer learning, comprising:

[0036] A data processing unit, configured to obtain the genomic data sets of the source population and the target population, and respectively preprocess each of the genomic data sets to obtain the first preprocessed data corresponding to the source population and the second preprocessed data corresponding to the target population;

[0037] A vector determination unit, configured to train and input the first preprocessed data into an initial polygenic risk score model to generate a weight vector of SNP sites;

[0038] A migration unit, configured to adopt an adversarial domain adaptation algorithm to align the feature distributions of the first preprocessed data and the second preprocessed data through a discriminator network, so as to dynamically adjust the model parameters of the initial polygenic risk score model to obtain a polygenic risk score model adapted across populations;

[0039] A correction unit, configured to screen the allele frequency vectors in the second preprocessed data and calculate the weight correction coefficient of the specific SNP sites of the target population;

[0040] A calculation unit, configured to fuse the weights of shared SNPs and target population-specific SNPs based on the weight correction coefficient to generate a dynamically calibrated polygenic risk score formula, and calculate the polygenic risk score value according to the polygenic risk score formula.

[0041] The beneficial effects of the present invention are as follows:

[0042] (1) By dynamically aligning the feature distributions of different populations through transfer learning and adversarial domain adaptation algorithms, the present invention effectively alleviates the problem of genetic heterogeneity, enabling the prediction performance (such as AUC, PPV, NPV) of the PRS model in the target population to be close to or exceed the level of the source population.

[0043] (2) The present invention supports the dynamic adjustment of model parameters and the incremental learning mechanism. When introducing new population data, there is no need to retrain the entire model, reducing the computing power and data requirements, and achieving continuous optimization of lightweight updates and multi-population generalization capabilities.

[0044] (3) Through a standardized preprocessing process (data quality control, frequency normalization, linkage disequilibrium pruning, and genomic coordinate alignment), the present invention ensures the high quality and consistency of cross-population data, providing a reliable basis for model training.

[0045] (4) The present invention fuses the weight correction of shared SNPs and the effect values of target population-specific SNPs, combines the frequency difference penalty mechanism, balances the cross-population shared genetic signals and local specific associations, and significantly improves the accuracy of disease risk prediction and the anti-interference ability of the model. Brief Description of the Drawings

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0047] Figure 1 It is the flowchart of the method provided by the embodiment of the present invention;

[0048] Figure 2 It is the preprocessing flowchart provided by the embodiment of the present invention;

[0049] Figure 3 It is the schematic diagram of the system structure provided by the embodiment of the present invention. Detailed implementation manners

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0051] The object of the present invention is to provide a multi-ethnic PRS dynamic calibration method and system based on transfer learning. Through a dynamic calibration mechanism driven by transfer learning, an optimized data preprocessing process, and a dual-path risk score fusion strategy, high-precision prediction, cross-ethnic generalization ability, and continuous scalability of the multi-ethnic PRS model are achieved, providing an efficient and reliable genetic risk assessment tool for the precision medicine application of complex diseases.

[0052] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the drawings and specific implementation manners.

[0053] Figure 1 It is the flowchart of the method provided by the embodiment of the present invention. As Figure 1 shown, the present invention provides a multi-ethnic PRS dynamic calibration method based on transfer learning, including:

[0054] Step 100: Obtain the genomic dataset of the source ethnic group and the genomic dataset of the target ethnic group, and perform preprocessing on each genomic dataset respectively to obtain the first preprocessed data corresponding to the source ethnic group and the second preprocessed data corresponding to the target ethnic group;

[0055] Step 200: Input the first preprocessed data into the initial multi-gene risk score model for training to generate the weight vector of SNP sites;

[0056] Step 300: Adopt an adversarial domain adaptation algorithm to align the feature distributions of the first preprocessed data and the second preprocessed data through a discriminator network, so as to dynamically adjust the model parameters of the initial polygenic risk score model and obtain a cross-ethnicity adapted polygenic risk score model;

[0057] Step 400: Screen the allele frequency vectors in the second preprocessed data and calculate the weight correction coefficients of the specific SNP sites of the target ethnic group;

[0058] Step 500: Based on the weight correction coefficients, fuse the weights of the shared SNPs and the target ethnic group-specific SNPs to generate a dynamically calibrated polygenic risk score formula, and calculate the polygenic risk score value according to the polygenic risk score formula.

[0059] Specifically, in step 100 of this embodiment, first, obtain the original genomic datasets of the source ethnic group and the target ethnic group through a genomic sequencing platform, and perform data quality control (QC) on the two datasets respectively, including removing low-quality samples, SNP sites with a missing rate > 5%, and sites violating the Hardy-Weinberg equilibrium to generate quality-controlled data; secondly, perform frequency normalization (normalizing allele frequencies to the reference genome) and Z-score normalization on the quality-controlled data to eliminate the dimensional difference, and then remove highly linked redundant SNP sites through a linkage disequilibrium (LD) pruning algorithm to obtain balanced data; finally, use a genomic coordinate conversion tool (such as LiftOver) to uniformly align the SNP sites of the source ethnic group and the target ethnic group to the same reference genome version (such as GRCh38), and fill in the missing sites according to the ethnic group-specific allele frequencies, and finally output the first preprocessed data and the second preprocessed data.

[0060] Preferably, it further includes:

[0061] Divide the second preprocessed data into a training set and a validation set, and evaluate the AUC, PPV, and NPV of the calibrated polygenic risk score value through ten-fold cross-validation.

[0062] Specifically, the process of dividing the second preprocessed data into a training set and a validation set and evaluating the calibrated polygenic risk score value through ten-fold cross-validation is as follows:

[0063] First, the second preprocessed data is divided into 10 non-overlapping subsets according to the proportion of the target ethnic group disease phenotype labels using the stratified random sampling method, ensuring that the distribution of positive and negative samples in each fold is consistent with the original dataset; subsequently, ten rounds of cross-validation are performed. In each round, one fold is selected as the validation set, and the remaining nine folds are combined as the training set. The parameters of the multi-gene risk score model adapted across ethnic groups (including the discriminator weights of the adversarial domain adaptation algorithm and the SNP weight correction coefficient) are dynamically adjusted using the training set data, and the calibrated PRS value of each sample is calculated on the validation set; finally, based on the PRS value and the true phenotype label of the validation set, the AUC value is calculated through ROC curve analysis. At the same time, according to the preset risk threshold (such as the PRS median or the clinically defined critical value), the numbers of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) are counted, and the PPV (TP / (TP + FP)) and NPV (TN / (TN + FN)) are calculated respectively, and the results of the ten rounds are summarized and their mean is taken as the final performance index. In addition, to reduce the influence of randomness, the ten-fold cross-validation is repeated three times and the standard deviation of the indicators is reported to ensure the robustness of the evaluation results.

[0064] Preferably, it further includes:

[0065] When a new ethnic group dataset is input, the multi-gene risk score model adapted across ethnic groups is updated based on the incremental learning algorithm.

[0066] Optionally, in this embodiment, first, the incremental learning process is initialized, the parameters of the current cross-ethnic adaptation model are loaded (including the weights of the adversarial domain adaptation discriminator network, the shared SNP weight vector, and the ethnic group-specific correction coefficient), and the genomic dataset of the newly added ethnic group is introduced; the same preprocessing process as that of the source ethnic group and the target ethnic group is performed on this dataset (including data quality control, frequency normalization, linkage disequilibrium pruning, and reference genome alignment) to generate the third preprocessed data. Subsequently, the Elastic Weight Consolidation (EWC) algorithm is used to constrain the model update: by calculating the Fisher information matrix of the existing model parameters on the historical data, the parameters crucial for the prediction performance of the existing ethnic groups are identified, and a regularization penalty term is imposed in the incremental training to suppress the drastic changes of the key parameters and prevent catastrophic forgetting. Then, the third preprocessed data of the newly added ethnic group is merged with the historical training data subset (retaining some representative samples through random sampling) and input into the model for fine-tuning: by the adversarial domain adaptation algorithm, the feature distributions of the newly added ethnic group and the existing ethnic groups are dynamically aligned, and at the same time, using the incremental SNP weight correction module, the weight correction coefficient of the shared SNP is recalculated based on the allele frequency vector of the newly added ethnic group (a dynamic decay factor is introduced in the formula to gradually reduce the influence of the historical ethnic group frequency difference), and the specific SNP effect value of the newly added ethnic group is integrated (obtained through incremental GWAS analysis). Finally, the AUC, PPV, and NPV of the updated model on the newly added ethnic group and the historical ethnic groups are evaluated through cross-validation. If the performance meets the standard, the incremental update is completed; otherwise, iterative optimization is triggered until convergence. This process supports multiple rounds of incremental expansion to ensure a balance between lightweight update and cross-ethnic generalization ability when introducing new ethnic groups.

[0067] Preferably, the genomic data includes the genotypes, allele frequencies, and phenotypic labels of single nucleotide polymorphism sites.

[0068] Preferably, as Figure 2 shown, in this embodiment, each of the genomic datasets is preprocessed separately to obtain the first preprocessed data corresponding to the source ethnic group and the second preprocessed data corresponding to the target ethnic group, including:

[0069] Data quality control is performed on each of the genomic datasets to obtain quality control data;

[0070] Frequency normalization and standardization are performed on the quality control data to obtain standardized data;

[0071] Linkage disequilibrium pruning is performed on the standardized data to obtain balanced data;

[0072] Align the SNP locus coordinates of the source population and the target population in the balanced data to the same reference genome version, and fill them using population-specific allele frequencies to obtain the first preprocessed data and the second preprocessed data.

[0073] Specifically, the process of obtaining and processing the genomic data is as follows: First, collect the original genomic data of the source population and the target population through high-throughput sequencing technology or genotyping chip platforms, ensuring that the data contains genotype information (such as AA, AG, GG types) of single nucleotide polymorphisms (SNPs), allele frequencies (calculated based on population statistics), and phenotypic labels associated with the target disease (such as cases labeled as 1 and controls labeled as 0). The genotype data needs to be verified through quality control to ensure that the genotyping accuracy rate is higher than 99%, and the phenotypic labels need to be strictly matched with the clinical diagnosis records to avoid mislabeling. The allele frequencies are calculated by statistically analyzing the distribution proportion of the minor allele at each SNP locus in the population and are cross-validated by associating with public databases (such as the 1000 Genomes Project).

[0074] For the preprocessing process of the genomic dataset, first perform data quality control: Use bioinformatics tools (such as PLINK) to filter low-quality samples (such as inconsistent gender, sample duplication, or abnormal heterozygosity) and SNP loci (loci with a missing rate higher than 5%, a p-value of the Hardy-Weinberg equilibrium test less than 1e-6, or a minor allele frequency lower than 1% are excluded) to generate quality control data. Subsequently, perform frequency normalization on the remaining SNP loci, adjust their allele frequencies to the standard distribution of the reference genome (such as GRCh38), and apply Z-score normalization to eliminate the dimensional differences between different loci. This step ensures the numerical stability of subsequent model training.

[0075] After completing the standardization, further remove highly redundant SNP loci through the linkage disequilibrium pruning algorithm (in this embodiment, LD pruning based on a sliding window is used), set the r² threshold to 0.2, the window size to 50 SNPs, and the step size to 5 SNPs to finally generate balanced data. Finally, use a genomic coordinate conversion tool (such as UCSC LiftOver) to uniformly map the SNP loci of the source population and the target population to the same reference genome version and align the coordinate information. If there are SNP loci in a certain population that are not covered by the reference version, they are filled according to the population-specific allele frequencies to ensure the complete comparability of the data of the two populations at the genomic coordinate level, thereby outputting the first preprocessed data and the second preprocessed data.

[0076] Preferably, the calculation formula for the weight correction coefficient of the specific SNP loci of the target population is:

[0077]

[0078] Among them, is the original weight of the th SNP in the source population, and the original weight comes from the initial polygenic risk score model; is the allele frequency of the th SNP in the target population; is the allele frequency of the th SNP in the source population; is the frequency difference penalty coefficient, which is used to suppress the weight contribution of SNPs with too large frequency differences, and its default value is 0.5; is the weight correction coefficient.

[0079] Optionally, in this embodiment, the weight correction coefficient of the target population-specific SNP locus is realized by dynamically fusing the original weight of the source population and the cross-population allele frequency difference. Specifically, based on the weights of each SNP in the initial model trained with the source population, first, proportional scaling is performed according to the ratio of the allele frequencies of the target population and the source population, which directly reflects the genetic background differences between populations; second, a frequency difference penalty mechanism is introduced to apply non-linear attenuation to SNPs with frequency differences exceeding the threshold, and the weight contribution is suppressed through a preset penalty coefficient (default 0.5) to avoid model bias caused by extreme frequency differences. The creativity of this method lies in the dynamic combination of frequency ratio adjustment and difference penalty, which not only retains the information of cross-population shared genetic effects but also can adaptively weaken population-specific noise, solving the problems of over-reliance on a single frequency or ignoring local genetic heterogeneity in traditional static weight calibration, and significantly improving the robustness and generalization ability of the multi-population PRS model.

[0080] Preferably, the determination criteria for the allele frequency vector in the second preprocessed data include:

[0081] In the second preprocessed data, the allele frequency > 1%;

[0082] The p-value of the genetic association with the disease phenotype of the target population in the first preprocessed data < 0.05.

[0083] Specifically, in this embodiment, allele frequency threshold filtering is performed on the SNP sites after preprocessing of the target population group, and only the sites with a frequency higher than 1% are retained to ensure that the selected sites have population representativeness in the target population group and avoid introducing statistical noise due to low-frequency variations. Secondly, based on the established SNP-disease association results in the preprocessed data of the source population group, the sites significantly related to the phenotypes of the target population group are screened out, and it is required that the genetic association p-value in the source population group is lower than 0.05, so as to strengthen the biological consistency of the cross-population genetic effect. The creativity of this method lies in the organic combination of the population genetic characteristics of the target population group and the disease association evidence of the source population group, which not only avoids the omission of functional sites by single frequency screening, but also improves the biological interpretability of the selected SNPs through cross-population association verification, effectively solving the problem of site selection bias caused by population heterogeneity in traditional methods, thereby enhancing the scientific rigor and prediction efficiency of the multi-population PRS model.

[0084] Preferably, the polygenic risk score formula is as follows:

[0085]

[0086] Wherein, is the polygenic risk score value, m is the total number of shared SNPs, is the genotype coding of the th shared SNP, is the total number of target population group-specific SNPs, is the th target population group-specific SNP effect value, which is calculated through independent association analysis (such as GWAS) of the target population group, is the th target population group-specific SNP genotype coding.

[0087] Optionally, the polygenic risk score formula of this embodiment achieves dynamic calibration by integrating the cross-population shared genetic effect and the collaborative contribution of target population group-specific sites. Specifically, the model first performs weighted summation on the genotype coding of each shared site based on the shared SNP weights adjusted by transfer learning to reflect the genetic risk characteristics shared across populations; at the same time, specifically selected SNPs screened out in the independent analysis (such as GWAS) of the target population group are additionally introduced, and their effect values are multiplied by the genotype coding and then accumulated to capture the genetic association signals unique to the target population group. The creativity of this method lies in breaking the dependence of the traditional PRS model on a single population or static weights, balancing the shared information and population group-specific differences through a dual-path fusion mechanism, not only retaining the improvement of the cross-population generalization ability by transfer learning, but also using the local data of the target population group to strengthen the adaptability of the model to the local genetic structure, thereby significantly improving the accuracy of disease risk prediction and cross-population robustness.

[0088] Corresponding to the above method, such as Figure 3 shown, this embodiment also provides a multi-ethnic PRS dynamic calibration system based on transfer learning, including:

[0089] A data processing unit, configured to obtain a genomic dataset of a source ethnic group and a genomic dataset of a target ethnic group, and perform preprocessing on each of the genomic datasets to obtain first preprocessed data corresponding to the source ethnic group and second preprocessed data corresponding to the target ethnic group;

[0090] A vector determination unit, configured to train and input the first preprocessed data into an initial polygenic risk score model to generate a weight vector of SNP sites;

[0091] A migration unit, configured to adopt an adversarial domain adaptation algorithm to align the feature distributions of the first preprocessed data and the second preprocessed data through a discriminator network, so as to dynamically adjust the model parameters of the initial polygenic risk score model to obtain a cross-ethnicity adapted polygenic risk score model;

[0092] A correction unit, configured to screen the allele frequency vectors in the second preprocessed data and calculate the weight correction coefficient of the specific SNP sites of the target ethnic group;

[0093] A calculation unit, configured to fuse the weights of shared SNPs and target ethnic group-specific SNPs based on the weight correction coefficient to generate a dynamically calibrated polygenic risk score formula, and calculate a polygenic risk score value according to the polygenic risk score formula.

[0094] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0095] In this article, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A multi-ethnic PRS dynamic calibration method based on transfer learning, characterized in that Including: Obtain the genomic datasets of the source population and the target population, and preprocess each of the genomic datasets to obtain the first preprocessed data corresponding to the source population and the second preprocessed data corresponding to the target population; Train and input the first preprocessed data into an initial polygenic risk score model to generate a weight vector of SNP loci; Adopt an adversarial domain adaptation algorithm to align the feature distributions of the first preprocessed data and the second preprocessed data through a discriminator network, so as to dynamically adjust the model parameters of the initial polygenic risk score model to obtain a cross-population adapted polygenic risk score model; Screen the allele frequency vectors in the second preprocessed data and calculate the weight correction coefficient of the specific SNP loci of the target population; Based on the weight correction coefficient, fuse the weights of shared SNPs and target population-specific SNPs to generate a dynamically calibrated polygenic risk score formula, and calculate the polygenic risk score value according to the polygenic risk score formula; The calculation formula for the weight correction coefficient of the specific SNP loci of the target population is: ； Among them, is the original weight of the -th SNP in the source population, and the original weight is from the initial polygenic risk score model; is the allele frequency of the -th SNP in the target population; is the allele frequency of the -th SNP in the source population; is the frequency difference penalty coefficient, which is used to suppress the weight contribution of SNPs with too large frequency differences, The default value of is 0.5; is the weight correction coefficient; The polygenic risk score formula is: ； wherein, is the polygenic risk score value, m is the total number of shared SNPs, is the genotype encoding of the th shared SNP, is the total number of target population-specific SNPs, is the effect value of the th target population-specific SNP, calculated by independent association analysis of the target population, is the genotype encoding of the th target population-specific SNP.

2. The multi-ethnic PRS dynamic calibration method based on transfer learning according to claim 1, wherein Also including: Divide the second preprocessed data into a training set and a validation set, and evaluate the AUC, PPV, and NPV of the calibrated polygenic risk score value through ten-fold cross-validation.

3. The multi-ethnic PRS dynamic calibration method based on transfer learning according to claim 1, wherein Also including: When a new population dataset is input, update the cross-population adapted polygenic risk score model based on an incremental learning algorithm.

4. The multi-ethnic PRS dynamic calibration method based on transfer learning according to claim 1, wherein The genomic data includes the genotypes, allele frequencies, and phenotype labels of single nucleotide polymorphism loci.

5. The multi-ethnic PRS dynamic calibration method based on transfer learning according to claim 1, wherein Preprocess each of the genomic datasets to obtain the first preprocessed data corresponding to the source population and the second preprocessed data corresponding to the target population, including: Perform data quality control on each of the genomic datasets to obtain quality control data; Perform frequency normalization and standardization on the quality control data to obtain standardized data; Perform linkage disequilibrium pruning on the standardized data to obtain balanced data; Align the SNP locus coordinates of the source population and the target population in the balanced data to the same reference genome version, and fill them with population-specific allele frequencies to obtain the first preprocessed data and the second preprocessed data.

6. The multi-ethnic PRS dynamic calibration method based on transfer learning according to claim 1, wherein The determination criteria for the allele frequency vectors in the second preprocessed data include: Allele frequencies in the second preprocessed data > 1%; The genetic association p-value with the disease phenotype of the target population in the first preprocessed data < 0.

05.

7. A multi-ethnic PRS dynamic calibration system based on transfer learning, characterized in that, Including: A data processing unit for obtaining the genomic datasets of the source population and the target population, and preprocessing each of the genomic datasets to obtain the first preprocessed data corresponding to the source population and the second preprocessed data corresponding to the target population; A vector determination unit for training and inputting the first preprocessed data into an initial polygenic risk score model to generate a weight vector of SNP loci; A migration unit, which is used to adopt an adversarial domain adaptation algorithm to align the feature distributions of the first preprocessed data and the second preprocessed data through a discriminator network, so as to dynamically adjust the model parameters of the initial polygenic risk score model and obtain a polygenic risk score model adapted across ethnic groups; A correction unit, which is used to screen the allele frequency vectors in the second preprocessed data and calculate the weight correction coefficient of the specific SNP sites of the target ethnic group; A calculation unit, which is used to fuse the weights of shared SNPs and target ethnic group-specific SNPs based on the weight correction coefficient, generate a dynamically calibrated polygenic risk score formula, and calculate the polygenic risk score value according to the polygenic risk score formula; The calculation formula for the weight correction coefficient of the specific SNP sites of the target ethnic group is: ； Among them, is the original weight of the -th SNP in the source population, and the original weight is from the initial polygenic risk score model; is the allele frequency of the -th SNP in the target population; is the allele frequency of the -th SNP in the source population; is the frequency difference penalty coefficient, which is used to suppress the weight contribution of SNPs with too large frequency differences, has a default value of 0.5; is the weight correction coefficient; The polygenic risk score formula is: ； Among them, is the polygenic risk score value, m is the total number of shared SNPs, is the genotype encoding of the th shared SNP, is the total number of target population-specific SNPs, is the th effect value of the target population-specific SNP, calculated through independent association analysis of the target population, is the th genotype encoding of the target population-specific SNP.

Images