Method and system for biological age prediction based on multi-omics fusion and covariate fusion

Abstract

This invention discloses a method and system for predicting biological age based on multi-omics fusion and covariate fusion, belonging to the interdisciplinary field of bioinformatics and artificial intelligence. The method includes: S1, acquiring metabolomics data, metagenomics data, metabolic pathway data, and host covariate data of the subject; S2, performing multi-level feature screening on the metabolomics data, metagenomics data, and metabolic pathway data respectively, selecting a predetermined number of core features to obtain a screened multi-omics feature set; S3, performing feature-level concatenation and fusion of the screened multi-omics feature set with the host covariate data to obtain a fused feature matrix; S4, using the fused feature matrix as input, training with the AdaBoost ensemble regression algorithm to obtain a biological age prediction model; and predicting the biological age of the subject. This invention can integrate multi-dimensional omics information, effectively control host confounding factors, and has good generalization ability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics and artificial intelligence, and in particular relates to a biological age prediction method and system based on multi-omics fusion and covariate fusion. Background Technology

[0002] Biological age is a quantitative indicator that measures the degree of decline in an individual's physiological functions. Compared to chronological age, biological age more accurately reflects an individual's true level of aging and health status. Accurate prediction of biological age is of great significance for assessing the degree of aging, providing early warning of age-related diseases, and evaluating the effectiveness of anti-aging interventions. In recent years, with the development of high-throughput sequencing and omics technologies, biological age prediction methods based on biomarkers such as gut microbiota have become a research hotspot.

[0003] However, existing technologies have the following shortcomings: First, relying solely on single-dimensional genomic data (metagenomics) provides limited information. The composition, function, and metabolic interactions of the gut microbiota with the host are multi-layered biological processes, and focusing only on species abundance cannot fully reflect the impact of "microbiota-host" interactions on aging.

[0004] Second, the confounding effects of host covariates are not considered. Numerous studies have shown that body mass index (BMI) is one of the most significant host factors influencing gut microbiota composition (see Vujkovic-Cvijin et al., 2020, Nature; Wilmanski et al., 2021, Nature Aging), and gender also has a significant impact on gut microbiota (see Rothschild et al., 2018, Nature). Current techniques ignore these confounding factors, potentially leading the model to learn "BMI-related microbiota differences" rather than "age-related microbiota differences," thus generating spurious associations.

[0005] Third, using a single regression algorithm can easily lead to overfitting and insufficient generalization ability when the sample size is limited.

[0006] Fourth, it only outputs the predicted age and cannot identify specific aging-related biomarkers, making it difficult to guide precise intervention. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a biological age prediction method and system based on multi-omics fusion and covariate fusion, which can integrate multi-dimensional omics information, effectively control host confounding factors, and has good generalization ability.

[0008] To achieve the above-mentioned objectives, the first objective of this invention is to provide a biological age prediction method based on multi-omics fusion and covariate fusion, comprising: S1. Data Acquisition and Preprocessing: Acquire metabolomics data, metagenomics data, metabolic pathway data, and host covariate data of the subjects; the host covariate data includes at least body mass index and sex. S2. Multi-level feature screening: Multi-level feature screening is performed on the metabolomics data, metagenomics data and metabolic pathway data respectively, and a predetermined number of core features are selected from each omics data to obtain the screened multi-omics feature set; S3. Feature Fusion: The selected multi-omics feature set is concatenated and fused with the host covariate data at the feature level to obtain a fused feature matrix; S4. Model Training and Prediction: The fused feature matrix is ​​used as input and trained using the AdaBoost ensemble regression algorithm to obtain a biological age prediction model; the biological age prediction model is then used to predict the biological age of the subjects.

[0009] Preferably, in S2, the metabolomics data and metabolic pathway data are screened using a strategy comprising the following four levels: The first level is variance threshold filtering, which removes features whose variance is lower than the first threshold. The second level is correlation filtering, which retains features whose absolute correlation coefficient with age is greater than the second threshold and removes redundant features whose cross-correlation coefficient is greater than the third threshold. The third level is based on the feature importance ranking of random forest, retaining the feature with the highest number of features in the importance ranking; The fourth level involves a selection process based on mutual information regression, retaining a predetermined number of features mentioned before the mutual information values ​​were sorted. The metagenomics were subjected to a three-level screening strategy, including the following steps: Level 1: Variance threshold filtering, which removes only features with zero variance; The second level is correlation filtering, which retains bacterial species features with an absolute correlation coefficient of age |corr| > 0.15, while removing redundant features with a cross-correlation coefficient between features > 0.9; The third level involves random forest importance ranking, using RandomForestRegressor to evaluate the importance of the remaining features.

[0010] Preferably, in S3, a horizontal splicing method is used to splice the screened metabolomics core features, metagenomics core features, and metabolic pathway core features with BMI features and gender features as independent dimensions to form the fused feature matrix; during the node splitting process of each base learner in the AdaBoost ensemble regression algorithm, the BMI features and gender features compete equally with other features in the multi-omics feature set for the splitting threshold.

[0011] Preferably, in S4, the AdaBoost ensemble regression algorithm uses a decision tree as a base learner and iterates through multiple rounds. In each round, the sample weights are dynamically updated based on the sample prediction error, so that subsequent base learners focus on samples that are difficult to predict. Finally, the prediction results of all base learners are integrated in the form of weighted median or weighted average.

[0012] Preferably, it also includes S5, biomarker classification: based on the feature importance attributes of the trained biological age prediction model, calculate the contribution of each feature in the fusion feature matrix to age prediction; and according to the correlation direction between each feature and age, identify the corresponding biomarker as an accelerated aging biomarker or an aging-mitigating biomarker.

[0013] Preferably, in S1, the preprocessing includes: Metabolomics data preprocessing: Non-targeted metabolomics detection was performed on fecal samples to obtain raw data; peak detection, retention time alignment, feature grouping and missing peak filling were performed on the raw data, and the data was matched and identified with the metabolomics database to obtain a peak area data matrix of metabolite characteristics; Metagenomics data preprocessing: Shotgun metagenomic sequencing was performed on fecal samples to obtain raw sequencing data; the raw sequencing data underwent quality control, host sequence removal, and species classification annotation to obtain a species relative abundance matrix; and metabolic pathway annotation was performed on the raw sequencing data to obtain a metabolic pathway abundance matrix. Host covariate data processing: Collect host BMI and gender information corresponding to fecal samples, treat BMI as a continuous numerical variable, encode gender as a binary variable, and fill in missing BMI values; Sample alignment: Metabolomics data, metagenomics data, metabolic pathway data, and host covariate data are aligned by sample ID, and samples shared by all four are taken as the final analysis sample set.

[0014] Preferably, in the metabolomics data preprocessing: The non-targeted metabolomics detection was performed using liquid chromatography-tandem technology; Peak detection, retention time alignment, feature grouping, and missing peak imputation were performed using the XCMS software tool. The metabolomics databases include HMDB, METLIN, MassBank and / or PubChem; The quality error of the matching identification is ≤10ppm, and the MS / MS matching score is ≥0.7.

[0015] Preferably, S4 includes performing regression mean correction on the prediction results, and calculating the corrected aging acceleration using the following formula: ; in, This is the predicted age output by the biological age prediction model, where Age_chron is the subject's actual age. and The slope and intercept are obtained by performing least-squares linear regression on the predicted age of all samples against their actual age; AgeAccel_corrected is the corrected aging acceleration after removing the regression mean bias, with a positive value indicating accelerated actual aging of the subjects and a negative value indicating delayed actual aging.

[0016] Preferably, the biomarker typing includes quantifying the aging contribution of each biomarker using the following formula: C_j = SHAP_j · (σ_age / σ_feature_j) Where SHAP_j is the mean SHAP value of feature j, σ_age is the standard deviation of age, and σ_feature_j is the standard deviation of feature j; C_j is in years / unit feature change, representing the predicted age change corresponding to each standard deviation change of feature j.

[0017] A second objective of this invention is to provide a biological age prediction system based on multi-omics fusion and covariate fusion, comprising: Data acquisition module: Acquires metabolomics data, metagenomics data, metabolic pathway data, and host covariate data of the subjects; the host covariate data includes at least body mass index and sex; Multi-level feature screening module: Performs multi-level feature screening on the metabolomics data, metagenomics data and metabolic pathway data respectively, and selects a predetermined number of core features from each omics data to obtain the screened multi-omics feature set; Feature fusion module: performs feature-level concatenation and fusion of the selected multi-omics feature set with the host covariate data to obtain a fused feature matrix; Model training and prediction module: The fused feature matrix is ​​used as input and trained using the AdaBoost ensemble regression algorithm to obtain a biological age prediction model; and the biological age prediction model is used to predict the biological age of the subjects.

[0018] A third objective of this invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned biological age prediction method based on multi-omics fusion and covariate fusion.

[0019] A fourth objective of the present invention is to provide an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described biological age prediction method based on multi-omics fusion and covariate fusion.

[0020] The advantages and positive effects of this application are: This invention integrates metabolomics, metagenomics, and metabolic pathway data simultaneously in biological age prediction, achieving information fusion across three levels: microbial community composition, microbial community function, and microbial-host metabolic interactions. Experimental data demonstrate that the three-omics fusion model achieves five-fold cross-validation. The score reached 0.4692, outperforming any single- or dual-omics combination. After incorporating BMI and gender covariates, the 8:2 stratified test set... The accuracy reaches 0.5516, and the MAE is only 4.33 years, which far exceeds the prediction accuracy of existing single-mathematics-based methods.

[0021] This invention employs a four-level / three-level multi-level feature selection strategy, effectively reducing feature redundancy and overfitting risks. Simultaneously, it utilizes the AdaBoost ensemble learning algorithm, employing a 200-round iterative adaptive sample weight update mechanism to ensure subsequent base learners focus on difficult-to-predict samples. Experiments demonstrate that the 8:2 stratified test... (0.5516) is better than five-fold cross-validation. (0.4269) indicates that the model has excellent generalization ability on unseen data and no serious overfitting.

[0022] This invention treats BMI and gender as independent features, directly concatenating them with the selected omics features during the feature fusion stage, and participating in decision tree splitting on an equal footing with the omics features during AdaBoost model training. This design enables the model to automatically learn the microbiota-age conditional association after controlling for BMI and gender, effectively eliminating the interference of host factors on the prediction results. This is the core innovation of this invention, distinguishing it from existing technologies (such as residual methods or hierarchical modeling methods).

[0023] Based on the feature importance attribute of the model, this invention can automatically identify key aging-related biomarkers and classify them into accelerated aging biomarkers (increasing in abundance with age) and slowed aging biomarkers (decreasing in abundance with age), providing clear target information for precise anti-aging intervention. For example, *Faecalibacterium prausnitzii* was identified as a slowed aging biomarker, and *Bacteroides fragilis* as an accelerated aging biomarker.

[0024] Unlike existing technologies that only utilize gut microbiota species abundance, this invention integrates information from three dimensions: metabolite molecules, microbiota species, and metabolic pathways. This enables a multi-level characterization of aging from molecular, species, and functional perspectives, resulting in a more solid biological foundation and more interpretable prediction results. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A first flowchart of a preferred embodiment of the present invention is shown; Figure 2 A second flowchart of a preferred embodiment of the present invention is shown; Figure 3 A flowchart of a multi-level feature filtering process in a preferred embodiment of the present invention is shown; Figure 4 The flowchart of the AdaBoost regression algorithm training process in a preferred embodiment of the present invention is shown. Figure 5 A 92-dimensional fusion feature matrix structure diagram is shown in a preferred embodiment of the present invention; Figure 6 The diagram shows a scatter plot of predicted values ​​versus actual values ​​and a residual analysis plot in a preferred embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1 and Figure 2A biological age prediction method based on multi-omics fusion and covariate fusion, comprising: Data Acquisition and Preprocessing: Acquire metabolomics data, metagenomics data, metabolic pathway data, and host covariate data of the subjects; the host covariate data includes at least body mass index (BMI) and gender; Multi-level feature screening: Multi-level feature screening is performed on the metabolomics data, metagenomics data, and metabolic pathway data respectively, and a predetermined number of core features are selected from each omics data to obtain a screened multi-omics feature set; wherein, the host covariate data does not participate in the multi-level feature screening; Feature fusion: The selected multi-omics feature set is concatenated and fused with the host covariate data at the feature level to obtain a fused feature matrix; Model training and prediction: The fused feature matrix is ​​used as input and trained using the AdaBoost ensemble regression algorithm to obtain a biological age prediction model; and the biological age prediction model is used to predict the biological age of the subjects.

[0029] In a preferred embodiment of the present invention, the multi-level feature screening step employs a strategy comprising the following four levels of screening for the metabolomics data and metabolic pathway data: The first level is variance threshold filtering, which removes features whose variance is lower than the first threshold. The second level is correlation filtering, which retains features whose absolute correlation coefficient with age is greater than the second threshold and removes redundant features whose cross-correlation coefficient is greater than the third threshold. The third level is based on the feature importance ranking of random forest, retaining the feature with the highest number of features in the importance ranking; The fourth level involves a selection process based on mutual information regression, retaining a predetermined number of features mentioned before the mutual information values ​​were sorted.

[0030] In a preferred embodiment of the present invention, during the feature fusion process, a horizontal splicing method is adopted to splice the screened metabolomics core features, metagenomics core features, and metabolic pathway core features with BMI features and gender features as independent dimensions to form the fused feature matrix; during the node splitting process of each base learner in the AdaBoost ensemble regression algorithm, the BMI features and gender features compete equally with other features in the multi-omics feature set for the splitting threshold.

[0031] In a preferred embodiment of the present invention, during the model training and prediction process, the AdaBoost ensemble regression algorithm uses a decision tree as a base learner and iterates through multiple rounds. In each round, the sample weights are dynamically updated based on the sample prediction error, so that subsequent base learners focus on samples that are difficult to predict. Finally, the prediction results of all base learners are integrated in the form of weighted median or weighted average.

[0032] In a preferred embodiment of the present invention, the method further includes biomarker typing: based on the feature importance attributes of the trained biological age prediction model, the contribution of each feature in the fusion feature matrix to age prediction is calculated; and according to the correlation direction between each feature and age, the corresponding biomarker is identified as an accelerated aging biomarker or an anti-aging biomarker.

[0033] In a preferred embodiment of the present invention, the data acquisition and preprocessing further includes: Metabolomics data preprocessing: Non-targeted metabolomics detection was performed on fecal samples to obtain raw data; peak detection, retention time alignment, feature grouping and missing peak filling were performed on the raw data, and the data was matched and identified with the metabolomics database to obtain a metabolite characteristic peak area data matrix; Metagenomics data preprocessing: Shotgun metagenomic sequencing was performed on fecal samples to obtain raw sequencing data; the raw sequencing data underwent quality control, host sequence removal, and species classification annotation to obtain a species relative abundance matrix; and metabolic pathway annotation was performed on the raw sequencing data to obtain a metabolic pathway abundance matrix. Host covariate data processing: Collect host BMI and gender information corresponding to fecal samples, treat BMI as a continuous numerical variable, encode gender as a binary variable, and fill in missing BMI values; Sample alignment: Metabolomics data, metagenomics data, metabolic pathway data, and host covariate data are aligned by sample ID, and samples shared by all four are taken as the final analysis sample set.

[0034] In a preferred embodiment of the present invention, during the metabolomics data preprocessing process: the non-targeted metabolomics detection is performed using liquid chromatography-tandem (LC-MS / MS) technology; peak detection, retention time alignment, feature grouping, and missing peak filling are performed using XCMS software tools; the metabolomics database includes HMDB, METLIN, MassBank, and / or PubChem; the quality error of the matching identification is ≤10ppm, and the MS / MS matching score is ≥0.7.

[0035] In a preferred embodiment of the present invention, during the metagenomic data preprocessing: quality control is performed using the Fastp tool, including the removal of adapter sequences, low-quality bases, and reads shorter than 50 bp; host sequence removal is performed using the Bowtie2 tool, aligning the quality-controlled reads to the human reference genome Hg38 and removing the aligned host-derived sequences; species classification annotation is performed using the MetaPhlAn tool to obtain a relative abundance matrix of species; and metabolic pathway annotation is performed using the HUMAnN tool to obtain an abundance matrix of gene families and metabolic pathways.

[0036] In a preferred embodiment of the present invention, the metagenomic data preprocessing process further includes pathway filtering of the metabolic pathway abundance matrix: removing unmapped or unintegrated pathway features that begin with UNMAPPED or UNINTEGRATED; and using a centered log-ratio (CLR) transformation to handle the closure effect of the metabolic pathway data.

[0037] In a preferred embodiment of the present invention, during the host covariate data processing: the gender is encoded as: male = 1, female = 0, and missing values ​​are filled with 0; the missing BMI values ​​are filled with the sample BMI mean or median.

[0038] The core of this invention lies in the deep integration of metabolomics, metagenomics (species composition of gut microbiota), and metabolic pathway data with host BMI and gender covariates, and the achievement of high-precision biological age prediction through multi-level feature screening and AdaBoost ensemble regression algorithm.

[0039] Three-omics integration (metabolomics + metagenomics + metabolic pathways): This invention integrates metabolomics, metagenomics, and metabolic pathway data simultaneously in age prediction, achieving information fusion across three levels: microbial community composition, microbial community function, and microbial-host metabolic interactions. Each of the three omics approaches retains 30 core features after multi-level feature screening, forming a 90-dimensional multi-omics fusion feature matrix.

[0040] Multi-level feature selection strategy: Existing technologies employ a single GBRT feature selection. This invention uses a four-level screening strategy (metabolicomics and pathways): Level 1 variance threshold filtering (threshold=0.01) → Level 2 Pearson correlation filtering (|corr|>0.2) → Level 3 Random Forest importance ranking (RF, n_estimators=200, max_depth=8) → Level 4 mutual information regression selection (mutual_info_regression); metagenomics employs a three-level screening (variance threshold=0 → |corr|>0.15 → RF importance). Each omics ultimately retains 30 features with the highest predictive value, effectively reducing feature redundancy and overfitting risks.

[0041] BMI and gender were concatenated and fused as covariate features: BMI, as a continuous numerical feature (1-dimensional), and gender, as a binary feature (male = 1, female = 0, 1-dimensional), are directly concatenated with the 90-dimensional feature matrix after tri-omics screening via np.hstack to form a 92-dimensional fused feature matrix. BMI and gender do not participate in the omics feature screening process; they are added directly during the feature fusion stage to ensure that covariate information is not lost during screening. During the node splitting process of each base learner (decision tree) in AdaBoost, BMI and gender features compete equally with the other 90 omics features for the splitting threshold, enabling the model to automatically learn and control the microbiome-age conditional association after BMI and gender are selected.

[0042] AdaBoost ensemble regression algorithm: The AdaBoost ensemble learning algorithm was employed, using DecisionTreeRegressor (max_depth=3) as the base learner, iterating for 200 rounds with a learning rate of 0.1 and a linear loss function. In each iteration, the sample weights were dynamically updated based on the sample prediction error, allowing subsequent base learners to focus on difficult-to-predict samples. Finally, the predictions from all base learners were integrated using a weighted median to obtain a robust age prediction.

[0043] Accelerated / Remission Classification of Gut Aging Markers: The feature importance is calculated based on the feature_importances_ attribute of the AdaBoost model, and the identified aging-related biomarkers are divided into accelerated aging type (abundance increases with age) and slowed aging type (abundance decreases with age), providing a basis for targeted intervention.

[0044] Optimal Model: An AdaBoost regression model with 92 fusion features based on three omics (30-dimensional metabolomics + 30-dimensional metagenomics + 30-dimensional pathways) + BMI covariate (1-dimensional) + gender covariate (1-dimensional). 1. Performance of 5-fold cross-validation: CV Mean = 0.4269, CV Mean MAE = 5.37 years, CV MeanRMSE = 6.58 years; 2. 8:2 tiered performance testing: Test = 0.5516, Test MAE = 4.33 years, Test RMSE = 5.64 years; 3. Tri-omics fusion effect: AdaBoost model CV with tri-omics fusion (90 dimensions) The value reached 0.4692, which is better than any single-omics or dual-omics combination, verifying the information complementarity value of multi-omics fusion; 4. Covariate integration effect: After adding BMI + gender covariates (92 dimensions), the 8:2 test set The model has a mean squared error of 0.5516 and a mean length of effect (MAE) of only 4.33 years, demonstrating good generalization ability on unseen data. 5. Biomarker identification effect: 30 key aging-related bacterial species and metabolites were identified, including Faecalibacterium prausnitzii (anti-inflammatory bacteria, anti-aging type) and Bacteroides fragilis (pro-inflammatory bacteria, accelerating aging type).

[0045] The following is in conjunction with the appendix Figures 3 to 6 The invention provides a detailed implementation method with specific data: AdaBoost biological age prediction (optimal model) based on tri-omics fusion + BMI + gender covariate.

[0046] Step 1: Data Collection Fecal samples and host background information were collected from 71 Chinese adults. The population characteristics are as follows: Age range: 35-65 years old, average age 50.2 years old Gender distribution: 44 males and 27 females BMI range: 15.0-51.1 kg / The average BMI is 24.8 kg / kg.

[0047] Inclusion criteria: Age 30-65 years, no serious gastrointestinal diseases, and no use of antibiotics in the past 3 months. Sample type: Fecal sample (for metabolomics and metagenomics testing) Step 2: Data Detection and Preprocessing Step 1: Metabolomics Data Acquisition Untargeted metabolomics analysis (LC-MS / MS) was performed on fecal samples. Raw data were converted to mzML / mzXML format using ProteoWizardMSConvert. XCMS was used for peak detection, retention time alignment, feature grouping, and missing peak imputation. Matching and identification were performed against databases such as HMDB, METLIN, MassBank, and PubChem (quality error ≤10ppm, MS / MS matching score ≥0.7). A peak area data matrix of 16875 metabolite features was finally obtained (72 samples × 16875 features).

[0048] Step 2: Acquisition of metagenomic data Shotgun metagenomic sequencing was performed on fecal samples (Illumina NovaSeq 6000 platform, PE150). Fastp (v0.23.2) was used for quality control (removal of adapter sequences, low-quality bases, and reads <50 bp). Bowtie2 was used to align the quality-controlled clean reads to the human reference genome Hg38, removing host-derived sequences. MetaPhlAn (v3.0) was used for species taxonomic annotation, obtaining a relative abundance matrix for 1304 species. HUMAnN (v3.0) was used for metabolic pathway annotation, obtaining abundance matrices for gene families and metabolic pathways.

[0049] Step 3: Metabolic pathway data preprocessing The original HUMAnN annotation contained 23,021 pathway features. After removing unmapped / unintegrated pathways starting with UNMAPPED or UNINTEGRATED, 22,552 valid pathway features were retained. The closure effect of the constitutive data was addressed using the Centered Log-Ratio (CLR) transformation.

[0050] Step 4: Organizing Host Covariate Data Host BMI and sex information were collected corresponding to fecal samples. BMI was a continuous numerical variable (unit: kg / cm²). Gender was coded as a binary variable (male = 1, female = 0). Missing BMI values ​​were represented by the mean BMI (24.8 kg / cm²). (Fill in the blanks.) BMI and gender are not involved in the feature selection process in step 5. They are directly added as additional dimensions in the feature fusion stage to ensure that covariate information is not discarded during the selection process.

[0051] Step 5: Sample Alignment Aligning metabolomics, metagenomics, metabolic pathway, and BMI / gender data by sample ID, and taking samples shared across all four categories, a total of 71 valid samples were obtained. The age range was 35-65 years, with a mean age of 50.2 years.

[0052] Step 3: Multi-level Feature Filtering Step 6: Metabolomics Feature Screening (Four-Level Screening) Level 1: Variance Threshold Filtering: VarianceThreshold (threshold=0.01), removes low-variance metabolite features with variance below 0.01; Level 2: Relevance filtering: Calculate the Pearson correlation coefficient between each feature and age, retain features with |correlation coefficient|>0.2; remove redundant features with cross-correlation coefficients>0.9 (retain the one with a stronger correlation to age); Level 3: Random Forest Importance Ranking: RandomForestRegressor(n_estimators=200,max_depth=8, random_state=42) is used to evaluate the importance of the remaining features, and the top 60 features in importance ranking are selected (max_features×2). Level 4: Mutual Information Regression Selection: The mutual information values ​​of 60 candidate features with age are calculated using mutual_info_regression(random_state=42), and the top 30 features with the highest mutual information are selected as the final metabolomics features.

[0053] The 30 representative metabolomics features selected include: Putrescine, Psilocin, Glycylvaline, O-feruloylquinate, N-Acetylanthranilic acid, 5-Acetylamino-6-formylamino-3-methyluracil, Spermidine, Ascorbicacid, L-Carnitine, etc.

[0054] Step 7: Metagenomic Feature Screening (Three-Level Screening) Level 1: Variance Threshold Filtering: VarianceThreshold(threshold=0), only removes zero-variance features (the number of metagenomic features is already small, so the threshold is 0). Level 2: Correlation filtering: retain bacterial species features with a correlation coefficient > 0.15 with age, and remove redundant features with a cross-correlation coefficient > 0.9; Level 3: Random Forest Importance Ranking: RandomForestRegressor(n_estimators=200,max_depth=8, random_state=42) is used to evaluate the importance of the remaining features, and the top 30 bacterial species in importance ranking are taken as the final metagenomic features.

[0055] The 30 metagenomic core strains selected include: 1. *Faecalibacterium prausnitzii*: Core anti-inflammatory bacterium, butyrate-producing. 2. Bacteroides fragilis: a pro-inflammatory bacterium associated with intestinal barrier damage. 3. Bifidobacterium pseudocatenulatum: a probiotic that produces short-chain fatty acids. 4. Bacteroides thetaiotaomicron 5. Bacteroides caccae 6. Blautia wexlerae 7. Eubacterium rectale: Butyrate-producing bacteria 8. Anaerostipes hadrus (Gluttonous Anaerobic Gluttonous Bacteria) 9. Gemmiger formicilis 10. Parabacteroides distasonis 11. Dialister hominis (human Dialister bacterium) 12. *Lachnospira eligens*: Butyrate-producing bacterium, anti-inflammatory. 13. Megamonas funiformis 14. Klebsiella pneumoniae: an opportunistic pathogen. 15. Butyricimonas virosa 16. *Intestinibacter bartlettii* (Pasteurella multocida) 17. Anaerotignum faecicola (fecal anaerobic bacteria) 18. Veillonella rogosae 19. Lacrimispora amygdalina (Tonsilicone lacrimalis) 20. Mediterraneibacter butyricigenes (Mediterranean butyrate-producing bacteria) 21. Lachnospira pectinoschiza (Split pectinus) 22. Clostridium SGB48024 23. Faecalibacterium SGB15346 24. Mogibacterium diversum 25. Parasutterella SGB9260 26. Hydrogenoanaerobacterium saccharovorans (Hydrogenoanaerobic bacteria) 27. *Eisenbergiella massiliensis* 28. Mediterraneibacter sp NSJ_151 (Mediterranean fungus NSJ_151) 29. GGB2998_SGB3988 (Uncultured Clostridium SGB3988) 30. GGB9359_SGB14334 (Uncultured Clostridium SGB14334) Step 8: Metabolic pathway feature screening (four-level screening) Using the same four-level screening strategy as metabolomics (variance threshold 0.01 → |corr|>0.2 → RF importance → mutual information), 30 core pathway features were selected from 22,552 valid pathway features. Representative pathways include: fatty acid β-oxidation pathway, butyrate metabolism pathway, branched-chain amino acid degradation pathway, and tryptophan metabolism pathway.

[0056] Step 4: Feature Fusion (92-dimensional fusion matrix) Step 9: Covariate Feature Preparation BMI feature: The BMI values ​​of the 71 samples are used to construct a 1-dimensional feature vector (71×1 matrix), and missing values ​​are filled with the mean of 24.8.

[0057] Gender features: Gender is encoded as a binary variable to form a 1-dimensional feature vector (71×1 matrix), male=1, female=0, and missing values ​​are filled with 0.

[0058] Step 10: Feature splicing and fusion (corresponding code: np.hstack) The selected omics features and covariate features are concatenated into a unified 92-dimensional fusion matrix using np.hstack (horizontal concatenation). X_fused = np.hstack([X_metabolomics(71×30), X_metagenomics(71×30),X_pathway(71×30), bmi_feature(71×1), gender_feature(71×1)]) The final result is a two-dimensional matrix X_fused = (71 samples, 92 features).

[0059] Explanation of BMI covariate control mechanism: BMI is used as the 91st feature in model training. During node splitting in each base learner (DecisionTreeRegressor, max_depth=3) of AdaBoost, the BMI feature competes equally with the other 90 omics features and one gender feature for the splitting threshold. This design is based on the following scientific principles: Vujkovic-Cvijin et al. (2020, Nature) confirmed in a large-scale cohort study of over 3400 people that BMI is one of the host factors most strongly associated with gut microbiota composition; Wilmanski et al. (2021, Nature Aging) found that the association between multiple gut microbiota species and age only became significant after controlling for BMI.

[0060] Therefore, by directly incorporating BMI into the feature matrix instead of performing residual regression first, AdaBoost can automatically learn at each split node: for a specific range of bacterial abundance, how much effect age variation has at that BMI level, i.e., learning the bacterial community-age conditional association with BMI as a conditional variable. This is the core innovation of this invention, distinguishing it from existing technologies (such as residual methods or hierarchical modeling methods).

[0061] Gender covariate control mechanism: Rothschild et al. (2018, Nature) found that while the effect of sex on gut microbiota composition was weaker than that of BMI, it was still significant. Sex, as the 92nd feature, participated in feature assembly and model training in the same way, enabling the model to automatically learn the differences in gut microbiota-age associations between different sexes.

[0062] Step 5: Model Training and Evaluation Step 11: Data Standardization The feature matrix is ​​standardized using Z-score (zero mean, unit variance) using StandardScaler. The standardization formula is: x_scaled = (x - μ_train) / σ_train. In cross-validation, each fold is independently fitted on the training set and transformed on the validation set. In the 8:2 stratified split, the data is fitted on the training set, and the mean and standard deviation of the training set are used to transform the test set data to avoid data leakage.

[0063] Step 12: AdaBoost Model Training AdaBoost (Adaptive Boosting) regression algorithm is an iterative ensemble learning method based on adaptive updates of sample weights. The training process is as follows: (1) Initialization: Each sample is assigned the same weight w_i = 1 / N (N=71 is the number of samples); (2) Iterative training (200 rounds in total, t=1,2,...,200): a. Train a base learner based on the current sample weight distribution: DecisionTreeRegressor(max_depth=3); b. Calculate the predictions of the base learner on all samples. ; c. Calculate the error for each sample: ,in This represents the maximum error. d. Calculate the weighted error: E_t = Σ w_i × e_i; e. Calculate the weights of the base learner: β_t = E_t / (1 - E_t); f. Update sample weights: w_i ← w_i × β_t^(1-e_i), increase the weight of samples with large prediction errors; g. Normalized weights: w_i ← w_i / Σ w_j; (3) Final prediction (loss function is linear): , where h_t(x) is the predicted value of the t-th base learner; (4) Learning rate shrinkage: The contribution of each base learner is multiplied by learning_rate=0.1 to prevent overfitting.

[0064] Complete parameter settings for the AdaBoost model (corresponding to age_prediction_v6.py L264): model = AdaBoostRegressor(n_estimators = 200), number of base learners / number of iterations learning_rate = 0.1, the learning rate, controls the contribution shrinkage coefficient of the base learner in each round. `estimator = DecisionTreeRegressor(max_depth=3)`, base learner (default value) loss = "linear", linear loss function Step 13: Dual-strategy evaluation Strategy 1: Five-fold cross-validation (key evaluation metric): KFold(n_splits=5, shuffle=True, random_state=42) randomly shuffles 71 samples and divides them into 5 equal parts. Each part is used as the validation set in turn, and the remaining 4 parts are used as the training set. This process is repeated 5 times. The mean and standard deviation of MAE and RMSE are used. This strategy is more stable and reliable when the sample size is small (N=71), and is the main evaluation method of this invention.

[0065] Strategy 2: 8:2 stratified independent testing (auxiliary evaluation indicators): StratifiedShuffleSplit(test_size=0.2, random_state=42) performs stratified sampling by dividing the data into three strata based on age quantiles. The training set contains 56 samples, and the test set contains 15 samples, ensuring that the age distribution of the training and test sets is consistent.

[0066] Step Six: Predicted Output and Validation (Best Model Results) Step 14: Age Prediction Output Final feature dimensions: 92 dimensions (30 metabolomics + 30 metagenomics + 30 metabolic pathways + 1 BMI + 1 gender) Optimal model: AdaBoost + tri-omics fusion + BMI covariate + gender covariate Five-fold cross-validation: CV Mean = 0.4269, CV Mean MAE = 5.37 years, CV Mean RMSE = 6.58 years 8:2 tiered testing: Test = 0.5516, Test MAE = 4.33 years, Test RMSE = 5.64 years Table 1 is a summary table of complete performance.

[0067] Comparison of predicted values ​​and actual ages for 15 samples in the test set: Actual age: [53, 39, 56, 46, 60, 55, 56, 41, 62, 59, 57, 42, 57, 42, 36] Predicted age: [54.7, 49.7, 52.3, 54.0, 56.4, 55.5, 54.9, 48.9, 56.1, 57.4, 57.0, 48.6, 55.8, 40.5, 47.0] Step 14A: Calculation of Aging Acceleration Corrected by Regression Mean In the field of biological age prediction, ΔAge is directly used to predict age. The use of actual age to determine whether aging is accelerated or delayed suffers from the classic regression to the mean problem: predictive models naturally tend to overestimate the age of younger individuals and underestimate the age of older individuals, causing a systematic shift in ΔAge across different age groups. This results in younger individuals being incorrectly labeled as "accelerated aging" and older individuals being incorrectly labeled as "delayed aging." To address this issue, this invention proposes a regression mean-corrected aging acceleration (AgeAccel_corrected), calculated using the following formula:

[0068] in, Age_chron represents the predicted age of the subject by the biological age prediction model, while Age_chron represents the subject's actual age. and The slope and intercept are obtained by performing ordinary least squares (OLS) linear regression on the predicted age and actual age of all 71 samples. The residual AgeAccel_corrected is the corrected aging acceleration after removing the regression mean bias. It measures "the degree to which the physiological age of a subject is higher or lower than the expected value of their peers, given the subject's actual age". If AgeAccel_corrected > 0, it indicates that the subject's physiological aging rate is faster than the average level of their peers (accelerated real aging); if AgeAccel_corrected < 0, it indicates that the subject's physiological aging rate is slower than the average level of their peers (delayed real aging); if AgeAccel_corrected ≈ 0, it indicates that the subject is at the average aging level of their peers.

[0069] In this embodiment, OLS linear regression was performed on the predicted age and actual age of all 71 samples to obtain... = 0.8093, = 9.9478 (R = 0.9569, Before correction, the original ΔAge was significantly negatively correlated with actual age. This means that younger individuals are systematically overestimated and older individuals are systematically underestimated; after correction, the correlation coefficient between AgeAccel_corrected and actual age decreases to r ≈ 0.0000 (p = 1.00), indicating that the regression mean bias has been completely eliminated.

[0070] Validation by age group: For the 35-45 age group (n=25), the original mean ΔAge was +2.18 years, and the corrected mean AgeAccel_corrected was [missing value]. 0.28 years; the original ΔAge mean for the 46-55 age group (n=18) was +0.98 years, adjusted to +0.80 years; the original ΔAge mean for the 56-65 age group (n=28) was... The original value was 1.63 years, which was corrected to 0.27 years. This shows that ΔAge exhibited a systematic shift in all age groups before correction, while the corrected values ​​for all age groups approached zero, validating the effectiveness of AgeAccel_corrected.

[0071] Taking the first sample in the test set in step 14 as an example: actual age 53 years, original predicted age 54.71 years, original ΔAge = +1.71 years; after regression mean correction, AgeAccel_corrected = 54.71. (0.8093×53 +9.9478) = 54.71 52.84 = +1.87 years. The difference before and after correction indicates that the original ΔAge (+1.71 years) slightly underestimates the true aging acceleration (+1.87 years) of this sample. Compared to directly using ΔAge, AgeAccel_corrected provides a more robust and comparable assessment of aging acceleration / delay, unaffected by the age distribution of the subjects, and suitable for cross-sectional comparisons and longitudinal tracking across different age groups.

[0072] Step 15: Model Validation (1) Residual analysis: The mean of the prediction residual (actual age - predicted age) is about 0 years (each fold in the 5-fold CV is calculated independently), indicating that the model has no systematic overestimation or underestimation; (2) Generalization ability: 8:2 stratified test (0.5516) and CV (0.4269) is on the same order of magnitude, and the test MAE (4.33 years) is better than the CV MAE (5.37 years), indicating that the model does not have serious overfitting on the unseen data and generalizes well; (3) Verification of tri-omics fusion: CV of tri-omics fusion (90 dimensions) (0.4692) is superior to any single-omics (except metabolomics 0.4825) or dual-omics combination, and the test results after adding BMI and sex are also superior. The improvement from 0.5125 to 0.5516 and the decrease in TestMAE from 4.61 years to 4.33 years indicate that multi-omics fusion and covariate control jointly improved model performance.

[0073] Step 7: Identification and Classification of Intestinal Aging Biomarkers Step 16: Marker Identification Based on Model Feature Importance Using the `feature_importances_` attribute of the AdaBoost model, the contribution of each input feature in the 92-dimensional feature matrix to age prediction is calculated. Based on the correlation and importance of features with age, the markers are divided into the following two categories: Biomarkers that accelerate aging (increase in abundance with age, promoting the aging process): 1. Bacteroides fragilis: a pro-inflammatory bacterium associated with intestinal barrier damage and systemic inflammation, with increased abundance in older individuals; 2. Klebsiella pneumoniae: an opportunistic pathogen, its abundance increases in older individuals and is associated with an increased risk of opportunistic infections; 3. Parabacteroides distasonis: Associated with altered bile acid metabolism, its abundance increases in older individuals; 4. Putrescine: A polyamine metabolite, high concentrations of which are associated with cytotoxicity and inflammation, and its levels are elevated in older individuals; 5. Megamonas funiformis: a gas-producing bacterium associated with intestinal bloating and decreased digestive function.

[0074] Anti-aging biomarkers (whose abundance decreases with age and have a protective effect): 1. Faecalibacterium prausnitzii: a core anti-inflammatory bacterium that produces butyrate and maintains the integrity of the intestinal barrier. Its abundance decreases in older individuals and is closely associated with inflammatory aging (Ghosh et al., 2020, Gut). 2. Bifidobacterium pseudocatenulatum: a probiotic that produces short-chain fatty acids, with reduced abundance in older individuals; 3. Lachnospira eligens: Butyrate-producing bacteria with anti-inflammatory effects, its abundance decreases in older individuals; 4. Gemmiger formicilis: Associated with a healthy gut microbiota, its abundance decreases in older individuals; 5. Spermidine: a polyamine metabolite with autophagy-inducing and anti-aging effects, its levels decrease in older individuals; 6. Ascorbic acid (vitamin C): An antioxidant, its levels decrease in older individuals and are associated with increased oxidative stress.

[0075] Step 16A: Measuring the contribution of biomarkers to aging To further quantify the specific contribution of each biomarker identified in step 16 to age prediction, this invention uses SHAP (SHapley Additive exPlanations) values ​​combined with scaling transformation to define the Biomarker Aging Contribution (BAC) index, calculated as follows: C_j = SHAP_j · (σ_age / σ_feature_j) Where SHAP_j is the mean absolute SHAP value of feature j across all 71 samples, σ_age is the standard deviation of the age label, and σ_feature_j is the standardized standard deviation of feature j. C_j is measured in years per unit of feature change. The physical meaning of C_j is: with other features remaining constant, a change of 1 standard deviation in feature j is expected to cause a change in predicted age of C_j years. Positive values ​​indicate that increased abundance of the biomarker is associated with increased predicted age (accelerated aging), while negative values ​​indicate that decreased abundance is associated with increased predicted age (mitigated aging).

[0076] Based on the above formula, the KernelExplainer was used to calculate the SHAP value for all 71 samples, yielding the BAC value for each feature. The top 5 biomarkers ranked by BAC value and their contributions are as follows: The top-ranked metabolomics feature (Metab_11) has a BAC of 5.41 years / SD, indicating that a 1 standard deviation change in this metabolite predicts an age change of approximately 5.41 years; the second-ranked metabolomics feature (Metab_16) has a BAC of 3.33 years / SD; the third-ranked metabolomics feature (Metab_26) has a BAC of 2.74 years / SD; the fourth-ranked metabolomics feature (Metab_15) has a BAC of 2.23 years / SD; and the fifth-ranked metabolomics feature (Metab_25) has a BAC of 1.76 years / SD. Among metagenomic features, the highest-ranked bacterial species feature (Meta_8) has a BAC of 1.42 years / SD, showing a negative correlation with age (aging-mitigating type). Among metabolic pathway features, the highest-ranking pathway feature (Path_14) has a BAC of 1.40 years / SD.

[0077] The BAC value for the BMI feature was 0.45 years / SD, and the BAC value for the gender feature was 0.30 years / SD, both significantly lower than the BAC values ​​for the core omics features, further validating that the model predictions are mainly driven by omics features rather than by covariate confounding.

[0078] The introduction of BAC values ​​upgrades biomarkers from qualitative classification to quantitative assessment, enabling this invention not only to identify which biomarkers are related to aging, but also to accurately quantify the aging contribution of each biomarker, providing data support for prioritizing precise interventions.

[0079] Step 17: Confounding analysis of biomarkers and BMI To verify the effectiveness of BMI covariate control, a BMI confounding analysis was performed on the identified biomarkers: (1) Calculate the partial correlation coefficient between each biomarker and BMI (after controlling for age); (2) Compare the changes in importance of the same biomarker in the model without BMI covariates (90 dimensions) and the model with BMI covariates (92 dimensions); (3) If a biomarker is highly important in the model without BMI but its importance decreases significantly in the model with BMI, then the aging association of the biomarker is actually driven by BMI and is a spurious association. (4) If the importance of a certain marker does not change much in the two models, then the association between the marker and age is not confounded by BMI and is a real aging signal.

[0080] A biological age prediction system based on multi-omics fusion and covariate fusion includes: Data acquisition module: used to acquire metabolomics data, metagenomics data, metabolic pathway data, and host covariate data including at least BMI and gender from the subjects; Multi-level feature screening module: used to perform multi-level feature screening on the metabolomics data, metagenomics data and metabolic pathway data respectively, and to screen out a predetermined number of core features from each omics data to obtain a screened multi-omics feature set; wherein, the host covariate data does not participate in the multi-level feature screening; Feature fusion module: used to perform feature-level concatenation and fusion of the selected multi-omics feature set with the host covariate data to obtain a fused feature matrix; Model training and prediction module: This module takes the fused feature matrix as input, trains it using the AdaBoost ensemble regression algorithm to obtain a biological age prediction model, and uses the biological age prediction model to predict the biological age of the subjects.

[0081] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned biological age prediction method based on multi-omics fusion and covariate fusion.

[0082] An electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described biological age prediction method based on multi-omics fusion and covariate fusion.

[0083] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented, in whole or in part, as a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line, or wireless (e.g., infrared, wireless, microwave, etc.) means). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0084] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A biological age prediction method based on multi-omics fusion and covariate fusion, characterized in that, include: S1. Data Acquisition and Preprocessing: Acquire metabolomics data, metagenomics data, metabolic pathway data, and host covariate data of the subjects; the host covariate data includes at least body mass index and sex. S2. Multi-level feature screening: Multi-level feature screening is performed on the metabolomics data, metagenomics data and metabolic pathway data respectively, and a predetermined number of core features are selected from each omics data to obtain the screened multi-omics feature set; S3. Feature Fusion: The selected multi-omics feature set is concatenated and fused with the host covariate data at the feature level to obtain a fused feature matrix; S4. Model Training and Prediction: The fused feature matrix is ​​used as input and trained using the AdaBoost ensemble regression algorithm to obtain a biological age prediction model; the biological age prediction model is then used to predict the biological age of the subjects. 2.The method of Claim 1, wherein, In S2, the metabolomics data and metabolic pathway data are screened using a four-level screening strategy: The first level is variance threshold filtering, which removes features whose variance is lower than the first threshold. The second level is correlation filtering, which retains features whose absolute correlation coefficient with age is greater than the second threshold and removes redundant features whose cross-correlation coefficient is greater than the third threshold. The third level is based on the feature importance ranking of random forest, retaining the feature with the highest number of features in the importance ranking; The fourth level involves a selection process based on mutual information regression, retaining a predetermined number of features mentioned before the mutual information values ​​were sorted. The metagenomics were subjected to a three-level screening strategy, including the following: Level 1: Variance threshold filtering, which removes only features with zero variance; The second level is correlation filtering, which retains bacterial species features with an absolute correlation coefficient of age |corr| > 0.15, while removing redundant features with a cross-correlation coefficient between features > 0.9; The third level involves random forest importance ranking, using RandomForestRegressor to evaluate the importance of the remaining features. 3.The method of Claim 1, wherein, In S3, a horizontal splicing method is used to splice the screened metabolomics core features, metagenomics core features, and metabolic pathway core features with BMI features and gender features as independent dimensions to form the fused feature matrix. During the node splitting process of each base learner in the AdaBoost ensemble regression algorithm, the BMI features and gender features compete equally with other features in the multi-omics feature set for the splitting threshold. 4.The method of Claim 1, wherein, In S4, the AdaBoost ensemble regression algorithm uses decision trees as base learners and iterates through multiple rounds. In each round, the sample weights are dynamically updated based on the sample prediction error, so that subsequent base learners focus on samples that are difficult to predict. Finally, the prediction results of all base learners are integrated in the form of weighted median or weighted average. 5.The method of Claim 1, wherein, It also includes S5, biomarker classification: based on the feature importance attributes of the trained biological age prediction model, calculate the contribution of each feature in the fusion feature matrix to age prediction; and according to the correlation direction between each feature and age, identify the corresponding biomarker as an accelerated aging biomarker or an aging-mitigating biomarker. 6.The method of Claim 1, wherein, In S1, preprocessing includes: Metabolomics data preprocessing: Non-targeted metabolomics detection was performed on fecal samples to obtain raw data; peak detection, retention time alignment, feature grouping and missing peak filling were performed on the raw data, and the data was matched and identified with the metabolomics database to obtain a peak area data matrix of metabolite characteristics; Metagenomics data preprocessing: Shotgun metagenomic sequencing was performed on fecal samples to obtain raw sequencing data; the raw sequencing data underwent quality control, host sequence removal, and species classification annotation to obtain a species relative abundance matrix; and metabolic pathway annotation was performed on the raw sequencing data to obtain a metabolic pathway abundance matrix. Host covariate data processing: Collect host BMI and gender information corresponding to fecal samples, treat BMI as a continuous numerical variable, encode gender as a binary variable, and fill in missing BMI values; Sample alignment: Metabolomics data, metagenomics data, metabolic pathway data, and host covariate data are aligned by sample ID, and samples shared by all four are taken as the final analysis sample set.

7. The method of Claim 6, wherein the method is based on multi-omics fusion and covariate fusion for biological age prediction. In the preprocessing of the metabolomics data: The non-targeted metabolomics detection was performed using liquid chromatography-tandem technology; Peak detection, retention time alignment, feature grouping, and missing peak imputation were performed using the XCMS software tool. The metabolomics databases include HMDB, METLIN, MassBank and / or PubChem; The quality error of the matching identification is ≤10ppm, and the MS / MS matching score is ≥0.

7. 8.The method of Claim 1, wherein, S4 includes regression mean correction of the prediction results, using the following formula to calculate the corrected aging acceleration: ； in, This is the predicted age output by the biological age prediction model, where Age_chron is the subject's actual age. and The slope and intercept are obtained by performing least-squares linear regression on the predicted age of all samples against their actual age; AgeAccel_corrected is the corrected aging acceleration after removing the regression mean bias, with a positive value indicating accelerated actual aging of the subjects and a negative value indicating delayed actual aging.

9. The biological age prediction method based on multi-omics fusion and covariate fusion according to claim 5, characterized in that, The biomarker genotyping includes quantifying the aging contribution of each biomarker using the following formula: C_j = SHAP_j · (σ_age / σ_feature_j) Where SHAP_j is the mean SHAP value of feature j, σ_age is the standard deviation of age, and σ_feature_j is the standard deviation of feature j; C_j is in years / unit feature change, representing the predicted age change corresponding to each standard deviation change of feature j. 10.A biological age prediction system based on multi-omics fusion and covariate fusion, characterized in that, include: Data acquisition module: Acquires metabolomics data, metagenomics data, metabolic pathway data, and host covariate data of the subjects; the host covariate data includes at least body mass index and sex; Multi-level feature screening module: Performs multi-level feature screening on the metabolomics data, metagenomics data and metabolic pathway data respectively, and selects a predetermined number of core features from each omics data to obtain the screened multi-omics feature set; Feature fusion module: performs feature-level concatenation and fusion of the selected multi-omics feature set with the host covariate data to obtain a fused feature matrix; Model training and prediction module: The fused feature matrix is ​​used as input and trained using the AdaBoost ensemble regression algorithm to obtain a biological age prediction model; and the biological age prediction model is used to predict the biological age of the subjects.

Images