A long-life digital twin construction method and system based on multi-omics data and deep learning

Abstract

The present application relates to the cross field of bioinformatics, artificial intelligence and precision medicine, and particularly relates to a longevity digital twin construction method and system based on multi-omics data and deep learning. The method comprises: obtaining multi-omics data and clinical variable data, and generating a unified input data set after preprocessing; extracting an omics core feature set; then performing multi-label joint training using time sequence age and biological indicators to obtain genomics, proteomics and metabolomics age, respectively, to form a single-omics age set; introducing an attention fusion network to dynamically allocate the contribution weight of each age and clinical variable score, and obtain a comprehensive biological age through weighted fusion; finally, a deep survival analysis model is used to model survival risk and calculate the probability of surviving to 100 years old to generate a longevity digital twin result. The present application significantly improves the accuracy and interpretability of individualized biological age assessment, and provides a reliable digital twin tool for longevity prediction and health management.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of bioinformatics, artificial intelligence and precision medicine, and in particular to a method and system for constructing longevity digital twins based on multi-omics data and deep learning. Background Technology

[0002] In the interdisciplinary field of bioinformatics, artificial intelligence and precision medicine, existing solutions typically involve collecting, preprocessing, extracting features and predicting models based on genomics data, proteomics data, metabolomics data and clinical variable data. These solutions suffer from limitations such as unstable construction of a unified input dataset, discontinuous generation pathways for comprehensive biological age, and fragmented output links for longevity digital twin results.

[0003] Existing technologies can be mainly categorized as follows: The first category consists of biological age models based on single omics, such as DNA methylation clocks and proteomics clocks. However, a single omics model can only reflect one aspect of aging and cannot comprehensively cover the multidimensional molecular mechanisms of aging. The second category comprises biological age assessment methods based on body composition data. While these achieve non-invasive detection, they struggle to delve into the molecular mechanisms, limiting predictive accuracy. The third category consists of multi-omics fusion models, but these often employ fixed weight allocations (such as simple weighted summation), failing to adapt to the dynamic differences in the contributions of various omics across different life stages and health states. The fourth category comprises aging assessment methods based on machine learning, primarily focusing on assessing the current state and lacking the ability to predict future longevity potential. In summary, existing technologies suffer from technical deficiencies such as incomplete information, static weight allocation, lack of longevity prediction capabilities, and insufficient interpretability.

[0004] Furthermore, existing methods often rely on single-omics modeling, fixed-weight fusion, or simple combinations of clinical variable data and model output. When inputting the single-omics age set and the clinical variable data into attention fusion networks and deep survival analysis models to achieve longevity digital twin results, distortions in dynamic contribution weight allocation and decoupling from the target age probability estimation are likely to occur, making it difficult to achieve stable longevity digital twin results.

[0005] For the joint processing of the single-omics age set and the clinical variable data input attention fusion network and deep survival analysis model, existing technologies generally have disconnects in the linking of heterogeneous multi-omics data, the synergy of comprehensive biological age and clinical variable data, and the modeling of survival risk and the probability estimation of target age. It is difficult to form a consistent process of data collection, preprocessing, feature extraction, dynamic contribution weight allocation, survival risk modeling and record retrieval in the longevity digital twin construction scenario, resulting in a lack of stable correspondence between comprehensive biological age and longevity digital twin results. Summary of the Invention

[0006] This invention provides a method and system for constructing a longevity digital twin based on multi-omics data and deep learning, in order to solve the problem of how to input the single-omics age set and the clinical variable data into an attention fusion network and a deep survival analysis model to achieve longevity digital twin results.

[0007] To address the aforementioned technical problems, this invention provides a method for constructing long-lived digital twins based on multi-omics data and deep learning, comprising: S100. Obtain multi-omics data and clinical variable data to obtain a unified input dataset; S200. Based on the unified input dataset, perform principal component analysis and stacked autoencoder dimensionality reduction to extract core features and obtain the omics core feature set. S300. Based on the core feature set of omics, perform multi-label joint training of time-series age and biological indicators to obtain a single-omics age set. S400. Based on the single-omics age set, perform dynamic contribution weight allocation and weighted fusion processing to obtain the comprehensive biological age; S500. Based on the comprehensive biological age, perform survival risk modeling and target age probability estimation to obtain longevity digital twin results.

[0008] Furthermore, the multi-omics data and clinical variable data include: The multi-omics data includes genomics data, proteomics data, and metabolomics data; The clinical variable data include gender, age, smoking status, alcohol consumption, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes.

[0009] Furthermore, the process of acquiring multi-omics data and clinical variable data to obtain a unified input dataset includes: After acquisition, preprocessing is performed, which includes quality control processing, standardization processing, missing value imputation processing, and batch correction processing. The quality control process includes performing stratified screening on the four types of data after access, and deleting records with insufficient sequencing depth, excessively high site missing ratio, abnormal sample duplication rate, abnormal intra-batch fluctuation, abnormal retention time drift, or conflicting value boundaries. The standardization process includes performing Z-score standardization on the data that has passed quality control. The missing value imputation process includes performing K-nearest neighbor imputation on the standardized data and simultaneously retaining the missing label field during the imputation process; The batch correction process includes performing ComBat batch correction on genomics, proteomics, and metabolomics data with missing value imputation, and performing consistency processing on clinical variable data according to collection time and source system; The unified input dataset includes sample number, collection time, batch label, quality control status field, missing marker field, and data version field.

[0010] Furthermore, the process of performing principal component analysis and stacked autoencoder dimensionality reduction to extract core features includes: The principal component analysis includes: splitting the genomics data subset, proteomics data subset, and metabolomics data subset from the unified input dataset, performing principal component analysis on each subset, retaining principal components according to the rule that the cumulative variance contribution rate is not less than 85%, and outputting the principal component score matrix. The stacked autoencoder for dimensionality reduction and core feature extraction includes: The principal component score matrix is ​​input into a stack-based autoencoder; The stacked autoencoder includes a multi-layer encoding structure and a multi-layer decoding structure, compressing each data subset layer by layer to output a compressed representation with a fixed dimension. The stacked autoencoder also includes dimension verification, which unifies the output dimension into 32 dimensions to generate genomics core feature vector, proteomics core feature vector and metabolomics core feature vector. The three core feature vectors mentioned above are encapsulated together with the sample number, batch label, collection time, and data version to generate the omics core feature set.

[0011] Furthermore, the process of multi-label joint training of time-series age and biological indicators includes: The multi-label joint training includes: constructing a genomics age network, a proteomics age network, and a metabolomics age network, respectively. Each network is a deep neural network and includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The genomic age network is trained using time-series age labels, receives the genomic core feature vector from the omics core feature set, and outputs the genomic biological age. The proteomics age network is trained using a combination of time-series age labels and inflammatory marker labels. The inflammatory marker labels include interleukin-6, tumor necrosis factor-α, and p16INK4a. It receives the core feature vector of proteomics and outputs the biological age of proteomics. The metabolomics age network is trained using a combination of time-series age labels and metabolic indicator labels. The metabolic indicator labels include glucose, fatty acids, glutathione, and malondialdehyde, receive a metabolomics core feature vector, and output metabolomics biological age. The three biological ages mentioned above are combined with the sample number, batch label, collection time, and data version to generate a single-omics age set.

[0012] Furthermore, the process of dynamically allocating contribution weights includes: Acquire clinical variable data and generate clinical variable scores using a logistic regression model; The genomic biological age, proteomics biological age, and metabolomics biological age in the single-omics age set are concatenated with the clinical variable scores to form a fusion input vector; The fused input vector is input into the attention fusion network, which includes an input layer, a weight calculation layer, a normalization processing layer, and an output layer. The weight calculation layer generates four initial weight values ​​based on the current sample, and the normalization processing layer converts the initial weight values ​​into dynamic contribution weights that sum to 1. The dynamic contribution weights change synchronously with the current sample's omics age characterization and clinical variable scores, and different samples correspond to different weight distributions.

[0013] Furthermore, the weighted fusion process includes: The genomic biological age, proteomics biological age, metabolomics biological age, and clinical variable scores are weighted and summed according to dynamic contribution weights to output the comprehensive biological age; at the same time, an attention weight heatmap is generated as the contribution distribution result.

[0014] Furthermore, the survival risk modeling process includes: Obtain clinical variable data corresponding to the same sample number. The clinical variable data includes gender, age, smoking status, drinking status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes. The data, which integrate biological age and clinical variables, are input into the DeepSurv model, which includes an input receiving layer, a risk characterization layer, a risk association layer, and an output layer.

[0015] Furthermore, the process of estimating the target age probability includes: The probability estimation unit, based on the risk status record and baseline risk parameters output by the DeepSurv model, continuously extrapolates from the actual age of the current sample to the target age of 100 years, generating the survival probability at the target age. The results of longevity digital twin are encapsulated together, including biological age, survival probability at target age, sample number, batch label, collection time, data version, model version, and contribution distribution.

[0016] Furthermore, a longevity digital twin construction system based on multi-omics data and deep learning includes: a data acquisition module, a preprocessing module, a feature extraction module, a single-omics age module, a dynamic fusion module, a survival prediction module, and a log and version management module; the modules are connected in sequence to implement the method described in any of the above-mentioned methods.

[0017] The key innovations of this invention include: (1) Genomics data, proteomics data, metabolomics data and clinical variable data are subjected to quality control, standardization, missing value imputation and batch correction to generate a unified input dataset. The unified input dataset is then input into principal component analysis and stacked autoencoder to extract genomics core feature vectors, proteomics core feature vectors and metabolomics core feature vectors. The core feature sets of omics are then input into genomics age network, proteomics age network and metabolomics age network respectively to generate single omics age sets.

[0018] (2) Input the single-omics age set and the clinical variable data into the attention fusion network, perform dynamic contribution weight allocation and weighted fusion processing to generate a comprehensive biological age, wherein the input objects of the attention fusion network are genomic biological age, proteomics biological age, metabolomics biological age and clinical variable score.

[0019] (3) Input the comprehensive biological age and the clinical variable data into the deep survival analysis model, perform survival risk modeling and target age probability estimation, generate longevity digital twin results, and make the comprehensive biological age a core input object in the deep survival analysis model to participate in the target age survival probability output.

[0020] The following are its main beneficial effects: (1) In view of the problems of scattered processing links, inconsistent input objects, and difficulty in connecting single omics results to subsequent modeling in the existing schemes, the present invention makes consistent transfer relationships between data objects in the feature extraction and age generation stages by unifying the continuous organization of input datasets, core omics feature sets and single omics age sets, thus forming a continuous processing link from multi-omics data to single omics age sets.

[0021] (2) In view of the problem that fixed weight fusion, simple splicing or static combination in the existing schemes are difficult to reflect the differences in genomic biological age, proteomics biological age, metabolomics biological age and clinical variable scores among different individuals, the present invention performs dynamic contribution weight allocation and weighted fusion processing through attention fusion network, so that the comprehensive biological age is based on the joint input of the single omics age set and the clinical variable data, thereby forming a dynamic weight link corresponding to the current sample in the fusion process, and preserving the corresponding relationship for the generation of subsequent contribution distribution results.

[0022] (3) In view of the lack of direct connection between biological age results and future outcome prediction in existing schemes, and the separation of survival risk modeling and longevity prediction links, this invention inputs comprehensive biological age and clinical variable data into a deep survival analysis model, performs survival risk modeling and target age probability estimation, so that the comprehensive biological age formed in the previous steps directly enters the longevity digital twin result generation link, and completes the output from comprehensive biological age to target age survival probability in the same processing flow.

[0023] (4) This invention achieves a leap from "static scoring" to "dynamic assessment" by introducing integrated biological age and DeepSurv survival analysis, expanding the research dimension from the current state to future trends. The attention weight changes dynamically with individual characteristics, which can adapt to the aging characteristics of different life stages: young individuals focus on genetic risk, while older individuals focus on metabolic and functional status.

[0024] (5) This invention breaks through the limitations of fixed weights and introduces a multi-head attention mechanism to adaptively learn the contribution weights of each omics. For individuals carrying longevity-related genes (such as FOXO3), the genomics attention weight will be significantly reduced; for individuals with metabolic disorders, the metabolomics attention weight will be significantly increased, making the evaluation results more consistent with the actual situation of the individual.

[0025] (6) This invention takes into account both the interpretability and interventionability of the model. It uses the attention weight heatmap to intuitively present the differences in the contribution of each omics to the overall biological age, locates the core driving factors of individual aging, and realizes the integrated design of "assessment-location-intervention".

[0026] (7) Model validation results show that the AUC value of the comprehensive biological age for predicting life expectancy of 100 years is 0.89 (95% confidence interval: 0.85-0.93), which is significantly better than that of single-atom biological age (genomics AUC=0.72, proteomics AUC=0.78, metabolomics AUC=0.76), and also better than traditional time-series age (AUC=0.65). In the 60-70 age group, the predictive AUC of the comprehensive biological age reaches 0.91, which is significantly higher than the best value of single-atom (0.79).

[0027] (8) This invention can be widely applied in fields such as health management, evaluation of anti-aging intervention effects, precision medicine, and insurance actuarial science. By assessing an individual's biological age and probability of living to 100, it can provide a scientific basis for personalized health management; by monitoring changes in biological age before and after intervention, it can evaluate the effectiveness of anti-aging intervention measures; and by identifying the driving factors of individual aging, it can provide target guidance for precision medicine. Attached Figure Description

[0028] Figure 1 A flowchart illustrating a method for constructing a long-lived digital twin based on multi-omics data and deep learning, provided for an embodiment of this application; Figure 2 This is a structural block diagram of a long-lived digital twin construction system based on multi-omics data and deep learning, provided for embodiments of this application. Detailed Implementation

[0029] The longevity digital twin construction method and system provided by this invention, based on multi-omics data and deep learning, can be widely applied in fields such as health management, evaluation of anti-aging intervention effects, precision medicine, and actuarial science. By assessing an individual's biological age and probability of living to 100, it can provide a scientific basis for personalized health management; by monitoring changes in biological age before and after intervention, it can evaluate the effectiveness of anti-aging interventions; and by identifying the driving factors of individual aging, it can provide target guidance for precision medicine. This invention has high industrial applicability and broad market prospects.

[0030] Example 1: Refer to Figure 1 This is a flowchart illustrating a method for constructing a long-lived digital twin based on multi-omics data and deep learning, provided by an embodiment of the present invention. The process may include at least steps S100-S500: S100. Obtain multi-omics data and clinical variable data to obtain a unified input dataset; S200. Based on the unified input dataset, perform principal component analysis and stacked autoencoder dimensionality reduction to extract core features and obtain the omics core feature set. S300. Based on the core feature set of omics, perform multi-label joint training of time-series age and biological indicators to obtain a single-omics age set. S400. Based on the single-omics age set, perform dynamic contribution weight allocation and weighted fusion processing to obtain the comprehensive biological age; S500. Based on the comprehensive biological age, perform survival risk modeling and target age probability estimation to obtain longevity digital twin results.

[0031] S100. Obtain multi-omics data and clinical variable data to obtain a unified input dataset; The multi-omics data includes genomics data, proteomics data, and metabolomics data; the clinical variable data includes gender, age, smoking status, alcohol consumption status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes.

[0032] The genomic data includes whole-genome sequencing or whole-exome sequencing data. Candidate gene loci associated with aging and longevity are extracted from an aging-related gene database, and a polygenic risk score (PRS) is calculated.

[0033] The proteomics data includes plasma protein abundance data detected by liquid chromatography-mass spectrometry, and key proteins related to inflammation, cellular senescence, and protein homeostasis are screened, including one or more of IL-6, TNF-α, and p16INK4a.

[0034] The metabolomics data includes plasma metabolite abundance data detected by ultra-high performance liquid chromatography-tandem mass spectrometry, focusing on metabolites related to energy metabolism, oxidative stress, and amino acid metabolism, including one or more of glucose, fatty acids, glutathione, and malondialdehyde.

[0035] Specifically, step S100 is completed collaboratively by a data acquisition module and a preprocessing module. The data acquisition module includes a genomics data access unit, a proteomics data access unit, a metabolomics data access unit, and a clinical variable data access unit. The preprocessing module includes a quality control unit, a standardization unit, a missing value imputation unit, a batch correction unit, and a record unit. This step is triggered by the data acquisition module when a new subject sample arrives, a predetermined sampling period ends, or clinical variable data is completed. When multiple rounds of submission or retesting records exist for the same sample, the record unit establishes a version order based on sample number, collection time, and batch label, and sends the latest qualified version to the preprocessing module. The genomic data refers to whole-genome sequencing data or whole-exome sequencing data, the former corresponding to whole-genome sequencing and the latter to whole-exome sequencing. The proteomics data refers to plasma protein abundance data; the metabolomics data refers to plasma metabolite abundance data; and the clinical variables include gender, age, smoking status, alcohol consumption status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes. Understandably, the genomic data, proteomics data, metabolomics data, and clinical variables constitute the minimum data set for this step, where sample number, batch label, collection time, and quality control status are fundamental management fields throughout subsequent steps.

[0036] Specifically, the genomics data access unit receives raw locus records, sample identity records, and sequencing batch records from the sequencing platform, and merges the locus table and identity table according to the sample number. In the merged data, sequencing depth, locus missing ratio, and sample duplication rate are checked first, and then candidate locus information is extracted from aging-related gene loci to generate a multi-gene risk score field, which is then incorporated into the genomics data. The proteomics data access unit receives protein name, abundance value, detection batch, and quality control well records from the plasma protein detection platform. Instrument alarm records are deleted first, and synonyms are merged according to the unified naming rules for protein names, while retaining the corresponding records for interleukin-6, tumor necrosis factor-α, and p16INK4a. The metabolomics data access unit receives metabolite name, abundance value, retention time, detection batch, and sample processing records from the plasma metabolite detection platform. Peak identification results are checked first, and then name normalization and abnormal peak deletion are performed on the corresponding records for glucose, fatty acids, glutathione, and malondialdehyde. The clinical variable data access unit receives data on gender, age, smoking status, alcohol consumption, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes from the physical examination information system and follow-up information system. During the receiving phase, it completes field mapping, timestamp unification, and missing item marking. Furthermore, in a scenario involving a joint deployment of a hospital physical examination center, sequencing laboratory, and clinical laboratory, after the subject completes blood collection, the sequencing platform, protein detection platform, metabolite detection platform, and physical examination information system send data packets of the same batch of samples back to the data acquisition module on the same day. The operation recording unit verifies the sample number, collection time, and batch label. If there are number conflicts, time inversions, or missing batches, an anomaly record is written and the data is transferred to the manual review queue. After review, the process continues.

[0037] Specifically, the quality control unit performs stratified screening on the four types of data after access. For the genomics data, the quality control unit deletes records with insufficient sequencing depth, excessively high site deletion ratios, and abnormal sample duplication rates; for the proteomics data, the quality control unit deletes records with excessively high deletion ratios, abnormal intra-batch fluctuations, and deviations from quality control wells; for the metabolomics data, the quality control unit deletes records with excessively high deletion ratios, abnormal retention time drift, and abnormal peak area fluctuations; for the clinical variable data, the quality control unit verifies the value boundaries, unit formats, and logical relationships of the medical history fields, and writes conflict markers for contradictory records. After completing the quality control, the standardization unit performs Z-score standardization on the genomics data, proteomics data, metabolomics data, and clinical variable data. Z-score standardization corresponds to Z-score Standardization, and the processing path is to first calculate the central and discrete statistics of each field, and then map each sample record under the same field to a uniform scale. Subsequently, the missing value imputation unit performs K-nearest neighbor imputation on the standardized records. The process involves first establishing a nearest neighbor set based on sample similarity, then imputing missing positions based on the same field values ​​within the nearest neighbor set, while simultaneously preserving the missing label field during the imputation process. Next, the batch correction unit performs ComBat batch correction on the imputed genomics, proteomics, and metabolomics data. The process involves first splitting the data by batch label, then identifying inter-batch system offsets, and finally outputting batch-corrected data blocks. For the clinical variable data, consistency processing is performed according to collection time and source system, and the processing results are incorporated into the corrected three-omics data. Understandably, the standardization, missing value imputation, and batch correction operate in a fixed order. If any step results in an empty field, missing batch label, or unresolved conflict markers, the preprocessing module pauses output and records the reason for the pause in the operation record unit.

[0038] Furthermore, the unified input dataset is not a simple concatenation result, but a data object encapsulated and generated by the preprocessing module according to a unified field structure. The unified field structure records sample number, collection time, batch label, genomics data field, proteomics data field, metabolomics data field, clinical variable data field, multigene risk score field, quality control status field, missing value marker field, and data version field. The genomics data field contains locus records and multigene risk scores after quality control, standardization, missing value imputation, and batch correction. The proteomics data field contains plasma protein abundance records processed using the same procedure. The metabolomics data field contains plasma metabolite abundance records processed using the same procedure. The clinical variable data field contains standardized physical examination and medical history records. When generating the unified input dataset, the operation recording unit synchronously writes the trigger source, processing order, anomaly correction record and output time, and sends the unified input dataset to S200's "input the unified input dataset into principal component analysis and stacked autoencoder", so that the unified input dataset becomes the direct input for S200 to extract genomics core feature vectors, proteomics core feature vectors and metabolomics core feature vectors.

[0039] The technical effect of this step can be summarized as follows: Step S100 organizes genomic data, proteomics data, metabolomics data, and clinical variable data from different sources, in different formats, and from different batches into a single processing base, ensuring consistency in the data objects received in subsequent steps. Step S100 places quality control, standardization, missing value imputation, and batch correction in the same operational chain, and retains the quality control status field, missing label field, and data version field, ensuring stable input for subsequent feature extraction steps. Step S100 also incorporates the multi-gene risk score field into the unified input dataset, enabling this invention to form a complete data foundation from the very beginning, focusing on the main theme of multi-omics aging characterization.

[0040] S200. Based on the unified input dataset, perform principal component analysis and stacked autoencoder dimensionality reduction to extract core features and obtain the omics core feature set. Specifically, step S200 is executed by the feature extraction module, which includes a data distribution unit, a principal component analysis unit, a stacked autoencoder unit, a dimension verification unit, a batch tracking unit, and a feature encapsulation unit. The feature extraction module receives the unified input dataset output from step S100 and establishes a current processing task based on sample number, collection time, batch label, and data version fields. When the quality control status field in the unified input dataset is qualified, the missing marker field has been completed, and the data version field is in the current version, the data distribution unit triggers this step. When the unified input dataset has an abnormal quality control status field, a missing batch label, or a conflicting data version field, the batch tracking unit writes an interruption record and suspends data delivery to the principal component analysis unit.

[0041] The principal component analysis retains principal components with a cumulative variance contribution rate of not less than 85%. The compressed representation dimension of the stacked autoencoder output is uniformly set to 32 dimensions, that is, the genomics core feature vector, proteomics core feature vector, and metabolomics core feature vector are all 32-dimensional real vectors.

[0042] Understandably, the unified input dataset is the only input object in this step. The feature extraction module does not directly read the original detection file, but extracts the content to be processed from the genomics data field, proteomics data field and metabolomics data field in the unified input dataset, so that the input boundary between S100 and S200 remains fixed.

[0043] Specifically, the data distribution unit splits genomics, proteomics, and metabolomics data subsets from the unified input dataset, and synchronously writes the sample number, batch label, collection time, and multi-gene risk score fields into their respective processing channels. The principal component analysis unit performs principal component analysis (PCA), a dimensionality reduction method that uses linear combination and variance ranking of high-dimensional detection fields. Its input is the standardized numerical matrix from the genomics, proteomics, and metabolomics data subsets, and its output is the principal component score matrix, cumulative variance contribution record, and retained dimension record. Further, the PCA unit first corrects the field order and verifies the numerical range of each data subset, then calculates the correlation structure between fields, and subsequently retains principal components according to the rule that the cumulative variance contribution rate is not less than 85%. For data subsets that do not reach the cumulative variance contribution threshold, the PCA unit writes a dimensionality reduction insufficiency record and sends the sample batch to the verification queue. After principal component analysis, the subset of genomics data outputs a genomics principal component field; the subset of proteomics data outputs a proteomics principal component field; and the subset of metabolomics data outputs a metabolomics principal component field. The cumulative variance contribution record and retained dimension record are stored by the batch tracking unit for use in the same batch of sample reproduction processing.

[0044] Furthermore, the stacked autoencoder unit performs stacked autoencoder (SAE) processing. The stacked autoencoder is a deep feature extraction network formed by cascading multi-layer coding and decoding structures. Its inputs are the genomics principal component field, the proteomics principal component field, and the metabolomics principal component field. Its outputs are a low-dimensional compressed representation and a reconstruction error record. The stacked autoencoder unit includes a genomics coding subunit, a proteomics coding subunit, a metabolomics coding subunit, and a reconstruction verification subunit. Specifically, the genomics coding subunit receives the genomics principal component field and performs layer-by-layer compression; the proteomics coding subunit receives the proteomics principal component field and performs layer-by-layer compression; the metabolomics coding subunit receives the metabolomics principal component field and performs layer-by-layer compression; and the reconstruction verification subunit decodes and reconstructs the compressed result and compares the deviation between the input and the reconstructed output. Specifically, in each processing channel, the stacked autoencoder unit first reads the retained dimension records of the current sample batch, then compresses the principal component fields layer by layer according to the preset network levels, and outputs a compressed representation of fixed dimensions after compression. When the reconstruction deviation exceeds the batch's allowable range, the reconstruction verification subunit marks the sample as a feature unstable sample and returns the corresponding record to the batch tracking unit. Understandably, the principal component analysis is responsible for compressing redundant fields, while the stacked autoencoder is responsible for extracting deep representation structures. Both run continuously in the same channel, with the former's output directly serving as the latter's input, forming a continuous processing chain from linear dimensionality reduction to deep feature extraction.

[0045] Specifically, the dimension verification unit performs field integrity checks and dimension consistency checks on the compressed representation output by the stacked autoencoder unit, and uniformly organizes the output dimensions into thirty-two dimensions according to the fixed constraints in the claims. The genomics coding subunit generates a genomics core feature vector after dimension verification, the proteomics coding subunit generates a proteomics core feature vector after dimension verification, and the metabolomics coding subunit generates a metabolomics core feature vector after dimension verification. The genomics core feature vector referred to herein is a low-dimensional combined representation extracted from the genomics data field and the multi-gene risk score field; the proteomics core feature vector referred to herein is a low-dimensional combined representation extracted from the proteomics data field; and the metabolomics core feature vector referred to herein is a low-dimensional combined representation extracted from the metabolomics data field. Furthermore, the feature encapsulation unit writes the genomics core feature vector, the proteomics core feature vector, and the metabolomics core feature vector into the same data object in a unified order, and simultaneously writes the sample number field, batch label field, collection time field, data version field, principal component retention record field, and reconstruction error record field, encapsulating and generating the omics core feature set; wherein, the output field names of the omics core feature set include "genomics core feature vector", "proteomics core feature vector", "metabolomics core feature vector", "sample number", "batch label", "collection time", "data version", "principal component retention record", and "reconstruction error record", and the omics core feature set serves as the direct input of S300 to "input the omics core feature set into the genomics age network, proteomics age network, and metabolomics age network".

[0046] Furthermore, in the engineering implementation scenario, after the physical examination center, sequencing laboratory, and clinical laboratory complete the testing of the same batch of subject samples, the unified input dataset generated by S100 enters the hospital's analysis server. The feature extraction module in the analysis server starts this step batch by batch, first distributing the genomics data fields, proteomics data fields, and metabolomics data fields of the same batch of subjects to three independent processing channels, and then sequentially executing the principal component analysis and the stacked autoencoder processing in each processing channel. For cases where there are retested samples within the same batch, the batch tracking unit calls the latest data version field to overwrite the old version record and retains the old version processing log; for cases where inconsistent dimensions, abnormal reconstruction deviations, or batch label mismatches occur during operation, the feature extraction module stops encapsulating the omics core feature set, writes the reason for the abnormality and the interruption time, and re-executes the current batch task after verification. Through this actual operation path, those skilled in the art can directly implement this step according to the order of data distribution, principal component analysis, stacked autoencoder compression, dimension verification, and feature encapsulation.

[0047] The technical effect of this step can be summarized as follows: Step S200 organizes the high-dimensional omics fields in the unified input dataset into low-dimensional feature representations with a unified structure, ensuring that the input objects received by the subsequent age network remain fixed. Step S200 connects the principal component analysis and the stacked autoencoder in the same processing chain, giving the genomics core feature vector, the proteomics core feature vector, and the metabolomics core feature vector a unified dimension and a unified encapsulation method. The omics core feature set output by Step S200 directly connects to Step S300, enabling the present invention to converge at the feature layer around the main line of multi-omics aging characterization.

[0048] S300. Based on the core feature set of omics, perform multi-label joint training of time-series age and biological indicators to obtain a single-omics age set. Specifically, step S300 is executed by a single-omics age module, which includes a feature routing unit, a network loading unit, a training management unit, an inference computation unit, a result verification unit, a log recording unit, and a result encapsulation unit. The single-omics age module receives the core omics feature set output by step S200 and reads the fields "Genomics Core Feature Vector," "Proteomics Core Feature Vector," "Metabolomics Core Feature Vector," "Sample Number," "Batch Label," "Collection Time," "Data Version," "Principal Component Retention Record," and "Reconstruction Error Record" as input for this step. Upon receiving the core omics feature set, the feature routing unit first verifies the dimensionality, sample number consistency, and data version status of the genomics core feature vector, proteomics core feature vector, and metabolomics core feature vector. When the dimensionality meets the fixed dimensionality requirement, the sample number matches, and the data version is the current version, this step is triggered. When there are abnormal reconstruction error records, mismatched sample numbers, or data version conflicts, the log recording unit writes the interruption reason and interruption time and suspends the distribution of the current sample to the network loading unit.

[0049] Understandably, the core omics feature set is the common input base of the genomics age network, the proteomics age network, and the metabolomics age network. The single-omics age module does not directly call the original detection records of S100, nor does it roll back to call the processing results before S200, so that the connection boundary between S200 and S300 remains fixed.

[0050] Specifically, the network loading unit loads the genomics age network, the proteomics age network, and the metabolomics age network. All three are Deep Neural Network (DNN) structures and each includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The genomics age network, as referred to here, is a dedicated network that receives the core genomics feature vector and outputs the genomics biological age; the proteomics age network, as referred to here, is a dedicated network that receives the core proteomics feature vector and outputs the proteomics biological age; and the metabolomics age network, as referred to here, is a dedicated network that receives the core metabolomics feature vector and outputs the metabolomics biological age. The training management unit stores the training label set, batch parameter set, iteration version set, and model activation status set for the corresponding networks and calls different training configurations according to network type. The genomic age network is trained using time-series age labels. During the training phase, the training management unit reads the core genomic feature vector and its corresponding time-series age label processed by the S200 from the training sample library. The core genomic feature vector is input into the input layer, mapped layer by layer through the first and second hidden layers, and an age representation value is generated at the output layer. This age representation value is then compared with the corresponding time-series age label, and the network parameters are updated accordingly. The proteomics age network is trained using time-series age labels and inflammatory marker labels. The inflammatory markers include interleukin-6, tumor necrosis factor-α, and p16INK4a. During the training phase, the training management unit simultaneously reads the core proteomics feature vector, time-series age label, and inflammatory marker label. The core proteomics feature vector is input into the proteomics age network, and an age-related representation value is generated at the output layer. The network parameters are then jointly corrected based on the time-series age label and the inflammatory marker label. The metabolomics age network is trained using time-series age labels and metabolic indicator labels, including glucose, fatty acids, glutathione, and malondialdehyde. During the training phase, the training management unit simultaneously reads the metabolomics core feature vector, time-series age labels, and metabolic indicator labels, and inputs the metabolomics core feature vector into the metabolomics age network. After processing at each layer, age-related representation values ​​are generated, and the network parameters are then jointly corrected based on the time-series age labels and metabolic indicator labels.

[0051] In the multi-label joint training of the proteomics age network, the loss weight for the time-series age label is 0.7, and the loss weight for the inflammatory marker label is 0.3; in the multi-label joint training of the metabolomics age network, the loss weight for the time-series age label is 0.6, and the loss weight for the metabolic marker label is 0.4. Mean squared error is used as the loss function for each network, and L2 regularization and Dropout layers are introduced to avoid overfitting.

[0052] Furthermore, the training management unit records the model version generated in each round of training. When new and old versions of the same network are stored in parallel, the network loading unit only loads the current version of the model that is in the effective state set. The old version of the model continues to be retained in the audit library for review and tracing.

[0053] Furthermore, after the model is loaded, the inference computing unit performs inference processing on the omics core feature set in a sample-level manner. For the genomics core feature vector, the inference computing unit calls the genomics age network, inputs the genomics core feature vector into the input layer, and continuously calculates through the first hidden layer and the second hidden layer to output the genomics biological age; for the proteomics core feature vector, the inference computing unit calls the proteomics age network, inputs the proteomics core feature vector into the input layer, and continuously calculates through the first hidden layer and the second hidden layer to output the proteomics biological age; for the metabolomics core feature vector, the inference computing unit calls the metabolomics age network, inputs the metabolomics core feature vector into the input layer, and continuously calculates through the first hidden layer and the second hidden layer to output the metabolomics biological age. The genomic biological age referred to herein is the age representation result generated by the genomic age network based on the genomic core feature vector; the proteomics biological age referred to herein is the age representation result generated by the proteomics age network based on the proteomics core feature vector; the metabolomics biological age referred to herein is the age representation result generated by the metabolomics age network based on the metabolomics core feature vector. The result verification unit performs range verification, null value verification, and sample correspondence verification on the three outputs. When all three outputs exist, the sample numbers are consistent, and there are no null value records, the result encapsulation unit writes the three outputs into a unified data object. When any output is missing, the sample relationship is broken, or the inference record is incomplete, the log recording unit writes a failure flag and returns the current sample to the manual review queue.

[0054] Furthermore, the result encapsulation unit encapsulates the genomic biological age, proteomics biological age, and metabolomics biological age together with the sample number, batch label, collection time, and data version to generate the single-omics age set. The output field names include "genomic biological age," "proteomics biological age," "metabolomics biological age," "sample number," "batch label," "collection time," and "data version." After the single-omics age set is generated, the result encapsulation unit sends it to S400's "input the single-omics age set and the clinical variable data into the attention fusion network, perform dynamic contribution weight allocation and weighted fusion processing, and generate a comprehensive biological age," making the genomic biological age, proteomics biological age, and metabolomics biological age direct inputs for subsequent dynamic contribution weight allocation and weighted fusion processing. Understandably, the three age results output in this step are not mutually exclusive but exist in parallel as three independent fields within the single-omics age set, and subsequent steps call them according to their field names.

[0055] Furthermore, in the engineering implementation scenario, after the hospital's physical examination center, sequencing laboratory, and laboratory department complete the testing of the same batch of subjects, S200 generates the core omics feature set and stores it in the hospital's analysis server. The single-omics age module automatically starts according to the night batch processing window. First, the feature routing unit reads the newly added core omics feature set for the day, and then the network loading unit loads the current effective versions of the genomics age network, the proteomics age network, and the metabolomics age network, respectively. Subsequently, the inference calculation unit outputs the genomics biological age, the proteomics biological age, and the metabolomics biological age for each sample in sequence, and the result encapsulation unit generates the single-omics age set. When there is a retest for the same subject, the training management unit reads the latest data version, and the log recording unit retains the old version inference log and the current version inference log. In the event of model version switching, sample interruption, or output abnormality during operation, the single-omics age module writes the switching time, the version before switching, the version after switching, and the reason for the abnormality into the log, thereby maintaining the continuous recording of the operation path of this step. Those skilled in the art can complete this step by following the processing sequence of feature routing, network loading, label training, inference computation, result verification, and result encapsulation.

[0056] The technical effect of this step can be summarized as follows: S300 transforms the core omics feature set into genomic biological age, proteomics biological age, and metabolomics biological age, so that the input objects received in the subsequent fusion step converge from the feature layer to the age representation layer. S300 sets up the genomic age network, proteomics age network, and metabolomics age network for the three types of omics respectively, and trains them with different label configurations, ensuring that the three age results retain their respective sources and semantics. The single-omics age set output by S300 directly connects to S400, allowing the present invention to continue along the main lines of multi-omics aging representation, dynamic fusion, and the construction of longevity digital twins.

[0057] S400. Based on the single-omics age set, perform dynamic contribution weight allocation and weighted fusion processing to obtain the comprehensive biological age; Specifically, step S400 is executed by a dynamic fusion module, which includes a data access unit, a scoring generation unit, a splicing unit, an attention fusion network, a weight recording unit, a fusion calculation unit, a result verification unit, and a version management unit. The data access unit receives the single-omics age set output by step S300 and reads the gender, age, smoking status, alcohol consumption status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes corresponding to the same sample number from the clinical variable data processed by step S100. The version management unit first checks the "sample number," "batch label," "collection time," and "data version" fields in the single-omics age set against the corresponding fields in the clinical variable data. If the sample number matches, the collection time is within the same analysis period, and the data version is the current version, this step is triggered. If there is a mismatch in sample number, a misalignment in time, or a version conflict, the version management unit writes an exception record and suspends the distribution of the current sample to the attention fusion network. Understandably, the single-omics age set is the main input object of this step, and the clinical variable data is the co-input object of this step. Only after both have completed the same sample merging in this step will they proceed to the subsequent dynamic contribution weight allocation and weighted fusion processing.

[0058] Specifically, the scoring generation unit performs structured processing and numerical transformation on the accessed clinical variable data to generate clinical variable scores. The clinical variable score referred to here is a single-value score result after uniformly converting continuous and categorical fields in the clinical variable data. The scoring generation unit uses a logistic regression model to generate this score. The logistic regression model receives gender, age, smoking status, alcohol consumption status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes, and forms input records according to a preset field order, then outputs clinical variable scores within the zero-to-one value range. Further, during runtime, the scoring generation unit first checks the field completeness, then calls the missing data marker field in S100 for default fields. If unclosed missing records exist, the scoring of the current sample is paused. After scoring is completed, the clinical variable score, along with the sample number and the data version, is written into an intermediate record for use by the splicing unit. The attention fusion network referred to here is a dedicated network that receives multiple age representations and clinical variable scores and outputs corresponding weight results. The attention fusion network includes an input layer, a weight calculation layer, a normalization processing layer, and an output layer. The input layer receives the genomic biological age, the proteomic biological age, the metabolomics biological age, and the clinical variable scores. The weight calculation layer generates initial weight values ​​corresponding to each input. The normalization processing layer organizes the initial weight values ​​into dynamic contribution weights. The output layer outputs the dynamic contribution weight results.

[0059] Further, the splicing unit extracts the fields "genomic biological age," "proteomics biological age," and "metabolomics biological age" from the single-omics age set, and extracts the "clinical variable score" field from the intermediate record, forming a fusion input vector in a fixed order, and then sends the fusion input vector into the attention fusion network. After receiving the fusion input vector, the attention fusion network first reads the numerical states of the genomic biological age, proteomics biological age, metabolomics biological age, and clinical variable score respectively by the weight calculation layer, and then generates four initial weight values ​​by combining the internal correlation strength of the same sample; subsequently, the normalization processing layer performs normalization weight calculation on the four initial weight values ​​to generate dynamic contribution weights corresponding to the genomic biological age, proteomics biological age, metabolomics biological age, and clinical variable score. The dynamic contribution weights referred to here are weight results that change synchronously with the multi-omics age representation and clinical variable score of the current sample. The four weights respectively reflect the contribution state of the genomic biological age, proteomics biological age, metabolomics biological age, and clinical variable score in the current sample. Understandably, the dynamic contribution weights are not fixed configuration parameters, but rather calculation results regenerated when each sample enters the attention fusion network. Therefore, different samples in the same batch may correspond to different weight distributions.

[0060] Specifically, the fusion calculation unit reads the dynamic contribution weights and performs weighted fusion processing on the genomic biological age, proteomics biological age, metabolomics biological age, and clinical variable score according to the field correspondence, outputting a comprehensive biological age. Here, comprehensive biological age refers to the unified age representation result after fusing the single-omics age set and the clinical variable score with the dynamic contribution weights. The result verification unit performs null value verification, range verification, and sample correspondence verification on the comprehensive biological age. When the comprehensive biological age exists, the sample numbers are consistent, and the dynamic contribution weights are complete, the weight recording unit writes the "genomic biological age weight," "proteomics biological age weight," "metabolomics biological age weight," and "clinical variable score weight" into the weight record, and simultaneously maps the dynamic contribution weights to an attention weight heatmap, generating a contribution distribution result. When there are missing weights, missing comprehensive biological age, or broken sample fields, the version management unit writes the failure reason and transfers the current sample to the review queue. Furthermore, the dynamic fusion module encapsulates "comprehensive biological age," "sample number," "batch label," "collection time," "data version," "genomic biological age weight," "proteomics biological age weight," "metabolomics biological age weight," "clinical variable score weight," and "contribution distribution result" as the output object of this step. Among them, "comprehensive biological age" is the core output field, which is directly called by S500 to "input the comprehensive biological age and the clinical variable data into the deep survival analysis model to perform survival risk modeling and target age probability estimation."

[0061] Furthermore, in the engineering implementation scenario, the hospital's analysis server receives the single-omics age set and clinical variable data of the same batch of subjects during a nighttime batch processing window. The dynamic fusion module first completes sample merging through the data access unit, then generates the clinical variable scores through the scoring generation unit, followed by the splicing unit forming a fusion input vector, which is then fed into the attention fusion network to generate the dynamic contribution weights. Finally, the fusion calculation unit outputs the comprehensive biological age, and the weight recording unit simultaneously generates the contribution distribution results. When there are retest records for the same subject, the version management unit reads the current version data and retains the old version's running log. When sample mismatch, missing weights, or scoring interruption occur during operation, the dynamic fusion module stops the current sample output and records the interruption time, sample number, and reason for the anomaly. Those skilled in the art can complete this step by following the processing sequence of data access, scoring generation, splicing, attention fusion, weight recording, fusion calculation, and result encapsulation.

[0062] In one specific embodiment, the dynamic contribution weight allocation specifically includes: concatenating the genomic biological age, proteomics biological age, metabolomics biological age from the single-omics age set with clinical variable scores into a fusion input vector. , recorded as

[0063] in, For genomic biological age, For proteomics biological age, For metabolomics biological age. For clinical variable scores, subscript Indicates the first One sample.

[0064] The contribution weights of each omics and clinical variable were calculated using an attention mechanism and normalized using the Softmax function.

[0065] in, For the first Attention weights for each input feature, Corresponding to , , and ; For the first The weight vector of each attention head. For the corresponding bias terms, all are adaptively learned through model training; It is a natural exponential function; represents the vector transpose; j represents the attention weight index of the input feature.

[0066] The weighted fusion process calculates the overall biological age using a weighted summation method:

[0067] in, For the first The overall biological age of each sample, To dynamically contribute weights, satisfying .

[0068] In this embodiment, the core feature vectors output by the stacked autoencoder are all set to 32 dimensions. The principal component analysis retains principal components with a cumulative variance contribution rate of not less than 85%.

[0069] The genomics age network, proteomics age network and metabolomics age network all adopt the structure of “input layer (32) - hidden layer 1 (64) - hidden layer 2 (32) - output layer (1)”, with ReLU as the activation function and Dropout ratio set to 0.2.

[0070] The joint loss function of the proteomics age network for:

[0071] Where MSE represents mean squared error, age is time-series age, and IL-6 is interleukin-6 level.

[0072] The joint loss function of the metabolomics age network is :

[0073] Among them, metabolic is a comprehensive metabolic indicator (blood glucose, blood lipids).

[0074] The attention fusion network employs a multi-head attention mechanism with 4 attention heads. The DeepSurv model uses a partial likelihood loss function, employs Adam as the optimizer, has a learning rate of 0.001, performs 200 training epochs, and includes 20 early stopping epochs.

[0075] Model validation was performed using five-fold cross-validation with a test set containing 200 independent samples. Results showed that the combined biological age predicted a 0.89 AUC (95% CI: 0.85–0.93) life expectancy, significantly better than single-agent biological age (genomics AUC = 0.72, proteomics AUC = 0.78, metabolomics AUC = 0.76). In the 60–70 age group (80 cases in the test set), the combined biological age predicted an AUC of 0.91, a significant improvement over the best single-agent result (0.79).

[0076] The technical effect of this step can be summarized as follows: S400 converges the single-omics age set and the clinical variable data into the same fusion chain, establishing the comprehensive biological age on the joint basis of single-omics age representation and clinical variable scores. S400 generates dynamic contribution weights through the attention fusion network, enabling age representations from different sources to form corresponding relationships within the same sample. The comprehensive biological age output by S400 directly connects to S500, allowing this invention to continue its development around the main theme of dynamic fusion and the construction of a longevity digital twin.

[0077] S500. Based on the comprehensive biological age, perform survival risk modeling and target age probability calculation to obtain longevity digital twin results; Specifically, S500 is executed by the survival prediction module, which includes a data access unit, a parameter calling unit, a deep survival analysis model, a probability calculation unit, a result verification unit, an output encapsulation unit, a log recording unit, and a version management unit. The data access unit receives the comprehensive biological age output by S400 and reads the gender, age, smoking status, alcohol consumption status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes corresponding to the same sample number from the clinical variable data processed by S100. The deep survival analysis model, referred to here, is a dedicated prediction model that receives the comprehensive biological age, the clinical variable data, and baseline risk parameters, and outputs the survival probability at the target age. The deep survival analysis model uses the DeepSurv model. Before the model runs, the version management unit checks the fields "sample number," "batch label," "collection time," "data version," and "comprehensive biological age," and compares them with the corresponding fields in the clinical variable data. This step is triggered when the sample numbers match, the collection time is within the same analysis cycle, the data version is the current version, and the comprehensive biological age exists. If there are sample number mismatches, data version conflicts, missing comprehensive biological age, or incomplete closure of missing clinical variable data, the log recording unit writes the reason for the anomaly, the anomaly time, and the sample number, and suspends the distribution of the current sample to the deep survival analysis model. Understandably, the comprehensive biological age is the core input field of this step, and the clinical variable data is a co-input field. Both undergo merging and consistency checks of the same sample before entering the deep survival analysis model.

[0078] Specifically, the parameter invocation unit retrieves the currently active version of the DeepSurv model from the model parameter library and simultaneously reads the baseline risk parameters. The baseline risk parameters, referred to here, are a set of basic risk parameters compiled from historical survival outcome records in the training sample library, used to describe the baseline risk status within different age ranges. When reading the baseline risk parameters, the parameter invocation unit simultaneously verifies the parameter version, training time, and active status. If an old version parameter or an invalid parameter exists, the current invocation is stopped and a version review process is initiated. The DeepSurv model includes an input receiving layer, a risk characterization layer, a risk association layer, and an output layer. The input receiving layer receives the comprehensive biological age, the clinical variable data, and the baseline risk parameters. The risk characterization layer jointly characterizes the comprehensive biological age and the clinical variable data. The risk association layer generates the survival risk status of the current sample, and the output layer outputs the survival probability at the target age. Furthermore, during the training phase, the DeepSurv model receives comprehensive biological age, clinical variable data, and historical follow-up outcome records from the training samples, and updates the model's internal parameters sequentially at the sample level. During the inference phase, the DeepSurv model no longer reads the outcome records from the training samples; instead, it directly receives the comprehensive biological age, clinical variable data, and baseline risk parameters of the current sample to generate the survival risk status corresponding to the current sample. Survival risk modeling, as referred to here, means the process of jointly processing the comprehensive biological age and clinical variable data within the same model path and outputting the survival risk status corresponding to the current sample.

[0079] Further, after the data access unit sends the comprehensive biological age and clinical variable data of the current sample to the input receiving layer, the risk characterization layer first reads the "comprehensive biological age" field, as well as the fields of gender, age, smoking status, drinking status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes, and performs field order sorting, numerical status verification, and sample-level binding. Subsequently, the risk association layer jointly calculates the comprehensive biological age and the clinical variable data to generate the risk ranking result and risk status record of the current sample. The target age probability estimation mentioned here refers to the process by which the probability estimation unit outputs the survival probability of the target age according to the risk path corresponding to the target age based on the risk status record and the baseline risk parameters. The target age is one hundred years old. Before outputting, the probability estimation unit first reads the actual age field of the current sample and performs continuous estimation from the risk starting point corresponding to the actual age, finally generating the survival probability of the target age. Specifically, after completing the calculation, the probability estimation unit generates a "target age survival probability" field and simultaneously writes it into the "risk status record," "model version," "parameter version," and "calculation time" fields. The result verification unit performs null value verification, range verification, and sample correspondence verification on the "target age survival probability" field. When the "target age survival probability" field exists, the sample number is consistent, and the model version and parameter version are in the current effective state, the output encapsulation process is allowed. When there are null values, sample mismatches, or model version mismatches, the log recording unit writes a failure flag and sends the current sample to the review queue.

[0080] Furthermore, the output encapsulation unit encapsulates the "Comprehensive Biological Age," "Target Age Survival Probability," "Sample Number," "Batch Label," "Collection Time," "Data Version," "Model Version," "Parameter Version," and "Estimation Time" into a longevity digital twin result. The longevity digital twin result referred to here is a unified result object at the current sample level, composed of the comprehensive biological age, the target age survival probability, and the model execution record. This unified result object corresponds to the longevity prediction result for each subject in the current analysis period. Further, when encapsulating the longevity digital twin result, the output encapsulation unit simultaneously associates it with the "Contribution Distribution Result" generated by S400 and writes both into the same sample record in the same result library, allowing subsequent calls to simultaneously read the comprehensive biological age, the target age survival probability, and the contribution distribution result. Understandably, the longevity digital twin result is the final output object of the main process of this invention. In this step, the comprehensive biological age is no longer further broken down into the genomic biological age, the proteomics biological age, and the metabolomics biological age, but rather serves as a unified age representation in survival risk modeling and target age probability estimation.

[0081] Furthermore, in the engineering implementation scenario, after the hospital's physical examination center, sequencing laboratory, and laboratory department complete the subject's testing, the S400 outputs the comprehensive biological age. The survival prediction module in the hospital's analysis server receives new samples in the daily batch processing window. The data access unit first merges the comprehensive biological age and clinical variable data into samples, then the parameter calling unit reads the currently effective DeepSurv model and the baseline risk parameters. Subsequently, the DeepSurv model performs survival risk modeling, and the probability estimation unit outputs the "target age survival probability" field. For cases where the same subject has retest records, the version management unit calls the current version's "comprehensive biological age" field and the clinical variable data, while retaining the old version's estimation log. In cases of model switching, parameter replacement, or sample interruption during operation, the log recording unit records the switching time, the version before switching, the version after switching, and the reason for the interruption. Those skilled in the art can complete this step by following the processing sequence of data access, parameter calling, survival risk modeling, target age probability estimation, result verification, and output encapsulation.

[0082] In one specific embodiment, the survival risk modeling employs the DeepSurv deep survival analysis model, where individuals... Risk function Defined as:

[0083] in, Represents age as a time variable; For individuals Age The instantaneous risk of death at that time (i.e., the risk function); The baseline risk function, which is only related to time. This information is obtained through nonparametric estimation of training samples; For comprehensive biological age Risk coefficient (scalar); This is a vector of risk coefficients for clinical variables, with dimensions similar to those of the clinical variables. same; The clinical variable vector includes gender, age, smoking status, alcohol consumption, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes.

[0084] In the target age probability calculation, for the current age... Individuals who survive to the target age The probability at age 1 is:

[0085] in Probability of survival at the target age (probability of living to 100 years old); Let be the integral variable, representing age from the current age. Continuous values ​​up to age 100; For age The definite integral; For age Baseline risk at that time.

[0086] The DeepSurv model employs a partial likelihood loss function during the training phase and optimizes its parameters using a stochastic gradient descent algorithm. , And the weights within the neural network.

[0087] The technical effect of this step can be summarized as follows: Step S500 converges the comprehensive biological age and the clinical variable data into the same survival prediction chain, allowing the aforementioned multi-omics aging characterization to enter the target age survival probability estimation stage. Step S500 generates the target age survival probability through the DeepSurv model and the baseline risk parameters, transforming the age characterization output from the preceding steps into a unified longevity digital twin result. As the final step of the main process, Step S500 encapsulates the comprehensive biological age, the target age survival probability, and the model execution record in the same result object, thus forming a closed loop around the technical main line of multi-omics aging characterization, dynamic fusion, and longevity prediction.

[0088] Example 2: Figure 2 This diagram illustrates a structural block diagram of a long-lived digital twin construction system based on multi-omics data and deep learning, according to an embodiment of the present invention. Figure 2 As shown, the structure may include: Data acquisition module 01 is used to acquire genomics data, proteomics data, metabolomics data, and clinical variable data. Specifically, the data acquisition module receives whole-genome sequencing data or whole-exome sequencing data, plasma protein abundance data, plasma metabolite abundance data, and data on gender, age, smoking status, alcohol consumption status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes corresponding to the same sample number, and merges them according to sample number, batch label, and collection time. The data acquisition module starts the receiving process when the collection time arrives, the sample test results arrive, or the clinical variable data is supplemented. Records lacking sample number, batch label, or collection time are written to the pending verification status. In the case of multiple records for the same sample number, the latest data version is retained and the source of the old version is registered. The data acquisition module outputs genomics data, proteomics data, metabolomics data, and clinical variable data, and sends the genomics data, proteomics data, metabolomics data, and clinical variable data to the preprocessing module as input for quality control, standardization, missing value imputation, and batch correction. At the same time, it sends the sample number, batch label, acquisition time, and data version to the log and version management module for registration.

[0089] Preprocessing module 02, connected to the data acquisition module, performs quality control, standardization, missing value imputation, and batch correction on the genomics data, proteomics data, metabolomics data, and clinical variable data to generate a unified input dataset. Specifically, the preprocessing module receives the genomics data, proteomics data, metabolomics data, and clinical variable data from the data acquisition module, first performs field verification according to sample number and batch label, and then performs screening or suspension processing on abnormal null values, duplicate records, and records with misaligned fields. During the quality control stage, the preprocessing module deletes site records, protein abundance records, metabolite abundance records, and clinical variable data records that do not meet the preset status. During the standardization stage, it performs uniform scaling on the retained records. During the missing value imputation stage, it performs neighbor sample completion on missing fields. During the batch correction stage, it performs batch offset adjustment on records from different batches, and retains missing markers, quality control status, and data version. The preprocessing module outputs a unified input dataset, which includes genomics data fields, proteomics data fields, metabolomics data fields, clinical variable data fields, sample number, batch label, collection time, and data version. The unified input dataset is then sent to the feature extraction module, while the quality control status, missing data markers, and data version are sent to the log and version management module for invocation.

[0090] Feature extraction module 03, connected to the preprocessing module, is used to input the unified input dataset into principal component analysis and a stacked autoencoder to extract genomics core feature vectors, proteomics core feature vectors, and metabolomics core feature vectors, generating a core omics feature set. Specifically, the feature extraction module receives the unified input dataset from the preprocessing module and splits it into genomics data fields, proteomics data fields, and metabolomics data fields, establishing three parallel processing links according to sample numbers. The feature extraction module first performs principal component analysis on the three types of fields respectively, retaining principal component records with a cumulative variance contribution rate of not less than 85%, and then inputs the principal component records into the stacked autoencoder to perform synchronous processing on the compressed representation and reconstruction state. When the reconstruction state is abnormal or the dimensionality state is inconsistent, the feature extraction module writes an interruption record and pauses the current sample flow. The feature extraction module outputs genomics core feature vectors, proteomics core feature vectors, and metabolomics core feature vectors. It then encapsulates these three core feature vectors along with sample number, batch label, collection time, and data version into an omics core feature set and sends it to the single-omics age module. Simultaneously, it sends the principal component retention record, reconstruction status, and data version to the log and version management module for registration.

[0091] The single-omics age module 04, connected to the feature extraction module, is used to input the core omics feature set into the genomics age network, proteomics age network, and metabolomics age network, and output the genomics biological age, proteomics biological age, and metabolomics biological age to generate a single-omics age set. Specifically, the single-omics age module receives the core omics feature set from the feature extraction module and reads the core genomics feature vector, the core proteomics feature vector, the core metabolomics feature vector, sample number, batch label, collection time, and data version. The single-omics age module sends the three types of core feature vectors into the genomics age network, proteomics age network, and metabolomics age network according to the network type. The genomics age network performs inference according to the training version corresponding to the time-series age label, the proteomics age network performs inference according to the training version corresponding to the time-series age label and the inflammatory marker label, and the metabolomics age network performs inference according to the training version corresponding to the time-series age label and the metabolic marker label, and performs sample consistency verification on the output results. The single-omics age module outputs genomic biological age, proteomics biological age, and metabolomics biological age. The three age results, along with the sample number, batch label, collection time, and data version, are encapsulated into a single-omics age set and sent to the dynamic fusion module. At the same time, the model version, inference time, and sample number are sent to the log and version management module for registration.

[0092] The dynamic fusion module 05, connected to the single-omics age module, is used to input the single-omics age set and the clinical variable data into the attention fusion network, perform dynamic contribution weight allocation and weighted fusion processing, and generate a comprehensive biological age. Specifically, the dynamic fusion module receives the single-omics age set from the single-omics age module and the clinical variable data from the preprocessing module. It first merges samples of the same type according to sample number, collection time, and data version, and then generates a clinical variable score from the clinical variable data. The dynamic fusion module concatenates the genomic biological age, proteomics biological age, metabolomics biological age, and clinical variable score in a fixed order as a fusion input object and sends it to the attention fusion network. The attention fusion network generates dynamic contribution weights for the four input objects, and then performs weighted fusion processing according to the dynamic contribution weights. When any input object is missing, the sample number is mismatched, or the dynamic contribution weight is missing, the dynamic fusion module writes an exception record and stops the current sample output. The dynamic fusion module outputs a comprehensive biological age and sends the comprehensive biological age, along with the sample number, batch label, collection time, data version, and dynamic contribution weight, to the survival prediction module. At the same time, the dynamic contribution weight and data version are sent to the log and version management module for registration.

[0093] The survival prediction module 06, connected to the dynamic fusion module, is used to input the comprehensive biological age and the clinical variable data into the deep survival analysis model, perform survival risk modeling and target age probability estimation, and generate a longevity digital twin result. Specifically, the survival prediction module receives the comprehensive biological age from the dynamic fusion module and the clinical variable data from the preprocessing module. It first completes the input binding according to sample number, collection time, and data version, and then calls the currently active deep survival analysis model and parameter version. The survival prediction module inputs the comprehensive biological age and the clinical variable data into the deep survival analysis model, performs survival risk modeling on the current sample, and generates the target age survival probability according to the estimation path of a target age of one hundred years. When the model version mismatch, parameter version mismatch, or input object is missing, the survival prediction module writes a failure record and stops the estimation of the current sample. The survival prediction module outputs longevity digital twin results, which include comprehensive biological age, target age survival probability, sample number, batch label, collection time, data version, model version, and parameter version. The longevity digital twin results are written into the result database, and the target age survival probability, model version, and parameter version are sent to the log and version management module for registration.

[0094] The log and version management module 07, connected to the survival prediction module, records sample number, batch label, collection time, data version, model version, and parameter version, and retrieves the current version data. Specifically, the log and version management module receives quality control status, missing data markers, and data version from the preprocessing module; principal component retention records, reconstruction status, and data version from the feature extraction module; model version, inference time, and sample number from the single-omics age module; dynamic contribution weights and data version from the dynamic fusion module; and target age survival probability, model version, and parameter version from the survival prediction module. It also establishes a full-process log based on sample number, batch label, and collection time. The log and version management module performs version comparison on the data version, model version, and parameter version uploaded by each module. It establishes a current version index for records in the current state, a historical version index for older version records, and a verification mark for conflicting records. When a module is called, it returns the current version data to the preprocessing module, the feature extraction module, the single-omics age module, the dynamic fusion module, and the survival prediction module. The log and version management module outputs the current version data call results and full process log records, enabling the data acquisition module, the preprocessing module, the feature extraction module, the single-omics age module, the dynamic fusion module, and the survival prediction module to run continuously under the same sample number, the same batch label, the same collection time, and the same data version.

Claims

1. A method for constructing long-lived digital twins based on multi-omics data and deep learning, characterized in that, include: S100. Obtain multi-omics data and clinical variable data to obtain a unified input dataset; S200. Based on the unified input dataset, perform principal component analysis and stacked autoencoder dimensionality reduction to extract core features and obtain the omics core feature set. S300. Based on the core feature set of omics, perform multi-label joint training of time-series age and biological indicators to obtain a single-omics age set. S400. Based on the single-omics age set, perform dynamic contribution weight allocation and weighted fusion processing to obtain the comprehensive biological age; S500. Based on the comprehensive biological age, perform survival risk modeling and target age probability estimation to obtain longevity digital twin results.

2. The method according to claim 1, characterized in that, The multi-omics data and clinical variable data include: The multi-omics data includes genomics data, proteomics data, and metabolomics data; The clinical variable data include gender, age, smoking status, alcohol consumption, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes.

3. The method according to claim 2, characterized in that, The process of acquiring multi-omics data and clinical variable data to obtain a unified input dataset includes: After acquisition, preprocessing is performed, which includes quality control processing, standardization processing, missing value imputation processing, and batch correction processing. The quality control process includes performing stratified screening on the four types of data after access, and deleting records with insufficient sequencing depth, excessively high site missing ratio, abnormal sample duplication rate, abnormal intra-batch fluctuation, abnormal retention time drift, or conflicting value boundaries. The standardization process includes performing Z-score standardization on the data that has passed quality control. The missing value imputation process includes performing K-nearest neighbor imputation on the standardized data and simultaneously retaining the missing label field during the imputation process; The batch correction process includes performing ComBat batch correction on genomics, proteomics, and metabolomics data with missing value imputation, and performing consistency processing on clinical variable data according to collection time and source system; The unified input dataset includes sample number, collection time, batch label, quality control status field, missing marker field, and data version field.

4. The method according to claim 3, characterized in that, The process of performing principal component analysis and stacked autoencoder dimensionality reduction to extract core features includes: The principal component analysis includes: splitting the genomics data subset, proteomics data subset, and metabolomics data subset from the unified input dataset, performing principal component analysis on each subset, retaining principal components according to the rule that the cumulative variance contribution rate is not less than 85%, and outputting the principal component score matrix. The stacked autoencoder for dimensionality reduction and core feature extraction includes: The principal component score matrix is ​​input into a stack-based autoencoder; The stacked autoencoder includes a multi-layer encoding structure and a multi-layer decoding structure, which compresses each data subset layer by layer and outputs a compressed representation with a fixed dimension. The stacked autoencoder also includes dimension verification, which unifies the output dimension into 32 dimensions to generate genomics core feature vector, proteomics core feature vector and metabolomics core feature vector. The three core feature vectors mentioned above, along with sample number, batch label, collection time, and data version, are encapsulated to generate an omics core feature set.

5. The method according to claim 4, characterized in that, The process of performing multi-label joint training of time-series age and biological indicators includes: The multi-label joint training includes: constructing a genomics age network, a proteomics age network, and a metabolomics age network, respectively. Each network is a deep neural network and includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The genomic age network is trained using time-series age labels, receives the genomic core feature vector from the omics core feature set, and outputs the genomic biological age. The proteomics age network is trained using a combination of time-series age labels and inflammatory marker labels. The inflammatory marker labels include interleukin-6, tumor necrosis factor-α, and p16INK4a. It receives the core feature vector of proteomics and outputs the biological age of proteomics. The metabolomics age network is trained using a combination of time-series age labels and metabolic indicator labels. The metabolic indicator labels include glucose, fatty acids, glutathione, and malondialdehyde, receive a metabolomics core feature vector, and output metabolomics biological age. The three biological ages mentioned above are combined with the sample number, batch label, collection time, and data version to generate a single-omics age set.

6. The method according to claim 5, characterized in that, The process of dynamic contribution weight allocation includes: Acquire clinical variable data and generate clinical variable scores using a logistic regression model; The genomic biological age, proteomics biological age, and metabolomics biological age in the single-omics age set are concatenated with the clinical variable scores to form a fusion input vector; The fused input vector is input into the attention fusion network, which includes an input layer, a weight calculation layer, a normalization processing layer, and an output layer. The weight calculation layer generates four initial weight values ​​based on the current sample, and the normalization processing layer converts the initial weight values ​​into dynamic contribution weights that sum to 1. The dynamic contribution weights change synchronously with the current sample's omics age characterization and clinical variable scores, and different samples correspond to different weight distributions.

7. The method according to claim 6, characterized in that, The weighted fusion process includes: The genomic biological age, proteomics biological age, metabolomics biological age, and clinical variable scores are weighted and summed according to dynamic contribution weights to output the comprehensive biological age; at the same time, an attention weight heatmap is generated as the contribution distribution result.

8. The method according to claim 7, characterized in that, The survival risk modeling process includes: Obtain clinical variable data corresponding to the same sample number. The clinical variable data includes gender, age, smoking status, drinking status, exercise frequency, blood pressure, blood glucose, blood lipids, history of hypertension, and history of diabetes. The data, which integrate biological age and clinical variables, are input into the DeepSurv model, which includes an input receiving layer, a risk characterization layer, a risk association layer, and an output layer.

9. The method according to claim 8, characterized in that, The process of calculating the probability of the target age includes: The probability extrapolation unit, based on the risk status record and baseline risk parameters output by the DeepSurv model, continuously extrapolates from the actual age of the current sample to the target age of 100 years, generating the survival probability at the target age. The results of longevity digital twin are encapsulated together, including biological age, survival probability at target age, sample number, batch label, collection time, data version, model version, and contribution distribution.

10. A long-lived digital twin construction system based on multi-omics data and deep learning, characterized in that, include: The system comprises a data acquisition module, a preprocessing module, a feature extraction module, a single-omics age module, a dynamic fusion module, a survival prediction module, and a log and version management module; these modules are connected in sequence to implement the method described in any one of claims 1-9.

Images