Machine learning based epigenetic age assessment system and method

Abstract

The application belongs to the technical field of bioinformatics, and discloses an epigenetic age assessment system and method based on machine learning; comprising: collecting methylation data of a user in real time; screening methylation feature data from the methylation data and compressing the methylation feature data into hidden feature data; obtaining historical feature data and constructing an age assessment model; inputting the hidden feature data obtained in real time into the age assessment model, predicting an epigenetic age, and marking the epigenetic age as a predicted age; systematically perturbing the predicted age, calculating a prediction error value of each simulated age, taking a simulated age corresponding to a minimum prediction error value as a corrected age; calculating a relative age, and taking a mean value of the relative age and the corrected age as an epigenetic age of the user; the application can improve the adaptability and accuracy of the model on different individuals and age groups, thereby improving the accuracy and timeliness of epigenetic age assessment.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and more specifically, to a machine learning-based epigenetic age assessment system and method. Background Technology

[0002] With the continuous development of biological and medical research, the molecular mechanisms of aging are gradually being revealed, among which epigenetics plays an important role. Epigenetic changes refer to changes in gene expression without alterations in gene sequence, and DNA methylation is one of the most widespread and important epigenetic modifications. Methylation mainly occurs at CpG island sites in DNA, and the methylation pattern of individuals changes significantly with age. These changes are closely related to aging-related physiological processes, such as gene expression regulation, cell repair mechanisms, and the decline of immune function. In recent years, researchers have discovered that by measuring an individual's DNA methylation level, their biological age, or epigenetic age, can be inferred. This age value may differ from the actual age, reflecting the rate of aging and health status of an individual. This helps to deepen the understanding of the aging process, predict the risk of age-related diseases (such as cardiovascular disease, cancer, and diabetes), and thus provide a scientific basis for health intervention and disease prevention.

[0003] With the development of high-throughput genomics technology, acquiring large-scale DNA methylation data has become easier and more efficient, making it possible to develop machine learning-based epigenetic age assessment systems. Machine learning technology can process large-scale methylation data, extracting potential and complex aging patterns to establish accurate epigenetic age prediction models. For example, patent CN119092140A discloses a method for assessing male biological age and predicting aging based on peripheral blood DNA methylation data. This method includes: screening DNA methylation sites related to male biological age and aging; downloading DNA methylation microarray sequencing data and screening male data; screening the modeling database and DNA methylation sites to obtain modeling samples and modeling sites; establishing an epigenetic clock model; comparing the epigenetic clock model with traditional models and evaluating the models; and performing biological age assessment and aging prediction. This invention can be effectively used to assess the age and degree of aging in a population.

[0004] However, while the aforementioned technologies can achieve epigenetic age assessment, they rely on screening for DNA methylation sites based on relevant literature or research findings. The screening process is relatively static and lacks real-time adaptability to different users, thus affecting the accuracy of the model on different individuals or age groups. At the same time, the aforementioned technologies lack a correction mechanism and cannot effectively correct the predicted results, thereby reducing the accuracy of epigenetic age assessment.

[0005] In view of this, the present invention proposes an epigenetic age assessment system and method based on machine learning to solve the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: a machine learning-based epigenetic age assessment method, comprising:

[0007] Real-time collection of users' methylation data;

[0008] Information gain analysis is performed on methylation data to calculate information randomness. Based on information randomness, methylation feature data is screened from the methylation data, and dimensionality reduction is performed on the methylation feature data to compress the methylation feature data into hidden feature data.

[0009] Acquire historical feature data, and construct an age assessment model based on the historical feature data and predefined machine learning algorithms;

[0010] The hidden feature data acquired in real time is input into the age assessment model to predict the corresponding epigenetic age and label it as the predicted age.

[0011] The predicted age is systematically perturbed to generate different simulated ages. The prediction error value corresponding to each simulated age is calculated, and the simulated age corresponding to the prediction error value with the smallest value is taken as the corrected age.

[0012] Calculate the relative age, and use the mean of the relative age and the corrected age as the user's epigenetic age.

[0013] Furthermore, the methylation data includes the methylation level at each site, where the site is a CpG site, and the methylation level is a continuous value between 0 and 1;

[0014] The method for calculating the randomness of information includes:

[0015] The real-time methylation data is labeled as real-time data, and historical data, including methylation data of users collected in previous years, is obtained. Methylation levels with the same locus in both real-time and historical data are grouped into a locus set, with each locus set corresponding to a specific locus. Methylation levels with the same value in each locus set are grouped into a numerical set. The number of methylation levels in each numerical set is counted and labeled as the level count. The number of numerical sets in each locus set is counted and labeled as the total level count. Each level count is multiplied by the corresponding total level count to obtain the level frequency for each numerical set. Logarithmic operations are performed with the natural constant as the base and each level frequency as the argument to obtain the level logarithm for each numerical set. Each level logarithm is multiplied by the corresponding level frequency to obtain the weighted level probability for each numerical set. The weighted level probabilities corresponding to numerical sets at the same locus set are summed sequentially, and the negative value is taken to obtain the information randomness for each locus set.

[0016] Furthermore, methods for screening methylation feature data include:

[0017] The number of loci is counted and labeled as the locus count. The randomness of each information metric is summed sequentially and divided by the locus count to obtain the information mean. The randomness of each information metric is subtracted from the information mean and squared to obtain the squared information difference. The squared information difference is summed sequentially and divided by the locus count to obtain the information variance. The square root of the information variance is taken to obtain the information standard deviation. The information mean plus the information standard deviation is used as the information threshold. Each information randomness is compared sequentially with the information threshold. If the information randomness is greater than the information threshold, the corresponding locus is labeled as a selected locus; if the information randomness is less than the information threshold, the corresponding locus is not labeled. The methylation levels corresponding to all selected loci are used as methylation feature data.

[0018] A pre-trained Boltzmann machine is used to compress methylation feature data into hidden feature data; each feature data in the hidden feature data represents a joint change pattern of methylation level corresponding to multiple sites in the methylation feature data.

[0019] Furthermore, the historical feature data consists of 'a' sets of hidden feature data acquired at historical moments, where 'a' is an integer greater than 1; the predefined machine learning algorithm is Lasso regression.

[0020] The method for constructing the age assessment model includes:

[0021] All hidden feature data in group a of the historical feature data are used as training data. A corresponding epigenetic age is set for each training data in group a. The training data and the corresponding epigenetic age are converted into a set of feature vectors. The model parameters are initialized, including regularization hyperparameters and regression coefficient set. The regression coefficient set includes the regression coefficient corresponding to each feature data in the hidden feature data. Each regression coefficient represents the weight of the corresponding feature data.

[0022] Each set of feature vectors is used as input to the age assessment model. The age assessment model outputs a predicted epigenetic age corresponding to each set of training data and uses the actual epigenetic age corresponding to each set of training data as the prediction target. The actual epigenetic age is the pre-set epigenetic age corresponding to the training data. The loss function value is calculated based on the feature vectors and the actual epigenetic age. The regression coefficient set is iteratively updated to minimize the loss function value. The loss change value corresponding to each iteration is calculated by subtracting the loss function value calculated in the previous iteration from the loss function value calculated in the current iteration.

[0023] A preset change threshold is set, and the loss change value calculated in each iteration is compared with the change threshold. When the loss change value is greater than or equal to the change threshold, the iteration continues. When the loss change value is less than the change threshold, the iteration ends, and the corresponding set of regression coefficients is used as the set of regression coefficients for the age assessment model, thus completing the construction of the age assessment model, which is a regularized linear regression model.

[0024] Furthermore, the method for calculating the loss function value includes:

[0025] Subtract the predicted epigenetic age from the actual epigenetic age corresponding to each training data set, and square the difference to obtain the squared difference for each training data set; sum the squared differences to obtain the total difference; take the absolute value of each regression coefficient in the regression coefficient set and sum them to obtain the absolute coefficient; multiply the absolute coefficient by the regularization hyperparameter and add it to the total difference to obtain the loss function value.

[0026] The method for iteratively updating the set of regression coefficients includes:

[0027] Preset coefficient range, where the coefficient range is the value range of the regression coefficient; randomly select c values from the coefficient range as a set of update sets, where c is the number of feature data in the hidden feature data; construct b sets of update sets in total, and the b sets of update sets are all different; set sequentially increasing digital labels for the b sets of update sets and mark them as set labels, and the range of the set labels is [1, b]; divide b + 1 by 2 as the initial window center and divide b - 1 by 2 as the initial window radius; define an iterative process, and the iterative process is: use the range of the set labels as the window range, generate d candidate solutions within the window range, 1 < d < b, and the candidate solutions correspond to the set labels one by one; sequentially use the update sets corresponding to each candidate solution as the regression coefficient sets of the age evaluation model, calculate the corresponding loss function values respectively, move the window center to the candidate solution with the smallest loss function value, and adjust the window radius.

[0028] The method for generating candidate solutions is: generate a random coefficient in the interval [-1, 1], multiply the window radius by the random coefficient and add the window center as a candidate solution; the method for adjusting the window radius is: randomly generate an adjustment coefficient in the interval [0, 1], multiply the window radius by the adjustment coefficient to obtain an adjustment radius, and adjust the window radius according to the adjustment radius.

[0029] Furthermore, the range of the simulated age is [20, 100], and the simulated age is the epigenetic age generated after systematic perturbation of the predicted age;

[0030] The method for calculating the prediction error value corresponding to each simulated age includes:

[0031] Obtain historical methylation data, where the historical methylation data is the methylation data corresponding to users with different epigenetic ages obtained at historical times; from the historical methylation data, screen out the methylation data whose epigenetic age is within the simulated age range and mark it as perturbed data; from each group of perturbed data, screen out the methylation level corresponding to each screening site respectively to form the perturbed feature data corresponding to each group of perturbed data; use the perturbed feature data with the same simulated age as a group of perturbed sets, and the perturbed sets correspond to the simulated ages one by one; according to the methylation feature data, screen out the corresponding matching feature data from each group of perturbed sets respectively; according to the matching feature data, calculate the mean feature data corresponding to each group of perturbed sets; use the pre-trained Boltzmann machine to compress each group of mean feature data into the corresponding hidden mean data; input each group of hidden mean data into the age evaluation model respectively to predict the corresponding epigenetic age and mark it as the perturbed age; subtract the perturbed age corresponding to each group of perturbed sets from the simulated age corresponding to each group of perturbed sets respectively and take the absolute value to obtain the prediction error value corresponding to each simulated age.

[0032] Further, the method for separately screening out corresponding matching feature data from each group of perturbation sets includes:

[0033] Count the number of methylation levels in the methylation feature data and label it as the matching number; take both the methylation feature data and the perturbation feature data as analysis feature data, successively add the methylation levels of each screening site in each group of analysis feature data, and then divide by the matching number to obtain the level mean corresponding to each group of analysis feature data; subtract the level mean corresponding to each methylation level in each group of analysis feature data, and then square it to obtain the squared level difference; successively add the squared level differences corresponding to each group of analysis feature data, and then take the square root to obtain the level standard deviation corresponding to each group of analysis feature data; subtract the level mean corresponding to each methylation level in each group of analysis feature data to obtain the level difference corresponding to each methylation level in each group of analysis feature data.

[0034] Take each group of perturbation feature data in each group of perturbation sets and the methylation feature data as a group of analysis sets, with the analysis sets corresponding to the perturbation feature data one by one; multiply the level standard deviations corresponding to each group of analysis sets to obtain the dispersion degree corresponding to each group of analysis sets; multiply the level differences with the same corresponding screening sites in each group of analysis sets and divide by the corresponding dispersion degree to obtain the correlation degree corresponding to each screening site in each group of analysis sets; successively add the correlation degrees corresponding to each group of analysis sets, and then divide by the matching number to obtain the total correlation degree corresponding to each group of analysis sets; sort the total correlation degrees corresponding to each group of perturbation sets from largest to smallest to generate a correlation ranking table corresponding to each group of perturbation sets; according to the forward order of the correlation ranking table, screen out y total correlation degrees from each correlation ranking table and label them as the screening degree, where y is the preset selection quantity, 1 < y < r, and r is the number of perturbation feature data in a group of perturbation sets; obtain the perturbation feature data corresponding to each screening degree and form the matching feature data.

[0035] Further, the method for calculating the mean feature data corresponding to each group of perturbation sets includes:

[0036] Successively add the methylation levels with the same corresponding screening sites in the matching feature data corresponding to each group of perturbation sets, and then divide by y to obtain the site level corresponding to each group of perturbation sets; the site levels corresponding to each group of perturbation sets constitute the mean feature data corresponding to the corresponding perturbation set.

[0037] Further, the method for calculating the relative age includes:

[0038] The process involves: acquiring previous feature data, specifically the methylation feature data obtained during the user's last epigenetic age assessment; labeling the real-time filtered methylation feature data as real-time feature data; subtracting the methylation level of the corresponding selected site from the previous feature data from the methylation level of each selected site in the real-time feature data to obtain the methylation level difference for each selected site; inputting the methylation level difference for each selected site into a trained fluctuation prediction model to predict the corresponding fluctuation age, which is the amount of fluctuation in epigenetic age caused by the methylation level difference; the training process of the fluctuation prediction model is consistent with that of the age assessment model, and both are regularized linear regression models; labeling the user's last assessed epigenetic age as the previous age; and adding the fluctuation age to the previous age to obtain the relative age.

[0039] An epigenetic age assessment system based on machine learning, implementing the aforementioned epigenetic age assessment method based on machine learning, includes:

[0040] The data acquisition module is used to collect users' methylation data in real time;

[0041] The data processing module is used to perform information gain analysis on methylation data, calculate information randomness, filter out methylation feature data from methylation data based on information randomness, and perform dimensionality reduction operation on methylation feature data to compress methylation feature data into hidden feature data.

[0042] The model building module is used to acquire historical feature data and build an age assessment model based on the historical feature data and predefined machine learning algorithms.

[0043] The age assessment module is used to input the real-time acquired hidden feature data into the age assessment model, predict the corresponding epigenetic age, and mark it as the predicted age.

[0044] The deviation correction module is used to systematically perturb the predicted age, generate different simulated ages, calculate the prediction error value corresponding to each simulated age, and take the simulated age corresponding to the prediction error value with the smallest value as the corrected age.

[0045] The secondary correction module is used to calculate the relative age, and the mean of the relative age and the corrected age is used as the user's epigenetic age.

[0046] The technical effects and advantages of the epigenetic age assessment system and method based on machine learning in this invention are as follows:

[0047] By collecting users' methylation data in real time and performing information gain analysis, key features related to epigenetic age can be dynamically screened from methylation data, avoiding the limitations of relying on static literature or research results and improving the model's adaptability and accuracy across different individuals and age groups. Employing machine learning-based modeling methods, complex aging patterns can be automatically extracted from large-scale methylation data to construct accurate epigenetic age prediction models, improving the accuracy and timeliness of epigenetic age assessment. Introducing systematic perturbations and prediction error correction mechanisms can effectively correct preliminary epigenetic age predictions, further improving the accuracy of assessment results and better reflecting an individual's actual epigenetic age. This not only provides a scientific basis for assessing individual aging rates and health status but also provides strong support for risk prediction and early intervention of age-related diseases. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of an epigenetic age assessment system based on machine learning, according to Embodiment 1 of the present invention.

[0049] Figure 2 This is a flowchart of a machine learning-based epigenetic age assessment method according to Embodiment 2 of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1

[0052] Please see Figure 1 As shown in this embodiment, an epigenetic age assessment system based on machine learning includes a data acquisition module, a data processing module, a model building module, an age assessment module, a bias correction module, and a secondary correction module; the modules are connected via wired and / or wireless means to realize data transmission between the modules.

[0053] The data acquisition module is used to collect users' methylation data in real time.

[0054] Methylation data includes the methylation level at each site, which is a CpG site. The methylation level is a continuous value between 0 and 1, representing the degree of methylation at the site; where 0 indicates that the site is completely unmethylated, i.e., 100% demethylation, and 1 indicates that the site is completely methylated, i.e., 100% methylation.

[0055] Methylation data provides information on methylation modifications in DNA, specifically the state of adding methyl groups (-CH3) to specific DNA bases (usually cytosine (C)). Methylation is an important epigenetic modification that can affect gene expression and regulation, thereby influencing cell function and organismal characteristics. Methylation data plays a central role in epigenetic age assessment, reflecting changes in gene expression regulation that occur with age. By analyzing methylation patterns at specific sites, an individual's biological age (i.e., epigenetic age) can be predicted. As the aging process progresses, the methylation levels of certain gene regions change, and these changes provide biomarkers of the aging process.

[0056] Methylation data is obtained by detecting methylation in a user's DNA using technologies such as methylation microarrays (e.g., Illumina BeadChip) and high-throughput sequencing (e.g., whole-genome sequencing). The user's DNA is extracted from biological samples such as blood, saliva, and tissue samples.

[0057] The data processing module is used to perform information gain analysis on methylation data, calculate information randomness, filter out methylation feature data from methylation data based on information randomness, and perform dimensionality reduction operation on methylation feature data to compress methylation feature data into hidden feature data.

[0058] Methods for calculating the randomness of information include:

[0059] Real-time methylation data is labeled as real-time data, and historical data is acquired, including methylation data collected from users in previous years, reflecting methylation characteristics at different ages. Historical data is obtained from the database built into the epigenetic age assessment system. Methylation levels at the same loci in both real-time and historical data are grouped into a locus set, with each locus set corresponding to a specific locus. Methylation levels with the same value in each locus set are grouped into a numerical set, and the number of methylation levels in each numerical set is counted and labeled as the level count. The number of numerical sets in each locus set is recorded and marked as the total number of levels. Each level number is multiplied by the corresponding total number of levels to obtain the level frequency for each numerical set. Using the natural constant (e, with a value greater than 2.71828) as the base and each level frequency as the argument, logarithmic operations are performed to obtain the level logarithm for each numerical set. Each level logarithm is multiplied by the corresponding level frequency to obtain the weighted level probability for each numerical set. The weighted level probabilities corresponding to the numerical sets of the same locus set are summed sequentially, and then the negative number is taken to obtain the information randomness for each locus set.

[0060] Methods for screening methylation feature data include:

[0061] The number of loci is counted and labeled as the locus count. The randomness of each information is summed sequentially and divided by the locus count to obtain the information mean. The randomness of each information is subtracted from the information mean and squared to obtain the squared information difference. The squared information difference is summed sequentially and divided by the locus count to obtain the information variance. The square root of the information variance is taken to obtain the information standard deviation. The information mean plus the information standard deviation is used as the information threshold. Each information randomness is compared sequentially with the information threshold. If the information randomness is greater than the information threshold, the corresponding locus is labeled as a selected locus; if the information randomness is less than the information threshold, the corresponding locus is not labeled. The methylation levels corresponding to all selected loci are used as methylation feature data.

[0062] A pre-trained Boltzmann machine is used to compress methylation feature data into hidden feature data. The Boltzmann machine is an existing technology, and its specific training process will not be detailed here. Hidden feature data consists of new variables extracted from the methylation feature data, which typically cannot be directly obtained from the methylation levels of loci. Each feature in the hidden feature data represents a joint change pattern in the methylation levels of multiple loci within the methylation feature data. In other words, the hidden feature data is a simplified representation meaningful for assessing epigenetic age, obtained by the Boltzmann machine by capturing the potential structures and relationships between different loci in the methylation feature data. Examples of hidden feature data include the degree of aging of the immune system (the immune system gradually declines with age, leading to methylation at relevant loci). The data includes features such as changes in methylation levels (this feature aggregates methylation levels at multiple sites associated with immune system aging, thus effectively reflecting the aging state of the immune system), DNA repair capacity (as age increases, the ability of cells to repair DNA damage gradually declines, and the methylation levels at related sites change; this feature represents methylation levels related to DNA repair mechanisms, allowing signals of decreased DNA repair capacity to be captured at a lower dimension), and hidden feature data (since hidden feature data is an optimized combination of methylation levels at a large number of sites based on methylation feature data (i.e., compressing the methylation levels of a large number of sites into a few features within hidden feature data), it can replace methylation feature data for subsequent accurate assessment of epigenetic age).

[0063] The model building module is used to acquire historical feature data and build an age assessment model based on the historical feature data and predefined machine learning algorithms.

[0064] Historical feature data consists of 'a' sets of hidden feature data acquired at historical moments, where 'a' is an integer greater than 1; historical feature data is obtained from the database built into the epigenetic age assessment system; the predefined machine learning algorithm is Lasso regression.

[0065] Methods for constructing age assessment models include:

[0066] All hidden feature data in group a of the historical feature data are used as training data. A corresponding epigenetic age is set for each training data in group a. The training data and the corresponding epigenetic age are converted into a set of feature vectors. The epigenetic age corresponding to the training data is determined by those skilled in the art during the historical evaluation of epigenetic age. Group a of training data is collected, and each group of training data is analyzed according to the actual situation to evaluate the corresponding epigenetic age. The corresponding epigenetic age is then set for each training data in group a. Model parameters are initialized, including regularization hyperparameters and a set of regression coefficients. The set of regression coefficients includes the regression coefficients corresponding to each feature data in the hidden feature data, and each regression coefficient represents the weight of the corresponding feature data. All model parameters are preset by those skilled in the art according to the actual situation, and each regression coefficient in the set of regression coefficients is preset in the interval [0,1].

[0067] Each set of feature vectors is used as input to the age assessment model. The age assessment model outputs a predicted epigenetic age corresponding to each set of training data and uses the actual epigenetic age corresponding to each set of training data as the prediction target. The actual epigenetic age is the pre-set epigenetic age corresponding to the training data. The loss function value is calculated based on the feature vectors and the actual epigenetic age. The regression coefficient set is iteratively updated to minimize the loss function value. The loss change value corresponding to each iteration is calculated by subtracting the loss function value calculated in the previous iteration from the loss function value calculated in the current iteration.

[0068] A preset change threshold is set, which is pre-set by those skilled in the art based on the actual situation. The loss change value calculated in each iteration is compared with the change threshold. When the loss change value is greater than or equal to the change threshold, the iteration continues. When the loss change value is less than the change threshold, the iteration ends, and the corresponding set of regression coefficients is used as the set of regression coefficients for the age assessment model, thus completing the construction of the age assessment model. The age assessment model is a regularized linear regression model.

[0069] The methods for calculating the loss function value include:

[0070] Subtract the predicted epigenetic age corresponding to each set of training data from the actual epigenetic age corresponding to it, square the result, and obtain the squared difference corresponding to each set of training data; sum up each squared difference in turn to obtain the total difference value; take the absolute value of each regression coefficient in the regression coefficient set and sum them up in turn to obtain the absolute coefficient; multiply the absolute coefficient by the regularization hyperparameter and then add the total difference value to obtain the loss function value.

[0071] The method for iteratively updating the regression coefficient set includes:

[0072] A preset coefficient range, where the coefficient range is the value range of the regression coefficient, and the coefficient range is preset by those skilled in the art according to the actual situation; randomly select c values from the coefficient range as a set of update sets, where c is the number of feature data in the hidden feature data; construct b groups of update sets in total, and the b groups of update sets are all different; set sequentially increasing digital labels for the b groups of update sets and mark them as set labels, and the range of the set labels is [1, b]; divide b + 1 by 2 as the initial window center and divide b - 1 by 2 as the initial window radius; define the iterative process, where the iterative process is: use the range of the set labels as the window range, generate d candidate solutions within the window range, 1 < d < b, and the candidate solutions correspond to the set labels one by one; sequentially use the update set corresponding to each candidate solution as the regression coefficient set of the age evaluation model, calculate the corresponding loss function values respectively, move the window center to the candidate solution with the smallest loss function value, and adjust the window radius.

[0073] The method for generating candidate solutions is: generate a random coefficient in the interval [-1, 1], multiply the window radius by the random coefficient and add the window center to obtain a candidate solution; the method for adjusting the window radius is: randomly generate an adjustment coefficient in the interval [0, 1], multiply the window radius by the adjustment coefficient to obtain the adjusted radius, and adjust the window radius according to the adjusted radius.

[0074] The age evaluation module is used to input the hidden feature data obtained in real time into the age evaluation model, predict the corresponding epigenetic age, and mark it as the predicted age.

[0075] The deviation correction module is used to perform systematic perturbation on the predicted age, generate different simulated ages, calculate the predicted error value corresponding to each simulated age, and use the simulated age corresponding to the smallest predicted error value as the corrected age.

[0076] The range of systematic perturbation is [20, 100], that is, the range of simulated ages is [20, 100], and the simulated age is the epigenetic age generated after performing systematic perturbation on the predicted age.

[0077] The method for calculating the predicted error value corresponding to each simulated age includes:

[0078] Historical methylation data is acquired, consisting of methylation data corresponding to users of different epigenetic ages at historical moments. This historical methylation data is obtained from the database built into the epigenetic age assessment system. From the historical methylation data, methylation data with epigenetic ages falling within the simulated age range are selected and marked as perturbation data. From each group of perturbation data, the methylation level corresponding to each selected site is extracted to constitute the perturbation feature data corresponding to each group of perturbation data. Perturbation feature data with the same simulated age are grouped into a perturbation set, and this perturbation set is compared with the simulated age. The simulated ages are mapped one-to-one; based on methylation feature data, corresponding matching feature data are selected from each perturbation set; based on the matching feature data, the mean feature data corresponding to each perturbation set is calculated; using a pre-trained Boltzmann machine, each set of mean feature data is compressed into the corresponding hidden mean data; each set of hidden mean data is input into the age assessment model to predict the corresponding epigenetic age and is marked as the perturbation age; the simulated age corresponding to each perturbation set is subtracted from the corresponding perturbation age, and the absolute value is taken to obtain the prediction error value corresponding to each simulated age.

[0079] Methods for selecting corresponding matching feature data from each set of perturbations include:

[0080] The number of methylation levels in the methylation feature data is counted and marked as the number of matches. Both methylation feature data and perturbation feature data are used as analytical feature data; that is, a set of analytical feature data is represented as a set of methylation feature data or a set of perturbation feature data. The methylation levels of each screening site in each set of analytical feature data are summed sequentially and then divided by the number of matches to obtain the mean level for each set of analytical feature data. The mean level is subtracted from each methylation level in each set of analytical feature data, and the result is squared to obtain the squared level difference. The squared level differences are summed sequentially and then squared to obtain the standard deviation of the level for each set of analytical feature data. The mean level is subtracted from each methylation level in each set of analytical feature data to obtain the level difference for each methylation level in each set of analytical feature data.

[0081] For each set of perturbation feature data in each set of perturbation sets, it is respectively combined with the methylation feature data as a set of analysis sets, and the analysis sets correspond one-to-one with the perturbation feature data; multiply the horizontal standard deviations corresponding to each set of analysis sets to obtain the degree of dispersion corresponding to each set of analysis sets; multiply the horizontal differences with the same corresponding screening sites in each set of analysis sets and divide by the corresponding degree of dispersion to obtain the correlation degree corresponding to each screening site in each set of analysis sets; add up the correlation degrees corresponding to each set of analysis sets in sequence and then divide by the matching quantity to obtain the total correlation degree corresponding to each set of analysis sets; sort the total correlation degrees corresponding to each set of perturbation sets from largest to smallest to generate a correlation ranking table corresponding to each set of perturbation sets; according to the positive order of the correlation ranking table, screen out y total correlation degrees from each correlation ranking table and mark them as the screening degree, where y is the preset selection quantity, and the selection quantity is preset by those skilled in the art according to the actual situation, 1 < y < r, and r is the number of perturbation feature data in a set of perturbation sets; obtain the perturbation feature data corresponding to each screening degree and form the matching feature data.

[0082] The method for calculating the mean feature data corresponding to each set of perturbation sets includes:

[0083] In the matching feature data corresponding to each set of perturbation sets, add up the methylation levels with the same corresponding screening sites in sequence and then divide by y to obtain the site level corresponding to each set of perturbation sets; the site levels corresponding to all sets of perturbation sets constitute the mean feature data corresponding to the corresponding perturbation sets.

[0084] It should be noted that the purpose of setting the bias correction module is to correct the systematic error of the model: since there may be systematic biases in the machine learning model, by perturbing within different age ranges, the age with the smallest prediction error can be found, thereby correcting the bias of the model and improving the reliability of epigenetic age assessment; reducing the influence of individual differences: there are individual differences in methylation data, that is, methylation data of individuals of the same age may be different, so by perturbing within different age ranges, the influence of individual differences on single prediction can be eliminated, thereby obtaining a more stable epigenetic age assessment result; improving the generalization ability of the model and avoiding overfitting: the machine learning model may overfit the training data, resulting in a decrease in the prediction ability for unseen data. Therefore, by perturbing within different age ranges, the model can make predictions over a wider range of ages, thereby enhancing the robustness of the model and improving the performance of the model on new data. <000,0176><000,0177>The secondary correction module is used to calculate the relative age and take the mean of the relative age and the corrected age as the epigenetic age of the user; <000,0178><000,0179>The method for calculating the relative age includes: <000,0180><000,0181>The process involves: acquiring previous feature data, specifically the methylation feature data obtained during the user's last epigenetic age assessment; labeling the real-time filtered methylation feature data as real-time feature data; subtracting the methylation level of the corresponding selected site from the previous feature data from the methylation level of each selected site in the real-time feature data to obtain the methylation level difference for each selected site; inputting the methylation level difference for each selected site into a trained fluctuation prediction model to predict the corresponding fluctuation age, which is the amount of fluctuation in epigenetic age caused by the methylation level difference; the training process of the fluctuation prediction model is consistent with that of the age assessment model, and both are regularized linear regression models; labeling the user's last assessed epigenetic age as the previous age; and adding the fluctuation age to the previous age to obtain the relative age.

[0088] This embodiment dynamically extracts key features related to epigenetic age from methylation data by collecting users' methylation data in real time and performing information gain analysis. This avoids the limitations of relying on static literature or research results and improves the model's adaptability and accuracy across different individuals and age groups. Employing machine learning-based modeling methods, it automatically extracts complex aging patterns from large-scale methylation data to construct an accurate epigenetic age prediction model, improving the accuracy and timeliness of epigenetic age assessment. The introduction of systematic perturbation and prediction error correction mechanisms effectively corrects the initial predicted epigenetic age, further improving the accuracy of the assessment results and better reflecting the individual's actual epigenetic age. This not only provides a scientific basis for assessing individual aging rate and health status but also offers strong support for risk prediction and early intervention of age-related diseases.

[0089] Example 2

[0090] Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A machine learning-based epigenetic age assessment method is provided, comprising:

[0091] Real-time collection of users' methylation data;

[0092] Information gain analysis is performed on methylation data to calculate information randomness. Based on information randomness, methylation feature data is screened from the methylation data, and dimensionality reduction is performed on the methylation feature data to compress the methylation feature data into hidden feature data.

[0093] Acquire historical feature data, and construct an age assessment model based on the historical feature data and predefined machine learning algorithms;

[0094] The hidden feature data acquired in real time is input into the age assessment model to predict the corresponding epigenetic age and label it as the predicted age.

[0095] The predicted age is systematically perturbed to generate different simulated ages. The prediction error value corresponding to each simulated age is calculated, and the simulated age corresponding to the prediction error value with the smallest value is taken as the corrected age.

[0096] Calculate the relative age, and use the mean of the relative age and the corrected age as the user's epigenetic age.

[0097] Example 3

[0098] This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code that, when executed by the one or more processors, can perform a machine learning-based epigenetic age assessment method as described above.

[0099] The methods or systems according to embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. Storage devices in the electronic device, such as ROM or hard disk, may store a machine learning-based epigenetic age assessment method provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components in the electronic device shown in this application may be omitted according to actual needs.

[0100] Example 4

[0101] One embodiment of this application discloses a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions. When executed by a processor, the computer-readable instructions can perform a machine learning-based epigenetic age assessment method according to an embodiment of this application, as described with reference to the above figures. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0102] Furthermore, according to embodiments of this application, the processes described in the above-referenced flowcharts can be implemented as computer software programs. For example, this application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided in this application, such as a machine learning-based epigenetic age assessment method. When this computer program is executed by a central processing unit (CPU), it performs the functions defined in the method of this application.

[0103] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0104] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based epigenetic age assessment method, characterized in that, include: Real-time collection of users' methylation data; Information gain analysis is performed on methylation data to calculate information randomness. Based on information randomness, methylation feature data is screened from the methylation data, and dimensionality reduction is performed on the methylation feature data to compress it into hidden feature data. The methylation feature data includes the methylation level of different screening sites. Acquire historical feature data, and construct an age assessment model based on the historical feature data and predefined machine learning algorithms; The hidden feature data acquired in real time is input into the age assessment model to predict the corresponding epigenetic age and label it as the predicted age. The predicted age is systematically perturbed to generate different simulated ages. The prediction error value corresponding to each simulated age is calculated, and the simulated age corresponding to the smallest prediction error value is taken as the corrected age. The range of the simulated ages is... The simulated age is the epigenetic age generated after systematically perturbing the predicted age; The method for calculating the prediction error value corresponding to each simulated age includes: Historical methylation data is acquired, consisting of methylation data corresponding to users of different epigenetic ages obtained at historical moments. From this historical methylation data, methylation data with epigenetic ages falling within the simulated age range are selected and marked as perturbation data. From each set of perturbation data, the methylation level corresponding to each selected site is selected, forming perturbation feature data for each set. Perturbation feature data with the same simulated age are grouped into a perturbation set, with each perturbation set corresponding to a simulated age. Based on the methylation feature data, corresponding matching feature data is selected from each perturbation set. Based on the matching feature data, the mean feature data corresponding to each perturbation set is calculated. Using a pre-trained Boltzmann machine, each set of mean feature data is compressed into corresponding hidden mean data. Each set of hidden mean data is input into the age assessment model to predict the corresponding epigenetic age, which is then marked as a perturbation age. The simulated age corresponding to each perturbation set is subtracted from the corresponding perturbation age, and the absolute value is taken to obtain the prediction error value for each simulated age. Calculate the relative age, and use the mean of the relative age and the corrected age as the user's epigenetic age; the method for calculating the relative age includes: The process involves: acquiring previous feature data, specifically the methylation feature data obtained during the user's last epigenetic age assessment; labeling the real-time filtered methylation feature data as real-time feature data; subtracting the methylation level of the corresponding selected site from the previous feature data from the methylation level of each selected site in the real-time feature data to obtain the methylation level difference for each selected site; inputting the methylation level difference for each selected site into a trained fluctuation prediction model to predict the corresponding fluctuation age, which is the amount of fluctuation in epigenetic age caused by the methylation level difference; the training process of the fluctuation prediction model is consistent with that of the age assessment model, and both are regularized linear regression models; labeling the user's last assessed epigenetic age as the previous age; and adding the fluctuation age to the previous age to obtain the relative age.

2. The epigenetic age assessment method based on machine learning according to claim 1, characterized in that, The methylation data includes the methylation level at each site, where the site is a CpG site, and the methylation level is a continuous value between 0 and 1; The method for calculating the randomness of information includes: The methylation data collected in real time is labeled as real-time data, and historical data is obtained, including methylation data of users collected in previous years; the methylation levels of the same sites in the real-time data and historical data are grouped into a set of sites, and the set of sites corresponds one-to-one with the sites. The methylation levels with the same value in each site set are grouped into a set of values. The number of methylation levels in each set of values ​​is counted and labeled as the level count. Count the number of sets of values ​​in each site set and label them as the total number of levels; Multiply the number of each level by the total number of levels to obtain the level frequency for each set of values; Using the natural constant as the base and each level frequency as the argument, logarithmic operations are performed to obtain the level logarithm corresponding to each set of values. Multiply each level logarithm by the corresponding level frequency to obtain the weighted level probability for each set of values; The weighted level probabilities corresponding to the numerical sets of the same site set are added together sequentially, and then the opposite number is taken to obtain the information randomness corresponding to each site set.

3. The machine learning-based epigenetic age assessment method according to claim 2, characterized in that, Methods for screening methylation feature data include: The number of loci is counted and labeled as the locus count. The randomness of each information metric is summed sequentially and divided by the locus count to obtain the information mean. The randomness of each information metric is subtracted from the information mean and squared to obtain the squared information difference. The squared information difference is summed sequentially and divided by the locus count to obtain the information variance. The square root of the information variance is taken to obtain the information standard deviation. The information mean plus the information standard deviation is used as the information threshold. Each information randomness is compared sequentially with the information threshold. If the information randomness is greater than the information threshold, the corresponding locus is labeled as a selected locus; if the information randomness is less than the information threshold, the corresponding locus is not labeled. The methylation levels corresponding to all selected loci are used as methylation feature data. A pre-trained Boltzmann machine is used to compress methylation feature data into hidden feature data; each feature data in the hidden feature data represents a joint change pattern of methylation level corresponding to multiple sites in the methylation feature data.

4. The machine learning-based epigenetic age assessment method according to claim 3, characterized in that, The historical feature data was acquired at historical moments. Group hidden feature data, It is an integer greater than 1; the predefined machine learning algorithm is Lasso regression; The method for constructing the age assessment model includes: From historical feature data All hidden feature data were used as training data. Each training data set is assigned a corresponding epigenetic age. The training data and the corresponding epigenetic age are converted into a set of feature vectors. The model parameters are initialized, including regularization hyperparameters and a set of regression coefficients. The set of regression coefficients includes the regression coefficients corresponding to each feature data in the hidden feature data. Each regression coefficient represents the weight of the corresponding feature data. Each set of feature vectors is used as input to the age assessment model. The age assessment model outputs a predicted epigenetic age corresponding to each set of training data and uses the actual epigenetic age corresponding to each set of training data as the prediction target. The actual epigenetic age is the pre-set epigenetic age corresponding to the training data. The loss function value is calculated based on the feature vectors and the actual epigenetic age. The regression coefficient set is iteratively updated to minimize the loss function value. The loss change value corresponding to each iteration is calculated by subtracting the loss function value calculated in the previous iteration from the loss function value calculated in the current iteration. A preset change threshold is set, and the loss change value calculated in each iteration is compared with the change threshold. When the loss change value is greater than or equal to the change threshold, the iteration continues. When the loss change value is less than the change threshold, the iteration ends, and the corresponding set of regression coefficients is used as the set of regression coefficients for the age assessment model, thus completing the construction of the age assessment model, which is a regularized linear regression model.

5. The machine learning-based epigenetic age assessment method according to claim 4, characterized in that, The method for calculating the loss function value includes: Subtract the predicted epigenetic age from the actual epigenetic age corresponding to each training data set, and square the difference to obtain the squared difference for each training data set; sum the squared differences to obtain the total difference; take the absolute value of each regression coefficient in the regression coefficient set and sum them to obtain the absolute coefficient; multiply the absolute coefficient by the regularization hyperparameter and add it to the total difference to obtain the loss function value. The method for iteratively updating the set of regression coefficients includes: A preset coefficient range is defined, representing the range of values ​​for the regression coefficients; random selection is then performed from this range. A set of values ​​is used as an update set. To determine the number of feature data points in the hidden feature data; a total of [number] features were constructed. Group update collection, The updated sets for each group are all different; The group updates the set with sequentially increasing numerical labels, which are then marked as set labels. The range of set labels is... ;Will Add 1 and divide by 2 to set the initial window center. Subtract 1 and divide by 2 to obtain the initial window radius; define the iteration process as follows: using the range of set labels as the window range, generate within the window range. There are 10 candidate solutions. Each candidate solution corresponds one-to-one with a set label; the updated set corresponding to each candidate solution is used as the regression coefficient set of the age assessment model in turn, and the corresponding loss function value is calculated respectively. The window center is moved to the candidate solution with the smallest loss function value, and the window radius is adjusted. The method for generating candidate solutions is as follows: in the interval A random coefficient is generated, and the window radius is multiplied by the random coefficient and added to the window center to obtain a candidate solution. The method for adjusting the window radius is as follows: within the interval... A random adjustment coefficient is generated, the window radius is multiplied by the adjustment coefficient to obtain the adjustment radius, and the window radius is adjusted according to the adjustment radius.

6. The machine learning-based epigenetic age assessment method according to claim 5, characterized in that, The method for selecting corresponding matching feature data from each set of perturbations includes: The number of methylation levels in the methylation feature data was counted and marked as the number of matches. Both methylation feature data and perturbation feature data were used as analytical feature data. The methylation levels of each screening site in each set of analytical feature data were summed sequentially, then divided by the number of matches to obtain the mean level for each set of analytical feature data. The mean level was subtracted from each methylation level in each set of analytical feature data, and the result was squared to obtain the squared level difference. The squared level differences were summed sequentially, and the square root was taken to obtain the standard deviation of the level for each set of analytical feature data. The mean level was subtracted from each methylation level in each set of analytical feature data to obtain the level difference for each methylation level in each set of analytical feature data. Each perturbation feature data set in each perturbation set is paired with methylation feature data to form an analysis set, with a one-to-one correspondence between the analysis sets and the perturbation feature data. The standard deviation of the levels corresponding to each analysis set is multiplied to obtain the dispersion of each analysis set. The level differences of the same screening sites in each analysis set are multiplied and divided by the corresponding dispersion to obtain the correlation of each screening site in each analysis set. The correlations of each analysis set are summed sequentially and then divided by the number of matches to obtain the total correlation of each analysis set. The total correlations of each perturbation set are sorted from largest to smallest to generate a correlation ranking table for each perturbation set. Based on the ascending order of the correlation ranking table, samples are selected from each correlation ranking table. The overall relevance level is determined and marked as the screening level. The preset selection quantity, , The number of perturbation feature data in a set of perturbations; obtain the perturbation feature data corresponding to each filtering level, and form matching feature data.

7. The machine learning-based epigenetic age assessment method according to claim 6, characterized in that, The method for calculating the mean feature data corresponding to each set of disturbances includes: In the matching feature data corresponding to each perturbation set, the methylation levels of the same screening sites are added sequentially, and then divided by . Obtain the locus level corresponding to each perturbation set; combine all the locus levels corresponding to each perturbation set to form the mean feature data corresponding to the perturbation set.

8. A machine learning-based epigenetic age assessment system, implementing the machine learning-based epigenetic age assessment method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect users' methylation data in real time; The data processing module is used to perform information gain analysis on methylation data, calculate information randomness, filter out methylation feature data from methylation data based on information randomness, and perform dimensionality reduction operation on methylation feature data to compress methylation feature data into hidden feature data. The model building module is used to acquire historical feature data and build an age assessment model based on the historical feature data and predefined machine learning algorithms. The age assessment module is used to input the real-time acquired hidden feature data into the age assessment model, predict the corresponding epigenetic age, and mark it as the predicted age. The deviation correction module is used to systematically perturb the predicted age, generate different simulated ages, calculate the prediction error value corresponding to each simulated age, and take the simulated age corresponding to the prediction error value with the smallest value as the corrected age. The secondary correction module is used to calculate the relative age, and the mean of the relative age and the corrected age is used as the user's epigenetic age.

Images