An age prediction method based on ensemble learning of intestinal flora prediction model

Abstract

The present invention discloses an age prediction method based on integrated learning of intestinal flora prediction model, including: obtaining raw data of human intestinal flora metagenomics, quality controlling the obtained genomic data, obtaining intestinal flora species composition and metabolic pathway composition abundance table, and constructing a sample data set; then performing feature selection on the sample data set; using the screened features combined with host regional information to construct a multi-class age prediction model, and using grid search to determine the hyperparameters that make the prediction model more accurate and train and predict each optimal prediction model to obtain an integrated age prediction method; finally, using the determined intestinal flora features and the integrated age prediction method to predict the age of the sample, and determining the key species and pathways related to age through feature interpretation. The present invention adopts an integrated learning method, which effectively improves the accuracy of age prediction; and can provide directional guidance for the adjustment of intestinal flora, thereby achieving anti-aging intervention.

Description

Technical Field

[0001] The present invention relates to an age prediction method based on ensemble learning of intestinal flora prediction models, belonging to the cross-technical field of microbiology and computer science. In particular, the present invention relates to an adult age prediction method based on ensemble learning of intestinal flora prediction models. Background Art

[0002] Human aging is a continuous process that leads to a decrease in the physiological functions of organs and then to the occurrence of diseases. With the development of intestinal flora sequencing projects such as the Human Microbiome Project, the understanding of the interaction between intestinal flora and host organisms continues to deepen. Studies have shown that the composition, function and metabolites of intestinal flora will continue to change with age, and the age-specific changes in intestinal flora will also further affect the intestinal nutritional status of the host, thereby affecting the body's chronic inflammation and weakening process. Therefore, the association between intestinal flora and aging plays an important role in regulating human health, and intervention in the aging process can be achieved based on intestinal flora. However, due to the complex species and functional composition of intestinal flora, and the high characteristic dimensions and multiple interactive relationships at the levels of species composition, gene family and metabolic pathways, there is still a lack of unified and reliable research methods in the existing technology to associate multi-angle intestinal flora characteristics with age representation, making it difficult to identify intestinal flora markers related to age.

[0003] With the development of sequencing technology, metagenomic sequencing enables comprehensive characterization of intestinal flora. Compared with the previous 16s rRNA sequencing, metagenomic sequencing can not only achieve species annotation at the species level, but also further study the metabolic pathways and gene family composition of intestinal flora, which is a favorable tool for in-depth research on intestinal flora. Metagenomic sequencing combined with bioinformatics analysis tools and machine learning model algorithms provides methodological support for exploring the relationship between intestinal flora species, functions and genes and host health status, and the continuously updated and expanded genome database provides a data basis for large-scale analysis.

[0004] With the increase in intestinal flora-related research, a huge amount of metagenomic sequencing data has been generated; at the same time, with the continuous innovation of machine learning algorithms, more and more methods and models are available. However, due to the independence and regionality of research cohorts in the field, it is difficult for researchers to make extensive use of the massive intestinal flora metagenomic sequencing data, and because of the intersection between microbiology and computer science, there is still a lag in the use of machine learning methods in the study of intestinal flora, which leads to the current research on the association between intestinal flora and age. There are problems such as low prediction method accuracy, low data utilization, redundant features, single model algorithm, and influence of host background factors.

[0005] Therefore, there is still a lack of reliable solutions to how to use machine learning methods to build an accurate age prediction model based on intestinal flora metagenomic sequencing data and then parse key intestinal flora biomarkers. The current single models of age prediction methods based on intestinal flora, such as lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosting regression tree, XGBoost and LightGBM models, all have the disadvantages of low accuracy, low data utilization, redundant features, single model algorithm and being affected by host background factors. Summary of the invention

[0006] [Technical issues]

[0007] The purpose of the present invention is to solve the problems of low accuracy, low data utilization, feature redundancy, single model algorithm and influence of host background factors in the current age prediction model and prediction method based on intestinal flora, and propose an adult age prediction method based on integrated learning of intestinal flora prediction model and its application.

[0008] [Technical solution]

[0009] The above object of the present invention is achieved through the following technical solutions:

[0010] A method for predicting adult age based on ensemble learning of intestinal flora prediction models, specifically comprising the following steps:

[0011] Step 1: Obtain intestinal flora data samples of adults of different ages as raw data, and collect background information of the hosts corresponding to the data samples, including age and regional characteristics;

[0012] Step 2: preprocessing the acquired raw data, including sorting out the host background information and quality control of the raw data, as well as annotation of the composition and relative abundance of intestinal flora species and metabolic pathways; wherein the relative abundance refers to the relative abundance of both intestinal flora species and metabolic pathways;

[0013] Step 3: Normalize the relative abundance information table of intestinal flora species and metabolic pathways obtained after annotation, and select the best algorithm for feature selection algorithm;

[0014] Step 4: Use the best algorithm to perform feature selection on the relative abundance information table of intestinal flora species and metabolic pathways; use the relative abundance information table of intestinal flora species and metabolic pathways after feature selection to build various age prediction models, and use the grid search method to screen out the best age prediction model and its hyperparameter combination corresponding to each type of model among various age prediction models;

[0015] Step 5: The best prediction model selected from various models is used for ensemble learning to construct the final ensemble age prediction method;

[0016] Step 6: Use the integrated age prediction method constructed in step 5 to predict the age of the data sample to be predicted.

[0017] According to the technical solution of the present invention, the relative abundance table can be understood as a quantitative description of the different components of the intestinal flora species, that is, the metabolic pathways, and the composition information can also be described qualitatively, that is, what components there are; at the same time, it can also be described quantitatively, such as how many of each component there are. However, relative abundance is a description method in quantitative expression, and there can also be other expressions in different units. It can be understood that the corresponding relative abundance information table contains other necessary features in the age prediction process except for the regions added later.

[0018] According to the adult age prediction method of the present invention, further, in step 1, the intestinal flora data of adults of different ages are obtained as raw data, and the host background information corresponding to the samples is collected, which specifically includes:

[0019] Step A1: Relevant literature was searched based on keywords to construct a gut flora dataset, where the keywords included human gut flora and metagenome; and the corresponding gut flora original metagenome sequencing data and relevant host background information, including age and regional characteristics at the national level, were obtained through literature review and screening combined with the NCBI SRA database;

[0020] Step A2: Use metagenomics whole genome sequencing technology for sequencing and analysis, collect adult fecal samples for sequencing and collect relevant host background information.

[0021] According to the adult age prediction method of the present invention, further, in the step 2, the raw data is preprocessed including the host background information collation and the quality control of the raw data, and the annotation of the composition and relative abundance of the species and metabolic pathways of the intestinal flora specifically includes:

[0022] Step B1: Filter the background information of the hosts corresponding to the collected data samples, select adults over 18 years old with regional labels as research subjects, and cluster them from the national level to sub-intervals according to the geographical location of regional factors;

[0023] Step B2: For the raw data of intestinal flora species obtained, use BWA and samtools combined with the human reference genome Hg38 to remove the host sequence in the sequencing data, and use Trimmomatic to perform quality control on the host-removed sequence, including removing adapters and removing low-quality sequences;

[0024] Step B3: For high-quality sequences after quality control, use MetaPhlAn and HUMAnN software to annotate the species composition and metabolic pathways of the intestinal flora, so as to obtain the composition and relative abundance information table of the species and metabolic pathways of the intestinal flora.

[0025] According to the adult age prediction method of the present invention, further, the step three specifically includes the following steps:

[0026] Step C1: Normalize the annotated intestinal flora species composition and metabolic pathway abundance table, map all data to a specific numerical range, and use zero-mean normalization:

[0027] z=(x-μ) / σ

[0028] In the formula, z is the normalized value, x is the unnormalized original abundance value, μ is the corresponding characteristic abundance mean, and σ is the standard deviation of the corresponding characteristic abundance;

[0029] Step C2: The normalized relative abundance information table of intestinal flora species and metabolic pathways was used for feature selection using univariate linear regression test, continuous variable mutual information estimation, random forest, gradient boosted regression tree, XGBoost and LightGBM, respectively. The age prediction performance before and after feature selection was compared using lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models, and the feature selection algorithm with the largest number of models with significant increase in prediction accuracy was selected as the best method. When the prediction difference changes were the same, the feature selection algorithm with the least number of features was selected as the best algorithm; and the age prediction performance was evaluated by the determination coefficient, and its calculation formula is as follows:

[0030]

[0031] Where r 2 is the coefficient of determination, y is the true value of the sample, is the regression fitting value, is the mean value of y;

[0032] According to the adult age prediction method of the present invention, further, in the step 4, feature selection is performed and an age regression prediction model is constructed, and the grid search method is used to screen out various types of age prediction models. The specific process of the best prediction model corresponding to each type of model is as follows:

[0033] Step D1: Use the selected optimal algorithm to perform feature selection on the relative abundance information table of intestinal flora species and metabolic pathways; and use the regional factor label as an additional feature to merge the intestinal flora species and metabolic pathway abundance table after feature selection, thereby reducing the impact of host background factors, namely regional factors, on the composition structure of the intestinal flora.

[0034] Step D2: The combined intestinal flora species composition and metabolic pathway relative abundance table were used to construct an age prediction model using the lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models, and the grid search method was used to screen the model hyperparameter combination with the best age prediction performance in each type of model. The model corresponding to the best hyperparameter combination of the model was the best prediction model of that type.

[0035] According to the adult age prediction method of the present invention, further, in the step 5, the best prediction model selected from various models is used for ensemble learning to construct a final ensemble age prediction method, which specifically includes the following steps:

[0036] Step E1: using the combined relative abundance information table of intestinal flora species and metabolic pathways obtained in step D1, the best prediction model of each type of model in the lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models is used for model training and prediction;

[0037] Among them, the specific steps of model training and prediction for each optimal prediction model are as follows:

[0038] Step (a), randomly divide the data set used for model training into five equal parts;

[0039] Step (b), select four of the data as training data to construct the best age prediction model f n , where the model's hyperparameters are set to the best hyperparameter combination determined by grid search;

[0040] Step (c) uses the remaining data as the verification data in the model building process, and uses the constructed prediction model to make predictions, thereby obtaining a corresponding age prediction result p n ;

[0041] Step (d), repeat the above steps (b) and (c) until all five points of data have completed training and prediction, and the prediction result p will be obtained. n Merge to obtain an age prediction result P of the full dataset size, P = {p1, p2, p3, p4, p5}, and save the best prediction model F obtained in each round of training, F = {f1, f2, f3, f4, f5};

[0042] Step E2: The best prediction model F among the various models obtained in step E1 is used to construct the best prediction model set G, where G = {F1, F2, ..., F k}, the best prediction model set G contains the model subset F obtained after each best prediction model is trained; and the constructed best prediction model set G is used to predict the test set data x t To make a prediction, the specific process is as follows:

[0043] For each best prediction model F in the prediction model set G, use each model f constructed in it to test the test set data x t Make a prediction and get five copies of x t The age prediction results are averaged to obtain the x corresponding to the best prediction model F. t Average age prediction result p t .

[0044] According to the technical solution of the present invention, the data set used for model training refers to intestinal flora data of different categories, including but not limited to the relative abundance information table of intestinal flora species and metabolic pathways involved in the content of the present invention. Test set data refers to independent data that is not included in the data set used for model training and needs to be used for age prediction, which contains intestinal flora data of the same category as the training data. Step E1 is performed for model training for each category of intestinal flora data, and the data set used for model training can be predicted using step E2 respectively; that is, for n different categories of intestinal flora data (i.e., intestinal flora species composition and metabolic pathway relative abundance information table), the above-mentioned k types of models are used for integration, and according to the aforementioned model training method, for each type of data, the best prediction model is selected through grid search for each type of model, and finally a total of n×k best prediction models are integrated for learning.

[0045] According to the technical solution of the present invention, the purpose of ensemble learning is to compensate for the deviation of prediction results of different models by integrating the prediction results of multiple different models and increasing the diversity of models. Finally, the prediction results of multiple models are weighted by the obtained weighted weights. The final integrated age prediction method is based on the aforementioned model training method. For each different category of intestinal flora data, the corresponding optimal prediction model is independently constructed for the data set used for model training, and then the final age prediction result is obtained by weighted calculation based on the prediction results of each optimal prediction model.

[0046] Step E3: Use the age prediction results obtained by each best prediction model in the above prediction model set G to build a linear regression model, so as to obtain the weighted value of the predicted age and the actual age of different best prediction models.

[0047] According to the adult age prediction method of the present invention, further, the step six specifically includes the following steps:

[0048] For the intestinal flora species composition and metabolic pathway relative abundance table of the data sample to be predicted, the best prediction model set obtained in step E2 is used for prediction, and the age prediction results of the data sample according to various models are obtained. The prediction results are weighted and calculated using the weighted weights obtained in step E3, and the final predicted value of the age of the adult corresponding to the data sample to be predicted is output.

[0049] According to the adult age prediction method of the present invention, further, the method also includes using the constructed integrated age prediction method to perform feature interpretation on all features of the intestinal flora species and their metabolic pathways after the feature selection. The purpose of the feature interpretation is to determine the features that are most relevant to age among all the features, so that they may have potential interactions related to age and aging.

[0050] Feature interpretation specifically includes the following steps:

[0051] First, the features of the prediction model input are divided into different intervals according to the data distribution. For each sample in the interval, the upper and lower limits of the interval are used to replace the original feature values, and the predicted difference is calculated.

[0052] Secondly, the cumulative local effect method was used according to the prediction results of the integrated age prediction method to determine the effect value of each feature, that is, the predicted difference calculated in each interval mentioned above, so as to obtain the key species and metabolic pathways of the intestinal flora that are most relevant to age and have the highest degree of influence;

[0053] Finally, the effect values ​​calculated for each interval were summed and centered to obtain the effect value of the corresponding feature. The calculated cumulative effect values ​​of each feature were sorted in descending order of absolute value to determine the biomarkers associated with age phenotypes in the species composition and metabolic pathways of the intestinal flora.

[0054] The beneficial effects of the present invention are:

[0055] The present invention provides an adult age prediction method based on ensemble learning of an intestinal flora prediction model, as described in step C2. The present invention comprehensively considers the metagenomic sequencing results of the samples, including the species composition and metabolic pathways of the flora, and combines the host's regional factors to reduce the differences between independent research cohorts, thereby reflecting the changes in the intestinal flora of people of different ages.

[0056] According to the prediction method of the present invention, as described in each step of step five, the present invention combines a variety of heterogeneous machine learning models and uses ensemble learning to make up for the shortcomings of various models, so that the age prediction method finally constructed can achieve better prediction performance, reduce prediction bias, and effectively improve the accuracy of model prediction; at the same time, the model is characterized and key intestinal flora biomarkers related to age phenotypic changes are determined, which has reference significance for age feature analysis based on intestinal flora, intervention in the aging process, etc.

[0057] The technical solution of the present invention adopts an integrated learning method and combines the species composition of the intestinal flora, metabolic pathways and regional factors, which effectively improves the accuracy of age prediction; and can be used as an evaluation indicator of the aging process to provide targeted guidance for the adjustment of the intestinal flora, thereby achieving anti-aging intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0059] Figure 1 is a schematic flow chart of the method for predicting adult age according to the present invention;

[0060] Figure 2 It is a schematic diagram of the prediction performance of different prediction models combined with various feature selections in the present invention;

[0061] Figure 3 is a schematic diagram of the prediction accuracy of the predicted age and the actual age in the second embodiment of the present invention;

[0062] Figure 4It is a key biomarker with a higher impact on age as explained in Example 2 of the present invention. DETAILED DESCRIPTION

[0063] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0064] Embodiment 1:

[0065] This embodiment provides an age prediction method based on ensemble learning of intestinal flora prediction models. Figure 1 As shown, the age prediction method comprises the following steps:

[0066] Step 1: Obtain intestinal flora data samples of adults of different ages as raw data, and collect background information of the hosts corresponding to the data samples, including age and regional characteristics;

[0067] Step 2: preprocessing the acquired raw data, including sorting out the host background information and quality control of the raw data, and annotation of the composition and relative abundance of intestinal flora species and metabolic pathways, wherein the relative abundance refers to the relative abundance of both intestinal flora species and metabolic pathways;

[0068] Step 3: Normalize the relative abundance information table of the annotated intestinal flora species and metabolic pathways, and select the best algorithm for feature selection;

[0069] Step 4: Use the best algorithm to perform feature selection on the relative abundance information table of intestinal flora species and metabolic pathways; use the relative abundance data of intestinal flora species and metabolic pathways after feature selection and add additional regional characteristics to construct an age regression prediction model, and use the grid search method to screen out the model hyperparameter combination with the best age prediction performance, and select the best prediction model based on the age prediction accuracy after the model parameters are adjusted;

[0070] Step 5: Use the selected best prediction model through ensemble learning to construct the final integrated age prediction method;

[0071] Step 6: Use the constructed integrated age prediction method to predict the age of adults corresponding to the data sample.

[0072] According to the adult age prediction method of this embodiment, further, in step 1, the intestinal flora data of adults of different ages are obtained as raw data, and the host background information corresponding to the samples is collected, specifically including:

[0073] Step A1: Relevant literature was searched based on keywords to construct a gut flora dataset, where the keywords included human gut flora and metagenome; and the corresponding gut flora original metagenome sequencing data and relevant host background information, including age and regional characteristics at the national level, were obtained through literature review and screening combined with the NCBI SRA database;

[0074] Step A2: Use metagenomics whole genome sequencing technology for sequencing and analysis, collect adult fecal samples for sequencing and collect relevant host background information.

[0075] Furthermore, in step 2, the preprocessing of the raw data includes the collation of host background information and the quality control of the raw data. The annotation of the composition and relative abundance of the species and metabolic pathways of the intestinal flora specifically includes:

[0076] Step B1: Filter the background information of the hosts corresponding to the collected data samples, select adults over 18 years old with regional labels as research subjects, and cluster them from the national level to sub-intervals according to the geographical location of regional factors;

[0077] Step B2: For the raw data of intestinal flora species obtained, use BWA and samtools combined with the human reference genome Hg38 to remove the host sequence in the sequencing data, and use Trimmomatic to perform quality control on the host-removed sequence, including removing adapters and removing low-quality sequences;

[0078] Step B3: For high-quality sequences after quality control, use MetaPhlAn and HUMAnN software to annotate the species composition and metabolic pathways of the intestinal flora, so as to obtain the composition and relative abundance information table of the species and metabolic pathways of the intestinal flora.

[0079] Furthermore, the step three specifically includes the following steps:

[0080] Step C1: Normalize the annotated intestinal flora species composition and metabolic pathway abundance table, map all data to a specific numerical range, and use zero-mean normalization:

[0081] z=(x-μ) / σ

[0082] In the formula, z is the normalized value, x is the unnormalized original abundance value, μ is the corresponding characteristic abundance mean, and σ is the standard deviation of the corresponding characteristic abundance;

[0083] Step C2: The normalized relative abundance information table of intestinal flora species and metabolic pathways was used for feature selection using univariate linear regression test, continuous variable mutual information estimation, random forest, gradient boosted regression tree, XGBoost and LightGBM, respectively. The age prediction performance before and after feature selection was compared using lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models, and the feature selection algorithm with the largest number of models with significant increase in prediction accuracy was selected as the best method. When the prediction difference changes were the same, the feature selection algorithm with the least number of features was selected as the best algorithm; and the age prediction performance was evaluated by the determination coefficient, and its calculation formula is as follows:

[0084]

[0085] Where r 2 is the coefficient of determination, y is the true value of the sample, is the regression fitting value, is the mean value of y;

[0086] Furthermore, in step 4, feature selection is performed and an age regression prediction model is constructed. The specific process of using a grid search method to screen out the best performance model is as follows:

[0087] Step D1: Use the selected optimal algorithm to perform feature selection on the relative abundance information table of intestinal flora species and metabolic pathways; and use the regional factor label as an additional feature to merge the intestinal flora species and metabolic pathway abundance table after feature selection, thereby reducing the impact of host background factors, namely regional factors, on the composition structure of the intestinal flora.

[0088] Step D2: Use the lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosting regression tree, XGBoost and LightGBM models to construct age prediction models for the merged intestinal flora species composition and metabolic pathway table, and use the grid search method to find the best hyperparameter combination for each model. The model hyperparameter combination with the highest age prediction accuracy is taken as the best parameter. The best prediction model is selected based on the age prediction accuracy after the model parameters are adjusted, and the best hyperparameter combination obtained through grid search is saved.

[0089] Furthermore, in step 5, the best prediction model selected from various models is used for ensemble learning to construct a final ensemble age prediction method, which specifically includes the following steps:

[0090] Step E1: using the combined relative abundance information table of intestinal flora species and metabolic pathways obtained in step D1, the best prediction model of each type of model in the lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models is used for model training and prediction;

[0091] Among them, the specific steps of model training and prediction for each optimal prediction model are as follows:

[0092] Step (a), randomly divide the data set used for model training into five equal parts;

[0093] Step (b), select four of the data as training data to construct the best age prediction model f n , where the model's hyperparameters are set to the best hyperparameter combination determined by grid search;

[0094] Step (c) uses the remaining data as the verification data in the model building process, and uses the constructed prediction model to make predictions, thereby obtaining a corresponding age prediction result p n ;

[0095] Step (d), repeat the above steps (b) and (c) until all five points of data have completed training and prediction, and obtain the prediction result p n Merge to obtain an age prediction result P of the full dataset size, P = {p1, p2, p3, p4, p5}, and save the best prediction model F = {f1, f2, f3, f4, f5} obtained in each round of training;

[0096] Step E2: The best prediction model F among the various models obtained in step E1 is used to construct the best prediction model set G = {F1, F2, ..., F k}, the best prediction model set G contains the model subset F obtained after each best prediction model is trained; and the constructed best prediction model set G is used to predict the test set data x t To make a prediction, the specific process is as follows:

[0097] For each best prediction model F in the prediction model set G, use each model f constructed in it to test the test set data x t Make a prediction and get five copies of x t The age prediction results are averaged to obtain the x corresponding to the best prediction model F. t Average age prediction result p t .

[0098] According to the technical solution of the present invention, the data set used for model training refers to intestinal flora data of different categories, including but not limited to the relative abundance information table of intestinal flora species and metabolic pathways involved in the content of the present invention. Test set data refers to independent data that is not included in the data set used for model training and needs to be used for age prediction, which contains intestinal flora data of the same category as the training data. Step E1 is performed for model training for each category of intestinal flora data, and the data set used for model training can be predicted using step E2 respectively; that is, for n different categories of intestinal flora data (i.e., intestinal flora species composition and metabolic pathway relative abundance information table), the above-mentioned k types of models are used for integration, and according to the aforementioned model training method, for each type of data, the best prediction model is selected through grid search for each type of model, and finally a total of n×k best prediction models are integrated for learning.

[0099] According to this embodiment, the purpose of ensemble learning is to compensate for the deviation of prediction results of different models by integrating the prediction results of multiple different models and increasing the diversity of models. Finally, the prediction results of multiple models are weighted and calculated by the obtained weighted weights. The final integrated age prediction method is based on the aforementioned model training method. For each different category of intestinal flora data, the corresponding optimal prediction model is independently constructed for the data set used for model training, and then the final age prediction result is obtained by weighted calculation based on the prediction results of each optimal prediction model.

[0100] Step E3: Use the age prediction results obtained by each best prediction model in the above prediction model set G to build a linear regression model, so as to obtain the weighted value of the predicted age and the actual age of different best prediction models.

[0101] According to the adult age prediction method of the present invention, further, the step six specifically includes the following steps:

[0102] For the intestinal flora species composition and metabolic pathway relative abundance table of the data sample to be predicted, the best prediction model set obtained in step E2 is used for prediction, and the age prediction results of the data sample according to various models are obtained. The prediction results are weighted and calculated using the weighted weights obtained in step E3, and the final predicted value of the age of the adult corresponding to the data sample to be predicted is output.

[0103] Furthermore, the method also includes using the constructed integrated age prediction method to perform feature interpretation on all features of the intestinal flora species and their metabolic pathways after the feature selection. The purpose of the feature interpretation is to determine the features that are most relevant to age among all the features, so that they may have potential interactions related to age and aging.

[0104] Feature interpretation specifically includes the following steps:

[0105] First, the features of the prediction model input are divided into different intervals according to the data distribution. For each sample in the interval, the upper and lower limits of the interval are used to replace the original feature values, and the predicted difference is calculated.

[0106] Secondly, the cumulative local effect method was used according to the prediction results of the integrated age prediction method to determine the effect value of each feature, that is, the predicted difference calculated in each interval mentioned above, so as to obtain the key species and metabolic pathways of the intestinal flora that are most relevant to age and have the highest degree of influence;

[0107] Finally, the effect values ​​calculated for each interval were summed and centered to obtain the effect value of the corresponding feature. The calculated cumulative effect values ​​of each feature were sorted in descending order of absolute value to determine the biomarkers associated with age phenotypes in the species composition and metabolic pathways of the intestinal flora.

[0108] The technical solution of this embodiment adopts an integrated learning method and combines the species composition of the intestinal flora, metabolic pathways and regional factors, which effectively improves the accuracy of age prediction; and can be used as an evaluation indicator of the aging process to provide targeted guidance for the adjustment of the intestinal flora, thereby achieving anti-aging intervention.

[0109] Embodiment 2

[0110] According to this embodiment, the age prediction method based on intestinal flora model ensemble learning of the present invention has the following specific implementation steps:

[0111] (1) The raw data obtained from gene databases such as NCBI (National Center for Biotechnology Information) and EMBL (European Molecular Biology Laboratory) were preprocessed, including the collation of host background information and quality control of the raw data, as well as the annotation of the composition and abundance of species and metabolic pathways.

[0112] The background information of the samples collected was clustered from the country level to the sub-region level using the geographical location of the geographical factors according to the United Nations subregions, standard country or area codes for statistical use.

[0113] For the obtained raw data of intestinal flora, BWA was used to align the sample sequences with the human genome sequence Hg38, and samtools was used to filter out the unaligned sequences in the raw data and sort and merge them, thereby removing the host sequence information in the raw data. Trimmomatic was also used to remove low-quality, unqualified double-end sequences, and adapters in the data after host removal to improve the sequence quality.

[0114] For high-quality sequences after quality control, MetaPhlAn2 and HUMAnN2 software were used to annotate species composition and metabolic pathways, and the relative abundance values ​​of the annotation results were calculated to obtain the composition and abundance information table of species and metabolic pathways;

[0115] (2) The annotated species composition and metabolic pathway abundance tables are high-dimensional data, with the number of features being 904 and 468, respectively. Before performing feature selection on the annotated species composition and metabolic pathway abundance tables, the data were normalized to map all the data into a specific numerical range.

[0116] This experiment uses zero-mean normalization:

[0117] z=(x-μ) / σ

[0118] Where z is the normalized value, x is the unnormalized original abundance value, μ is the corresponding characteristic abundance mean, and σ is the standard deviation of the corresponding characteristic abundance.

[0119] The basic process of feature selection from the intestinal flora dataset is:

[0120] The normalized species and metabolic pathway tables were subjected to feature selection using univariate linear regression test, continuous variable mutual information estimation, random forest, gradient boosted regression tree, XGBoost and LightGBM models, respectively.

[0121] After feature selection, the age prediction performance of the original data and the data after feature selection was calculated using the lasso algorithm, elastic net, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models using 10-fold 5-fold cross validation.

[0122] The age prediction performance was evaluated by determining the following statement:

[0123]

[0124] In the formula, r2 is the determination coefficient, y is the true value of the sample, is the regression fitting value, is the mean value of y.

[0125] The T test was used to compare the performance of age prediction of the data after feature selection with that of the original unselected data. The feature selection method with the largest number of models that significantly improved the prediction performance compared with the original data was selected as the best method; when the number of models with the same performance improvement was the same, the feature selection method with the smallest number of features was selected as the best method.

[0126] like Figure 2 As shown in the figure, the age prediction performance of different prediction models combined with various feature selection algorithms was comprehensively compared, where "*" indicates that there is a significant difference between the prediction performance after feature selection and without feature selection. The results show that the feature selection method based on gradient boosting regression tree has the best feature selection performance and can effectively reduce the feature dimensions of species composition (224) and metabolic pathways (223);

[0127] (3) The lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosting regression tree, XGBoost and LightGBM models were used to construct age prediction models for the screened species composition and metabolic pathway table, and the grid search method was used to find the optimal hyperparameter combination of each model.

[0128] The grid search method is used to find the best hyperparameters for each model. At the same time, a 5-fold 10-fold cross validation is performed on the basis of the grid search to avoid random fluctuations in the model prediction performance during the grid search process. For the lasso algorithm, the optimal combination of penalty term alpha and maximum number of iterations is found; for the elastic network, the optimal combination of penalty term constant alpha, penalty term mixing parameter l1_ratio and maximum number of iterations is found; for Bayesian Ridge regression, the optimal combination of maximum number of iterations, alpha_1, alpha_2, lambda_1 and lambd_2 is found; for support vector machine, the optimal combination of regularization parameter C and kernel function is found; for random forest, the optimal combination of number of trees, node splitting conditions and minimum number of node samples is found; for gradient boosted decision tree, the optimal combination of loss function, learning rate, number of boosts and minimum number of leaf nodes is found; for XGBoost, the optimal combination of number of boosts, maximum tree depth and learning rate is found; for LightGBM, the optimal combination of maximum number of leaves, maximum tree depth, learning rate and number of boosts is found.

[0129] After grid search, the species prediction accuracy of the lasso algorithm was 0.296 (alpha = 0.2, maximum number of iterations = 1000), and the pathway prediction accuracy was 0.326 (alpha = 0.2, maximum number of iterations = 1000); the species prediction accuracy of the elastic network algorithm was 0.313 (alpha = 0.2, l1_traio = 0.1, maximum number of iterations = 1000), and the pathway prediction accuracy was 0.333 (alpha = 0.2, l1_traio = 0.2, maximum number of iterations = 1000); the species prediction accuracy of the Bayesian Ridge regression algorithm was 0.303 (alpha_1 = 5e-7, alpha _2=5e-6, lambda_1=5e-6, lambda_2=5e-7, maximum number of iterations=300), pathway prediction accuracy is 0.331(alpha_1=5e-7, alpha_2=5e-6, lambda_1=5e-6, lambda_2=5e-7, maximum number of iterations=300); support vector machine species prediction accuracy is 0.321(C=5.0, kernel function=rbf), pathway prediction accuracy is 0.336(C=5.0, kernel function=rbf); random forest species prediction accuracy is 0.444(node ​​segmentation condition=mse, minimum node sample size=1 , number of iterations = 1000), the pathway prediction accuracy is 0.446 (node ​​segmentation condition = mse, node minimum sample size = 1, number of iterations = 1000); the species prediction accuracy of the gradient boosting decision tree is 0.505 (learning rate = 0.1, loss function = ls, number of boosts = 500, minimum sample size of leaf nodes = 10), and the pathway prediction accuracy is 0.450 (learning rate = 0.05, loss function = ls, number of boosts = 500, minimum sample size of leaf nodes = 10); the species prediction accuracy of XGBoost is 0.507 (learning rate = 0.1, maximum tree depth = 6, number of boosts = 500), and the pathway prediction accuracy is 0 .445 (learning rate = 0.5, maximum tree depth = 10, number of boosts = 300); XGBRF species prediction accuracy is 0.417 (learning rate = 1, maximum tree depth = 10, number of boosts = 500), and pathway prediction accuracy is 0.439 (learning rate = 1, maximum tree depth = 10, number of boosts = 500); LightGBM species prediction accuracy is 0.530 (learning rate = 0.05, maximum tree depth = -1, maximum number of leaves = 15, number of boosts = 500), and pathway prediction accuracy is 0.462 (learning rate = 0.05, maximum tree depth = -1, maximum number of leaves = 15, number of boosts = 500);

[0130] (4) Multi-model ensemble learning, and ensemble age prediction based on the best model after hyperparameter adjustment; prediction model ensemble learning is divided into two steps:

[0131] Step 1: For a single optimal prediction model, the training data set is randomly divided into five parts, and four of them are used to build an age prediction model, where the hyperparameters are set to the optimal parameter values ​​determined by grid search. The remaining part and the test set are predicted using the constructed model. The above process is repeated until all five training sets have completed the prediction, thereby obtaining the age prediction data of one training set, and the model obtained in each training process is saved;

[0132] Step 2: Use the optimized lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosting regression tree, XGBoost and LightGBM models as base learners to complete step 1 on species composition and metabolic pathway data; then use the age prediction results obtained by various models, and use the linear regression model as the weighting method to obtain the weighted values ​​of age predictions of different models and the actual age, as well as the final prediction results.

[0133] Through the integrated learning of multiple heterogeneous machine learning prediction models and different metagenomics annotation data, the deviation in the model prediction process can be effectively reduced, thereby achieving comprehensive utilization of intestinal flora metagenomics data. The final age prediction performance of the integrated method is 0.578; the regional factors after clustering are further added as additional features for correction, such as Figure 3 There is an obvious linear relationship between the predicted age value and the actual age. The age prediction performance of the method based on ensemble learning is 0.598, indicating that the age prediction method constructed by the present invention has good prediction performance.

[0134] (5) Predictions are made based on the final prediction method obtained by model ensemble learning, and the effect value of each feature is determined based on the prediction results, thereby obtaining the key species and pathways related to age. The input features are divided into different intervals according to the data distribution. For each sample in each interval, the upper and lower limits of the interval are used to replace the original feature values, and the predicted differences are calculated; finally, the effect indices calculated for each interval are summed and centered to obtain the effect value of the corresponding feature. The calculated cumulative feature effect values ​​are sorted in descending order of absolute value to determine the key species and metabolic pathways related to age.

[0135] like Figure 4As shown, biomarkers with different degrees of influence on age were determined by ranking the average cumulative effect values. Markers with average effect values ​​greater than zero indicate that as their relative abundance increases, the predicted age will increase in the prediction model, which may be a potential unfavorable factor; conversely, as the abundance increases, the predicted value of the prediction model can be reduced, which may have potential anti-aging ability; finally, the adult age prediction and feature interpretation based on ensemble learning of the intestinal flora model were completed.

[0136] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0137] 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 principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An age prediction method based on ensemble learning of intestinal flora prediction model, characterized in that: The prediction method comprises the following steps: Step 1: Obtain intestinal flora data samples of adults of different ages as raw data, and collect background information of the hosts corresponding to the data samples, including age and regional characteristics; Step 2: Preprocessing the raw data obtained in step 1, including sorting out the host background information and quality control of the raw data, and annotating the intestinal flora species and metabolic pathways to obtain a relative abundance information table; wherein the relative abundance refers to the relative abundance of both the intestinal flora species and the metabolic pathways; Step 3: Normalize the relative abundance information table obtained after annotation and select the best algorithm for feature selection; Step 4: Use the best algorithm to perform feature selection on the relative abundance information table of intestinal flora species and metabolic pathways; use the relative abundance information table after feature selection to build multiple age prediction models; and use the grid search method to screen the best prediction model in each category of models; Step 5: The best prediction model selected from various models is used for ensemble learning to construct the final ensemble age prediction method; Step 6: Use the integrated age prediction method constructed in step 5 to predict the age of the data sample to be predicted; The step 2 also includes: Step B1: Filter the background information of the hosts corresponding to the collected data samples, select adults over 18 years old with regional labels as research subjects, and cluster them from the national level to sub-intervals according to the geographical location of regional factors; Step B2: For the raw data of intestinal flora species obtained, use BWA and samtools combined with the human reference genome Hg38 to remove the host sequence in the sequencing data, and use Trimmomatic to perform quality control on the host-removed sequence, including removing adapters and removing low-quality sequences; Step B3: For high-quality sequences after quality control, use MetaPhlAn and HUMAnN software to annotate species composition and metabolic pathways, thereby obtaining the composition and relative abundance information table of intestinal flora species and metabolic pathways; The step three also includes: Step C1: Normalize the composition and relative abundance information table of the annotated intestinal flora species and metabolic pathways, map all data into a specific numerical range, and use zero-mean normalization: z=(x-μ) / σ In the formula, z is the normalized value, x is the unnormalized original abundance value, μμ is the mean abundance of the feature corresponding to x, and σ is the standard deviation of the corresponding feature abundance; Step C2: The normalized relative abundance information table of intestinal flora species and metabolic pathways was used for feature selection using univariate linear regression test, continuous variable mutual information estimation, random forest, gradient boosted regression tree, XGBoost and LightGBM. The age prediction performance before and after feature selection was compared using lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models, and the feature selection algorithm with the most significant increase in the number of models with the most prediction accuracy was selected as the best method; when the prediction difference changes were the same, the feature selection algorithm with the least number of features was selected as the best algorithm; The step 4 also includes: Step D1: Use the selected optimal algorithm to perform feature selection on the relative abundance information table of intestinal flora species and metabolic pathways, and use the regional factor label as an additional feature to merge with the relative abundance information table of intestinal flora species and metabolic pathways after feature selection; Step D2: The combined intestinal flora species composition and metabolic pathway relative abundance information table was used to construct an age prediction model using the lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models, and the grid search method was used to screen the model hyperparameter combination with the best age prediction performance in each of the above models. The model corresponding to the best hyperparameter combination was the best prediction model in this type of model; The step five also includes: Step E1: using the combined relative abundance information table of intestinal flora species and metabolic pathways obtained in step D1, the best prediction model of each type of model in the lasso algorithm, elastic network, Bayesian ridge regression, support vector machine, random forest, gradient boosted regression tree, XGBoost and LightGBM models is used for model training and prediction; Step E2: The best prediction model F among the various models obtained in step E1 is used to construct the best prediction model set G, where G = {F1, F2, ..., F k }; and use the optimal prediction model set G to predict the test set data that needs age prediction; Step E3: constructing a linear regression model using the age prediction results obtained by each best prediction model in the prediction model set G, thereby obtaining weighted values ​​of the predicted age and the actual age of different best prediction models; In step E1, the model training and prediction steps for each optimal prediction model include: Step (a): randomly divide the dataset used for model training into five equal parts; Step (b): Select four of the data as training data to construct the best age prediction model f n , where the model's hyperparameters are set to the best hyperparameter combination determined by grid search; Step (c): Use the remaining data as the verification data in the model building process, and use the constructed prediction model to make predictions, so as to obtain a corresponding age prediction result p n ; Step (d): Repeat the process of step (b) and step (c) until all five points of data have completed training and prediction, and obtain the prediction result p n Merge to obtain an age prediction result P of the full data set size, P = {p1, p2, p3, p4, p5}, and save the best prediction model F obtained in each round of training, F = {f1, f2, f3, f4, f5}.

2. The prediction method according to claim 1, characterized in that: The step one also includes: Step A1: Relevant literature was searched based on keywords to construct a gut flora dataset, where the keywords included human gut flora and metagenome; and the corresponding gut flora original metagenome sequencing data and corresponding host background information, including age and regional characteristics at the national level, were obtained through literature review and screening combined with the NCBI SRA database; Step A2: Use metagenomics whole genome sequencing technology for sequencing and analysis, collect adult fecal samples for sequencing and collect relevant host background information.

3. The prediction method according to claim 1, characterized in that: The step 5 further includes: The final integrated age prediction method is to independently construct the corresponding best prediction model for each different category of intestinal flora data according to the model training method of step E1, and then obtain the final age prediction result through weighted calculation based on the prediction results of each best prediction model.

4. The prediction method according to claim 1, characterized in that: The step six also includes: For the intestinal flora species composition and metabolic pathway relative abundance table of the data sample to be predicted, use the best prediction model set obtained in step E2 in step five to make predictions, obtain the age prediction results of the data sample according to various models, and use the weighted weights obtained in step E3 to perform weighted calculations on the prediction results, and output the final predicted value of the age of the adult corresponding to the data sample to be predicted.

5. The prediction method according to claim 1, characterized in that: The prediction method further comprises: All features of the intestinal flora species and their metabolic pathways after feature selection were interpreted. The purpose of feature interpretation was to determine the feature that was most relevant to age among all features, so as to determine the possible age-related interactions of the feature.

Images