Metabolic disease onset risk prediction model and construction method thereof

Abstract

The invention discloses a metabolic disease onset risk prediction model and a construction method thereof. Based on serum proteome data of a large-scale longitudinal population queue, a health aging score including 22 serum proteins is developed, and the onset risk of metabolic diseases of middle-aged and elderly individuals within 9 years can be measured. The health aging score is helpful for preventing and treating aging-related metabolic diseases and promoting health aging.

Description

Technical Field

[0001] The present invention belongs to the field of biotechnology, and in particular relates to a metabolic disease risk prediction model and a construction method thereof. Background Art

[0002] Aging is a gradual and irreversible physiological process, which is manifested by the general degeneration of multi-dimensional functions of cells, tissues, organs, etc., and is usually accompanied by the occurrence and development of chronic diseases, especially metabolic diseases. Chronic metabolic diseases related to aging mainly include hyperlipidemia (fasting plasma total cholesterol (TC) ≥ 6.2mmol / L and / or low-density lipoprotein cholesterol (LDL-C) ≥ 4.1mmol / L and / or triglycerides (TG) ≥ 2.3mmol / L and / or high-density lipoprotein cholesterol (HDL-C) ≤ 1.0mmol / L), hypertension (systolic blood pressure (SBP) ≥ 140mmHg and / or diastolic blood pressure (DBP) ≥ 90mmHg), type 2 diabetes (fasting blood glucose (FBG) ≥ 7.0mmol / L and / or) and fatty liver (ultrasound diagnosis).

[0003] As people age, the risk of these metabolic diseases increases significantly. Their complications and the cardiovascular diseases they cause pose serious health threats to the elderly, causing a heavy medical and public health burden. Therefore, it is particularly important to effectively identify and intervene in aging-related metabolic diseases, which will help reduce disease damage and medical burden.

[0004] At present, clinical anthropometry, blood biochemical testing, ultrasound and other diagnostic and treatment methods can effectively screen for metabolic diseases. However, these diagnostic and treatment methods can only reflect the patient's current health and disease status, and cannot predict the long-term risk of individual metabolic diseases in advance. Therefore, there is an urgent need to develop biomarker tools that can measure the health of individuals and predict the risk of metabolic diseases. Through this tool, the risk of metabolic diseases during the aging process can be predicted in advance, which is conducive to early dietary, lifestyle and even drug interventions, thereby preventing the occurrence of metabolic diseases and their complications.

[0005] The invention application with publication number CN111599476A discloses a prediction model for hypertension, a method for establishing the same, and a biomarker for predicting hypertension. By screening the differential metabolites between normal people and hypertensive people, three groups of differential metabolites with significant differences are obtained, namely adenine nucleotide, 1-methylnicotinamide, and inosine riboside; the differential metabolites are analyzed; and a prediction model is established.

[0006] The invention application with publication number CN112786204A discloses a machine learning diabetes risk prediction method and application, the method includes: data acquisition module: acquiring metabolomics data; data preprocessing module: preprocessing the acquired data; machine learning module: for the purpose of predicting diabetes risk, building a prediction model based on machine learning algorithm and metabolomics data; display output module: testing the obtained samples to be predicted and outputting the prediction results. The prediction result of 1 indicates that there is a risk of diabetes, and when the result is 0, it indicates that there is no risk of diabetes.

[0007] The above two existing technologies both predict the risk of a single metabolic-related disease through metabolites.

[0008] Protein homeostasis is crucial to maintaining the normal functioning of the body, and disturbances in proteome homeostasis across tissues and organs are closely related to the progression of aging-related metabolic diseases. The blood proteome contains proteins from various cells, extracellular fluids, tissues, and organs, and is an ideal biomarker for identifying the health status of various tissues and organs. Summary of the invention

[0009] In view of the above-mentioned deficiencies in the prior art, the present invention provides a metabolic disease risk prediction model and a construction method thereof.

[0010] The present invention first provides a combination of protein markers related to the prediction of the risk of metabolic disease, which consists of P04114_APOB, P55056_APOC4, P01009_A1AT, P02654_APOC1, P02750_A2GL, P10643_CO7, P02748_CO9, Q9Y613_FHOD1, P01023_A2MG, Q99684_GFI1, P3 It is composed of 22 serum proteins: 6955_PEDF, P18428_LBP, P02774_VTDB, P02751_FINC, P08697_A2AP, Q9UPU7_TBD2B, A6NFN9_ANKUB, P06312_KV401, P27169_PON1, P23142_FBLN1, P02787_TRFE, and P00740_FA9.

[0011] The present invention further provides the use of a reagent for detecting the metabolic disease onset risk prediction-related protein marker combination in the preparation of a metabolic disease onset risk prediction product.

[0012] Preferably, the reagent is a reagent for performing proteomic analysis on a serum sample of an individual.

[0013] The present invention further provides a metabolic disease risk prediction product, comprising a reagent for detecting a combination of protein markers related to the metabolic disease risk prediction.

[0014] The present invention also provides a method for constructing a metabolic disease risk prediction model, comprising the following steps:

[0015] (1) Sample processing: serum samples were collected as training set, internal validation set, and external validation cohort, and each serum sample was subjected to proteomic analysis;

[0016] Among them, the individuals from whom the serum samples were derived were divided into healthy and unhealthy. Individuals with any of the 14 aging-related chronic diseases were considered unhealthy, and otherwise they were considered healthy.

[0017] The 14 chronic diseases related to aging are: hyperlipidemia, type 2 diabetes, hypertension, stroke, heart disease, fatty liver, cirrhosis, hepatitis, kidney disease, cancer, gout, rheumatoid arthritis, cataracts, and Parkinson's disease;

[0018] (2) The training set was used to construct a random forest model to distinguish between healthy and unhealthy samples, using age-related serum proteins as feature variables;

[0019] (3) According to the importance of variables in the model evaluated by mean accuracy decline, the random forest model with 22 feature variables was selected as the final model. The 22 feature variables were: P04114_APOB, P55056_APOC4, P01009_A1AT, P02654_APOC1, P02750_A2GL, P10643_CO7, P02748_CO9, Q9Y613_FHOD1, P01023_A 2MG, Q99684_GFI1, P36955_PEDF, P18428_LBP, P02774_VTDB, P02751_FINC, P08697_A2AP, Q9UPU7 _TBD2B, A6NFN9_ANKUB, P06312_KV401, P27169_PON1, P23142_FBLN1, P02787_TRFE, P00740_FA9.

[0020] Preferably, in step (2), the initial input variables of the random forest model are 86 age-related serum proteins.

[0021] Preferably, in step (2), in the random forest model, the number of decision trees trained is set to 1000, and the number of variables randomly selected for each decision tree is 4.

[0022] Preferably, in step (3), the variables are sorted in descending order of importance, the top n most important variables are selected to retrain the random forest model, the average error of each model is calculated by a ten-fold cross validation method, and the model with the smallest average error is selected as the final model;

[0023] The value of n is a positive integer from 1 to the total number of all variables.

[0024] The present invention also provides a metabolic disease risk prediction model constructed by the construction method.

[0025] Preferably, when the metabolic disease risk prediction model is used, a serum sample from the individual to be predicted is subjected to proteomic analysis to obtain relative abundance information of the 22 characteristic variables, which are then input into the metabolic disease risk prediction model to generate a healthy aging score, and the probability of being classified as healthy is calculated as the healthy aging score. The higher the healthy aging score, the lower the predicted risk of metabolic diseases, especially the risk of new-onset hyperlipidemia, hypertension, diabetes and fatty liver.

[0026] Beneficial effects of the present invention:

[0027] Based on serum proteome data from a large-scale longitudinal population cohort, the present invention develops a healthy aging score that includes 22 serum proteins, which can measure the risk of developing metabolic diseases in middle-aged and elderly individuals within 9 years. The healthy aging score will help prevent and treat aging-related metabolic diseases and promote healthy aging.

[0028] When proteomic analysis is performed on individual serum samples, only a small amount (e.g. 1 μl) of blood sample is needed to complete the test. In addition, the test is fast, with results available in about 2 to 3 hours. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is an experimental flow chart in the embodiment of the present invention.

[0030] Figure 2 22 serum proteins in the random forest model for this application.

[0031] Figure 3 The figure shows the validation results of the model in the internal validation set and the external validation cohort.

[0032] Figure 4 This is a graph showing the relationship between the baseline protein healthy aging score and the risk of metabolic diseases in the next 9 years. DETAILED DESCRIPTION

[0033] Example 1

[0034] We propose a healthy aging score (including 22 serum proteins) that can measure an individual's current health and indicate the risk of metabolic diseases in the next 9 years.

[0035] Among them, health is defined by 14 aging-related chronic diseases, and samples suffering from any of the diseases are defined as unhealthy. The 14 diseases are: hyperlipidemia, type 2 diabetes, hypertension, stroke, heart disease, fatty liver, cirrhosis, hepatitis, kidney disease, cancer, gout, rheumatoid arthritis, cataracts, and Parkinson's disease.

[0036] (1) First step: serum proteomic analysis

[0037] We performed proteomic analysis on serum samples from 1785 participants in the discovery cohort (542 healthy samples and 1243 disease samples as training set), 1629 participants in the internal validation cohort (508 healthy samples and 1121 disease samples as internal validation set), and 124 participants in the external validation cohort (28 healthy samples and 96 disease samples as external validation cohort) ( Figure 1 ). The specific proteomics analysis method is as follows:

[0038] 1) 1 μL serum sample was lysed with 20 μL lysis buffer (8 M urea, 100 mM ammonium bicarbonate (ABB)) at 32°C for 30 minutes, and then reduced and alkylated with 10 mM tris(2-carboxyethyl)phosphine (TCEP) and 40 mM iodoacetamide (IAA), respectively. Before enzymatic hydrolysis, 70 μL 100 mM ABB was added to the sample to dilute urea. The protein extract was digested with two-step enzyme-substrate ratio of 1:60 (final ratio of 1:30) overnight trypsin at 32°C for 4 h and 12 h, respectively. The digestion was then terminated by adjusting the pH to 2-3 with 1% trifluoroacetic acid (TFA). The peptides were cleaned with HRP SOLAu columns (purchased from Thermo Fisher Scientific) before mass spectrometry analysis.

[0039] 2) Microliter liquid chromatography-high resolution mass spectrometry analysis

[0040] The peptide samples were injected into the Eksigent NanoLC 400 liquid phase system and combined with the TripleTOF 5600 system for satellite scanning technology (SWATH-MS) analysis. Briefly, 0.5 μg of peptide was analyzed along the analytical column (3 μm, ChromXP C18CL, 150*0.3mm) separation, 5%-30% buffer B liquid phase gradient separation for 20min, flow rate of 5μL / min. Buffer A contains 2% acetonitrile (ACN) and 0.1% formic acid (FA) in HPLC water, and buffer B contains 98% ACN and 0.1% formic acid in HPLC water. The SWATH-MS method uses 55 variable isolation window settings. The MS1 scan and MS / MS scan accumulation times are set to 100ms and 31ms, respectively. Each cycle time is 1.8s.

[0041] 3) DIA-NN (1.8) was used to analyze the mass spectrometry result files. In the DIA-NN settings, the retention time extraction window was automatically set by the software, and the m / z extraction windows for MS1 and MS2 were 20 ppm and 50 ppm, respectively. The false discovery rates (FDRs) of protein and peptide identification were controlled below 1%.

[0042] (2) Step 2: Construction of healthy aging score

[0043] We used the training set to build a random forest model to distinguish healthy and unhealthy samples. The initial input variables of the model were 86 serum proteins significantly associated with age (see Table 1, from the UniProtKB / Swiss-Prot database (https: / / www.uniprot.org / )), specifically the relative abundance information of proteins. Before entering the model, the relative abundance of all proteins was logarithmically transformed (log 10 (X), where X is the relative abundance of the protein).

[0044] Table 1

[0045]

[0046]

[0047]

[0048] The training set included 542 healthy samples and 1243 disease samples. The sample was classified as healthy or not as a classification label (0 / 1, healthy as 1, unhealthy as 0), and 86 age-related serum proteins were used as feature variables to train the initial random forest model. In all random forest models, the number of decision trees trained was set to 1000, and the number of variables randomly selected for each decision tree was 4.

[0049] The importance of variables (proteins) in the model was evaluated based on the mean decrease in accuracy (MDA). MDA is the decrease in accuracy of the model after deleting each feature variable (protein), reflecting the influence of each feature on the accuracy of the model.

[0050] Select the appropriate number of variables to determine the final random forest model. Sort the 86 feature variables (serum proteins) in descending order of importance, select the top n (1 to 86) important variables to retrain the random forest model, and use a ten-fold cross-validation method to calculate the average error of each model. Select the model with the smallest average error. In this study, when the top 22 feature variables were selected, the trained model had the smallest error. The random forest model using 22 feature variables (proteins) was used as the final model.

[0051] Finally, we selected the top 22 serum proteins (P04114_APOB, P55056_APOC4, P01009_A1AT, P02654_APOC1, P02750_A2GL, P10643_CO7, P02748_CO9, Q9Y613_FHOD1, P01023_A2MG, Q99684_GFI1, P36955_PEDF, P 18428_LBP, P02774_VTDB, P02751_FINC, P08697_A2AP, Q9UPU7_TBD2B, A6NFN9_ANKUB, P06312_KV401, P27169_PON1, P23142_FBLN1, P02787_TRFE, P00740_FA9) were used as input variables in the training set to construct the final random forest model ( Figure 2 ). The names of the above serum proteins are derived from the UniProtKB / Swiss-Prot database. The part before “_” is the entry identifier in the Uniprot database, and the part after “_” is the corresponding entry name.

[0052] We validated the model in the internal validation set and obtained an area under the receiver operating characteristic curve (AUC) of 0.70; at the same time, we validated the model in the external validation cohort and obtained an area under the curve (AUC) of 0.71 ( Figure 3 ).

[0053] Based on the trained random forest model containing the top 22 proteins in terms of importance, we can calculate the probability of a sample being classified as "healthy". This probability is estimated by the tree voting (TV) aggregation method, that is, calculating the proportion of decision trees that voted as "healthy" to the total number of decision trees. The calculation formula is as follows:

[0054]

[0055] Where X represents the matrix of 22 serum proteins in the sample; T represents the total number of trees in the random forest model, which is set to 1000 in this study; the function ht(X, “healthy”) represents the voting result of whether the participant is classified as “healthy” in the tth tree (t = 1, 2, 3, …, T).

[0056] In the tree voting aggregation method, the voting results of different decision trees may represent whether the sample is healthy or not at different levels. The more decision trees that vote for "healthy", the greater the probability that the sample is judged as "healthy" overall. Therefore, we use the probability output by the model as an indicator to measure the health of the sample, and use it as the healthy aging score of the sample. The higher the healthy aging score of the sample, the better the health expectancy and the lower the risk of related metabolic diseases.

[0057] (3) Step 3: Serum protein healthy aging score predicts the risk of metabolic diseases

[0058] Using the above random forest model containing 22 serum proteins and the tree voting aggregation method, we calculated the healthy aging score using the serum proteome data of 3414 participants in the Guangzhou Nutrition and Health Cohort (GNHS) at baseline, and integrated the information on new-onset hyperlipidemia, hypertension, diabetes, and fatty liver in the 9-year follow-up of the GNHS cohort. We used the Cox proportional hazards model adjusted for age, sex, and BMI to calculate the relationship between the baseline protein healthy aging score of the cohort population and the risk of new-onset metabolic diseases in the next 9 years.

[0059] We found that for every 1 standard deviation (SD) increase in baseline protein healthy aging score, the risk of developing the four metabolic diseases in the next 9 years increased by 53% (hazard ratio (HR) and 95% confidence interval (CI): 0.47, 0.26-0.85), the risk of developing hyperlipidemia increased by 32% (HR and 95% CI: 0.68, 0.49-0.95), the risk of developing hypertension increased by 40% (HR and 95% CI: 0.60, 0.40-0.91), the risk of developing type 2 diabetes increased by 53% (HR and 95% CI: 0.47, 0.27-0.82), and the risk of developing fatty liver increased by 53% (HR and 95% CI: 0.47, 0.36-0.60) ( Figure 4 ), indicating that the protein healthy aging score, as a biomarker score, can effectively predict the risk of metabolic diseases in the next 9 years.

Claims

1. A combination of protein markers related to the prediction of metabolic disease risk, characterized in that: P04114_APOB, P55056_APOC4, P01009_A1AT, P02654_APOC1, P02750_A2GL, P10643_CO7, P02748_CO9, Q9Y613_FHOD1, P01023_A2MG, Q99684_GFI1, P36955_PEDF, P18428_LBP, P02774_VTDB, P02751_FINC, P08697_A2AP, Q9UPU7_TBD2B, It is composed of 22 serum proteins: A6NFN9_ANKUB, P06312_KV401, P27169_PON1, P23142_FBLN1, P02787_TRFE, and P00740_FA9.

2. Use of a reagent for detecting the combination of protein markers related to the prediction of metabolic disease risk as described in claim 1 in the preparation of a product for predicting the risk of metabolic disease.

3. The use according to claim 2, characterized in that: The reagent is a reagent for performing proteomic analysis on a serum sample of an individual.

4. A metabolic disease risk prediction product, characterized in that: It comprises a reagent for detecting the combination of protein markers related to the prediction of the risk of metabolic disease according to claim 1.

5. A method for constructing a metabolic disease risk prediction model, characterized in that: The following steps are involved: (1) Sample processing: serum samples were collected as training set, internal validation set, and external validation cohort, and each serum sample was subjected to proteomic analysis; Among them, the individuals from whom the serum samples were derived were divided into healthy and unhealthy. Individuals with any of the 14 aging-related chronic diseases were considered unhealthy, and otherwise they were considered healthy. The 14 chronic diseases related to aging are: hyperlipidemia, type 2 diabetes, hypertension, stroke, heart disease, fatty liver, cirrhosis, hepatitis, kidney disease, cancer, gout, rheumatoid arthritis, cataracts, and Parkinson's disease; (2) The training set was used to construct a random forest model to distinguish between healthy and unhealthy samples, using age-related serum proteins as feature variables; (3) According to the average accuracy decline evaluation model, the random forest model with 22 feature variables was selected as the final model. The 22 feature variables are: P04114_APOB, P55056_APOC4, P01009_A1AT, P02654_APOC1, P02750_A2GL, P10643_CO7, P02748_CO9, Q9Y613_FHOD1, P01023_A2MG, Q99684_GFI1, P36955_PEDF, P18428_LBP, P02774_VTDB, P02751_FINC, P08697_A2AP, Q9UPU7_T BD2B, A6NFN9_ANKUB, P06312_KV401, P27169_PON1, P23142_FBLN1, P02787_TRFE, P00740_FA9.

6. The method for constructing a metabolic disease risk prediction model according to claim 5, characterized in that: In step (2), the initial input variables of the random forest model are 86 age-related serum proteins.

7. The method for constructing a metabolic disease risk prediction model according to claim 5, characterized in that: In step (2), in the random forest model, the number of decision trees trained is set to 1000, and the number of variables randomly selected for each decision tree is 4.

8. The method for constructing a metabolic disease risk prediction model according to claim 5, characterized in that: In step (3), the variables are sorted in descending order of importance, and the top n most important variables are selected to retrain the random forest model. The average error of each model is calculated by a ten-fold cross validation method, and the model with the smallest average error is selected as the final model. The value of n is a positive integer from 1 to the total number of all variables.

9. The metabolic disease risk prediction model constructed by the construction method according to any one of claims 5 to 8.

10. The metabolic disease risk prediction model according to claim 9, characterized in that: When in use, the serum sample from the individual to be predicted is subjected to proteomic analysis to obtain the relative abundance information of the 22 characteristic variables, which are then input into the metabolic disease risk prediction model to generate a healthy aging score, and the probability of being classified as healthy is calculated as the healthy aging score. The higher the healthy aging score, the lower the predicted risk of metabolic disease.

Images