Proteomic-based method, apparatus and medium for predicting future health status of an individual

Abstract

The present application relates to a kind of individual future health state prediction method, device and medium based on proteomics, wherein the method comprises the following steps: obtaining proteomics data and preprocessing;Shared network construction: construct individual past and future comorbidity condition prediction neural network based on twin network framework;Trunk network construction: construct health-specific outcome prediction neural network based on multilayer perceptron method;Health assessment network integrates the shared network and trunk network architecture, and extracts features thereof to fuse and fine-tune in latent space, updates fine-tuning network parameters by further training, and outputs the risk assessment probability of multiple health-specific outcomes in the future.Compared with prior art, the present application focuses on proteomics data processing and modeling method, uses past and future health estimates as prior information, and realizes individual health condition assessment with multiple diseases and death as outcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics and biomedical engineering, and in particular to a method and device for predicting the future health status of an individual based on proteomics and a medium. BACKGROUND

[0002] Early health level prediction is of great significance for individual health management and treatment. By determining the potential risk of disease and assessing the health of individuals, early intervention and treatment can be provided to maximize the protection and improvement of individual health. Current technical routes and methods have the following limitations: first, most studies are based on cross-sectional data, which greatly limits the application of developed models in real application scenarios; in addition, future disease and death events are important indicators of future health status, but current technologies only focus on predicting a single outcome (such as death, specific diseases, aging, cognitive or motor function decline, etc.), which cannot assess the future health level in a panoramic manner.

[0003] Proteins are important functional molecules in cells. As the final product of gene expression, proteomics data can directly reflect the function of cells or tissues, reveal the internal mechanism of biological systems, and better reflect the actual state of individuals. Previous studies have determined that abnormal changes in the expression levels of certain specific proteins can be associated with the occurrence and progression of diseases such as cancer, cardiovascular disease, and diabetes, and can be used as potential drug targets and biomarkers. Therefore, developing health assessment models based on proteomics data has important clinical significance and application prospects. However, due to the complexity of proteomics data and the massive growth of data volume, there is no research on health assessment models for multiple diseases and death outcomes based on proteomics data. SUMMARY

[0004] The purpose of the present application is to provide a method and device for predicting the future health status of an individual based on proteomics and a medium.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A method for predicting the future health status of an individual based on proteomics, comprising the following steps:

[0007] Step 1) Obtain proteomics data and perform preprocessing;

[0008] Step 2) Shared network construction: construct an individual's past and future co-morbidity status prediction neural network based on a twin network framework;

[0009] Step 3) Main network construction: construct a health-specific outcome prediction neural network based on a multi-layer perceptron method;

[0010] Step 4) The health assessment network integrates the shared network and the backbone network architecture, extracts features, and fine-tunes them in the latent space. The network parameters are updated by further training, and the risk assessment probability of multiple health-specific outcomes in the future is output.

[0011] The input of the shared network is the pre-processed plasma protein level vectorized numerical value, and the output is the number of past disease categories and the number of future disease categories. The determination method of the number of past disease categories and the number of future disease categories is: the time point of collecting plasma protein is set as the baseline time, and the number of multiple diseases and death categories in the individual electronic medical record information is determined according to the baseline time point, which is divided into baseline before and baseline after.

[0012] The shared network contains two branch networks with the same architecture, and each encoding branch network includes four fully connected layers and an output layer. Each fully connected layer contains 512, 256, 128, and 64 nodes, respectively. The activation function between the fully connected layers is ReLu, and the activation function between the last fully connected layer and the output layer is a linear function.

[0013] The input of the backbone network is the pre-processed plasma protein level vectorized numerical value, and the output is the health assessment information after the baseline plasma protein collection time point. This information covers the future risk probability estimation of multiple health-specific outcomes.

[0014] The encoder of the backbone network includes four fully connected layers and an output layer. Each fully connected layer contains 512, 256, 128, and 64 nodes, respectively. The activation function between the fully connected layers is ReLu, and the activation function between the last fully connected layer and the output layer is a Sigmoid function.

[0015] The input of the health assessment network is the vector numerical value of plasma protein level, and the output is the same as the backbone network, which is the risk probability estimation of multiple future health-specific outcomes.

[0016] The health assessment network fuses the processed protein features in the two branch encoders of the shared network for past and future comorbidity outcomes and the protein features extracted by the encoding part of the backbone network.

[0017] The health assessment network combines the 64 pre-training weights of the last full connection layer of the two encoder branches of the network and the 64 pre-training weights of the last full connection layer of the backbone network in a series form; the output after the combination is input into full connection layers with node numbers of 128 and 64 in sequence and then output through an output layer, wherein the series combination output is connected with the full connection layers by a ReLu activation function, the full connection layers are connected by a ReLu activation function, and a Sigmoid function is used as the activation function between the last full connection layer and the output layer.

[0018] A device for predicting future health status of an individual based on proteomics, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the method as described above when executing the program.

[0019] A storage medium having a program stored thereon, wherein the program implements the method as described above when executed.

[0020] Compared with the prior art, the present application has the following beneficial effects:

[0021] (1) Previous prediction researches based on proteomics mainly based on cross-sectional data, and the lack of longitudinal clinical data greatly limits the application of the developed model in real application scenarios. The present application uses the data of more than 14 years of health outcome tracking follow-up of participants as input, and the longitudinal cohort data enables the model to match the real application scenario when constructed. The prediction model is built using proteins and clinical indicators obtained at the baseline to evaluate future outcomes.

[0022] (2) The future disease and death events of an individual are important evaluation indicators of the future health status of the individual, but current models only focus on a single outcome. The present application classifies and organizes the electronic medical records of the participants during the follow-up period in primary care records, hospitalization records and death registration, which covers multiple clinical health outcomes, so that the present application can systematically evaluate the future health from multiple dimensions and all perspectives.

[0023] (3) The present application realizes the estimation of the overall disease burden of an individual and the risk prediction of a total of 45 key health outcomes in the future by constructing two important components of the shared network and the health-specific outcome network, realizing the comprehensive and detailed evaluation of the future health status of an individual. DETAILED DESCRIPTION

[0024] Figure 1 The method flowchart of the present application;

[0025] Figure 2 The health assessment network model structure;

[0026] Figure 3Forest plots are presented for the prediction of 45 health-specific outcomes for different feature sets. DETAILED DESCRIPTION

[0027] The present application will be described in detail below with reference to the drawings and specific examples. The present embodiment is implemented on the premise of the technical solution of the present application, and gives a detailed implementation and specific operation process, but the protection scope of the present application is not limited to the following examples.

[0028] Example 1

[0029] The present application processes and processes proteomics data of 1461 plasma proteins, and constructs a future health assessment model based on a deep learning method, wherein the health assessment model integrates a twin shared network and a specific outcome prediction network two parts; the health assessment covers 45 kinds of longitudinal future health specific outcomes, including all-cause death, 4 disease-specific deaths, 26 new diseases and 14 disease categories (including infectious, blood, endocrine, mental, neurological, circulatory, respiratory, digestive, skin, musculoskeletal and urogenital system diseases and cancer).

[0030] Specifically, as shown in Figure 1 , the following steps are included:

[0031] Step 1) Obtain proteomics data and pre-process.

[0032] Proteomics data acquisition and processing: blood samples were collected in EDTA (9ml) vacuum containers, and divided into 850ul EDTA plasma, buffy coat and red blood cell aliquots. The plasma samples to be analyzed were stored in a refrigerator at -80℃, and were used for parallel measurement of 1461 unique proteins by combining adjacent extension assay and next-generation sequencing. The proteomics data include proteins measured on four panels of cardiometabolic, inflammatory, neurological and oncological proteins. Protein expression measurement data is obtained based on the standardized measurement process of OLINK Company in Sweden.

[0033] The standardized process is used for pre-processing of proteomics expression data (Normalized Protein eXpression, NPX) for modeling as follows: divide the count of each sample and each assay by the count of the extended control, and further logarithmically transform the ratio. By considering the median of the extended control standardized count, batch-specific NPX median and the difference of assay-specific NPX median of each batch, the intra-plate and inter-plate differences are minimized.

[0034] Health outcome acquisition: The training data outcomes are divided into three parts. Two encoding branches in the shared network correspond to the number of past disease categories and the number of future disease categories. The number of health categories is defined as 14 health categories (0-14 values), and the health categories include infectious diseases, cancer, blood immunity, endocrine, mental behavior, nervous system, eye, ear, circulation, respiration, digestion, skin, muscle and skeleton, urogenital); the definition of past and future is before and after the individual collects plasma proteomics data. Secondly, the outcome defined in the main network is the corresponding 45 outcomes in the ICD-10 code (including 14 disease categories, 26 specific diseases, and all-cause death and 4 specific deaths).

[0035] Step 2) Shared network construction: Construct an individual past and future comorbidity condition prediction neural network based on the twin network framework.

[0036] The input of the shared network is 1461 pre-processed plasma protein level vectorized numerical values, and the output (model training outcome) is the number of past disease categories and the number of future disease categories. The determination method of the number of past disease categories and the number of future disease categories is: the time point of collecting plasma proteins is set as the baseline time, and the 14 diseases and death categories in the individual electronic medical record information (including infectious diseases, cancer, blood immunity, endocrine, mental behavior, nervous system, eye, ear, circulation, respiration, digestion, skin, muscle and skeleton, urogenital and all-cause death) are divided into baseline before and baseline after the statistical number (0-14) according to the baseline time point.

[0037] As shown in Figure 2 , the shared network contains two branch networks with the same architecture, and each encoding branch network includes four fully connected layers and an output layer, wherein each fully connected layer contains 512, 256, 128 and 64 nodes respectively, and the fully connected layers are connected through the activation function of ReLu. The activation function between the last fully connected layer and the output layer adopts the linear (Linear) function.

[0038] Shared network training: First, train the two encoders of past disease and future disease separately, then migrate the weights of the two pre-trained encoders to the two branches of the shared network, and update the weights by re-tuning. Both separate training and combined training use the Adam optimizer, the learning rate is set to 1x10 -5 , the training batch is 128, the iteration number is 1000, in order to reduce the risk of overfitting, the iteration node defined for early termination training is set to 25 consecutive iterations of the validation set loss function without reduction.

[0039] Step 3) Main network construction: Construct a health-specific outcome prediction neural network based on the multilayer perceptron method.

[0040] The input of the backbone network is 1461 pre-processed blood plasma protein level vectorized numerical values, and the output (model training result) is health assessment information after the baseline blood plasma protein collection time point, which covers 45 health-specific outcomes, including all-cause death, 4 disease-specific deaths, 26 new diseases, and 14 disease categories; the output information is the risk probability estimate of 45 specific health outcomes.

[0041] As shown in Figure 2 The encoder of the backbone network includes four fully connected layers and an output layer, each of which contains 512, 256, 128, and 64 nodes, respectively, and is connected by ReLu activation functions between the fully connected layers. The activation function between the last fully connected layer and the output layer uses a Sigmoid function.

[0042] Backbone network training: using Adam optimizer, learning rate set to 1x10 -5 , training batch is 128, iteration number is 1000, to reduce the risk of overfitting, the iteration node definition for early termination training is set to continuous 25 iterations of the validation set loss function not reduced.

[0043] Step 4) The health assessment network integrates the shared network and the backbone network architecture, and extracts features in the latent space for fusion and fine-tuning. Through further training, the fine-tuning network parameters are updated, and the risk assessment probability of multiple health-specific outcomes in the future is output.

[0044] The health assessment network includes a shared network and a backbone network, so the same 1461 blood plasma protein level vector values are input into the health assessment network as two input information, and the output is the same as the backbone network, which is the risk probability estimate of 45 future health-specific outcomes.

[0045] The health assessment network fuses the processed protein features in the two branch encoders of the shared network for past and future comorbidity outcomes and the protein features extracted by the encoding part of the backbone network. Specifically, as shown in Figure 2 The 64 pre-training weights of the last fully connected layer of the two encoder branches of the shared network are merged in series with the 64 pre-training weights of the last fully connected layer of the backbone network (i.e. 64+64+64 = 192 neural node weight values). The merged output is input into the fully connected layers with node numbers of 128 and 64, respectively, and then output through the output layer, where the series-merged output is connected to the fully connected layers by ReLu activation functions, and the activation function between the last fully connected layer and the output layer uses a Sigmoid function.

[0046] Health assessment model training: The previous shared network and backbone network have obtained the optimal weight parameters through pre-training. After the health assessment network is constructed, all pre-trained parameters are first frozen during training, and only the last 128 and 64 node fully connected layers are trained; after the preliminary training is completed, all backbone network weight parameters are unfrozen and participate in fine-tuning update together, and note that the shared network weight parameters are always frozen in this step. The Adam optimizer is used for training, the learning rate is set to 1x10 -5 , the training batch is 128, the iteration number is 1000, and the iteration node for early termination of training is defined as the validation set loss function that has not been reduced for 25 consecutive iterations.

[0047] Figure 3 The forest plot of the prediction performance of different feature sets on 45 health-specific outcomes is shown. The prediction performance of proteomics data (leftmost column) and the prediction performance of proteomics data + clinical indicators (rightmost column). In order to compare the prediction performance of proteomics data, different feature sets are compared, including age + gender, blood indicators (25 kinds), clinical indicators (54 kinds, including demographic information, blood indicators, living habits, medical history, medication history, etc.). In each comparison group, the vertical line represents the prediction performance of proteomics data only, the dot represents the prediction performance of each feature set only, and the triangle represents the prediction performance of each feature set + proteomics data.

[0048] In this embodiment, the prediction risk probability output by the health assessment network model is calculated using the Cox proportional hazards regression model, and the Hazard ratio (HR) and significance p value are calculated. Table 1 reports the HR value of the output risk probability under different covariate feature sets, and the corresponding p value of the HR value is statistically significant.

[0049]

[0050]

[0051]

[0052] Embodiment 2

[0053] The embodiment provides a proteomics-based individual future health state prediction device, which comprises a memory, a processor, and a program stored in the memory, and the processor implements the method as described above when executing the program.

[0054] Embodiment 3

[0055] The embodiment provides a storage medium having a program stored thereon, and the program is executed to implement the method as described above.

[0056] If the above functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0057] The preferred embodiments of the present application are described in detail above. It should be understood that those skilled in the art can make many modifications and changes without creative labor based on the concept of the present application. Therefore, any technical solutions obtained by logical analysis, reasoning, or limited experiments based on the prior art according to the concept of the present application shall be within the protection scope determined by the claims.

Claims

1. A method for predicting future health status of an individual based on proteomics, characterized by, The method comprises the following steps: Step 1) obtaining proteomics data and preprocessing; Step 2) shared network construction: constructing an individual past and future comorbidity condition prediction neural network based on a twin network framework; Step 3) backbone network construction: constructing a health-specific outcome prediction neural network based on a multilayer perceptron method; the input of the backbone network is a preprocessed plasma protein level vectorized numerical value, and the output is health assessment information after the baseline plasma protein collection time point, which covers the future risk probability estimation of multiple health-specific outcomes; the encoder of the backbone network comprises four fully connected layers and an output layer, wherein each fully connected layer comprises 512, 256, 128 and 64 nodes respectively, the fully connected layers are connected through a ReLu activation function, and the activation function between the last fully connected layer and the output layer adopts a Sigmoid function; Step 4) the health assessment network integrates the shared network and the backbone network architecture, extracts features in the latent space, and updates the fine-tuning network parameters through further training, and outputs the future risk assessment probability of multiple health-specific outcomes; The health assessment network fuses the processed protein features in the two branches of the shared network and the protein features extracted by the encoding part of the backbone network; the health assessment network combines the 64 pre-training weights of the last fully connected layer of the two encoder branches of the shared network and the 64 pre-training weights of the last fully connected layer of the backbone network in a series form; the output of the combined output is input into the fully connected layers with node numbers of 128 and 64 in turn, and then output through the output layer, wherein the series combined output is connected with the fully connected layers through a ReLu activation function, and the activation function between the last fully connected layer and the output layer adopts a Sigmoid function.

2. The proteomic-based method of predicting future health status of an individual according to claim 1, wherein, The input of the shared network is a preprocessed plasma protein level vectorized numerical value, and the output is the number of past disease categories and the number of future disease categories; the determination method of the number of past disease categories and the number of future disease categories is as follows: the time point of collecting plasma protein is set as the baseline time, and the statistical number of multiple diseases and death categories in the electronic medical record information of the individual is determined according to the baseline time point.

3. The proteomic-based method of predicting future health status of an individual of claim 1, wherein, The shared network comprises two branch networks with the same architecture, and each encoding branch network comprises four fully connected layers and an output layer, wherein each fully connected layer comprises 512, 256, 128 and 64 nodes respectively, the fully connected layers are connected through a ReLu activation function, and the activation function between the last fully connected layer and the output layer adopts a linear function.

4. The proteomic-based method of predicting future health status of an individual of claim 1, wherein, The input of the health assessment network is a vector numerical value of plasma protein level, and the output is the same as that of the backbone network, which is the risk probability estimation of multiple future health-specific outcomes.

5. A proteomics-based individual future health state prediction apparatus comprising a memory, a processor, and a program stored in the memory, wherein The processor implements the method of any one of claims 1-4 when executing the program.

6. A storage medium having stored thereon a program, characterized by The program is executed to implement the method of any one of claims 1-4.

Images