Biological age prediction model training method, age prediction method, device and electronic equipment

Abstract

The invention relates to a biological age prediction model training method, an age prediction method, an age prediction device and electronic equipment, and the method comprises the steps: calling a data reading interface, and loading biological sample data; the biological sample data comprises demographic data and protein data; the protein data comprises expression quantity data of various proteins; preprocessing the biological sample data to obtain a preprocessed data set; constructing a multi-layer perceptron model, and performing initial training on the multi-layer perceptron model based on the preprocessed data set to obtain a preliminary training model; based on weight parameters of the preliminary training model, screening high-contribution characteristic proteins related to biological age to generate a characteristic protein set; and performing feature optimization training on the preliminary training model based on the feature protein set, and outputting the model after the feature optimization training as a biological age prediction model. According to the method, the training accuracy and the model generalization ability of the biological age prediction model can be improved, and the technical application cost is reduced.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of bioinformatics analysis technology, and in particular to a training method, apparatus, electronic device and storage medium for a biological age prediction model. Background Technology

[0002] Aging, as a core systemic biological phenomenon in the life process, is essentially a dynamic process of coordinated degeneration of multiple organs and molecular networks in the body. Biological age, as a key indicator reflecting an individual's true physiological state, more accurately reveals health status, disease susceptibility, and the rate of aging than chronological age alone, and has become a core target for precision health management and anti-aging research. With the scientific consensus that aging is assessable, quantifiable, and interventional, the development of efficient biological age assessment tools has become a common need for both academia and industry.

[0003] Current bioage assessment technologies are mainly developed around three core pathways: First, assessment methods based on clinical records and lifestyle habits, which build predictive models by integrating demographic information, physical examination indicators, lifestyle questionnaires, and other data; second, DNA methylation clock based on epigenetics, which quantitatively predicts the bioage of adults by screening the methylation level of specific CpG sites, and is currently the most widely used technical approach; and third, predictive models based on multi-omics data, including transcriptomics, metabolomics, and proteomics-related models.

[0004] Existing technologies still have significant limitations: First, methods based on clinical records and lifestyle habits rely on subjectively reported information and fragmented medical data, resulting in insufficient data completeness and objectivity, limited prediction accuracy, and difficulty in capturing early aging signals at the molecular level. Second, while DNA methylation clocks have a certain predictive accuracy, the technology is costly to detect, complex to operate, and highly tissue-specific, applicable only to a few sample types such as blood and brain tissue. Furthermore, it cannot effectively explain the biological mechanisms behind aging, limiting its generalization and clinical accessibility. Third, existing proteomics-related models mostly rely on small sample data or single tissue samples for training, failing to fully explore the complex relationships within protein networks. This leads to problems such as redundant feature selection and large prediction errors. Moreover, some models only focus on the adult population, lacking adaptability to aging characteristics throughout the entire life cycle. These models are inefficient when processing complete proteomics data and lack generalization ability in key feature input scenarios, making it difficult to balance prediction accuracy and ease of application.

[0005] Therefore, how to solve the problems of high testing costs, strong tissue specificity, and insufficient adaptability throughout the entire life cycle, and how to construct a training scheme for an accurate, efficient, and clinically accessible biological age prediction model, are urgent problems that need to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide at least one training method, device, electronic device and storage medium for a biological age prediction model, which can improve the training accuracy and generalization ability of the biological age prediction model, reduce the cost of technology application, have a wide range of sample adaptation, and have strong accessibility and practicality in clinical practice.

[0007] To address the aforementioned technical problems, at least one embodiment of this application provides a training method for a biological age prediction model, comprising: The data reading interface is invoked to load biological sample data; the biological sample data includes: demographic data and protein data; the demographic data includes: participant identifiers and actual age labels; the protein data includes expression level data of various proteins; The biological sample data is preprocessed to obtain a preprocessed dataset; A multilayer perceptron model is constructed, and initial training is performed on the multilayer perceptron model based on the preprocessed dataset to obtain a preliminary trained model; Based on the weight parameters of the preliminary training model, high-contribution feature proteins related to biological age are screened to generate a feature protein set. Based on the set of characteristic proteins, the initial training model is subjected to feature optimization training, and the feature-optimized model is output as a biological age prediction model.

[0008] In one embodiment, constructing the multilayer perceptron model includes: A network hierarchy structure is constructed, comprising an input layer, at least three hidden layers, and an output layer connected in sequence; wherein the input layer, hidden layers, and output layer are all fully connected layer structures. Set the number of neurons in the input layer to be consistent with the number of features in the protein data; The number of neurons in each hidden layer is set to decrease sequentially according to a preset feature compression ratio, and a batch normalization component, a leakage correction linear unit nonlinear activation function, and a random deactivation component are sequentially embedded in each hidden layer. Set the number of neurons in the output layer to 1.

[0009] In one embodiment, initial training of the multilayer perceptron model based on the preprocessed dataset includes: Configure the Adam optimizer by setting its learning rate to a preset learning rate and its weight decay coefficient to a first weight decay coefficient. The mean squared error loss function and the smoothing L1 loss function are combined into a hybrid loss function according to a preset weight ratio. The preprocessed dataset is loaded in batches according to a preset batch size and input into the multilayer perceptron model to perform forward propagation calculation. The loss value corresponding to each batch of data is calculated through the hybrid loss function. The loss value includes: training loss value and validation loss value. Initiate iterative training by calling the model training function. After each iteration, update the model parameters through backpropagation using the Adam optimizer based on the training loss value and the weight decay coefficient, and adjust the learning rate based on the validation loss value. Based on the validation loss value and the number of iterations, determine whether the training termination condition is met. If it is met, stop training and output the current model as the initial training model.

[0010] In one embodiment, feature optimization training is performed on the initial trained model based on the set of featured proteins, including: The number of neurons in the input layer of the multilayer perceptron model is adjusted to the number of features in the feature protein set, the loss function is adjusted to the smoothed L1 loss function, and the weight decay coefficient is adjusted to the second weight decay coefficient. Extract the dataset that matches the set of feature proteins from the preprocessed dataset, and use it as the target dataset; The target dataset is loaded in batches according to a preset batch size, and the forward propagation calculation is performed by inputting it into the multilayer perceptron model to obtain the model output after training stops.

[0011] In one embodiment, based on the weight parameters of the initially trained model, high-contribution feature proteins related to biological age are screened, including: Extract the weight matrix from the input layer to the first hidden layer of the initially trained model; Based on the weight matrix, the features corresponding to all protein-related columns are sorted in descending order of absolute weight value; The proteins corresponding to the highest-ranked features are selected as the high-contribution feature proteins.

[0012] In one embodiment, preprocessing operations are performed on the biological sample data, including: The invalid data in the biological sample data is cleaned, and samples and protein-related features that do not meet the data quality requirements are removed. For missing values ​​in the cleaned data, determine the corresponding statistical feature values ​​of the feature columns and fill them in; Convert all filled data to the preset numerical format; Standardize the protein expression data after format conversion; Based on the actual age label, the standardized biological sample data is split into training set, validation set and test set using stratified sampling. Feature matrices and target vectors are separated from each dataset; the feature matrices are standardized protein expression data, and the target vectors are numerical vectors corresponding to actual age labels; the datasets include: training set, validation set, and test set. The training set, the validation set, the test set, and the corresponding feature matrix and target vector are integrated to form the preprocessed dataset.

[0013] At least one embodiment of this application also provides a method for predicting biological age, based on a training method for the biological age prediction model, the prediction method comprising: Receive biological sample data to be predicted input by the user; the biological sample data to be predicted includes: full protein expression data or characteristic protein expression data; the characteristic protein expression data is matched with the protein types of the characteristic protein set; Preprocessing operations are performed on the biological sample data to be predicted to obtain preprocessed data; If the biological sample data to be predicted is the characteristic protein expression data, the preprocessed data is input into the trained biological age prediction model, and the biological age prediction result is output. If the biological sample data to be predicted is the full protein expression data, protein expression level data that are consistent with the types of the feature protein set are selected from the preprocessed data to generate a target feature matrix; The target feature matrix is ​​input into the biological age prediction model, and the biological age prediction result is output.

[0014] At least one embodiment of this application also provides a training apparatus for a biological age prediction model, comprising: The sample loading unit is used to call the data reading interface and load biological sample data; the biological sample data includes: demographic data and protein data; the demographic data includes: participant identifiers and actual age labels; the protein data includes expression level data of various proteins; A sample preprocessing unit is used to perform preprocessing operations on the biological sample data to obtain a preprocessed dataset. An initial training unit is used to construct a multilayer perceptron model and perform initial training on the multilayer perceptron model based on the preprocessed dataset to obtain a preliminary trained model. The protein screening unit is used to screen high-contribution characteristic proteins related to biological age based on the weight parameters of the preliminary training model, and generate a set of characteristic proteins. An optimization training unit is used to perform feature optimization training on the initial training model based on the set of feature proteins, and output the feature-optimized model as a biological age prediction model.

[0015] At least one embodiment of this application also provides a biological age prediction device, and a training device based on the biological age prediction model, the prediction device comprising: A prediction data receiving unit is used to receive biological sample data to be predicted input by a user; the biological sample data to be predicted includes: full protein expression data or characteristic protein expression data; the characteristic protein expression data is matched with the protein types of the characteristic protein set; The data preprocessing unit is used to perform preprocessing operations on the biological sample data to be predicted to obtain preprocessed data; The feature protein processing unit is used to input the preprocessed data into the trained biological age prediction model and output the biological age prediction result if the biological sample data to be predicted is the feature protein expression data. A full protein extraction unit is used to, if the biological sample data to be predicted is the full protein expression data, filter out the protein expression level data that is consistent with the type of the feature protein set from the preprocessed data, and generate a target feature matrix. The matrix prediction unit is used to input the target feature matrix into the biological age prediction model and output the biological age prediction result.

[0016] At least one embodiment of this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described training method for a biological age prediction model.

[0017] At least one embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described training method for the biological age prediction model.

[0018] The training method for the biological age prediction model provided in this application loads biological sample data containing demographic data and protein expression level data through a standardized data reading interface. Then, systematic preprocessing operations are performed to improve data quality, meeting the model's input requirements. A multilayer perceptron (MLP) model adapted to high-dimensional protein data is constructed. The multilayer fully connected structure of the MLP model effectively captures the complex nonlinear relationship between protein expression and biological age, overcoming the limitations of traditional linear models. Initial training is conducted based on the preprocessed high-quality dataset, enabling the MLP model to fully learn the potential correlation patterns of all protein features, forming a preliminary training model with basic predictive capabilities. Then, based on the weight parameters of the initial training model, feature proteins are selected to quantify their contribution to biological age prediction. Redundant and low-correlation noise features are precisely eliminated, achieving efficient dimensionality reduction of features. The generated feature protein set focuses on core associations. Based on the core feature protein set, the initial training model is further optimized, which reduces the computational complexity of training and avoids interference from invalid features. This provides clear core input for optimized training, strengthening the association between core features and biological age without reconstructing the model. The model parameters are more focused on key patterns, and the final output biological age prediction model is both efficient and reliable, meeting the needs of practical application scenarios.

[0019] This method achieves a synergistic improvement in data quality, model adaptability, training efficiency, and prediction performance, ultimately enabling the efficient output of a biological age prediction model with high accuracy, strong generalization ability, and controllable computational cost, providing a scientific and practical technical solution for biological age prediction based on protein expression data. Attached Figure Description

[0020] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0021] Figure 1 This is a flowchart illustrating a training method for a biological age prediction model provided in one embodiment of this application; Figure 2 This is a schematic diagram illustrating the changing trend of the loss value during model training, provided in one embodiment of this application. Figure 3 This is a flowchart illustrating a method for predicting biological age, provided in one embodiment of this application. Figure 4 This is a schematic diagram of a training device for a biological age prediction model provided in one embodiment of this application; Figure 5 This is a schematic diagram of a biological age prediction device provided in one embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0023] As the core executors of life activities, proteins form dynamic networks that directly regulate physiological homeostasis. Imbalances in protein homeostasis are a hallmark of the aging process. Compared to biomarkers such as the genome and epigenome, dynamic changes in protein expression levels can better reflect the instantaneous state of physiological functions and age-related functional alterations in real time. In recent years, the establishment of large-scale biobanks (such as the UK Biobank) and breakthroughs in high-throughput protein detection technologies have provided massive data support for mining aging-related protein biomarkers and building accurate predictive models, making proteomics-based biological age assessment a crucial direction for overcoming traditional technological bottlenecks. Simultaneously, the development of artificial intelligence and machine learning algorithms has further enabled the in-depth mining of aging characteristics from complex protein data, laying the technological foundation for training models to build high-precision, highly generalizable biological age prediction models, and propelling aging assessment from traditional single indicators to a multi-dimensional, systematic molecular level.

[0024] This invention proposes a training method for a biological age prediction model based on protein expression data. The implementation details of the training method for the biological age prediction model in this embodiment are described below. The following content is only for the convenience of understanding and is not necessary for implementing this solution.

[0025] Example 1: The specific process of training the biological age prediction model in this embodiment can be as follows: Figure 1 As shown, it includes: Step 101: Call the data reading interface to load biological sample data.

[0026] The data reading interface is called to read biological sample data from a specified data source (such as a biological sample database, data storage server, etc.). The biological sample data mainly includes two types of core data: demographic data (such as the demographic column) and protein data.

[0027] The demographic data includes participant identifiers and actual age labels, and may also include date of birth, gender, etc. Participant identifiers are used to distinguish different individual samples, and actual age labels record the individual's true physiological age for supervised training.

[0028] The protein data should include quantitative expression levels of multiple proteins, i.e., the expression level of each protein in the sample. In this method, the degree of aging is determined based on the protein expression level, ensuring that each sample contains feature information for model learning.

[0029] Step 102: Perform preprocessing operations on the biological sample data to obtain a preprocessed dataset.

[0030] Preprocessing operations, including a series of standardized and normalized quality optimization and format adaptation operations, are performed on biological sample data to eliminate invalid information, unify data format, and improve data usability, ultimately forming a preprocessed dataset with a well-structured and high-quality dataset that can be directly used for model training.

[0031] In this embodiment, the specific type of preprocessing operation is not limited. It can be set according to the actual data format type. For example, outlier correction, duplicate data removal, data format normalization, feature dimension filtering and other data processing operations can be configured.

[0032] Step 103: Construct a multilayer perceptron model and perform initial training on the multilayer perceptron model based on the preprocessed dataset to obtain a preliminary trained model.

[0033] Biological age prediction essentially involves using algorithms to uncover the correlation between protein expression and age, building a multilayer perceptron (MLP) neural network structure adapted to the needs of biological age prediction, and conducting the first round of systematic training using a preprocessed standardized dataset. The multilayer fully connected structure of the MLP model can capture the complex nonlinear relationship between the two, enabling the model to initially learn the correlation between protein expression data and biological age, and finally outputting a preliminary training model with basic prediction capabilities that can support subsequent optimization.

[0034] In this embodiment, the specific model type and model structure of the multilayer perceptron model are not limited. For example, a fully connected MLP integrating batch normalization and random deactivation, a deep MLP with leakage correction linear units (Leaky ReLU), or a feature-level compression fully connected MLP can be selected.

[0035] Step 104: Based on the weight parameters of the initially trained model, screen for high-contribution feature proteins related to biological age and generate a set of feature proteins.

[0036] The weight parameters of the initial training model are data reflecting the importance of each protein feature formed by the model in the initial training phase after learning the association between the full protein expression data and biological age. Specifically, this can be achieved by obtaining the weight matrix from the input layer to the first hidden layer. This matrix is ​​a direct quantitative representation of the association between each protein feature and biological age during the model learning process. Each protein feature corresponds to a column of weight values ​​in the matrix. The magnitude of the weight value directly reflects the degree of influence of the protein on the model output (biological age prediction result). The larger the absolute value of the weight, the stronger the influence of the protein feature in the model learning and the higher its contribution to biological age prediction.

[0037] By utilizing the weight parameters learned in the initial training model, high-contribution feature proteins related to biological age are accurately identified and screened. High-contribution feature proteins refer to those proteins that have a significant impact on biological age prediction results and have a strong correlation. Their contribution is directly reflected by the quantitative indicators of the weight parameters. Redundant and low-correlation protein features are eliminated, and finally integrated to form a set of feature proteins focusing on core correlations (such as INSL3, ELN, EDA2R, GDF15), realizing the value extraction of the initial training results and focusing on the core effective components from the full set of protein features.

[0038] The number of high-contribution characteristic proteins selected is not limited, and their contribution to age prediction is the core criterion. The high-contribution proteins corresponding to different protein feature sets (10, 20, 50, 100, 2858 proteins) are different. The table below shows a performance comparison table of a biological age prediction model divided by protein feature sets.

[0039]

[0040] The number of protein features is positively correlated with model performance; the coefficient of determination (R²) increases with the number of features. 2 The performance of the model gradually improved, and the mean absolute error (MAE) continued to decrease. The model corresponding to 2858 full proteins had the best performance, with R... 2 The expression level reached 0.8684, and the MAE was only 2.35 years. In practical applications, it is theoretically recommended to input the expression data of the top 500 high-contribution proteins. However, since Olink detection usually outputs 2858 protein data, an integrated model based on full protein training can be used. The training process of this basic training model must be completed first, and then a prediction model is built based on it. Finally, the biological age result is output through the prediction model.

[0041] Initial training is conducted based on the full protein data set. While this captures all potential associations, it inevitably includes redundant features with weak correlations to biological age, and even noisy features that interfere with the model's judgment. These features increase the computational complexity of subsequent training and may lead to model overfitting. Changes in biological age are only closely related to changes in the expression levels of a few key proteins. In this step, the weight parameters of the initially trained model are used as the selection criterion to ensure that the selected high-contribution proteins are features that the model has actually learned and are strongly correlated with biological age. Selecting high-contribution protein features allows subsequent optimization training to focus on these core associations, avoiding wasting computational resources on redundant features and making the training more targeted.

[0042] Step 105: Perform feature optimization training on the preliminary trained model based on the feature protein set, and output the feature-optimized model as the biological age prediction model.

[0043] Using the selected set of high-contribution feature proteins as input, the preliminary training model generated after initial training is subjected to targeted parameter optimization and deep training. This strengthens the weight ratio and association capture ability of core features, further enhancing the model's learning of the association between core proteins and biological age. This upgrades the model from having basic prediction capabilities to having efficient and accurate prediction capabilities. The association patterns of core feature proteins are more universal and stable. Optimized training based on this set can reduce the model's dependence on redundant noise features, avoid overfitting problems, and make the model more adaptable and predictive stable on new data to be predicted. Ultimately, it outputs a biological age prediction model with better performance, higher accuracy, and stronger generalization ability.

[0044] The effectiveness of the biological age prediction model was validated using data from 44,179 individuals, covering an age range of 18 to 90 years. In the initial training phase, the model was trained based on expression data for approximately 2,858 proteins. At this stage, the model's coefficient of determination (R²) was... 2 The coefficient of determination (R²) reached 0.8684, indicating a relatively excellent fit. In the advanced training phase, the model focused on training the top 100 key proteins (such as GDF15 and INSL3), with corresponding R² values ​​reaching [value missing]. 2 The correlation coefficient was 0.8372, and the mean absolute error (MAE) was 2.40. To verify the model's performance and stability, a test set containing 10,000 samples was used for testing. The results showed that the Pearson correlation coefficient of the test set was 0.88, which verifies that the model has good robustness.

[0045] Based on the above introduction, the training method for the biological age prediction model provided in this embodiment loads biological sample data containing demographic data and protein expression data through a standardized data reading interface. Then, systematic preprocessing operations are performed to improve data quality, meeting the model's input requirements. A multilayer perceptron (MLP) model adapted to high-dimensional protein data is constructed. The multilayer fully connected structure of the MLP model effectively captures the complex nonlinear relationship between protein expression and biological age, overcoming the limitations of traditional linear models. Initial training is conducted based on the preprocessed high-quality dataset, enabling the MLP model to fully learn the potential correlation patterns of all protein features, forming a preliminary training model with basic predictive capabilities. The model is then used to select feature proteins based on the weight parameters of the initial training model. The contribution of proteins to biological age prediction is quantified, and redundant and low-correlation noise features are accurately removed. This achieves efficient dimensionality reduction of features. The resulting feature protein set focuses on core associations. Based on the core feature protein set, the initial training model is further optimized. This reduces the computational complexity of training and avoids interference from invalid features. It provides clear core inputs for optimized training, strengthens the association learning between core features and biological age without reconstructing the model, and makes the model parameters more focused on key patterns. The final output biological age prediction model is both efficient and reliable, and can meet the needs of practical application scenarios.

[0046] This method achieves a synergistic improvement in data quality, model adaptability, training efficiency, and prediction performance, ultimately enabling the efficient output of a biological age prediction model with high accuracy, strong generalization ability, and controllable computational cost, providing a scientific and practical technical solution for biological age prediction based on protein expression data.

[0047] Example 2: In biological age prediction scenarios, protein expression data is characterized by high dimensionality and complex feature associations. Traditional models or simple neural networks often face challenges such as difficulty in capturing the nonlinear relationship between proteins and biological age, susceptibility to data noise interference leading to overfitting during training, and insufficient training stability, directly affecting the efficiency of subsequent model training and the final prediction accuracy. To construct a prediction model architecture that adapts to the characteristics of high-dimensional biological data and combines learning depth with training stability, this embodiment proposes a method for constructing a multilayer perceptron model. By optimizing the network layer design, accurately matching the data dimension, and embedding multifunctional enhancement components, it can efficiently mine the complex relationship between protein expression and biological age, suppress data noise interference, avoid overfitting, ensure the stability of the training process, and lay a high-performance architectural foundation for subsequent initial training and optimization training.

[0048] Step 103 involves constructing a multilayer perceptron model, which can be performed as follows: Step 31: Construct a network hierarchy consisting of an input layer, at least three hidden layers, and an output layer connected in sequence.

[0049] A hierarchical network structure is constructed, consisting of an input layer, at least three hidden layers, and an output layer connected sequentially. The network is built in a serial manner, with no fewer than three hidden layers, overcoming the limitation of shallow networks in learning nonlinear relationships and ensuring that the model has sufficient depth to learn nonlinear associations. The input layer, hidden layer, and output layer are all fully connected layer structures, with each neuron establishing connections with all neurons in the adjacent layers. This can fully capture the cross-correlation between protein features and adapt to the feature extraction needs of one-dimensional protein expression data.

[0050] Step 32: Set the number of neurons in the input layer to match the number of features in the protein data.

[0051] The number of features in the protein data corresponds to the number of protein types involved in training. The number of neurons in the input layer corresponds to this number, with each neuron specifically responsible for receiving the expression level data of one protein. This ensures that every feature of the protein data (i.e., the expression level of each protein) can be accurately received and processed by the model.

[0052] Step 33: The number of neurons in each hidden layer is set to decrease sequentially according to a preset feature compression ratio, and batch normalization components, leakage correction linear unit nonlinear activation functions, and random deactivation components are embedded sequentially in each hidden layer.

[0053] First, set the feature compression ratio, such as an initial hidden layer with 128 neurons, and then decreasing the number of neurons in each subsequent layer by 50% to 64, 32, etc. Configure the number of neurons in each hidden layer according to this ratio. Then, in the computation process of each hidden layer, add a batch normalization component, a Leaky ReLU activation function (with a slope of 0.1), and a Dropout component (with a probability of 0.2) in sequence to ensure that the components work together in the order of normalization → non-linear activation → regularization. The batch normalization component can stabilize the distribution of input data in each layer and reduce gradient fluctuations; Leaky ReLU is used to introduce non-linear transformations, which can avoid negative gradient vanishing compared to traditional ReLU and can improve the model's ability to capture complex non-linear correlations; the Dropout component can randomly shield some neurons, which can prevent the model from over-relying on specific features and effectively suppress overfitting. The collaboration of multiple components ensures a stable and efficient training process.

[0054] Among them, high-dimensional protein data contains redundant information, and the design of decreasing number of neurons can gradually filter out invalid features and focus on core information related to biological age.

[0055] Step 34: Set the number of neurons in the output layer to 1.

[0056] The number of neurons in the output layer corresponds to the output dimension of the model. It is set to 1, meaning that the model will only output a continuous value, namely the prediction result of biological age, in order to meet the output requirements of regression tasks.

[0057] In the multilayer perceptron model with the above architecture, the input layer, hidden layer, and output layer all adopt a fully connected layer structure. This maximizes the capture of global cross-correlation between protein features, avoids the limitations of non-fully connected structures in adapting to one-dimensional protein sequence data, and fits the morphological characteristics of protein expression data. The number of neurons in the input layer is consistent with the number of features in the protein data, ensuring that the expression level features of each protein can be individually received and participate in model learning. The number of neurons in the hidden layer decreases sequentially according to a preset ratio, which can gradually compress redundant information from high-dimensional protein data and focus on low-dimensional core features strongly correlated with biological age, adapting to multiple biological processes in the aging process. Interactive modeling requirements balance model complexity and generalization ability; batch normalization components stabilize the distribution of input data at each layer, reduce gradient fluctuations from interfering with training, and solve the problem of easy divergence in training with high-dimensional data; leakage correction linear units can replace traditional activation functions, effectively avoid negative gradient vanishing, and significantly improve the model's ability to capture complex nonlinear relationships between proteins and biological age; random inactivation components suppress the model's over-reliance on specific features by randomly shielding some neurons, which can reduce the risk of overfitting from the architectural level and enhance the model's robustness on new data; the number of neurons in the output layer is set to 1, directly adapting to the regression task objective of biological age prediction.

[0058] This collaborative architecture enables the model to efficiently process protein expression data from different sources and tissues, and to achieve high prediction accuracy through precise feature extraction and stable training.

[0059] Example 3: In the initial training phase of biological age prediction models, two major technical challenges are faced: First, protein expression data is high-dimensional, noisy, and may contain outliers. A single loss function is prone to the contradiction of pursuing accuracy but being weak against outliers, or emphasizing robustness but lacking accuracy. Second, traditional optimizers are prone to slow convergence and parameter update oscillations in high-dimensional data training, and the lack of effective training process control may lead to model overfitting or insufficient training, directly affecting the basic predictive ability of the initially trained model and the reliability of subsequent feature selection. To ensure that the initial training can efficiently mine the correlation between all protein features and biological age, while also ensuring training stability and model generalization ability, this embodiment proposes a targeted training strategy configuration.

[0060] Specifically, in step 103, the initial training of the multilayer perceptron model is performed based on the preprocessed dataset, which can be carried out in the following steps: Step 35: Configure the Adam optimizer by setting the learning rate of the Adam optimizer to the preset learning rate and the weight decay coefficient to the first weight decay coefficient.

[0061] In this embodiment, the Adam optimizer is configured to perform parameter iterative updates and regularization control during the initial training phase of the multilayer perceptron model. The Adam optimizer combines the stability of momentum gradient descent with the flexibility of adaptive learning rate, and can adapt to the training requirements of high-dimensional data.

[0062] In this embodiment, based on the characteristics of the biological age prediction task, the learning rate is set to a specific value of the preset learning rate. This specific value is not limited in this embodiment. The preset learning rate determines the initial step size of the parameter updates; for example, it can be set to 0.001 to balance convergence speed and update stability. The weight decay coefficient is set as the first weight decay coefficient, which is used to implement L2 regularization and suppress model overfitting. Specifically, it can be set to 1e-4 to control the growth rate of the parameter weights. These parameters are embedded into the optimizer, forming the core tool for model parameter updates.

[0063] Traditional optimizers are prone to problems such as slow convergence and parameter oscillation during training with high-dimensional data. Adaptive momentum optimizers can dynamically adjust the update step size according to the parameter gradient, thereby improving convergence efficiency. The reasonable setting of the learning rate avoids training divergence caused by an excessively large initial step size or training inefficiency caused by an excessively small step size. The weight decay coefficient, by penalizing excessively large parameter weights, suppresses the model's over-reliance on redundant features from the source, thereby reducing the risk of overfitting.

[0064] Step 36: Combine the mean squared error loss function and the smoothing L1 loss function into a hybrid loss function according to a preset weight ratio.

[0065] The Mean Squared Error (MSE) loss function provides more accurate quantification of prediction bias, which is beneficial for improving model prediction accuracy; the Smooth L1 loss function is insensitive to outliers in the data and has stronger robustness. By combining the advantages of the two loss functions, the prediction accuracy of regular samples can be guaranteed by MSE, while the interference of outliers on training can be reduced by Smooth L1, thus achieving a balance between prediction accuracy and outlier resistance.

[0066] The preset weight ratio of the combination, such as MSE weight 0.5 and SmoothL1 weight 0.5, can be adjusted according to the task requirements to determine the contribution of the two losses. The specific values ​​are not limited in this embodiment.

[0067] Step 37: Load the preprocessed dataset in batches according to the preset batch size, input it into the multilayer perceptron model to perform forward propagation calculation, and calculate the loss value corresponding to each batch of data through the hybrid loss function; The preset batch size refers to the number of samples input to the model each time. Batch loading can avoid memory overflow problems caused by loading the full amount of data at once, while improving the parallel efficiency of data processing and computation.

[0068] Forward propagation is the process by which the model extracts features and outputs predicted values ​​from the input data based on the current parameters. The training loss reflects the model's fit on the training data, while the validation loss reflects the model's generalization ability. The loss values ​​include both training and validation loss, which can provide data support for subsequent learning rate adjustments and termination condition determination, thus avoiding model overfitting.

[0069] In this step, a preset batch size is set (e.g., 64, to balance memory usage and training efficiency). The training and validation sets in the preprocessed dataset are loaded in batches using a data loading tool (e.g., Data Loader). Each batch of training and validation data is input into the multilayer perceptron model, which performs forward propagation calculations and outputs the predicted age value for the corresponding batch. The predicted value and the actual age label (target vector) are input into the hybrid loss function to calculate the training loss value and validation loss value for that batch, which are used for subsequent parameter updates and training monitoring.

[0070] Step 38: Call the model training function to start iterative training. After each iteration, the model parameters are updated by backpropagation through the Adam optimizer based on the training loss value and the weight decay coefficient, and the learning rate is adjusted based on the validation loss value.

[0071] Call the preset model training function to start multiple rounds of iterative training, such as... Figure 2 The diagram shows the trend of loss value during model training. In the early stage of training (when the number of rounds is 0), the training loss is about 2500 and the validation loss is about 2000. As training progresses, both types of loss decrease rapidly and tend to stabilize when the number of rounds is about 2.5. Finally, the training loss drops to about 60 and the validation loss stabilizes in the range of 50-70.

[0072] After each iteration, the parameter gradient is calculated based on the training loss value of that iteration. Combined with the first weight decay coefficient, the parameters of each layer of the model (such as the weight matrix from the input layer to the hidden layer) are updated by backpropagation through the adaptive momentum optimizer. Backpropagation refers to the process of deriving the parameter gradient and updating it based on the loss value. At the same time, the validation loss value of each iteration is monitored. If the validation loss value does not decrease for a preset number of iterations (such as 5 iterations), the learning rate is decayed by a preset ratio (such as 0.5), and the parameter update step size is adjusted. The learning rate is dynamically adjusted based on the validation loss. In the early stage of training, a larger step size can be maintained to improve the convergence speed. After the validation loss stabilizes, the step size is reduced for fine optimization, while avoiding model overfitting caused by the increase of validation loss.

[0073] Step 39: Determine whether the training termination condition is met based on the validation loss value and the number of iterations. If it is met, stop training and output the current model as the initial training model.

[0074] Pre-set training termination conditions, such as setting dual training termination conditions: first, the number of iterations reaches a preset maximum number of rounds (e.g., 200 rounds); second, the loss value is verified to have not decreased for a preset number of rounds (e.g., 20 rounds) (early stop condition). In this embodiment, the specific setting of the training termination conditions is not limited, and only the above conditions are used as examples for introduction.

[0075] After each iteration, synchronously check whether any termination condition is met; if it is met, immediately stop iterative training, save all parameters of the current model (such as weight matrix, layer configuration, etc.), and use the model output as the initial training model.

[0076] Based on the above introduction, the initial training method of this model uses an adaptive momentum optimizer combined with weight decay to achieve efficient regularization, a hybrid loss function to balance prediction accuracy and outlier resistance, batch loading to improve training efficiency and dual loss monitoring, dynamic learning rate adjustment and dual termination conditions to precisely control the training process, and multiple strategies to adapt to the characteristics of high-dimensional protein data, achieving fast convergence and effectively preventing overfitting. This ensures that the initial trained model has both solid learning ability and stable generalization performance.

[0077] Example 4: Based on Example 3, after completing the initial training based on the full set of protein data, in order to fully leverage the core value of the feature protein set and further improve the model training efficiency and prediction accuracy, this example proposes a feature optimization training configuration method based on the initial training process described in Example 3. By selectively adjusting the core parameters of the model to match and refine feature characteristics, extracting target datasets that focus on core features, and continuing the efficient training process, the model can deeply focus on the core association between high-contribution proteins and biological age, reduce redundant calculations and noise interference, and achieve simultaneous improvement in training efficiency, prediction accuracy, and generalization ability.

[0078] Step 105 involves performing feature optimization training on the initially trained model based on the feature protein set. This can be done by following these steps: Step 51: Adjust the number of neurons in the input layer of the multilayer perceptron model to the number of features in the feature protein set, adjust the loss function to the smoothed L1 loss function, and adjust the weight decay coefficient to the second weight decay coefficient.

[0079] First, count the total number of features in the feature protein set (e.g., 500), then update the number of neurons in the model's input layer to this value to ensure that each core protein feature is accurately received and that the model is precisely matched with the refined core feature data.

[0080] The initial training hybrid loss function is replaced with a smooth L1 loss function, focusing on the robustness requirements of refined data and leveraging its resistance to outliers and computational stability to adapt to low-noise core datasets.

[0081] Adjust the weight decay coefficient from the first weight decay coefficient (e.g., 1e-4) to the second weight decay coefficient (e.g., 1e-3) to enhance the regularization constraint.

[0082] Step 52: Extract the dataset that matches the set of feature proteins from the preprocessed dataset as the target dataset.

[0083] Based on the protein type list of the feature protein set, the feature matrix of the preprocessed dataset is traversed to select all feature columns whose column names match the protein types in the list; the sample data corresponding to these feature columns (including the feature matrices and target vectors of the training set and validation set) are retained and integrated according to the original data structure to form the target dataset, ensuring that the format and labels of the target dataset are consistent with the preprocessed dataset, with only the feature dimensions simplified.

[0084] Step 53: Load the target dataset in batches according to the preset batch size, execute the step of inputting it into the multilayer perceptron model to perform forward propagation calculation, and obtain the model output after training stops.

[0085] Using the same preset batch size as the initial training (e.g., 64), the training and validation sets of the target dataset are loaded in batches using a data loading tool. Then, the process jumps to step 37 to execute forward propagation calculations and subsequent steps by inputting the data into the multilayer perceptron model. Specifically, each batch of data is input into the adjusted multilayer perceptron model, and forward propagation calculations are performed to obtain the predicted values. The training loss and validation loss are calculated based on the smooth L1 loss function. The iterative update logic of the initial training (updating parameters through backpropagation of the Adam optimizer), the dynamic learning rate adjustment strategy (based on validation loss decay), and the dual termination conditions (upper limit of iterations + early stopping due to validation loss) are used until the termination conditions are met, and the model output after training is stopped is obtained.

[0086] Based on the above introduction, the feature optimization training method provided in this embodiment achieves deep adaptation between the model and core features by precisely adjusting the number of neurons in the input layer to match the dimension of the feature protein set, switching the smooth L1 loss function to match the low noise characteristics of the refined dataset, and increasing the second weight decay coefficient to enhance regularization and prevent overfitting. It also selectively extracts the target dataset to remove redundant interference, uses batch loading to balance training efficiency and memory control, and adopts mature forward propagation and training management logic to ensure process stability. Overall, it focuses on the core association between high-contribution proteins and biological age, reducing computational complexity while deepening the model's learning of key patterns, ultimately achieving a synergistic improvement in training efficiency, prediction accuracy, and generalization ability.

[0087] Example 5: After the initial training of a multilayer perceptron model is completed, it is necessary to screen high-contribution protein features based on the model's weight parameters. However, traditional feature screening methods often suffer from vague selection criteria and a lack of understanding of the model's actual learning patterns. Some methods rely on independent statistical features (such as variance and correlation), failing to fully utilize the protein-age association information learned during model training. This may lead to a mismatch between the selected high-contribution features and the core associations actually learned by the model, resulting in redundant features or missing key features, which in turn affects the efficiency and accuracy of subsequent optimization training. To ensure that the feature screening process is scientific, objective, and aligned with the model's learning logic, this embodiment proposes a method for screening high-contribution protein features. By extracting the core weight matrix from the model's input layer to the first hidden layer, the feature contribution is quantified by the absolute value of the weights. After sorting in descending order, key features are selected in a targeted manner. This method can accurately identify core proteins strongly associated with biological age, avoiding biases caused by external statistical indicators. It ensures that the screening results are highly consistent with the model's learning patterns, providing a well-targeted and high-quality feature foundation for subsequent optimization training.

[0088] Step 104: Based on the weight parameters of the initially trained model, screen for high-contribution characteristic proteins related to biological age. This can be performed as follows: Step 41: Extract the weight matrix from the input layer to the first hidden layer of the initial training model.

[0089] The weight matrix directly maps the influence of each protein feature on the initial feature extraction of the model. The weight matrix from the input layer to the first hidden layer is the core parameter for the first feature transformation after the protein features enter the model. Each protein corresponds to a column of weights in the matrix. The magnitude of the weight value directly reflects the initial contribution potential of the protein in biological age prediction. It can retain the original contribution information of the protein features to the greatest extent, avoid the distortion of contribution caused by subsequent layer transformations, and provide the most direct and reliable basis for screening.

[0090] Step 42: Based on the weight matrix, sort the features corresponding to all protein-related columns in descending order of absolute weight value.

[0091] The magnitude of the absolute value of the weight directly corresponds to the strength of the feature's contribution. The larger the absolute value, the stronger the influence of the protein on the model output (biological age prediction result) and the closer its correlation with biological age. For the extracted weight matrix, the absolute value of the weight value of each column (corresponding to a single protein) is calculated, and a correspondence table between protein name and absolute weight value is established. Using the absolute weight value as the sorting key, all protein features are sorted in descending order from largest to smallest, generating a feature importance ranking list. After sorting, high-contribution and low-contribution features can be intuitively distinguished.

[0092] Step 43: Select the proteins corresponding to the preset number of features that rank highly as high-contribution feature proteins.

[0093] From the descendingly sorted list, a predetermined number of proteins at the top of the sorting are extracted and integrated to form a structured set of high-contribution feature proteins. The specific number of proteins extracted is not limited in this embodiment and can be set according to the computational cost of model training and the accuracy requirements of subsequent optimization training; for example, it can be set to 500 or 100.

[0094] The high-contribution protein screening method provided in this embodiment uses the weight matrix from the input layer to the first hidden layer of the initial training model as the core basis to directly quantify the initial contribution of proteins to biological age prediction. The screening results are highly consistent with the actual learning pattern of the model, avoiding the correlation bias caused by the reliance on independent statistical features in traditional methods. A unified and reproducible screening criterion is established by sorting the absolute values ​​of the weights in descending order, clearly distinguishing the importance level of features and eliminating subjective judgment errors. High-contribution proteins are selected in a predetermined number to precisely control the feature dimension and effectively eliminate redundant and low-correlation noise features, which reduces the computational complexity of subsequent optimization training and reduces the risk of overfitting. The final generated feature protein set is highly targeted, providing a high-quality feature foundation for the subsequent model to focus on core correlations and deepen learning.

[0095] Example 6: In the above embodiments, there are no restrictions on the type of operation or the execution order of the preprocessing operations for the read biological sample data. In order to systematically improve the integrity, consistency and applicability of biological sample data and eliminate the interference of various data defects on subsequent MLP model training, this embodiment proposes a preprocessing method. By sequentially performing the entire process of data cleaning, missing value imputation, format conversion, standardization, stratified sampling and splitting, and feature and target vector separation, invalid interference data can be effectively removed, sample age distribution can be balanced, and dimensional differences can be eliminated. This ensures that the data input to the model has high-quality and standardized characteristics, laying a solid foundation for the subsequent model to accurately explore the correlation between protein expression and biological age.

[0096] Step 102 involves preprocessing the biological sample data, which can be performed as follows: Step 21: Clean the invalid data in the biological sample data and remove samples and protein-related features that do not meet the data quality requirements.

[0097] Invalid data refers to information that cannot reflect true biological characteristics or may interfere with model training. This includes samples with too many missing values, feature columns with an excessive proportion of missing protein expression data, outliers that deviate significantly from the reasonable range, and meaningless duplicate records.

[0098] Data quality non-compliance refers to the preset data quality requirements, which can be set based on preset standards (such as the percentage of missing values ​​in the sample, the percentage of missing values ​​in protein features, and the threshold for outlier determination).

[0099] First, set data quality thresholds (e.g., the percentage of missing values ​​in samples should not exceed 20%, and the percentage of missing values ​​in protein feature columns should not exceed 50%). Then, use data cleaning tools to traverse the biological sample data, such as using the clean_data function to clean the data, remove samples with more than 20% missing values, filter protein columns with more than 50% missing values, convert them to numeric type using pd.to_numeric, identify and delete samples that meet the invalid data criteria, and remove protein-related feature columns that do not meet the quality requirements, retaining only data units that meet the completeness and reliability standards.

[0100] Step 22: For missing values ​​in the cleaned data, determine the statistical feature values ​​of the corresponding feature columns and fill them in.

[0101] The statistical characteristic value of the corresponding characteristic column refers to a representative value (such as median, mean, mode, etc.) calculated based on the valid data in the characteristic column. The appropriate statistical measure can be selected according to the data distribution characteristics, but this embodiment does not limit it.

[0102] Identify the location of all missing values ​​and their corresponding feature columns in the cleaned data. For each column of protein features or key demographic information, calculate the statistical feature value of its effective data (e.g., the median is often used to fill in protein expression data to avoid the influence of extreme values). Then, use the statistical feature value to fill in the missing positions of the corresponding columns to ensure that there are no data gaps in each row of samples and each feature column.

[0103] Step 23: Convert all the filled data to a preset numerical format.

[0104] Define the target numerical format that the model is compatible with. Use data conversion tools to traverse all the filled data and convert non-numerical information such as text data and categorical data into corresponding numerical codes (e.g., map categorical variables to integers and convert text tags to floating-point numbers). Ensure that all data columns are stored in the preset numerical format without any format conflicts. Specifically, Standard Scaler can be used to standardize protein expression values ​​using Z-score, but it is not limited to this.

[0105] Step 24: Standardize the protein expression data after format conversion.

[0106] Using a pre-defined standardization algorithm (such as Z-score standardization, Min-Max standardization, etc.), all protein expression data after format conversion are processed. Statistical parameters (such as mean, standard deviation, maximum, and minimum) of each protein expression class are calculated. Each data point is converted into a standardized value through an algorithm formula, mapping protein expression data of different magnitudes to a unified numerical range (such as mean of 0 and variance of 1). This avoids the model from being overly biased towards high-magnitude features due to differences in the expression magnitude of proteins themselves, and ensures that all protein expression data are at the same magnitude level.

[0107] Step 25: Based on the actual age label, the standardized biological sample data is split into training set, validation set and test set using stratified sampling.

[0108] First, divide the data into age ranges based on actual age labels (e.g., 10 years per range) and count the proportion of samples in each range in the original data. Then, according to a preset ratio (e.g., 64% for training set, 16% for validation set, and 20% for test set), extract samples from each age range and integrate them into training set, validation set, and test set respectively. This ensures that the age distribution of the three datasets is highly consistent with the original data. The three datasets respectively play the core role of monitoring the model learning and training process, adjusting parameters, and finally evaluating performance.

[0109] The specific preset ratio is not limited in this embodiment, and can be set according to the actual training and verification needs.

[0110] Step 26: Separate the feature matrix and target vector from each dataset.

[0111] The feature matrix is ​​the core basis for the model to learn the correlation between biological ages. It consists of standardized protein expression data, with each row representing a sample and each column representing the expression level of a protein.

[0112] The target vector is the standard answer for model training. It consists of the actual age label corresponding to each sample, with each element corresponding to the true age of a sample.

[0113] The training, validation, and test sets are decomposed separately. Standardized protein expression levels for all samples in each dataset are extracted and organized into a feature matrix based on the sample-feature dimension. Simultaneously, the actual age labels for all samples in each dataset are extracted and arranged in sample order to form a target vector, ensuring a one-to-one correspondence between the rows of the feature matrix and the elements of the target vector for each sample. By separating the feature matrix and the target vector, the structured dataset is converted into an input format directly recognizable by the model, providing a suitable data format for subsequent model training and parameter optimization.

[0114] Step 27: Integrate the training set, validation set, test set, and corresponding feature matrices and target vectors as a preprocessing dataset.

[0115] A data association index is established to bind the training set with the corresponding training feature matrix and training target vector, the validation set with the corresponding validation feature matrix and validation target vector, and the test set with the corresponding test feature matrix and test target vector. At the same time, the data storage format is unified to ensure that the three types of datasets and their corresponding matrices and vectors can be quickly accessed through the index, forming a structured preprocessed dataset.

[0116] The preprocessing method provided in this embodiment improves data quality and consistency through systematic cleaning, filling, and standardization operations. Through stratified sampling and feature-target vector separation, it not only meets the supervised training requirements of MLP models but also ensures the representativeness of the dataset distribution, ultimately integrating to form a structured preprocessed dataset. It should be noted that this embodiment only uses the above preprocessing process as an example for introduction. Different settings are made according to different data processing needs and application scenarios. This embodiment is not limited to this. Other processing methods can refer to the introduction of this embodiment, and will not be repeated here.

[0117] Example 7: The above embodiments have constructed a biological age prediction model that focuses on core features and possesses both high accuracy and strong generalization ability. This embodiment proposes a biological age prediction method based on the biological age prediction model proposed in the above embodiments, such as... Figure 3The diagram shows a flowchart of a method for predicting biological age, which includes the following steps: Step 106: Receive the biological sample data to be predicted input by the user.

[0118] Receive protein expression-related data of the organism whose age is to be predicted, i.e., the sample data of the organism to be predicted, provided by the user.

[0119] The biological sample data to be predicted includes two legitimate types: one is full protein expression data, which is the expression level data of all detected proteins that have not been filtered by the user; the other is feature protein expression data, which is the expression level data of the feature protein set that has been pre-selected by the user and is completely corresponding to the protein types of the feature protein set generated during the training process.

[0120] If the number of protein types in the data is the same as the total number of proteins during training (or if no feature selection markers are explicitly marked), it can be identified as full protein expression data; if the protein types in the data completely match the list of feature protein sets (which can be compared by unique identifiers such as protein name and number), it can be identified as feature protein expression data.

[0121] This step is compatible with both data types, avoiding forcing users to perform additional feature processing and lowering the barrier to entry.

[0122] Step 107: Perform preprocessing operations on the biological sample data to be predicted to obtain preprocessed data.

[0123] Perform the same preprocessing operations as during model training on the user-input biological sample data (full protein expression data or feature protein expression data) to transform the raw input data into preprocessed data with a uniform format, meeting quality standards, and having identifiable features. Establish a quality and format alignment mechanism between the data to be predicted and the training data to ensure that the data input to the model is consistent with the preprocessed dataset used during training in terms of distribution characteristics, numerical range, and structural regularity.

[0124] The preprocessing operations will not be described in detail in this embodiment. Please refer to the description in the above embodiments.

[0125] Step 108: If the biological sample data to be predicted is characteristic protein expression data, input the preprocessed data into the trained biological age prediction model and output the biological age prediction result.

[0126] When the user-input data to be predicted has been clearly identified as feature protein expression data (and meets the requirement of matching the types of feature protein sets), and after preprocessing in step 107, the preprocessed data is of high quality and uniform format, the preprocessed data is directly input into the biological age prediction model that has undergone initial training and feature optimization training. The model outputs the final biological age prediction result through calculation and inference. The model can accurately call the learned core feature association rules, avoid inference bias caused by data differences, and quickly complete the prediction inference.

[0127] Step 109: If the biological sample data to be predicted is full protein expression data, select protein expression level data that are consistent with the types of the feature protein set from the preprocessed data and generate the target feature matrix.

[0128] The number of neurons in the input layer of a trained biological age prediction model is adapted to the number of features in the feature protein set (e.g., 500). However, the number of columns in the full protein expression data (e.g., thousands) is far greater than this. Direct input would lead to incompatibility between the model's input dimensions, causing calculation errors or prediction biases. When the user-input data to be predicted is determined to be full protein expression data, and after preprocessing in step 107 (cleaning, filling, standardization, etc.) to form preprocessed data of acceptable quality, expression level data that is completely consistent with the protein types of the feature protein set generated during model training is precisely selected from this preprocessed data. Finally, these are integrated to form a target feature matrix with a standardized format and matching dimensions. The target feature matrix contains only a two-dimensional numerical matrix of feature protein expression level data, with rows corresponding to the samples to be predicted and columns corresponding to feature proteins. The matrix structure is completely adapted to the input feature matrix during model training. Through selection, the data dimension can be precisely compressed to the model-adaptive dimension, achieving data-model dimensional alignment.

[0129] Specifically, a list of solidified characteristic protein sets (containing key information such as protein name, unique identifier, and feature sequence number) can be extracted from the storage module of the biological age prediction model. This list serves as the core reference standard for screening. The column names of the preprocessed full protein expression data (each column corresponds to a protein) are parsed. The unique identifier of each protein is compared with the list of characteristic protein sets one by one to identify all columns that match successfully, i.e., the expression level data columns that are consistent with the type of characteristic protein. All columns that match successfully are retained, and redundant protein columns that do not match the characteristic protein set in the preprocessed data are removed to ensure that the remaining data only contains the expression level information corresponding to the core features.

[0130] Step 110: Input the target feature matrix into the biological age prediction model and output the biological age prediction result.

[0131] The target feature matrix, which contains only the expression level data of characteristic proteins and is generated after screening in step 109, is input into the biological age prediction model that has undergone initial training and feature optimization training. Through the calculation and reasoning of the model, the biological age prediction result of the corresponding sample to be predicted is finally output.

[0132] Based on the above introduction, the biological age prediction method provided in this embodiment supports two input types: full protein expression data and feature protein expression data. This satisfies the needs of ordinary users who do not have feature screening capabilities, and is also suitable for professional users who have pre-processed core features. It does not require mandatory data format uniformity, which can significantly reduce the usage threshold. For the two input types, the two data types are processed separately. Feature data is directly connected to the model to achieve rapid prediction, while full data is screened to complete accurate inference. This ensures both the efficiency of batch prediction scenarios and the traceability and stability of the results. The prediction process relies entirely on the mature model formed by initial training and optimization training. The core feature association rules and weight parameters are fully reused. Combined with data adaptation and screening design, the model can maintain high robustness under different data input scenarios.

[0133] It should be noted that the biological age prediction method provided in this embodiment and the training method of the biological age prediction model provided in the above embodiments can be referred to each other, and the repeated parts will not be described again in this embodiment.

[0134] Example 8: This embodiment relates to a training device for a biological age prediction model. A schematic diagram of the training device for this biological age prediction model is shown below. Figure 4 As shown, it includes: a sample loading unit 201, a sample preprocessing unit 202, an initial training unit 203, a protein screening unit 204, and an optimized training unit 205.

[0135] The sample loading unit 201 is used to call the data reading interface to load biological sample data. The biological sample data includes: demographic data and protein data. The demographic data includes: participant identification and actual age label. The protein data includes expression level data of various proteins. The sample preprocessing unit 202 is used to perform preprocessing operations on biological sample data to obtain a preprocessed dataset. The initial training unit 203 is used to build a multilayer perceptron model and perform initial training on the multilayer perceptron model based on the preprocessed dataset to obtain a preliminary trained model. Protein screening unit 204 is used to screen high-contribution feature proteins related to biological age based on the weight parameters of the preliminary trained model, and generate a set of feature proteins. The optimization training unit 205 is used to perform feature optimization training on the initial training model based on the feature protein set, and output the feature-optimized model as a biological age prediction model.

[0136] It should be noted that the content of the training device for the biological age prediction model provided in this embodiment can be referred to in conjunction with the training method for the biological age prediction model provided in the above embodiments, and the repeated parts will not be described again in this embodiment.

[0137] In the training device of the biological age prediction model provided in this embodiment, the sample loading unit focuses on demographic data and protein expression data, using protein as an objective biomarker as the core training basis, thus eliminating the dependence of traditional models on methylation data and subjective clinical indicators; the sample preprocessing unit generates a high-quality preprocessed dataset; the MLP model constructed by the initial training unit can efficiently capture the complex nonlinear relationship between protein expression and biological age, quickly establishing a robust preliminary training model; the protein screening unit selects high-contribution feature proteins based on the weight matrix of the preliminary training model, eliminating redundant and low-correlation components by quantifying feature importance, which reduces the computational complexity and overfitting risk of subsequent training, while ensuring that the model focuses on core biological associations and improves generalization ability; the optimization training unit dynamically adjusts the model input layer dimension, loss function, and weight decay coefficient for the selected feature protein set, allowing the model to deeply focus on the association rules of core features, and the final output model has both high prediction accuracy (R²) and high performance. 2 Up to 0.8684) and low mean absolute error (MAE as low as 2.35 years).

[0138] This training device ensures the traceability and stability of the training process, supports efficient training with batch samples, and is suitable for application scenarios with large-scale biological sample databases, providing high-performance model support for subsequent accurate predictions.

[0139] Furthermore, it should be noted that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units do not exist in this embodiment.

[0140] Example 9: This embodiment relates to a biological age prediction device. A schematic diagram of the biological age prediction device in this embodiment is shown below. Figure 5As shown, it includes: a prediction data receiving unit 206, a data preprocessing unit 207, a feature protein processing unit 208, a full protein extraction unit 209, and a matrix prediction unit 210.

[0141] The prediction data receiving unit 206 is used to receive the biological sample data to be predicted input by the user; the biological sample data to be predicted includes: full protein expression data or characteristic protein expression data; the characteristic protein expression data is matched with the protein types of the characteristic protein set; Data preprocessing unit 207 is used to perform preprocessing operations on the biological sample data to be predicted to obtain preprocessed data; The feature protein processing unit 208 is used to input preprocessed data into a trained biological age prediction model and output biological age prediction results if the biological sample data to be predicted is feature protein expression data. The full protein extraction unit 209 is used to filter out protein expression data that are consistent with the types of the feature protein set from the preprocessed data if the biological sample data to be predicted is full protein expression data, and generate a target feature matrix. The matrix prediction unit 210 is used to input the target feature matrix into the biological age prediction model and output the biological age prediction result.

[0142] It should be noted that the content of the biological age prediction device provided in this embodiment can be referred to in conjunction with the biological age prediction method provided in the above embodiments, and the repeated parts will not be described again in this embodiment.

[0143] In the biological age prediction device provided in this embodiment, the prediction data receiving unit is compatible with both full protein expression data and feature protein expression data. Users do not need to uniformly format data or be forced to perform feature preprocessing. This satisfies the precise data input needs of professional users and also adapts to the full data submission scenarios of ordinary users, significantly improving the device's versatility and practicality. The data preprocessing unit continues the standardized processing logic of the biological age prediction model training phase to ensure data quality. The feature protein processing unit directly connects to the trained model for matched feature data, eliminating redundant screening steps and achieving rapid prediction. The full protein extraction unit generates a dimensionally adapted target feature matrix by accurately screening core data consistent with the feature protein set. This eliminates low-contribution redundant features, reduces computational complexity, and ensures that the core associations of the input model are consistent with the training logic, achieving dual-path adaptation of rapid prediction and accurate inference.

[0144] This device relies on a biological age prediction model optimized by the training device. It fully reuses the correlation patterns and weight parameters between high-contribution proteins and biological age learned by the model. Combined with data adaptation and core feature focusing design, the model can maintain high robustness under different data input scenarios. The prediction results continue the high performance indicators of the training stage, while inheriting the cross-tissue applicability of the model and adapting to the prediction needs of multiple types of biological samples.

[0145] Furthermore, it should be noted that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units do not exist in this embodiment.

[0146] Example 10: Another embodiment of this application relates to an electronic device, such as... Figure 6 As shown, it includes: at least one processor 301; and a memory 302 communicatively connected to at least one processor 301; wherein the memory 302 stores instructions executable by at least one processor 301, the instructions being executed by at least one processor 301 to enable at least one processor 301 to perform the steps of the training method and / or the prediction method of the biological age prediction model in the above embodiments.

[0147] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0148] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0149] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A training method for a biological age prediction model, characterized in that, include: Call the data reading interface to load biological sample data; The biological sample data includes: demographic data and protein data; the demographic data includes: participant identifiers and actual age labels; the protein data includes expression level data for multiple proteins; The biological sample data is preprocessed to obtain a preprocessed dataset; A multilayer perceptron model is constructed, and initial training is performed on the multilayer perceptron model based on the preprocessed dataset to obtain a preliminary trained model; Based on the weight parameters of the preliminary training model, high-contribution feature proteins related to biological age are screened to generate a feature protein set. Based on the set of characteristic proteins, the initial training model is subjected to feature optimization training, and the feature-optimized model is output as a biological age prediction model.

2. The training method for the biological age prediction model according to claim 1, characterized in that, The construction of the multilayer perceptron model includes: A network hierarchy structure is constructed, comprising an input layer, at least three hidden layers, and an output layer connected in sequence; wherein the input layer, hidden layers, and output layer are all fully connected layer structures. Set the number of neurons in the input layer to be consistent with the number of features in the protein data; The number of neurons in each hidden layer is set to decrease sequentially according to a preset feature compression ratio, and a batch normalization component, a leakage correction linear unit nonlinear activation function, and a random deactivation component are sequentially embedded in each hidden layer. Set the number of neurons in the output layer to 1.

3. The training method for the biological age prediction model according to claim 1, characterized in that, The multilayer perceptron model is initially trained based on the preprocessed dataset, including: Configure the Adam optimizer by setting its learning rate to a preset learning rate and its weight decay coefficient to a first weight decay coefficient. The mean squared error loss function and the smoothing L1 loss function are combined into a hybrid loss function according to a preset weight ratio. The preprocessed dataset is loaded in batches according to a preset batch size and input into the multilayer perceptron model to perform forward propagation calculation. The loss value corresponding to each batch of data is calculated through the hybrid loss function. The loss value includes: training loss value and validation loss value. Initiate iterative training by calling the model training function. After each iteration, update the model parameters through backpropagation using the Adam optimizer based on the training loss value and the weight decay coefficient, and adjust the learning rate based on the validation loss value. Based on the validation loss value and the number of iterations, determine whether the training termination condition is met. If it is met, stop training and output the current model as the initial training model.

4. The training method for the biological age prediction model according to claim 3, characterized in that, Based on the set of featured proteins, feature optimization training is performed on the initial trained model, including: The number of neurons in the input layer of the multilayer perceptron model is adjusted to the number of features in the feature protein set, the loss function is adjusted to the smoothed L1 loss function, and the weight decay coefficient is adjusted to the second weight decay coefficient. Extract the dataset that matches the set of feature proteins from the preprocessed dataset, and use it as the target dataset; The target dataset is loaded in batches according to a preset batch size, and the forward propagation calculation is performed by inputting it into the multilayer perceptron model to obtain the model output after training stops.

5. The training method for the biological age prediction model according to claim 1, characterized in that, Based on the weight parameters of the preliminary trained model, high-contribution feature proteins related to biological age are screened, including: Extract the weight matrix from the input layer to the first hidden layer of the initially trained model; Based on the weight matrix, the features corresponding to all protein-related columns are sorted in descending order of absolute weight value; The proteins corresponding to the highest-ranked features are selected as the high-contribution feature proteins.

6. The training method for the biological age prediction model according to claim 1, characterized in that, Perform preprocessing operations on the biological sample data, including: The invalid data in the biological sample data is cleaned, and samples and protein-related features that do not meet the data quality requirements are removed. For missing values ​​in the cleaned data, determine the corresponding statistical feature values ​​of the feature columns and fill them in; Convert all filled data to the preset numerical format; Standardize the protein expression data after format conversion; Based on the actual age label, the standardized biological sample data is split into training set, validation set and test set using stratified sampling. Feature matrices and target vectors are separated from each dataset; the feature matrices are standardized protein expression data, and the target vectors are numerical vectors corresponding to actual age labels; the datasets include: training set, validation set, and test set. The training set, the validation set, the test set, and the corresponding feature matrix and target vector are integrated to form the preprocessed dataset.

7. A method for predicting biological age, characterized in that, The training method based on the biological age prediction model according to any one of claims 1 to 6, the prediction method comprising: Receive biological sample data to be predicted input by the user; the biological sample data to be predicted includes: full protein expression data or characteristic protein expression data; the characteristic protein expression data is matched with the protein types of the characteristic protein set; Preprocessing operations are performed on the biological sample data to be predicted to obtain preprocessed data; If the biological sample data to be predicted is the characteristic protein expression data, the preprocessed data is input into the trained biological age prediction model, and the biological age prediction result is output. If the biological sample data to be predicted is the full protein expression data, protein expression level data that are consistent with the types of the feature protein set are selected from the preprocessed data to generate a target feature matrix; The target feature matrix is ​​input into the biological age prediction model, and the biological age prediction result is output.

8. A training device for a biological age prediction model, characterized in that, include: The sample loading unit is used to call the data reading interface and load biological sample data. The biological sample data includes: demographic data and protein data; the demographic data includes: participant identifiers and actual age labels; the protein data includes expression level data for multiple proteins; A sample preprocessing unit is used to perform preprocessing operations on the biological sample data to obtain a preprocessed dataset. An initial training unit is used to construct a multilayer perceptron model and perform initial training on the multilayer perceptron model based on the preprocessed dataset to obtain a preliminary trained model. The protein screening unit is used to screen high-contribution characteristic proteins related to biological age based on the weight parameters of the preliminary training model, and generate a set of characteristic proteins. An optimization training unit is used to perform feature optimization training on the initial training model based on the set of feature proteins, and output the feature-optimized model as a biological age prediction model.

9. A device for predicting biological age, characterized in that, The training device based on the biological age prediction model of claim 8, the prediction device comprising: A prediction data receiving unit is used to receive biological sample data to be predicted input by a user; the biological sample data to be predicted includes: full protein expression data or characteristic protein expression data; the characteristic protein expression data is matched with the protein types of the characteristic protein set; The data preprocessing unit is used to perform preprocessing operations on the biological sample data to be predicted to obtain preprocessed data; The feature protein processing unit is used to input the preprocessed data into the trained biological age prediction model and output the biological age prediction result if the biological sample data to be predicted is the feature protein expression data. A full protein extraction unit is used to, if the biological sample data to be predicted is the full protein expression data, filter out the protein expression level data that is consistent with the type of the feature protein set from the preprocessed data, and generate a target feature matrix. The matrix prediction unit is used to input the target feature matrix into the biological age prediction model and output the biological age prediction result.

10. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the training method of the biological age prediction model as described in any one of claims 1 to 6 and / or the biological age prediction method as described in claim 7.

Images