Chicken deterioration marker screening and AICAR prediction method and system based on multi-task learning and transfer learning

Abstract

The invention provides a chicken deterioration marker screening and AICAR prediction method and system based on multi-task learning and transfer learning. The method comprises the following steps: selecting chicken samples with different storage conditions and shelf lives, collecting metabolite data of multiple batches, and carrying out quantitative analysis on the metabolite by utilizing high performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), liquid chromatography-mass spectrometry (LC-MS) and other technologies. Through the multi-task learning model, multiple differential metabolites are deeply mined, and the prediction precision and the calculation efficiency are improved. And providing chicken storage period optimization suggestions according to the AICAR predicted value. The innovative method and system can significantly improve scientificity and accuracy of chicken quality evaluation, have wide application prospects, and are especially suitable for the fields of food quality monitoring, supply chain management and intelligent storage and transportation.

Description

Technical Field

[0001] The present invention relates to the application field of food quality monitoring and spoilage assessment, and particularly to a screening system and method for chicken spoilage markers and AICAR prediction based on multi-task learning and transfer learning. Background Art

[0002] In the field of food safety, the spoilage assessment of meat products has long relied on visual inspection and sensory evaluation. However, these traditional methods have disadvantages such as strong subjectivity, long assessment cycles, and cumbersome operations, and are difficult to meet the needs of modern and automated food safety monitoring. Especially for perishable foods such as chicken, its spoilage process is affected by various factors, including storage temperature, humidity, and storage duration. Most of the existing spoilage assessment methods focus on sensory evaluation and simple microbiological detection, unable to deeply analyze the molecular-level changes during spoilage, and lacking an efficient and real-time warning system.

[0003] In recent years, the rapid development of metabolomics technology has provided new ideas for food spoilage detection. Through techniques such as high-performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), and liquid chromatography-mass spectrometry (LC-MS), metabolite characteristic data in chicken can be extracted in a high-throughput manner to analyze the spoilage process of chicken. However, how to screen out key markers related to chicken spoilage from the huge metabolite data and accurately predict spoilage under different storage conditions remains an urgent problem to be solved.

[0004] The application of artificial intelligence, especially deep learning technology, can process complex data and extract potential rules and patterns from it. By using multi-task learning and transfer learning algorithms and combining cross-sample data of different storage conditions and spoilage stages, a more accurate and adaptable spoilage prediction model can be constructed. Therefore, combining metabolomics analysis with a multi-task learning model provides unprecedented possibilities for chicken spoilage assessment, enabling real-time and accurate spoilage monitoring and shelf-life prediction. Summary of the Invention

[0005] The purpose of the present invention is to provide a screening method and system for chicken spoilage markers and AICAR prediction based on multi-task learning and transfer learning to solve problems such as the inability of traditional spoilage detection methods to real-time and accurately evaluate the spoilage state of chicken and predict the remaining shelf life.

[0006] To achieve the above purpose, the present invention proposes the following technical solution: A screening system for chicken spoilage markers and AICAR prediction based on multi-task learning and transfer learning, including a sample collection module, a metabolite extraction module, a metabolomics analysis module, a multi-task learning model module, and an AICAR prediction module connected in sequence;

[0007] Sample collection module: This module is responsible for collecting chicken samples under different storage conditions and shelf lives, selecting chicken samples at different storage time periods, fresh, 3 days, 7 days, and 10 days, covering all stages of chicken from fresh to spoiled; the sample collection at each stage ensures the consistency between batches and the diversity of data;

[0008] Metabolite extraction module: During the metabolite extraction process, polar metabolites such as amino acids and acids use methanol, acetonitrile, or water-acetonitrile mixed solvents, while lipophilic metabolites use chloroform and dichloromethane solvents; the chicken samples are processed by cryogenic grinding technology to maintain the stability of metabolites and maximize the uniformity of the samples, providing accurate and high-quality samples for metabolomics analysis;

[0009] Metabolomics analysis module: High-throughput analysis techniques such as high-performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), and liquid chromatography-mass spectrometry (LC-MS) are used to perform metabolomics analysis on chicken samples; amino acids, sugars, and organic acids, polar and water-soluble metabolites are analyzed by HPLC; then, lipids, fatty acids, and carotenoids, volatile and non-volatile metabolites are analyzed by GC-MS and LC-MS; during the analysis, special attention is paid to the concentration changes of AICAR (5-aminoimidazole-4-carboxamide riboside) and its related metabolites to reveal the key metabolic markers during the spoilage process of chicken;

[0010] Multi-task learning model module: This module uses a multi-task learning (MTL) model to simultaneously process multi-dimensional data of chicken spoilage; through multi-task training on the collected metabolite data; through shared learning representations, the model can improve the ability to identify spoilage markers and enhance the generalization ability of the system on different data sets;

[0011] AICAR prediction module: This module uses a transfer learning algorithm, combined with the output of the multi-task learning model, to specifically predict the concentration changes of AICAR and its related metabolites; through transfer learning, the system can quickly transfer the data characteristics of different storage conditions and spoilage stages to new samples to achieve accurate prediction of AICAR; the core goal of this module is to real-time monitor the quality changes of chicken, predict its spoilage stage, and finally generate optimization suggestions for the storage period of chicken, so as to provide a scientific basis for the processing, transportation, and sales of chicken.

[0012] A method for screening chicken spoilage markers and predicting AICAR based on multi-task learning and transfer learning of the present invention includes the following steps:

[0013] Step 1: Sample collection

[0014] In the experiment of screening chicken spoilage markers and predicting AICAR, the sampling should cover fresh samples and chicken samples stored for 3 days, 7 days, and 10 days under different storage conditions. These samples can comprehensively reflect the biochemical changes of chicken at different stages from fresh to spoiled. At the same time, ensure that the samples under each storage condition and storage time come from multiple batches. In addition, the environmental factors affecting chicken spoilage during storage, including storage temperature and humidity variables, need to be recorded in real time during the collection process to ensure accurate data support for subsequent metabolite analysis and spoilage process modeling.

[0015] Step 2: Metabolite extraction

[0016] During the metabolite extraction process, polar metabolites such as amino acids and acids use methanol, acetonitrile, or a water-acetonitrile mixed solvent, while lipophilic metabolites use chloroform and dichloromethane solvents. The chicken samples are processed by cryogenic grinding technology to maintain the stability of metabolites and maximize the uniformity of the samples, providing accurate and high-quality samples for metabolomics analysis.

[0017] Step 3: Metabolomics analysis

[0018] High-throughput analysis techniques such as high-performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), and liquid chromatography-mass spectrometry (LC-MS) are used to perform metabolomics analysis on chicken samples. Amino acids, sugars, and organic acids, polar and water-soluble metabolites are analyzed by HPLC. Then, lipids, fatty acids, and carotenoids, volatile and non-volatile metabolites are analyzed by GC-MS and LC-MS. During the analysis, special attention is paid to the concentration changes of AICAR (5-aminoimidazole-4-carboxamide riboside) and its related metabolites to reveal the key metabolic markers during the chicken spoilage process.

[0019] Step 4: Multi-task learning and transfer learning model analysis

[0020] This module uses a multi-task learning (MTL) model to simultaneously process multi-dimensional data of chicken spoilage. Through multi-task training of the collected metabolite data, the model can improve the ability to identify spoilage markers and enhance the generalization ability of the system on different data sets through shared learning representations.

[0021] Step 5: AICAR prediction and spoilage assessment

[0022] Based on multi-task learning and transfer learning models, the AICAR prediction module plays a key role. This module conducts real-time spoilage prediction based on the extracted metabolite data and the learned spoilage model, and evaluates the spoilage status of chicken under different storage conditions. Through the prediction of AICAR, the system can output the spoilage stage of chicken and predict the remaining shelf life, providing accurate warning information for users. This process can help consumers or food processors optimize the storage conditions and consumption decisions of chicken according to the spoilage warning, avoid food waste, and ensure food safety. By continuously optimizing the AICAR model, the system can continuously improve the prediction accuracy, making the entire food storage and consumption process more intelligent and scientific.

[0023] Furthermore, step one includes:

[0024] Fresh samples: Collect chicken samples shortly after slaughter to ensure they are in an unspoiled state.

[0025] Stored samples: Select chicken samples stored for 3 days, 7 days, and 10 days according to experimental requirements. The storage conditions include:

[0026] Refrigeration condition: 0 - 4°C, suitable for short-term storage, simulating the preservation conditions in common households or supermarkets.

[0027] Freezing condition: -18°C, used to simulate long-term storage, usually for preservation.

[0028] For each storage condition, samples must be taken from multiple batches of chicken to avoid biases that may be brought by a single batch and ensure the comprehensiveness and representativeness of the data.

[0029] During the sampling process, the environmental factors during the storage of chicken must be recorded in real time, especially:

[0030] Temperature change T: The storage temperature is one of the key factors affecting the spoilage rate of chicken, usually monitored in real time through a temperature recorder or sensor.

[0031] Humidity change H: Humidity affects the water evaporation and bacterial reproduction of chicken and is one of the important factors in spoilage, which needs to be monitored through a humidity sensor.

[0032] Storage duration T s : The storage duration is the time variable of chicken spoilage, directly affecting its spoilage speed.

[0033] The deterioration rate (R) is a function jointly determined by storage conditions; temperature, humidity, and storage time have a significant impact on the deterioration process of chicken; generally, the higher the temperature, the faster the deterioration rate; the higher the humidity, the more rapid the deterioration process may be; the longer the storage time, the more significant the deterioration process becomes; this formula provides data support for subsequent metabolite analysis and deterioration model modeling. The recording formula for environmental factors is:

[0034]

[0035] Where: k: proportionality coefficient related to the reference conditions of the sample (usually a constant determined experimentally). α: influence coefficient of temperature on the deterioration rate. β: influence coefficient of humidity on the deterioration rate. γ: influence coefficient of storage duration on the deterioration rate. The combined influence of temperature, humidity, and storage time on the deterioration rate. R is the deterioration rate of chicken, T is the storage temperature, H is the storage humidity, and T s is the storage duration; represents the comprehensive result of the signal intensity of metabolites during the deterioration process; Intensity i represents the signal intensity of the i-th metabolite. RetentionTime is the retention time of the i-th metabolite. n is the number of metabolites;

[0036] Batch identification: Each sample should be labeled with a batch number, collection date, storage conditions, and storage time at the time of collection to ensure traceability of the specific information of each sample;

[0037] Sample processing: Ensure that the collected chicken samples are processed and stored as quickly as possible after sampling to prevent premature deterioration due to temperature fluctuations during transportation and storage;

[0038] To ensure the comprehensiveness and representativeness of experimental data, the collected samples should cover chicken under different batches, different breeding environments, and different processing conditions; the number of samples in each batch should be sufficient to ensure that the samples can comprehensively reflect different stages of deterioration;

[0039] Collecting multiple batches of samples under different storage conditions can ensure the extensiveness of data, avoid biases that may be brought by relying on only a certain batch, and improve the representativeness of data. According to the complexity of the deterioration process, the formula for the number of batches under each storage condition and storage time is

[0040]

[0041] Where N is the total required sample size, m is the number of different types of storage conditions (different storage days, different temperatures), and B i is the number of batches under each condition.

[0042] Furthermore, step two includes:

[0043] Before metabolite extraction, sample processing is carried out using cryogenic grinding technology; cryogenic grinding can rapidly freeze the sample through liquid nitrogen or in a low-temperature environment and grind the sample at low temperature; during cryogenic grinding, in order to avoid metabolite degradation, the temperature of the sample is usually controlled between -20°C and -80°C. This temperature range is generally considered "low temperature", which can effectively prevent the degradation and volatilization of most metabolites (especially heat-sensitive metabolites). The specific controlled temperature depends on the stability of the metabolites to be extracted and the experimental requirements;

[0044] Temperature change formula: During the grinding process, the temperature change of the sample is usually determined by the temperature of the external environment and the heat conduction characteristics of the sample; to ensure that the temperature during cryogenic grinding remains within the low-temperature range, the following formula is used to describe the temperature change. Through this formula, the heat distribution during the grinding process can be calculated, thereby controlling the cryogenic environment and avoiding metabolite degradation:

[0045]

[0046] Among them, ΔT represents the temperature change, Q is the amount of heat transfer during the grinding process, m is the mass of the sample, and c is the specific heat capacity of the sample;

[0047] Basis for solvent selection: For polar metabolites such as amino acids and acids, methanol, acetonitrile, or a water-acetonitrile mixed solvent is used, while for lipophilic metabolites, solvents such as chloroform and dichloromethane are more suitable;

[0048] Methanol has a strong polarity and can effectively dissolve these compounds containing polar functional groups; it is used to extract polar metabolites such as amino acids, peptides, organic acids, and nucleotides; Acetonitrile has a slightly lower polarity than methanol and is used to extract metabolites with relatively lower polarity such as small-molecule amino acids, phenolic compounds, and drug metabolites as a solvent in liquid chromatography analysis; The water-acetonitrile mixed solvent is a mixed solvent system of water and acetonitrile, and the polarity is adjusted according to the ratio of water and acetonitrile. It is used to extract small-molecule metabolites, sugars, certain drugs, and metabolites with polar and semi-polar properties; This mixed solvent is used for high-performance liquid chromatography (HPLC) analysis;

[0049] Chloroform is a low-polarity solvent that can effectively dissolve lipophilic metabolites such as fatty acids, steroids, and fat-soluble vitamins. It is usually used for liquid-liquid extraction and can extract lipids and other lipophilic compounds from the aqueous phase, and is widely used in the lipid extraction and mass spectrometry analysis of biological samples. Dichloromethane has a low polarity and is suitable for extracting fatty acids, steroids, lipids, and volatile compounds. Similar to chloroform, dichloromethane is commonly used for liquid-liquid extraction, has low toxicity and good volatility, can rapidly remove solvent residues, and is suitable for the extraction of lipophilic metabolites in mass spectrometry and gas chromatography analysis.

[0050] Solvent optimization formula: An important goal in the solvent selection process is to improve the extraction efficiency of metabolites; the following formula is used to estimate the extraction efficiency of different solvent systems:

[0051]

[0052] where C 溶液 is the concentration of metabolites extracted from the sample, and C 初样 is the concentration of metabolites in the initial sample;

[0053] Standard operating procedure for metabolite extraction:

[0054] Sample preparation: Place the frozen and ground chicken sample in an extraction container, add an appropriate amount of solvent to ensure that the solvent completely covers the sample;

[0055] Extraction process: Promote the full contact between the solvent and the metabolites in the chicken sample through ultrasonic oscillation, centrifugation or standing methods to enhance the extraction efficiency;

[0056] Centrifugal separation: After the extraction is completed, separate the mixture by a centrifuge, and set the centrifugation speed and time to separate the dissolved metabolites from the solvent;

[0057] Filtration and concentration: Remove large particle impurities through a filter, and use a vacuum concentrator to concentrate the solution to an appropriate volume for subsequent analysis;

[0058] Evaluation of extraction effect: Calculate the recovery rate of metabolites during the extraction process through the recovery rate calculation formula to evaluate the efficiency and effect of the extraction method;

[0059]

[0060] where the extract mass refers to the mass of metabolites extracted from the solvent, and the initial sample mass refers to the mass of the chicken sample put into the solvent system.

[0061] Furthermore, step three includes:

[0062] High performance liquid chromatography HPLC: Analyze amino acids, sugars and organic acids, polar and water-soluble metabolites by HPLC;

[0063] Analyze lipids, fatty acids and carotenoids, volatile and non-volatile metabolites by GC-MS and LC-MS; during the analysis, pay special attention to the concentration changes of AICAR (5-aminoimidazole-4-carboxamide riboside) and its related metabolites;

[0064] Data collection and analysis: During metabolomics analysis, the instrument generates a large amount of data, which includes the retention time, abundance, and mass spectrometry information of each metabolite; in order to extract useful metabolite information, the data needs to be preprocessed, calibrated, and integrated;

[0065] Data calibration: Each metabolite has a specific mass / charge ratio in the chromatogram and mass spectrometry; by comparing the retention time or mass spectrometry of known standards, the metabolites in each sample can be calibrated;

[0066] The calibration process can be expressed as:

[0067]

[0068] where m / z is the mass-to-charge ratio of the metabolite, z is the charge number of the metabolite, and q is the mass of the metabolite; through this formula, the mass spectrometry information of each metabolite can be accurately calibrated;

[0069] Quantitative analysis: The abundance of each metabolite is quantitatively analyzed by the area or height of the chromatographic peak; for HPLC, LC-MS, and GC-MS, the peak area method is used to calculate the metabolite concentration:

[0070]

[0071] where C x is the concentration of metabolite x, A x is the peak area of metabolite x in the chromatogram, A internal is the peak area of the internal standard substance, C internal is the known concentration of the internal standard substance; this method can compensate for instrument fluctuations and experimental errors;

[0072] Data integration and noise reduction: Since background noise is generated during the experiment, noise reduction processing needs to be performed on the chromatographic and mass spectrometry data; the methods used are smoothing filtering and denoising algorithms; after processing the signal, the characteristic information of the main metabolites in each sample can be extracted, and the filtering can be expressed by the following formula:

[0073]

[0074] where y i is the smoothed data point, x j is the original data point, and n is the size of the sliding window; this method helps to remove noise and improve data quality;

[0075] Multidimensional data integration: Metabolomics analysis generates data in different dimensions, including time, temperature, and storage status; therefore, it is necessary to integrate this multidimensional data together to form a unified metabolite dataset; this process can be carried out through the principal component analysis (PCA) multivariate analysis method;

[0076] Principal Component Analysis (PCA): PCA achieves dimensionality reduction by mapping the original data to a new coordinate system (principal components);

[0077] X' = XW

[0078] where X is the original data matrix, W is the principal component matrix, and X′ is the data after dimensionality reduction; this process can simplify multi-dimensional data into a few important components for subsequent analysis;

[0079] Marker Screening: In metabolomics data analysis, marker screening is a very crucial step, especially for markers related to chicken spoilage; statistical method ANOVA and machine learning-based feature selection method Lasso regression are used for screening;

[0080] Analysis of Variance (ANOVA) is used to compare whether there are significant differences in the means of multiple groups; ANOVA determines whether there are statistically significant differences by comparing the variability between groups and within groups; if the F value is less than the significance level of 0.05, it is considered that there are significant differences in the means of at least two groups; formula expression:

[0081]

[0082] where MS between : Mean Square Between Groups, representing the variation between different groups; MS within : Mean Square Within Groups, representing the degree of variation within the groups; k: number of groups of samples, chicken samples at different storage times; n i : number of samples in the i-th group; Mean of the i-th group; Overall mean of all groups; X ij : Observation value of the j-th sample in the i-th group; N: total number of samples; The F value represents the ratio of the between-group difference to the within-group difference; a larger F value indicates a significant between-group difference, which may be due to spoilage differences caused by different storage conditions or times;

[0083] Lasso Regression Formula:

[0084]

[0085] where β0: intercept term, representing the bias of the regression model, i.e., the baseline value of the prediction result; β: regression coefficient, a vector containing the weight values of each feature variable in the model; each β j represents the influence degree of this feature on the target variable; n: number of samples, i.e., the total number of observations in the dataset; y i : Observation value, representing the actual value corresponding to the i-th sample, the concentration of a certain characteristic spoilage marker related to chicken spoilage, the target variable; Xi : Input feature, representing the feature vector of the i-th sample, including the concentrations of various input data metabolites related to chicken spoilage, storage time, temperature, β0 + X i β: Predicted value, that is, the value calculated by the model through regression coefficients and input features, which is the prediction of y i ; λ: Regularization parameter, used to control the regularization strength in Lasso regression; a larger λ value will cause more regression coefficients to tend to zero, thus achieving the effect of feature selection. This parameter balances the fitting ability and complexity of the model, Sum of squared residuals, representing the difference between the model predicted value and the true observed value. The goal is to minimize this value to improve the prediction accuracy, L1 regularization term, used to penalize overly large regression coefficients and achieve feature selection by approaching some coefficients to zero; this helps reduce the complexity of the model and avoid overfitting, Solve for the regression coefficients by minimizing the above objective function.

[0086] Furthermore, step four includes:

[0087] In step three, a large amount of metabolite feature data was obtained through metabolomics analysis, including concentration data of various metabolites such as lipids, amino acids, and sugars; next, multi-task learning (MTL) and transfer learning techniques are used to process this data. The goal is to further improve the performance of the model through spoilage marker identification, spoilage stage prediction, and spoilage rate and degree assessment, and at the same time provide data input for AICAR prediction; the key to this step is to train a multi-task learning model (MTL) using these metabolite feature data and optimize the model effect by combining transfer learning methods;

[0088] Perform chicken spoilage analysis and prediction by combining multi-task learning and transfer learning with metabolomics data; through the identification of spoilage markers, prediction of spoilage stages, and assessment of spoilage rate and degree, the model provides accurate data outputs, which will directly serve as the input for step five, helping to achieve efficient prediction of the AICAR prediction module and providing data support for storage optimization and quality control;

[0089] The core of the input data comes from the metabolomics analysis in step three, specifically including the following types of data:

[0090] Metabolite feature vector: The metabolite concentration information of each sample obtained through metabolomics analysis, including lipid, amino acid, and sugar components; there are a total of n metabolites, and the concentration of each metabolite is x i (i = 1, 2,..., n), and the metabolite feature vector at temperature T can be expressed as:

[0091] X = [x1, x2, …, x n T

[0092] Storage conditions and storage time: Storage conditions (refrigeration, freezing) and storage time (3 days, 7 days, 10 days) are important factors affecting the spoilage of chicken; in order to incorporate storage condition information into the model, it can be represented as a vector s j , and the vector of storage conditions and storage time at temperature T can be expressed as:

[0093] S = [s1, s2, s3] T

[0094] where s j ∈ {0, 1} represents whether under a specific condition (refrigeration or not), and each element of s j corresponds to a storage condition;

[0095] Batch and environmental factors: Batch and environmental factors (temperature, humidity) also affect the spoilage of chicken; there are k environmental factors, and each factor is represented by e j , and the representation of environmental factor E at temperature T is:

[0096] E = [e1, e2, …, e k T

[0097] Multi-task learning (MTL) aims to improve the generalization ability of the model by sharing some model parameters while learning multiple related tasks; in this step, the main tasks include:

[0098] Identification of spoilage markers: The goal is to identify the key markers of chicken spoilage from metabolite feature data;

[0099] Prediction of the storage stage: The goal is to predict the spoilage stage of chicken at a certain moment;

[0100] Assessment of spoilage rate and degree: The goal is to assess the spoilage rate and degree of chicken;

[0101] Identification of spoilage markers: It is a classification problem, and the goal is to identify which metabolites are the markers of chicken spoilage; for each metabolite xi, it is necessary to predict whether it is a spoilage marker;

[0102] Label y j ∈ {0, 1} represents whether the i-th metabolite is a marker, and the output is the probability p j predicted by the model, that is, the probability that the metabolite is a spoilage marker; the loss function uses the cross-entropy loss function to measure the error of the classification problem; the cross-entropy loss function is as follows:

[0103] ​​

[0104] where m is the number of samples, y i is the true label of the i-th sample, and p i is the predicted probability of the i-th sample;

[0105] Prediction in the storage stage: It is a regression problem, and the goal is to predict the spoilage stage of chicken; the spoilage stage label of each sample is a discrete integer value (1 represents fresh, 2 represents slightly spoiled, 3 represents moderately spoiled, 4 represents severely spoiled); the output is the spoilage stage predicted by the model

[0106] The loss function for this task uses the mean squared error loss function (MSE) to measure the prediction error in the regression problem; the loss function is as follows:

[0107]

[0108] where m is the number of samples, is the predicted spoilage stage of the i-th sample, and y stage,i is the true spoilage stage label;

[0109] Assessment of spoilage speed and degree: It is a regression problem, and the goal is to evaluate the spoilage speed and degree of chicken; the label y rate,i represents the spoilage speed and degree, and the model output is the predicted value A regression model can be used to predict the spoilage speed and degree; its loss function also uses the mean squared error (MSE):

[0110]

[0111] where m is the number of samples, is the predicted spoilage speed and degree, and y rate,i is the true label;

[0112] Model structure and optimization

[0113] The multi-task learning model structure is as follows:

[0114] Shared layer: The shared layer extracts the general features of metabolite feature data X, storage conditions S, and environmental factors E through a neural network to enhance the generalization ability of the model; the output of the shared layer is an intermediate representation h shared ;

[0115] h shared = f shared (X, S, E)

[0116] where f shared is the function of the shared network, and h sharedis a shared representation;

[0117] Task-specific layer: Each task has a dedicated output layer that generates corresponding prediction results based on the shared representation h shared ;

[0118] For the spoilage marker recognition task, the output is p marker , and the classification activation function (sigmoid) is adopted:

[0119] p marker = σ(W marker h shared + b marker )

[0120] For the storage stage prediction task, the output is and the linear activation function is adopted:

[0121]

[0122] For the spoilage rate and degree assessment task, the output is and the linear activation function is also adopted:

[0123]

[0124] Optimization objective: The final loss function L total is the weighted sum of the losses of multiple tasks, where λ1, λ2, λ3 are the weight coefficients of the tasks:

[0125] L total = λ1L marker + λ2L stage + λ3L rate

[0126] The model parameters are updated by optimizing the loss function to achieve the comprehensive optimization of multiple tasks;

[0127] Transfer learning

[0128] Transfer learning is used to solve the problem of data scarcity, especially when dealing with new storage conditions or new batches of chicken data; The process is as follows:

[0129] Pretraining: Pretrain based on data from other storage conditions to learn general spoilage laws and metabolite characteristics;

[0130] Fine-tuning: Fine-tune on the data of the target task to adjust the model parameters to adapt to the new data environment; During the fine-tuning process, the weights of some shared layers are kept fixed, while the weights of the output layers specific to the tasks are updated;

[0131] Output Data and Subsequent Applications: Through the combination of multi-task learning and transfer learning, the output of the finally obtained model will be used as the input for the subsequent AICAR prediction module; these output data include:

[0132] Recognition Results of Spoilage Markers: Whether each metabolite is a spoilage marker and its concentration changes at different spoilage stages;

[0133] Prediction of Chicken Spoilage Stage: Based on the marker data and storage conditions output by the model, the spoilage stage of chicken is predicted in real time;

[0134] Assessment of Spoilage Rate and Degree: Output the assessment of the spoilage rate and degree of chicken;

[0135] These outputs will be used as the input for Step Five, further analyzed and generate the prediction results of the spoilage state, providing decision support for storage optimization and quality control.

[0136] Furthermore, Step Five includes:

[0137] In Step Four, relevant information on chicken spoilage, recognition results of spoilage markers, prediction of spoilage stage, and assessment of spoilage rate and degree are obtained through multi-task learning (MTL) and transfer learning techniques; this information will be used as input data into the AICAR prediction module in Step Five to accurately predict the spoilage process of chicken based on the existing metabolite characteristics and spoilage state, and provide decision support for storage optimization and quality control;

[0138] In Step Five, the AICAR prediction module comprehensively analyzes the recognition results of spoilage markers, prediction results of spoilage stages, and assessment data of spoilage rate and degree, accurately predicts the spoilage time range and confidence level of the spoilage state of chicken, and gives corresponding suggestions for optimizing storage conditions; this process relies on multiple calculation formulas, and the values obtained through these formulas provide reliable data support for storage optimization and quality control;

[0139] The input data includes the following aspects:

[0140] Recognition Results of Spoilage Markers: The marker recognition probability p of each metabolite from Step Four marker ;

[0141] These probability values reflect whether each metabolite is a spoilage marker and their spoilage degree;

[0142] Prediction Results of Spoilage Stage: The spoilage stage prediction value from Step Four Indicates the current spoilage stage of the chicken; this value is a discrete integer representing the freshness level of the chicken (1 represents fresh, 2 represents slightly spoiled, 3 represents moderately spoiled, 4 represents severely spoiled);

[0143] Assessment results of deterioration rate and degree: Deterioration rate from Step 4 and degree of deterioration;

[0144] The deterioration rate (unit: days) describes the speed at which chicken moves from its current state to the next deterioration stage;

[0145] The degree of deterioration (range: 0 to 1) represents the current deterioration level of the chicken, from completely fresh (0) to completely deteriorated (1);

[0146] Core objective of the AICAR prediction model: The objective of the AICAR (i.e., "Advanced Integrated Condition Assessment and Rating") prediction model is to accurately predict the deterioration state of chicken in the future based on data such as input metabolite characteristics, biomarker identification results, deterioration stage, and deterioration rate, and to give the corresponding deterioration time range, confidence level of the deterioration state, and suggestions for optimizing storage conditions;

[0147] AICAR prediction model structure

[0148] The main components of the AICAR model include the following parts:

[0149] Prediction of deterioration stage time: Based on the current deterioration stage, deterioration rate, and biomarker identification results, the model will calculate the time prediction for the chicken to move from the current stage to the next stage;

[0150] Time prediction formula:

[0151]

[0152] Where: S target : Identification of the target deterioration stage (the target is the "slightly deteriorated" stage, S target = 2); S current : Identification of the current deterioration stage (the current is the "fresh" stage, S current = 1); v rate,i : Deterioration rate (unit: days), i.e., the rate of the deterioration process; t pred,i : Predicted time to reach the target deterioration stage;

[0153] Definition of output values:

[0154] t pred,i : Represents the predicted time from the current deterioration stage to the target deterioration stage, in days;

[0155] Deterioration stages are usually discrete integers:

[0156] 1: Fresh

[0157] 2: Slightly deteriorated

[0158] 3: Moderately deteriorated

[0159] 4: Severely deteriorated

[0160] Prediction result t pred,i The defined range of:

[0161] A negative prediction result indicates that the chicken is in good condition, with a very slow spoilage rate or no spoilage occurring;

[0162] When the prediction time is 0 days, it means that the chicken has reached the target spoilage stage and immediate measures need to be taken;

[0163] When the prediction time is positive, it represents the time required from the current spoilage stage to the target spoilage stage, usually a positive integer;

[0164] Spoilage state confidence assessment: The AICAR model will also give the confidence c of the spoilage state prediction based on the identification probability of spoilage markers and the spoilage rate assessment result confidence,i , indicating the reliability of the model's prediction result;

[0165] Confidence calculation formula:

[0166]

[0167] Where: p marker,j : The identification probability of each marker; v rate,i : Spoilage rate (unit: days); α: Coefficient, controlling the influence degree of spoilage rate on confidence; c confidence,i : Confidence of spoilage prediction, ranging from 0 to 1;

[0168] Definition of output values:

[0169] c confidence,i : Confidence value, indicating the reliability of the prediction result, with a value range of 0 to 1;

[0170] 0 indicates that the model is very uncertain about the prediction result and the prediction reliability is low;

[0171] 1 indicates that the model is extremely certain about the prediction result and the prediction reliability is very high;

[0172] Suggestions for optimizing storage conditions: Based on the predicted spoilage stage and rate, the AICAR model will give suggestions for optimizing storage conditions; these suggestions include whether to change the storage temperature or humidity conditions to delay the spoilage of chicken;

[0173] Logic for generating optimization suggestions: Spoilage rate v rate,iHigh, and the deterioration stage is close to severe deterioration. It is recommended to transfer the chicken to a lower storage temperature; the deterioration stage is low and the deterioration speed is slow. It is recommended to maintain the current storage conditions;

[0174] Definition of output values:

[0175] The numerical values of the storage condition recommendations adopt the following several common storage schemes:

[0176] Refrigeration: Suitable for cases of mild deterioration or slow deterioration speed, with a temperature of 0°C to 4°C;

[0177] Freezing: Suitable for cases with a fast deterioration speed and a short expected deterioration time, with a temperature of -18°C to -20°C;

[0178] Normal temperature: Suitable for cases with a very slow deterioration speed, with a temperature of 20°C to 25°C;

[0179] Output results of the AICAR prediction module

[0180] The output of the AICAR prediction module will include the following items:

[0181] Deterioration time prediction: Output the time (unit: days) required for the chicken to reach the next deterioration stage from the current deterioration stage;

[0182] Deterioration confidence: Output the confidence of the deterioration state prediction, indicating the reliability of the model for the result;

[0183] Optimization suggestions for storage conditions: Output optimization suggestions for storage conditions based on the deterioration prediction results.

[0184] Compared with the prior art, the beneficial effects of the present invention are:

[0185] 1. Application of the multi-task learning model to improve the accuracy and comprehensiveness of biomarker recognition: Existing chicken deterioration detection methods usually focus on the detection of a single biomarker and rely on traditional chemical analysis or simple statistical analysis, making it difficult to comprehensively evaluate the deterioration state of chicken from a multi-dimensional perspective. The present invention realizes the joint analysis of multiple metabolites related to deterioration by introducing multi-task learning (MTL), and can simultaneously identify and screen multiple deterioration biomarkers. This method enhances the learning ability of the model in the multi-dimensional feature space by constructing a shared feature learning framework, making the screening of deterioration biomarkers more accurate and comprehensive, and significantly improving the reliability of prediction.

[0186] Through a multi-task learning network, the present invention separately processes different classification tasks of metabolites (such as classification tasks of amino acids, lipids, carbohydrates, etc.) and shares the feature hierarchy, thereby optimizing the learning effects of various tasks. This integrated learning method can effectively utilize the correlation between different markers and enhances the model's ability to analyze complex data structures. Compared with traditional methods, this multi-dimensional and multi-marker analysis enables the system to more accurately identify and predict the spoilage process of chicken.

[0187] 2. Transfer learning improves the adaptability and prediction accuracy of the model: In practical applications, the spoilage process of chicken is affected by different storage conditions (such as temperature, humidity) and different shelf lives (such as 3 days, 7 days, 10 days, etc.). Traditional methods usually analyze based on fixed experimental conditions and cannot effectively adapt to changes in storage conditions and time. The present invention solves this problem by introducing transfer learning.

[0188] The transfer learning algorithm optimizes cross-sample data, enabling the system to transfer the knowledge learned from one sample to another, thus quickly adapting to new storage conditions and spoilage stages. Specifically, the system trains a preliminary model using existing labeled data and then applies this model to new, unlabeled sample data. Through this transfer, the system can effectively improve the prediction accuracy under different storage conditions and different spoilage stages. For example, when comparing fresh chicken with chicken stored for several days, transfer learning can adaptively adjust the model parameters, enabling the system to accurately evaluate spoilage between different shelf lives.

[0189] 3. The AICAR prediction module realizes accurate spoilage stage prediction and remaining shelf life estimation: The present invention innovatively uses the AICAR prediction module, which can, based on metabolite data and spoilage models, real-time predict the spoilage state of chicken under different storage conditions. The core advantage of this module is its ability to meticulously estimate the current spoilage stage of chicken and predict the remaining shelf life. This accurate spoilage stage prediction and shelf life estimation can provide scientific storage and consumption decision-making support for users, helping to reduce waste and ensure food safety.

[0190] The AICAR prediction module constructs a prediction model based on multi-dimensional data by comprehensively considering various factors such as metabolite data, storage temperature and humidity, and time of chicken samples. By learning the spoilage laws of historical data, this model can real-time output information about the spoilage stage of chicken and the remaining shelf life. This enables the present invention to provide accurate time warnings, helping users to take appropriate measures before the chicken spoils, such as adjusting storage conditions or consuming in advance, thereby improving the efficiency of food management.

[0191] 4. Improve the accuracy and reliability of data: The system of the present invention ensures high quality and high accuracy of chicken metabolite data through high-throughput metabolomics analysis techniques (such as HPLC, GC-MS, and LC-MS). In traditional methods, data collection and analysis may be affected by factors such as the subjective judgment of experimenters and the performance of instruments, resulting in poor reliability of the results. Through the technology of the present invention, the quantitative analysis of metabolites adopts standardized high-throughput techniques, ensuring the repeatability and reliability of the data, thereby providing accurate input data for the subsequent multi-task learning model and AICAR prediction module. Description of the Drawings

[0192] Figure 1 : Overall flowchart of the present invention;

[0193] Figure 2 : Flowchart of the multi-task learning (MTL) algorithm introduced in the present invention;

[0194] Figure 3 : Flowchart of the transfer learning algorithm introduced in the present invention; Detailed Description of the Invention

[0195] As Figures 1 - 3 shown, the present invention provides a method and system for screening chicken spoilage markers and predicting AICAR based on multi-task learning and transfer learning. Through intelligent algorithms, in-depth analysis is carried out on the metabolite data of chicken samples, combined with multi-task learning and transfer learning, to accurately predict the spoilage process of chicken, thereby optimizing food storage and consumption decisions. The following is a detailed description of each module to illustrate the specific implementation manner of the present invention:

[0196] Sample Collection Module

[0197] The sample collection module is the basic module of the present invention, responsible for collecting data from chicken samples under different storage conditions and shelf lives, ensuring the diversity and representativeness of the data, and providing high-quality raw data for subsequent metabolite extraction and analysis.

[0198] Sampling scheme: This module collects chicken samples under various storage conditions, including fresh samples, chicken samples stored for 3 days, 7 days, 10 days, and other different time periods. To ensure the representativeness of the experimental data, the system will collect at least three batches of samples at each storage stage to avoid the influence of deviations in a single batch on the results.

[0199] Storage condition record: The storage environment (such as temperature, humidity, light, etc.) of each batch of samples will be recorded in real time and stored together with the sample data. These storage environment variables have an important impact on the spoilage process of chicken, so they must be accurately recorded for subsequent analysis.

[0200] Dynamic data collection: During the sample collection process, the system also monitors and records environmental changes to ensure the consistency of sample storage conditions and capture environmental fluctuations that may affect sample deterioration.

[0201] Metabolite extraction module

[0202] The metabolite extraction module is responsible for processing the collected chicken samples to extract metabolites in the samples, especially the chemical components related to chicken deterioration. The design of this module ensures the maximum retention of metabolite information in the samples, providing high-quality data for subsequent analysis.

[0203] Extraction efficiency control: By adjusting parameters such as extraction time and solvent volume, ensure the maximum retention of various metabolites in chicken, especially the metabolite markers related to the deterioration process. The extracted liquid metabolite samples will be prepared for subsequent analysis.

[0204] Metabolomics analysis module

[0205] The metabolomics analysis module is responsible for the quantitative and qualitative analysis of the extracted metabolite samples. Through high-throughput analysis techniques such as high-performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), and liquid chromatography-mass spectrometry (LC-MS), comprehensively analyze the metabolites in the samples and extract the characteristic data related to chicken deterioration.

[0206] Metabolite qualitative analysis: Use techniques such as LC-MS to analyze the complex metabolites in chicken and identify specific markers related to the deterioration process, such as fatty acids, amino acids, and their metabolites. Through mass spectrometry data, accurately determine the chemical structures of various metabolites and their change rules during the deterioration process.

[0207] Metabolite quantitative analysis: Through techniques such as HPLC, accurately determine the concentration levels of various metabolites, especially the changes in deterioration markers, and further provide data support for subsequent model establishment.

[0208] Multi-task learning model module

[0209] This module combines multi-task learning and transfer learning algorithms to jointly analyze metabolite data. By analyzing the metabolite characteristics of chicken samples under different storage conditions, identify multiple markers related to deterioration and generate metabolite feature vectors, providing basic data for AICAR prediction and deterioration assessment.

[0210] Multi-task learning: This module can simultaneously process multiple tasks within the same model framework, identify metabolites related to deterioration, and improve the accuracy of deterioration marker identification by learning the relationships between tasks. For example, the system simultaneously learns the metabolite changes under different storage conditions to help identify the characteristics shown in different environments.

[0211] Transfer learning: Utilizing cross-sample data, transfer learning algorithms enable the system to transfer from one spoilage stage (such as fresh to 3-day storage) to another spoilage stage (such as 7-day or 10-day storage). Through the optimization of cross-stage data, the system can improve the prediction accuracy of the spoilage process and ensure adaptation to changes in different storage conditions and spoilage stages.

[0212] AICAR Prediction Module

[0213] The AICAR Prediction Module makes real-time predictions on the spoilage process of chicken under different storage conditions based on metabolite data and the learned spoilage model. The system can output the spoilage stage of the chicken and the expected remaining shelf life according to the change of AICAR, helping users make reasonable storage and consumption decisions.

[0214] AICAR Prediction: The system evaluates the current spoilage state of the chicken by analyzing the key changes in metabolite data and combining with the AICAR molecular marker. As a key metabolite, the change of AICAR can reflect the health state of the chicken and predict the remaining shelf life of the chicken.

[0215] Spoilage Stage Evaluation: Based on the AICAR prediction results, the system can output the specific spoilage stage of the chicken, such as "fresh", "slightly spoiled", "spoiled approaching critical", etc., thus helping users determine whether the chicken is suitable for consumption or whether other treatments are needed.

[0216] Early Warning System: Combining the AICAR prediction results, the system can give an early warning to users about the spoilage trend of the chicken, prompting users to consume it as soon as possible or adopt other storage methods, thereby reducing food waste and avoiding food safety problems.

[0217] System Integration and Application

[0218] The system of the present invention is realized through the collaborative work of each module, by building a software programming and calling relevant general algorithm models (such as calling MATLAB algorithm models, which belongs to well-known technologies), and can achieve accurate prediction and real-time monitoring of the chicken spoilage process. The system can not only identify the spoilage markers of chicken under different storage conditions, but also provide accurate early warnings for users based on the AICAR prediction results, helping users make reasonable food management decisions. The system can be widely applied to fields such as quality control, food safety monitoring, and intelligent storage management in the food industry, improving the efficiency and accuracy of food quality monitoring.

[0219] Specific Numerical Analysis

[0220] Step 1: Sample Collection

[0221] 1.1 Sample Type

[0222] Fresh chicken samples: Collected within 2 hours after slaughter.

[0223] Stored samples:

[0224] Refrigeration conditions: 3 days (T = 3°C, H = 80%), 7 days (T = 3°C, H = 80%), 10 days (T = 3°C, H = 80%).

[0225] Freezing conditions: 3 days (T = -18°C, H = 65%), 7 days (T = -18°C, H = 65%), 10 days (T = -18°C, H = 65%).

[0226] 1.2 Sample quantity and batches

[0227] For refrigerated and frozen samples, the number of batches under each storage condition is 3.

[0228] Total sample quantity: 3 (batches) × 6 (storage durations) = 18 samples.

[0229] Step 2: Metabolite extraction

[0230] Freezing and grinding: Freezing and grinding is a crucial step in the metabolite extraction process.

[0231] Quick-freezing samples: Use liquid nitrogen or a low-temperature environment (-80°C to -20°C) to quickly freeze the samples. Liquid nitrogen can rapidly reduce the temperature of the samples to extremely low temperatures, thereby preventing damage to the integrity of cell walls and cell membranes and ensuring that metabolites are not oxidized or degraded. Low-temperature treatment can stabilize polar and lipophilic metabolites.

[0232] Temperature change control: During the grinding process, the heat generated will cause the temperature of the samples to rise, affecting the stability of metabolites. To ensure that the temperature change of the samples during grinding does not exceed 1°C, it is necessary to precisely control the operating parameters of the grinding equipment (such as grinding speed, operating time, etc.) and use an effective cooling system to maintain temperature stability. For example, a cooling circuit or intermittent cooling can be used to prevent excessive temperature, ensure that the temperature change during the grinding process does not exceed 1°C, and avoid metabolite degradation or volatilization.

[0233] Solvent selection and extraction: Solvent selection is a key link in the metabolite extraction process. The polarity of the solvent is closely related to the properties of the metabolites to be extracted. Reasonable selection of the solvent can not only improve the extraction efficiency but also ensure the integrity of the metabolites. According to the polarity of the metabolites, select a suitable solvent for extraction. The solvent systems required for the extraction of polar metabolites and lipophilic metabolites are different.

[0234] Extraction of polar metabolites: For polar metabolites (such as amino acids, peptides, organic acids, sugars and nucleotides, etc.), commonly used solvents include methanol, acetonitrile, and water-acetonitrile mixed solvents. Methanol: It has a strong polarity and can effectively dissolve compounds containing polar functional groups (such as amino acids, peptides and organic acids), and is suitable for extracting metabolites with strong polarity. Acetonitrile: Compared with methanol, acetonitrile has a slightly lower polarity and can dissolve small molecule compounds with relatively low polarity, such as phenolic compounds and certain drug metabolites. Water-acetonitrile mixed solvent: By adjusting the ratio of water to acetonitrile, the polarity of the solvent system can be flexibly adjusted, which is used to extract small molecule metabolites, sugars and certain drugs and metabolites.

[0235] Extraction of fat-soluble metabolites: For fat-soluble metabolites (such as fatty acids, steroids, fat-soluble vitamins, etc.), a solvent with lower polarity is required for extraction. Chloroform: As a low-polarity solvent, chloroform can efficiently dissolve fat-soluble substances such as fatty acids and steroids, and is widely used in lipid extraction and mass spectrometry analysis. Dichloromethane: Similar to chloroform, dichloromethane is also suitable for the extraction of fatty acids, steroids and fat-soluble compounds. Compared with chloroform, dichloromethane has lower toxicity and good volatility. It is suitable for the rapid removal of solvent residues and is often used in gas chromatography analysis.

[0236] Extraction efficiency optimization:

[0237] When selecting a solvent, it is also necessary to optimize the extraction process by calculating the extraction efficiency:

[0238]

[0239] Among them, C 溶液 represents the concentration of metabolites after extraction, C 初样 is the concentration of metabolites in the initial sample. By calculating the extraction efficiency of different solvent systems, the most suitable solvent can be selected for metabolite extraction.

[0240] Extraction process: The extraction process is a key step in the separation and enrichment of metabolites. It is necessary to ensure that the solvent is in full contact with the metabolites in the sample to achieve efficient extraction. Commonly used extraction methods include ultrasonic oscillation, centrifugation and static method.

[0241] Ultrasonic oscillation: Using the high-frequency oscillation of ultrasound, the solvent and the metabolites in the sample collide and rub strongly, promoting the dissolution and release of metabolites. Ultrasonic oscillation usually requires controlling the operation time (30 minutes to 1 hour) and temperature (0-4°C) to avoid overheating of the sample and causing metabolite degradation. Ultrasonic extraction can effectively improve the extraction rate of metabolites, especially for water-soluble and semi-polar compounds.

[0242] Centrifugation method: The centrifugal separation method can separate solid impurities in the solution by high-speed rotation, retaining the metabolites dissolved in the solvent. During centrifugation, a rotation speed of 3000 - 5000 rpm is usually used for 10 - 20 minutes to ensure clear separation of impurities and metabolites. Centrifugal separation can quickly and effectively remove insoluble substances from the extraction solution, obtaining a pure metabolite solution.

[0243] Static method: The static method enhances the solubility of metabolites by prolonging the contact time between the solvent and the sample, and is applicable to some metabolites that are relatively stable to the solvent. The static time is generally 1 - 2 hours and needs to be carried out in a low-temperature environment to prevent metabolite degradation due to high temperature.

[0244] Centrifugal separation: The purpose of centrifugal separation is to remove impurities and insoluble substances generated during the extraction process, retaining the metabolites dissolved in the solvent. By controlling the centrifugation conditions, impurities in the sample can be effectively separated to obtain a pure metabolite solution.

[0245] Centrifugation operation conditions: The centrifugation speed is usually set at 3000 - 5000 rpm and optimized according to the nature and particle size of the sample. The centrifugation time is set to 10 - 20 minutes to ensure thorough separation. In this way, there will be basically no solid impurities in the extracted solution, providing a pure sample for subsequent analysis steps.

[0246] Filtration and concentration: Filtration and concentration are necessary steps before subsequent analysis. Large particle impurities are removed by filtration to ensure the purity of the extraction solution. Concentration can remove the excess solvent and concentrate the metabolite concentration, providing a basis for high-sensitivity analysis.

[0247] Filtration: A microporous membrane with a pore size of 0.45 μm or 0.22 μm is used to remove solid substances such as large particle impurities and cell debris in the solution. The filtration operation can improve the accuracy and stability of subsequent analysis and avoid impurities affecting the analysis results.

[0248] Concentration: A vacuum concentrator is used to concentrate the extraction solution to an appropriate volume, generally concentrated to 1 / 5 to 1 / 10 of the original volume. The concentration process needs to be carried out under low-temperature conditions to prevent metabolite loss due to volatilization. The concentrated solution is ready to enter subsequent mass spectrometry or liquid chromatography analysis for qualitative and quantitative analysis.

[0249] Evaluation of extraction effect: Evaluation of the extraction effect is a very important step in the metabolite extraction process. By calculating the recovery rate, the efficiency of different solvent systems and extraction methods can be evaluated to ensure a high-quality metabolite sample is finally obtained.

[0250] Recovery rate calculation formula: The recovery rate is calculated by the following formula:

[0251]

[0252] Among them, the extract quality refers to the mass of metabolites extracted from the solvent, and the initial sample quality refers to the mass of the chicken sample put into the solvent system.

[0253] A high recovery rate indicates less loss during the extraction process and better extraction effect of metabolites. Through the evaluation of the recovery rate, the solvent selection and extraction conditions can be optimized to ensure the efficiency and reliability of the extraction process.

[0254] 1. Extraction of polar metabolites:

[0255] Experimental condition assumptions:

[0256] Initial metabolite concentration C of the sample 初样 = 5 mg / mL.

[0257] Metabolite concentration C after extraction 溶液 Three solvents were used respectively:

[0258] Methanol: C 溶液 = 3.5 mg / mL

[0259] Acetonitrile: C 溶液 = 3.0 mg / mL

[0260] Water - acetonitrile mixed solvent (50:50, v / v): C 溶液 = 4.0 mg / mL

[0261] Calculation of extraction efficiency: Using the formula:

[0262]

[0263] Extraction efficiency of methanol: η = 0.7 (70%)

[0264] Extraction efficiency of acetonitrile: η = 0.6 (60%)

[0265] Extraction efficiency of water - acetonitrile mixed solvent: η = 0.8 (80%)

[0266] Result: The water - acetonitrile mixed solvent showed the highest extraction efficiency and is suitable for extracting polar and semi - polar metabolites.

[0267] 2. Extraction of lipophilic metabolites:

[0268] Initial lipophilic metabolite concentration C of the sample 初样 = 10 mg / mL

[0269] Metabolite concentration C after extraction 溶液 Two solvents were used respectively:

[0270] Chloroform: C 溶液 = 7.5 mg / mL

[0271] Dichloromethane C 溶液 = 8.0 mg / mL

[0272] Calculation of extraction efficiency: Use the same formula:

[0273]

[0274] Chloroform extraction efficiency: η = 0.75 (75%)

[0275] Dichloromethane extraction efficiency: η = 0.8 (80%)

[0276] Result: The extraction efficiency of dichloromethane is slightly higher than that of chloroform, which is suitable for experiments that require rapid extraction and analysis of lipophilic metabolites.

[0277] Step 3: Metabolomics analysis

[0278] 3.1 Data calibration and quantitative analysis of HPLC, LC-MS and GC-MS

[0279] 1. HPLC data calibration and quantitative analysis

[0280] Sample analysis:

[0281] Analysis target: Polar and water-soluble metabolites such as amino acids, sugars and organic acids.

[0282] Calibration method: Establish the relationship between concentration and peak area through a standard solution with a known concentration.

[0283] Relationship between peak area and concentration: Assume that the peak area of the target metabolite is A x , and the peak area of the internal standard substance with a known concentration is A internal , the concentration of the internal standard is C internal , then the concentration C of the target metabolite x is calculated by the following formula:

[0284]

[0285] Sample concentration calculation: Assume that the peak area of the target metabolite is 5000, and the peak area of the internal standard substance is 10000, and the known concentration of the internal standard substance is 0.1 μM, then the concentration of the target metabolite is: C x = 0.05 μM;

[0286] 2. LC-MS data calibration and quantitative analysis

[0287] Sample analysis:

[0288] Analysis targets: Lipids, fatty acids, carotenoids and their related metabolites, with special attention paid to the concentration changes of AICAR.

[0289] Calibration method: Combining liquid chromatography (LC) and mass spectrometry (MS) analysis. After separating metabolites by liquid chromatography, accurate ion intensity data are provided by mass spectrometry.

[0290] Taking AICAR as an example, specific ion intensities will appear in its mass spectrum. For example, [M+H]+ is m / z 259 (molecular ion). By using standard solutions with known concentrations, the relationship between concentration and ion intensity (or peak area) is established.

[0291] Relationship between peak area and concentration: For example, when the concentration of the AICAR standard solution is 1 μM, the ion peak intensity is 5000; if the ion peak intensity of the sample to be measured is 2000, then its concentration C x is: C x = 0.4 μM

[0292] 3. GC-MS Data Calibration and Quantitative Analysis

[0293] Sample analysis:

[0294] Analysis targets: Volatile and non-volatile metabolites such as lipids and fatty acids.

[0295] Calibration method: Separate metabolites by gas chromatography and conduct qualitative and quantitative analysis through mass spectra.

[0296] Suppose there is a fatty acid (such as linoleic acid with m / z 279) in the sample to be measured. By comparing with known standard solutions, the relationship between concentration and peak area is established.

[0297] Relationship between peak area and concentration: Suppose the peak area in the standard solution is 8000, corresponding to a concentration of 1 μM; if the peak area in the sample to be measured is 4000, then its concentration is: C x = 0.5 μM;

[0298] 3.2 Data Denoising and Integration

[0299] 1. Data Denoising

[0300] In metabolomics analysis, experimental data often contain background noise, which may affect the accuracy of signals and the results of subsequent analysis. To reduce noise, a smoothing filtering method can be used for denoising. Suppose concentration data of a certain metabolite are obtained in LC-MS analysis, and these data have a certain amount of noise.

[0301] Suppose data:

[0302]

[0303] To reduce noise, we use the sliding window smoothing algorithm. Assuming the sliding window size \(n = 3\), the concentration at each time point will be replaced by the average concentration of one time point before and after this time point.

[0304] Smoothing process: At time 0: Smoothed value=\(\frac{1.2 + 1.5}{2}=1.35\); At time 1: Smoothed value=\(\frac{1.2 + 1.5+2.3}{3}=1.67\); At time 2: Smoothed value=\(\frac{1.5 + 2.3+2.5}{3}=2.10\); At time 3: Smoothed value=\(\frac{2.3 + 2.5+3.1}{3}=2.63\); At time 4: Smoothed value=\(\frac{2.5 + 3.1+3.3}{3}=2.97\); At time 5: Smoothed value=\(\frac{3.1 + 3.3+3.0}{3}=3.13\); At time 6: Smoothed value=\(\frac{3.3 + 3.0+2.8}{3}=3.03\); At time 7: Smoothed value=\(\frac{3.0 + 2.8+2.9}{3}=2.90\); At time 8: Smoothed value=\(\frac{2.8 + 2.9+2.6}{3}=2.77\); At time 9: Smoothed value=\(\frac{2.9 + 2.6}{2}=2.75\);

[0305] The data after smoothing:

[0306]

[0307] By this method, the small fluctuations and noises in the original data have been smoothed, thus obtaining a smoother concentration change curve, which provides more reliable data for subsequent analysis.

[0308] 2. Data integration and dimensionality reduction

[0309] Since metabolomics analysis involves multiple experimental dimensions (such as time, temperature, metabolite types, etc.), we need to integrate these multi-dimensional data together for analysis. The commonly used dimensionality reduction method is principal component analysis (PCA). Suppose we have a data set containing the concentrations of different metabolites. The columns of the data set include different metabolites, and the rows represent different time points or samples.

[0310] Suppose the data set:

[0311]

[0312]

[0313] PCA dimensionality reduction process:

[0314] Standardize the data: Since the concentration ranges of different metabolites are different, first standardize each column of data so that the mean of each variable is 0 and the standard deviation is 1. For example, the standardization process of AICAR is as follows:

[0315]

[0316] Among them, \(X\)AICAR is the original concentration data, μ AICAR and σ AICAR are the mean and standard deviation of the AICAR concentration, respectively.

[0317] Calculate the covariance matrix: Calculate the covariance matrix of the standardized data to evaluate the correlation between different metabolites.

[0318] Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix to obtain the principal components and the corresponding eigenvalues.

[0319] Select principal components: Select some principal components that explain the largest variance in the data (e.g., the first two principal components).

[0320] Assume that after PCA analysis, we obtain the scores of the first two principal components as follows:

[0321]

[0322]

[0323] Through PCA analysis, the concentration data of the original three metabolites are reduced to two principal components, which can explain most of the variance in the data. Through this dimensionality reduction method, the changing trends of metabolites at different time points can be observed more clearly.

[0324] 3. Biomarker screening and regression analysis

[0325] After dimensionality reduction using PCA, biomarker screening can be further performed. Assume that metabolites related to chicken spoilage are screened out through analysis of variance (ANOVA) or Lasso regression analysis. Through these methods, the key metabolites affecting chicken spoilage can be further identified, thus providing more reliable spoilage biomarkers.

[0326] Step 4: Multi-task learning and transfer learning

[0327] 1. Input of metabolite feature vectors

[0328] Assume there is the following metabolite data (concentrations of lipids, amino acids, sugars). Each sample includes the concentration information of these metabolites under different storage conditions and times:

[0329]

[0330]

[0331] This dataset contains the concentration information of multiple metabolites (e.g., AICAR, linoleic acid, glucose, etc.) in chicken samples under different storage conditions.

[0332] 2. Storage Conditions and Storage Time Vector

[0333] Storage conditions and time have a significant impact on spoilage and can be used as part of the input data. Storage conditions include refrigeration (C) and freezing (F), and the storage time can be 3 days, 7 days, or 10 days.

[0334]

[0335] 3. Environmental Factors (Temperature and Humidity)

[0336] Assume that environmental factors have an impact on the chicken spoilage process. The environmental factors for each sample include temperature and humidity. We can represent them as a vector, and the values of temperature and humidity are as follows:

[0337] Sample ID Temperature (°C) Humidity (%) 1 4 85 2 4 80 3 4 75 4 -18 90 5 -18 85 6 4 80 7 4 75 8 -18 80 9 -18 85 10 4 90

[0338] 4. Multi-Task Learning Model Training

[0339] Using the above metabolite data (lipids, amino acids, carbohydrates, etc.), storage conditions and time vector, and environmental factor vector as input, train a multi-task learning (MTL) model. This model will include multiple tasks:

[0340] Spoilage Marker Identification: Identify which metabolites are spoilage markers.

[0341] Storage Stage Prediction: Predict the current spoilage stage of the chicken.

[0342] Spoilage Rate and Degree Assessment: Assess the spoilage rate and degree of the chicken.

[0343] For each task, calculate the corresponding loss function:

[0344] Spoilage Marker Identification: Calculate the error using the cross-entropy loss function:

[0345]

[0346] where m is the number of samples, y i is the true label of the i-th sample, and p i is the predicted probability of the i-th sample;

[0347] Storage Stage Prediction: Use the mean squared error (MSE) loss function:

[0348]

[0349] where m is the number of samples, is the predicted spoilage stage of the i-th sample, and y stage,i is the true spoilage stage label;

[0350] Evaluation of deterioration rate and degree: The mean squared error loss function is also used:

[0351]

[0352] where m is the number of samples, is the predicted deterioration rate and degree, and y rate,i is the true label;

[0353] 5. Transfer learning

[0354] During the transfer learning process, we first use data under other storage conditions (such as refrigeration, freezing, etc.) for pre-training to obtain a basic model. Then, we fine-tune it on the target task (such as a new storage condition).

[0355] Fine-tuning dataset:

[0356] The fine-tuning dataset may contain the following data:

[0357]

[0358] When fine-tuning on the target task data, the model adjusts its parameters according to the existing pre-training information to adapt to the new storage condition.

[0359] Step 5: AICAR prediction module

[0360] 1. Prediction of the time of the deterioration stage

[0361] It is assumed that the prediction of the deterioration stage is calculated using the following formula:

[0362]

[0363] where: S target : Identification of the target deterioration stage (the target is the "mild deterioration" stage, S target = 2); S current : Identification of the current deterioration stage (the current is the "fresh" stage, S current = 1); v rate,i : Deterioration rate (unit: days), that is, the rate of the deterioration process; t pred,i : Predicted time to reach the target deterioration stage;

[0364] Assume that the current deterioration stage is fresh (1), the target deterioration stage is mild deterioration (2), and the deterioration rate is 1.5 days:

[0365] tpred,i = (2 - 1) / 1.5 = 0.67 days

[0366] 2. Confidence in the deterioration state

[0367] The confidence level is calculated using the following formula:

[0368]

[0369] where: p marker,j : the recognition probability of each marker; v rate,i : the deterioration rate (unit: days); α: the coefficient that controls the influence degree of the deterioration rate on the confidence level; c confidence,i : the confidence level of deterioration prediction, ranging from 0 to 1;

[0370] Assume that the marker recognition probability p marker,j = 0.8, the deterioration rate v rate,i = 1.5, and the coefficient α = 0.5: then c confidence,i = 0.6

[0371] 3. Suggestions for optimizing storage conditions

[0372] Based on the deterioration stage and the deterioration rate, the AICAR model gives optimization suggestions:

[0373] If the deterioration stage is close to severe deterioration (stage 3 or 4) and the deterioration rate is fast, it is recommended to transfer to frozen storage conditions (-18°C).

[0374] If the deterioration stage is low and the deterioration rate is slow, maintain the refrigerated storage conditions (4°C).

[0375] According to the above values, the deterioration stage is mild deterioration, the deterioration rate is 1.5 days, and it is recommended to maintain the refrigerated storage conditions (4°C).

[0376] Experiment 1: Verify the effectiveness of multi-task learning (MTL) and transfer learning in step four

[0377] 1. Experimental objectives

[0378] Verify the effectiveness of multi-task learning (MTL) and transfer learning techniques in processing metabolomics data, identifying chicken deterioration markers, predicting deterioration stages, and evaluating deterioration rates and degrees.

[0379] 2. Experimental design

[0380] Dataset preparation: Use the metabolomics data during the chicken storage process, including metabolite concentration data such as lipids, amino acids, and sugars, as well as storage conditions (refrigerated, frozen) and environmental factors (temperature, humidity). Select multiple sample datasets, including:

[0381] Fresh chicken (stage 1)

[0382] Mildly deteriorated chicken (stage 2)

[0383] Moderately deteriorated chicken (stage 3)

[0384] Severely deteriorated chicken (Stage 4)

[0385] Experimental data table:

[0386]

[0387] Task definition:

[0388] Deterioration marker identification task: Based on metabolite concentration data, identify which metabolites are deterioration markers (output: 0 or 1).

[0389] Storage stage prediction task: Predict which deterioration stage the chicken is in (output: 1, 2, 3, 4).

[0390] Deterioration rate and degree assessment task: Evaluate the deterioration rate (in days) and the degree of deterioration (a value from 0 to 1).

[0391] Model design: Use a multi-task learning (MTL) model and train the model by combining metabolomics data, storage conditions, and environmental factors. The loss function for each task is as follows:

[0392] Deterioration marker identification: Cross-entropy loss function

[0393] Storage stage prediction: Mean squared error loss function (MSE)

[0394] Deterioration rate and degree assessment: Mean squared error loss function (MSE)

[0395] At the same time, use the pre-training strategy of transfer learning, pre-train on the existing dataset, and fine-tune on the target task dataset.

[0396] 3. Experimental steps

[0397] Data preprocessing:

[0398] Standardize or normalize the metabolite concentration, storage conditions, storage time, environmental temperature, and humidity of each sample.

[0399] Handle missing values or outliers.

[0400] Multi-task learning model training:

[0401] Use metabolite concentration, storage conditions, and environmental factors as input data.

[0402] Use the cross-entropy loss function and the mean squared error loss function to train the deterioration marker identification, storage stage prediction, and deterioration rate and degree assessment tasks respectively.

[0403] Transfer learning fine-tuning:

[0404] Based on the pre-trained model, the model is fine-tuned using new storage conditions and environmental data.

[0405] When fine-tuning, the method of freezing some shared layers is used, and only the weights of the output layer for specific tasks are adjusted.

[0406] Performance evaluation:

[0407] The cross-validation method is used to evaluate the performance of the multi-task learning model.

[0408] The corresponding evaluation metrics are used for each task:

[0409] Identification of deterioration markers: Accuracy, Recall, Precision.

[0410] Prediction of the storage stage: Mean Squared Error (MSE).

[0411] Evaluation of deterioration rate and degree: Mean Squared Error (MSE).

[0412] 4. Experimental data and results

[0413]

[0414] Task of identifying deterioration markers: The model accurately identified the deterioration markers (accuracy was 90%).

[0415] Task of predicting the storage stage: The Mean Squared Error (MSE) for predicting the deterioration stage was 0.3, indicating that the model could accurately predict the deterioration stage.

[0416] Task of evaluating deterioration rate and degree: The Mean Squared Error (MSE) for the deterioration rate was 0.15, and the Mean Squared Error for the deterioration degree was 0.1.

[0417] 5. Experimental conclusion: This experiment verified the effectiveness of multi-task learning and transfer learning in predicting chicken deterioration. The experimental results showed that using metabolomics data and transfer learning methods could accurately identify deterioration markers, predict the deterioration stage, and evaluate the deterioration rate and degree.

[0418] Experiment 2: Verify the effectiveness of the step five - AICAR prediction module

[0419] 1. Experimental objectives

[0420] Verify the effectiveness of the AICAR prediction module in predicting the chicken deterioration status based on the output data of step four (identification results of deterioration markers, prediction of the deterioration stage, evaluation of deterioration rate and degree).

[0421] 2. Experimental design

[0422] Dataset Preparation: Use the experimental dataset in Step 4, including the identification probability of deterioration markers, the prediction results of deterioration stages, and the evaluation data of deterioration speed and degree for each sample.

[0423] Model Input:

[0424] Identification Probability of Deterioration Markers: The predicted probability of whether each metabolite is a deterioration marker.

[0425] Prediction Results of Deterioration Stages: The deterioration stage (from 1 to 4) that the chicken is in.

[0426] Evaluation Results of Deterioration Speed and Degree: The speed (days) and degree (from 0 to 1) of deterioration.

[0427] 3. Experimental Procedures

[0428] Model Input and Calculation:

[0429] Use the output of Step 4 as the input to the AICAR prediction module.

[0430] Calculate the time prediction from the current deterioration stage to the target deterioration stage using the deterioration stage time prediction formula:

[0431]

[0432] Calculate the confidence level of the deterioration state using the confidence level calculation formula:

[0433]

[0434] Optimization Suggestions for Storage Conditions:

[0435] Based on the predicted deterioration stage and speed, give optimization suggestions for storage conditions. If the deterioration speed is fast and the deterioration stage is close to severe deterioration, it is recommended to use freezing conditions; otherwise, maintain refrigeration conditions.

[0436] 4. Experimental Data and Results

[0437]

[0438] 5. Experimental Conclusions

[0439] Through this experiment, the AICAR prediction module can accurately predict the deterioration state of chicken based on the identification results of deterioration markers, the prediction of deterioration stages, and the evaluation of deterioration speed and degree, and give optimization suggestions for storage conditions. The experimental results show that the AICAR prediction module has high prediction accuracy and application value, and can provide effective data support for the optimization of chicken storage and quality control.

[0440] Although the embodiments of the present invention provide a detailed description, those skilled in the art can still make appropriate modifications or equivalent replacements to the specific implementation manners without departing from the core idea of the present invention, and all such modifications should be included within the protection scope of the present invention.

Claims

1. A chicken spoilage biomarker screening and AICAR prediction system based on multi-task learning and transfer learning, characterized in that It includes a sample collection module, a metabolite extraction module, a metabolomics analysis module, a multi-task learning model module, and an AICAR prediction module that are connected in sequence; Sample collection module: This module is responsible for collecting chicken samples under different storage conditions and shelf lives. Chicken samples at different storage time periods are selected, covering all stages from fresh to spoiled chicken. The sample collection for each stage ensures consistency between batches and data diversity; Metabolite extraction module: During the metabolite extraction process, polar metabolites such as amino acids and acids use methanol, acetonitrile, or a water-acetonitrile mixed solvent, while lipophilic metabolites use chloroform or dichloromethane solvents. The chicken samples are processed using cryogenic grinding technology to maintain the stability of metabolites; Metabolomics analysis module: High-throughput analysis techniques such as high-performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), and liquid chromatography-mass spectrometry (LC-MS) are used to perform metabolomics analysis on chicken samples. Amino acids, sugars, and organic acids, as well as polar and water-soluble metabolites, are analyzed by HPLC. Lipids, fatty acids, and carotenoids, as well as volatile and non-volatile metabolites, are analyzed by GC-MS and LC-MS. During the analysis, special attention is paid to the concentration changes of AICAR (5-aminoimidazole-4-carboxamide riboside) and its related metabolites to reveal key metabolic markers during the spoilage process of chicken; Multi-task learning model module: This module uses a multi-task learning (MTL) model to simultaneously process multi-dimensional data on chicken spoilage through multi-task training of the collected metabolite data. Through shared learning representations, the model can improve the ability to identify spoilage markers; AICAR prediction module: This module uses a transfer learning algorithm, combined with the output of the multi-task learning model, to specifically predict the concentration changes of AICAR and its related metabolites; Through transfer learning, the system can quickly transfer the data characteristics of different storage conditions and spoilage stages to new samples, achieving accurate prediction of AICAR.

2. A method for screening chicken spoilage markers and AICAR prediction based on multi-task learning and transfer learning, characterized in that, It includes the following steps: Step 1: Sample collection In the experiment for screening chicken spoilage markers and predicting AICAR, the sampling should cover fresh samples and chicken samples stored for 3 days, 7 days, and 10 days under different storage conditions. In addition, environmental factors affecting chicken spoilage during storage, including storage temperature and humidity variables, need to be recorded in real time during the collection process; Step 2: Metabolite extraction During the metabolite extraction process, polar metabolites such as amino acids and acids use methanol, acetonitrile, or a water-acetonitrile mixed solvent, while lipophilic metabolites use chloroform or dichloromethane solvents. The chicken samples are processed using cryogenic grinding technology to maintain the stability of metabolites and maximize the uniformity of the samples, providing accurate and high-quality samples for metabolomics analysis; Step 3: Metabolomics analysis Metabolomics analysis of chicken samples was performed using high-performance liquid chromatography (HPLC), gas chromatography-mass spectrometry (GC-MS), and liquid chromatography-mass spectrometry (LC-MS) high-throughput analysis techniques; amino acids, sugars, and organic acids, polar and water-soluble metabolites were analyzed by HPLC; then, lipids, fatty acids, and carotenoids, volatile and non-volatile metabolites were analyzed by GC-MS and LC-MS; during the analysis, special attention was paid to the concentration changes of AICAR (5-aminoimidazole-4-carboxamide riboside) and its related metabolites to reveal key metabolic markers during the spoilage process of chicken; Step Four: Multi-task Learning and Transfer Learning Model Analysis This module uses a multi-task learning (MTL) model to simultaneously process multi-dimensional data of chicken spoilage; through multi-task training of the collected metabolite data; through shared learning representations, the model can improve the recognition ability of spoilage markers and enhance the generalization ability of the system on different datasets; Step Five: AICAR Prediction and Spoilage Assessment Based on the multi-task learning and transfer learning model, the AICAR prediction module plays a key role; this module performs real-time spoilage prediction based on the extracted metabolite data and the learned spoilage model, and evaluates the spoilage status of chicken under different storage conditions; through the prediction of AICAR, the system can output the spoilage stage of chicken and predict the remaining shelf life, providing accurate warning information for users; this process can help consumers or food processors according to the spoilage warning.

3. The method according to claim 2, wherein: Step One specifically includes: Fresh samples: Chicken samples were collected shortly after slaughter to ensure they were in an unspoiled state; Stored samples: Chicken samples stored for 3 days, 7 days, and 10 days were selected; During the sampling process, the environmental factors during the storage of chicken must be recorded in real time, especially: Temperature change (T): Monitored in real time by a temperature recorder or sensor; Humidity change (H): Monitored by a humidity sensor; Storage duration T s : The storage duration is a time variable for the spoilage of chicken, which directly affects its spoilage rate; The spoilage rate (R) is a function jointly determined by the storage conditions; temperature, humidity, and storage time have a significant impact on the spoilage process of chicken; the higher the temperature, the faster the spoilage rate usually is; the higher the humidity, the more rapid the spoilage process may be; the longer the storage time, the more significant the spoilage process becomes; the following formula provides data support for subsequent metabolite analysis and spoilage model modeling, and the formula for recording environmental factors is: Where: k: proportionality coefficient related to the sample reference condition, α: influence coefficient of temperature on the deterioration rate, β: influence coefficient of humidity on the deterioration rate, γ: influence coefficient of storage duration on the deterioration rate The combined influence of temperature, humidity and storage time on the deterioration rate, R is the deterioration rate of chicken, T is the storage temperature, H is the storage humidity, T s is the storage duration; represents the comprehensive result of the signal intensity of metabolites during the deterioration process; Intensity i represents the signal intensity of the i-th metabolite, RetentionTime is the retention time of the i-th metabolite, and n is the number of metabolites; Batch identification: Each sample should be labeled with a batch number, collection date, storage conditions, and storage time at the time of collection to ensure that the specific information of each sample can be traced; Sample processing: Ensure that the collected chicken samples are processed and stored as quickly as possible after sampling to prevent premature spoilage due to temperature fluctuations during transportation and storage; To ensure the comprehensiveness and representativeness of the experimental data, the collected samples should cover chicken from different batches, different breeding environments, and different processing conditions; the number of samples in each batch should be sufficient to ensure that the samples can comprehensively reflect different stages of spoilage; Collecting multiple batches of samples under different storage conditions can ensure the comprehensiveness of the data, avoid biases that may be brought about by relying on only a certain batch, and improve the representativeness of the data. According to the complexity of the deterioration process, the formula for the number of batches under each storage condition and storage time is Among them, N is the total required sample size, m is the number of types under storage conditions (different storage days, different temperatures), and B i is the number of batches under each condition.

4. The method according to claim 2, wherein: Step 2 specifically includes: Before metabolite extraction, use cryogenic grinding technology for sample processing; cryogenic grinding can quickly freeze samples through liquid nitrogen or a low-temperature environment and grind samples at low temperature; during cryogenic grinding, control the temperature of the samples between -20°C and -80°C; Temperature change formula: During the grinding process, the temperature change of the sample is usually determined by the temperature of the external environment and the heat conduction characteristics of the sample; Basis for solvent selection: For polar metabolites such as amino acids and acids, use methanol, acetonitrile, or a water-acetonitrile mixed solvent, while for lipophilic metabolites, chloroform and dichloromethane solvents are more suitable; Solvent optimization formula: An important goal in the solvent selection process is to improve the extraction efficiency of metabolites; use the following formula to estimate the extraction efficiency of different solvent systems: Among them, C 溶液 is the metabolite concentration extracted from the sample, and C 初样 is the concentration of the metabolite in the initial sample.

5. The method according to claim 2, characterized in that: Step 3 specifically includes: High-performance liquid chromatography (HPLC): Analyze amino acids, sugars, and organic acids, polar and water-soluble metabolites by HPLC; analyze lipids, fatty acids, and carotenoids, volatile and non-volatile metabolites by GC-MS and LC-MS; Data collection and analysis: Preprocess, calibrate, and integrate data during metabolomics analysis; Data calibration: Each metabolite has a specific mass / charge ratio in the chromatogram and mass spectrum; calibrate the metabolites in each sample by comparing the retention time or mass spectrum of known standards; Quantitative analysis: Quantitatively analyze the abundance of each metabolite through the area or height of the chromatographic peak; for HPLC, LC-MS, and GC-MS, use the peak area method to calculate the metabolite concentration; Data integration and noise reduction: Since background noise is generated during the experiment, it is necessary to perform noise reduction on chromatographic and mass spectrometry data; the methods used are smoothing filtering and denoising algorithms; after processing the signal, extract the characteristic information of the main metabolites in each sample; Multidimensional data integration: Metabolomics analysis generates data in different dimensions, including time, temperature, and storage status; therefore, it is necessary to integrate these multidimensional data together to form a unified metabolite dataset; this process can be carried out through the principal component analysis (PCA) multivariate analysis method; Principal component analysis (PCA): PCA realizes dimensionality reduction by mapping the original data to a new coordinate system; X' = XW where X is the original data matrix, W is the principal component matrix, and X′ is the data after dimensionality reduction processing; this process can simplify the multidimensional data into a few important components for subsequent analysis; Marker screening: In metabolomics data analysis, marker screening is a very crucial step, especially for markers related to chicken meat deterioration; use statistical methods such as analysis of variance and the feature selection method Lasso regression based on machine learning for screening; Analysis of variance is used to compare whether there are significant differences in the means of multiple groups; analysis of variance determines whether there are statistically significant differences by comparing the variability between groups and within groups; if the F value is less than the significance level of 0.05, it is considered that there are significant differences in the means of at least two groups. Lasso regression formula: Among them, β0: intercept term, representing the bias of the regression model, i.e., the baseline value of the prediction result; β: regression coefficient, a vector containing the weight values of each feature variable in the model; each β j represents the influence degree of this feature on the target variable; n: number of samples, i.e., the total number of observations in the dataset; y i : observed value, representing the actual value corresponding to the i-th sample, the concentration of a certain feature deterioration marker related to chicken spoilage, the target variable; X i : input feature, representing the feature vector of the i-th sample, containing the concentrations of various input data metabolites related to chicken spoilage, storage time, temperature, β0 + X i β: predicted value, i.e., the value calculated by the model through the regression coefficient and input feature, which is the prediction of y i ; λ: regularization parameter, used to control the regularization strength in Lasso regression; a larger λ value will cause more regression coefficients to tend to zero, thus achieving the effect of feature selection. This parameter balances the fitting ability and complexity of the model, sum of squared residuals, representing the difference between the model predicted value and the true observed value. The goal is to minimize this value to improve the prediction accuracy, regularization term, used to penalize overly large regression coefficients and achieve feature selection by approaching some coefficients to zero; this helps reduce the complexity of the model and avoid overfitting, Solve for the regression coefficient by minimizing the above objective function.

6. The method according to claim 2, characterized in that: Step four specifically includes: In step three, a large amount of metabolite characteristic data was obtained through metabolomics analysis, including concentration data of various metabolites such as lipids, amino acids, and sugars; next, multi-task learning MTL and transfer learning techniques are used to process this data. The core of the input data comes from the metabolomics analysis in step three, specifically including the following types of data: Metabolite feature vector: The metabolite concentration information of each sample obtained through metabolomics analysis, including lipid, amino acid, and sugar components; there are a total of n metabolites, and the concentration of each metabolite is x i (i = 1, 2, …, n), the metabolite feature vector at temperature T can be expressed as: X = [x1, x2, …, x n T ​ Storage conditions and storage time: Storage conditions and storage time are important factors affecting the spoilage of chicken; in order to incorporate storage condition information into the model, it can be represented as a vector s j , and the vector of storage conditions and storage time at temperature T can be expressed as: S = [s1, s2, s3] T where s j ∈ {0, 1} indicates that under specific conditions (refrigerated or not), each element of s j corresponds to a storage condition; Batch and environmental factors: Batch and environmental factors also have an impact on the spoilage of chicken; there are k environmental factors, and each factor is represented by e j The representation of environmental factor E at temperature T is: E = [e1, e2, …, e k T ​ Multi-task learning MTL aims to simultaneously learn multiple related tasks by sharing some model parameters, thereby improving the generalization ability of the model; in this step, the main tasks include: Identification of spoilage markers: The goal is to identify the key markers of chicken spoilage from the metabolite characteristic data. Prediction of storage stage: The goal is to predict the spoilage stage of chicken at a certain moment. Assessment of spoilage speed and degree: The goal is to evaluate the spoilage speed and degree of chicken. Identification of spoilage markers: It is a classification problem, and the goal is to identify which metabolites are the markers of chicken spoilage; for each metabolite xi, it is necessary to predict whether it is a spoilage marker. Label y j ∈{0, 1} indicates whether the i-th metabolite is a biomarker, and the output is the probability p predicted by the model j , that is, the probability that the metabolite is a deteriorated biomarker; the loss function uses the cross-entropy loss function to measure the error of the classification problem; the cross-entropy loss function is as follows: where m is the number of samples, and y i is the true label of the i-th sample, and p i is the predicted probability of the i-th sample; Storage stage prediction: It is a regression problem, and the goal is to predict the spoilage stage of chicken; the spoilage stage label of each sample is a discrete integer value: where 1 represents fresh, 2 represents slightly spoiled, 3 represents moderately spoiled, and 4 represents severely spoiled; the output is the spoilage stage predicted by the model The loss function of this task uses the mean squared error loss function MSE, which is used to measure the prediction error in the regression problem; the loss function is as follows: where m is the number of samples, is the predicted deterioration stage of the i-th sample, y stage,i is the true deterioration stage label; Evaluation of spoilage rate and degree: It is a regression problem, and the goal is to evaluate the spoilage rate and degree of chicken; the label is y rate,i represents the spoilage rate and degree, and the model output is the predicted value A regression model can be used to predict the spoilage rate and degree; its loss function also uses the mean square error MSE: where m is the number of samples, is the predicted deterioration rate and degree, and y rate,i is the true label; The multi-task learning model structure is as follows: Shared layer: The shared layer extracts the general features of metabolite feature data X, storage conditions S, and environmental factors E through a neural network to enhance the generalization ability of the model; the output of the shared layer is an intermediate representation h shared ; h shared = f shared (X, S, E) where f shared is a function for the shared network, and h shared is the shared representation; Task-specific layer: Each task has a dedicated output layer that generates corresponding prediction results based on the shared representation h shared ; For the deteriorated marker recognition task, the output is p marker , and the classification activation function sigmoid is adopted: p marker = σ(W marker h shared + b marker ) For the storage stage prediction task, the output is using a linear activation function: For the task of evaluating the deterioration speed and degree, the output is The linear activation function is also used: Optimization objective: the final loss function L total is the weighted sum of the losses of multiple tasks, where λ1, λ2, λ3 are the weight coefficients of the tasks: L total = λ1L marker + λ2L stage + λ3L rate Update the model parameters by optimizing the loss function to achieve the comprehensive optimization of multiple tasks. Transfer learning: Transfer learning is used to solve the problem of data scarcity, especially when dealing with data of new storage conditions or new batches of chicken; its process is as follows: Pre-training: Based on data under other storage conditions, pre-train to learn general spoilage rules and metabolite characteristics. Fine-tuning: Fine-tune on the data of the target task, adjust the model parameters to make it adapt to the new data environment; during the fine-tuning process, the weights of some shared layers remain fixed, while the weights of the output layers for specific tasks are updated. Output data and subsequent applications: Through the combination of multi-task learning and transfer learning, the final model output will be used as the input of the subsequent AICAR prediction module; these output data include: Identification results of spoilage markers: Whether each metabolite is a spoilage marker and its concentration changes at different spoilage stages. Prediction of chicken spoilage stage: Based on the marker data and storage conditions output by the model, real-time predict the spoilage stage of chicken. Assessment of spoilage speed and degree: Output the assessment of the spoilage speed and degree of chicken spoilage. These outputs will be used as the input of step five, further analyzed and generate the prediction results of the spoilage state, providing decision support for storage optimization and quality control.

7. The method according to claim 2, wherein: Step five includes: In step four, relevant information on chicken spoilage, spoilage marker identification results, spoilage stage prediction, and spoilage rate and degree assessment are obtained through multi-task learning (MTL) and transfer learning techniques; this information will be used as input data and enter the AICAR prediction module in step five to accurately predict the spoilage process of chicken based on the existing metabolite characteristics and spoilage status, and provide decision support for storage optimization and quality control; In step five, the AICAR prediction module comprehensively analyzes the spoilage marker identification results, spoilage stage prediction results, and spoilage rate and degree assessment data to accurately predict the spoilage time range and confidence level of the spoilage status of chicken, and gives corresponding suggestions for optimizing storage conditions; The input data includes the following aspects: Results of identifying deterioration markers: Marker identification probability p of each metabolite from Step 4 marker ; These probability values reflect whether each metabolite is a spoilage marker and their degree of spoilage; Prediction result of the spoilage stage: the predicted value of the spoilage stage from Step 4 Indicates the current spoilage stage of the chicken; this value is a discrete integer representing the freshness level of the chicken: 1 represents fresh, 2 represents slightly spoiled, 3 represents moderately spoiled, and 4 represents severely spoiled; Assessment results of deterioration rate and degree: Deterioration rate from Step 4 and deterioration degree; The spoilage rate describes the speed at which chicken progresses from its current state to the next spoilage stage; The spoilage degree represents the current spoilage level of chicken, ranging from completely fresh (0) to completely spoiled (1); The core objective of the AICAR prediction model: is to accurately predict the spoilage status of chicken in the future based on the input metabolite characteristics, marker identification results, spoilage stage, and spoilage rate data, and give the corresponding spoilage time range, confidence level of the spoilage status, and suggestions for optimizing storage conditions; AICAR prediction model structure: The main components of the AICAR model include the following parts: Spoilage stage time prediction: Based on the current spoilage stage, spoilage rate, and marker identification results, the model calculates the time prediction for chicken to progress from the current stage to the next stage; Time prediction formula: Where: S target : Identification of the target deterioration stage; S current : Identification of the current deterioration stage; v rate,i : Deterioration rate, i.e., the rate of the deterioration process; t pred,i : Predicted time to reach the target deterioration stage; Definition of output values: t pred,i : represents the predicted time from the current deterioration stage to the target deterioration stage, in days; Spoilage stages are usually discrete integers: 1: fresh; 2: slightly spoiled; 3: moderately spoiled; 4: severely spoiled; Prediction result t pred,i defined range: A negative prediction result indicates that the chicken is in good condition, with a very slow spoilage rate or no spoilage; When the predicted time is 0 days, it means that the chicken has already reached the target spoilage stage and immediate measures need to be taken; When the predicted time is positive, it represents the time required for chicken to progress from the current spoilage stage to the target spoilage stage, usually a positive integer; Assessment of confidence in the deterioration state: The AICAR model will also give the confidence level c of the deterioration state prediction based on the recognition probability of deterioration markers and the assessment results of the deterioration rate, confidence,i indicating the reliability of the model's prediction results; Confidence level calculation formula: where: p marker,j : the recognition probability of each marker; v rate,i : the deterioration rate; α: a coefficient that controls the degree of influence of the deterioration rate on the confidence level; c confidence,i : the confidence level of deterioration prediction, ranging from 0 to 1; Definition of output values: c confidence,i : Confidence value, indicating the reliability of the prediction result, with a value range from 0 to 1; 0 indicates that the model is very uncertain about the prediction result and the reliability of the prediction is low; 1 indicates that the model is extremely certain about the prediction result and the reliability of the prediction is very high; Suggestions for optimizing storage conditions: Based on the predicted spoilage stage and rate, the AICAR model will give suggestions for optimizing storage conditions; these suggestions include whether to change the storage temperature or humidity conditions to delay chicken spoilage; Optimization suggestion generation logic: spoilage rate v rate,i High, and the spoilage stage is close to severe spoilage. It is recommended to transfer the chicken to a lower storage temperature; the spoilage stage is low and the spoilage rate is slow. It is recommended to maintain the current storage conditions; Definition of output values: The numerical values of storage condition suggestions adopt the following common storage schemes: Refrigeration: applicable to slightly spoiled or slow spoilage rate cases; Freezing: applicable to cases with a fast spoilage rate and a short predicted spoilage time; Room temperature: applicable to cases with a very slow spoilage rate; Output results of the AICAR prediction module: The output of the AICAR prediction module will include the following items: Spoilage time prediction: Output the time required for chicken to progress from the current spoilage stage to the next spoilage stage; Deterioration confidence level: Output the confidence level of the predicted deterioration state, indicating the reliability of the model for the result; Storage condition optimization suggestions: Output storage condition optimization suggestions based on the deterioration prediction results.

Images