Method and device for predicting flavor of low-sodium myofibrillar protein

Abstract

The invention discloses a low-sodium myofibrillar protein flavor prediction method and a low-sodium myofibrillar protein flavor prediction device. The method comprises the following steps: firstly, constructing a protein-flavor interaction system containing myofibrillar protein, potassium chloride, kappa-carrageenan and representative aldehyde flavor compounds; then, by means of spectroscopy, rheometer combination, an electronic tongue and the like, protein conformational change, typical aldehyde molecule binding rate and taste features are represented in a multi-dimensional mode. And establishing a random forest prediction model, taking protein conformation parameters as input, and taking a flavor binding rate and effective taste intensity as output. The model can reveal cross correlation between protein conformations and functions and identify key conformation parameters that affect flavor. And in combination with quantification and prediction capabilities of the model, a conformation-flavor relationship can be systematically analyzed, a basis is provided for optimization of a low-sodium formula, flavor is enhanced while salty taste is increased and bitter taste is inhibited, and healthy and sustainable development of low-sodium meat products is promoted.

Description

Technical Field

[0001] This invention belongs to the field of food processing technology, specifically relating to a method and apparatus for predicting the flavor of low-sodium myofibrillar proteins. Background Technology

[0002] The flavor and quality of meat products are determined by a variety of biochemical processes, including lipid oxidation, Maillard reactions, and protein degradation. Among these, the interaction between flavor compounds and protein molecules (including non-covalent and covalent bonds) is a key factor influencing the release behavior of volatile components and the intensity of sensory perception. Myofibrillar protein (MP), as the most abundant protein component in meat (accounting for approximately 55%–65% of total protein), significantly impacts the texture formation and flavor retention of meat products through its conformational state and structural changes. Existing research indicates that MP exhibits varying adsorption capacities for different volatile compounds, particularly showing a strong binding capacity for aldehydes. The carbon chain length and functional group structure of the compounds also affect the binding strength. In meat processing, sodium chloride (NaCl) is widely used to improve product texture and flavor, primarily by promoting MP dissolution and gel network formation. However, the health risks associated with excessive sodium intake make the development of low-sodium meat products an inevitable trend. Potassium chloride (KCl), a common substitute for sodium chloride, tends to produce a distinct bitter taste and metallic aftertaste at high substitution ratios, which limits its application in low-sodium meat products.

[0003] To address the aforementioned issues, the natural polysaccharide κ-carrageenan (KC), due to its porous network structure and sulfate groups on its polymer chains, can bind with potassium ions (K+). + Selective complexation occurs, reducing the concentration of free potassium ions and thus inhibiting bitterness to some extent. Simultaneously, potassium phosphate (KC) can synergistically form a relatively dense and stable gel structure with potassium phosphate (MP), which is beneficial for the retention of flavor compounds. However, the mechanism by which the conformational changes of MP affect the adsorption behavior of aldehyde flavor compounds under reduced sodium conditions remains unclear; the introduction of KC further complicates the relationship between protein structure, aldehyde binding capacity, and flavor retention in the system. Current technology lacks systematic research and effective control methods for protein-aldehyde interactions and their effects on flavor retention and bitterness inhibition in multi-component low-sodium meat product systems.

[0004] Traditional methods have significant limitations in accurately characterizing such low-sodium, multi-component protein-flavor systems. For example, cross-interference from multiple signals makes data interpretation difficult, and the accuracy for verifying complex mechanisms is insufficient. Random Forest (RF) has shown advantages in handling high-dimensional nonlinear data, being not only robust and reliable but also capable of quantifying the importance of each input feature, providing an effective tool for analyzing the conformational evolution of myofibril proteins and their aldehyde binding dynamics. However, a scheme combining RF methods with polysaccharide-regulated low-sodium meat product systems for systematically studying protein-aldehyde binding mechanisms, flavor retention, and bitterness inhibition has not yet been reported.

[0005] Therefore, developing a novel flavor modulation technology for low-sodium meat products is particularly urgent. The core of this solution lies in utilizing KC to target and regulate the conformation of myofibrillar proteins and their aldehyde adsorption behavior, thereby achieving flavor enhancement and maintenance while suppressing unpleasant taste sensations. Simultaneously, by integrating Fourier transform spectral characteristics with random forest modeling, a model capable of predicting and optimizing product flavor and taste is constructed, providing an innovative technical pathway for the precise design and molecular-level regulation of low-sodium meat products. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method and apparatus for predicting the flavor of low-sodium myofibrillar proteins, thereby solving the problem that the flavor of low-sodium meat products is difficult to accurately predict and optimize in the existing technologies.

[0007] According to one aspect of this application, a method for predicting the flavor of low-sodium myofibrillar proteins is disclosed, the method comprising: First, multi-dimensional data collection and feature construction were carried out. By constructing a simulated low-sodium meat product system (including myofibrillar protein, sodium and potassium salt mass, KC mass and one or more aldehyde flavor molecules), multi-spectral technology, electronic tongue technology and other means were used to collect multi-source heterogeneous data such as protein conformation, flavor molecule binding rate and taste characteristics to form the original dataset. The high-dimensional nonlinear data in the original dataset are extracted and dimensionality reduced using the Random Fourier Feature (RFF) method. Gaussian noise is injected into the feature data for data augmentation to improve the robustness and generalization ability of the model training and construct a fusion feature matrix. Based on the fused feature matrix, a multi-objective output random forest model is constructed. The model consists of multiple independent random forest regressors, which are used to simultaneously predict multiple target variables such as saltiness intensity, bitterness intensity, and flavor molecule binding rate. The model is trained and validated through cross-validation and permutation test. Based on the model analysis method, feature contribution analysis is performed on the trained random forest model to quantify the influence of each protein conformation feature on the flavor prediction results. This is used to guide the virtual screening and optimization of low-sodium product formulations, and to achieve precise control of the flavor characteristics of low-sodium meat products.

[0008] In some embodiments, the present invention provides a method for constructing a machine learning dataset, the method comprising: A controllable simulated low-sodium meat product system was constructed, using myofibrillar protein as a matrix, and the ratio of sodium to potassium salts, the amount of potassium chloride (KC) added, and the concentration of aldehyde flavor compounds were adjusted. Multi-dimensional data acquisition is performed on simulation systems with different component configurations. The acquisition employs at least one non-invasive detection technology, including multispectral imaging technology and electronic tongue analysis technology. Multispectral imaging technology is used to acquire spectral data characterizing protein conformation and flavor molecule binding state, and electronic tongue technology is used to acquire electrochemical signals characterizing taste profile, which are then converted into quantified taste feature data. The multi-source heterogeneous data is integrated and structured to form a dataset for training and validating machine learning models.

[0009] In some embodiments, the method further includes preprocessing and feature engineering the dataset, and using the RFF method to map and reduce the dimensionality of high-dimensional nonlinear features, thereby reducing the feature dimension while retaining key information. In some embodiments, to enhance model robustness, the method includes a data augmentation step that expands the training sample size by injecting small-amplitude Gaussian noise into the feature data; In some embodiments, the data after feature extraction, dimensionality reduction and data augmentation are used to construct a fusion feature matrix to support the training and prediction of a multi-objective machine learning model. In some embodiments, a multi-objective output random forest model is constructed based on the fused feature matrix, where each random forest regressor corresponds to a target variable, which is used to achieve parallel prediction of multiple flavor-related indicators. To verify the model performance, this invention uses the Leave-One-System-Out Cross-Validation (LOSO-CV) method to evaluate the model; In some embodiments, a permutation test is further used to verify the statistical significance of the model prediction results.

[0010] By comparing the performance of the original model with this benchmark, it can be determined whether the model's predictive ability is significantly better than random guessing, thus confirming that the associations learned by the model are real. In some embodiments, model performance is quantified using a preset evaluation metric, such as the coefficient of determination (R²). 2 ) and mean squared error (MSE), used to measure the consistency and deviation between predicted and actual values: ; in, This represents the physicochemical true value of the i-th sample in the dataset; This represents the model's prediction for the i-th sample; The represents the average of all true values ​​in the dataset; m represents the total number of samples.

[0011] ; in, The true value is obtained from physical and chemical experiments. is the target prediction result, and m is the number of training data samples.

[0012] In some embodiments, feature importance is calculated for each of the four target variables corresponding to a random forest sub-model. The importance scores of each sub-model are then normalized and summed or weighted to obtain a global feature importance ranking. Based on this ranking, the top 10 original physicochemical features are extracted from all input features to form a core feature set, which is used for interpretability analysis and mechanism research, thereby achieving feature compression while retaining key information.

[0013] According to any of the aforementioned methods, the input features include at least two types of information: the original low-dimensional conformation and physicochemical features, such as the amount of KC added, the amount and ratio of NaCl / KCl, the ratio of protein secondary structure, surface hydrophobicity, rheological parameters, etc. Based on the RFF method, high-dimensional nonlinear features were extracted from eight high-dimensional modal data, including synchronous fluorescence, Trp fluorescence, UV-Vis spectroscopy, storage modulus G', loss modulus G'', loss tangent tanδ, and apparent viscosity. These two types of features were included in the evaluation when summarizing feature importance, so as to achieve synergistic quantification of conformational information and nonlinear spectroscopic information. In some embodiments, the Tree Explainer is used to perform precise analysis on the trained random forest sub-model, applying the model to any sample: ; in, For all the true average values ​​corresponding to the trials, The SHAP value of the i-th input feature is used to quantify the marginal contribution effect of each input feature to the prediction result. The SHAP interpretable step according to any of the foregoing methods is characterized in that: the single feature SHAP value Based on the Shapley value theory, the mathematical form of the weighted summation of the marginal contributions of all feature subsets S that do not contain feature i is: ; Where N is the set of all features, For the size of the subset, This represents the model's output for sample x when only a subset of features S is used. This represents the model output after adding feature i to subset S, thereby obtaining the marginal effect of feature i on targets such as saltiness intensity, bitterness intensity, and aldehyde adsorption rate.

[0014] According to any of the aforementioned methods, calculate the corresponding SHAP value for all samples on the training set, and construct a sample × feature SHAP matrix, where each row of the matrix corresponds to a sample and each column corresponds to an input feature; The SHAP matrix can be generated for each of the four target variables, thereby quantitatively characterizing the positive and negative contributions and contribution intensity of KC addition, ionic strength, protein conformation parameters and RFF features to each flavor index in different samples, realizing a unified interpretation framework under multiple objectives and multiple modalities.

[0015] According to any of the aforementioned methods, the SHAP value of each input feature across all samples is statistically analyzed; A SHAP interaction contribution analysis method for quantifying the interaction effect between protein conformation and formulation features, characterized in that: for any two features x j With x k Define its interactive SHAP value on sample x. This is used to characterize the additional second-order contribution to the model output when both features coexist, and its calculation satisfies: ; in, Considered as a characteristic of collaborative consideration The model output, The contribution of each feature, The baseline output is used when there is no characteristic input, thereby quantitatively assessing the synergistic or antagonistic effects on flavor and aldehyde adsorption rate.

[0016] The model adopts a multi-output regression framework, using a random forest regressor as the base learner. The four target variables are jointly modeled by the Multi OutputRegressor to achieve simultaneous prediction of multiple taste attributes and aldehyde adsorption rate.

[0017] In some embodiments, constructing the random forest model includes: The parameters of the random forest regressor are set as follows: number of trees n_estimators = 200, maximum depth max_depth = 20, minimum number of split samples min_samples_split = 5, minimum number of leaf samples min_samples_leaf = 2, and n_jobs = −1 to enable multi-core parallel computing; A multi-output random forest model is constructed based on a multi-output regressor. The single random forest regressor is encapsulated into a multi-output structure. Independent regression sub-models are established for the four target variables, and the corresponding prediction results are output based on the same input feature vector. Model training and evaluation employed a five-fold cross-validation approach, using mean squared error (MSE) and coefficient of determination (R²) as performance evaluation metrics. The input features are processed by the feature preprocessing module, and the standardization method is used to unify features of different dimensions to zero mean and unit variance. Kernel approximation was performed on high-dimensional modal data such as synchronous fluorescence, Trp fluorescence, UV-Vis spectroscopy, storage modulus G', loss modulus G'', loss tangent tanδ, and apparent viscosity using random Fourier features (RFF). After RFF mapping, feature selection is performed on each modality feature, and the top 10 RFF components in terms of importance are extracted as the input features of that modality. The importance is determined based on model importance, variance contribution rate or correlation. During model training, a data augmentation method based on Gaussian noise injection is introduced to apply random perturbations to the input features in order to improve the robustness and generalization ability of the model. The model interpretation adopts the SHAP method, which calculates the contribution of input features at the sample level and the global level based on Tree Explainer, and is used to quantify the influence of conformational features, spectroscopic features and formulation parameters on multi-objective prediction results.

[0018] In some embodiments, the low-sodium myofibrillar protein data includes: The formulation and environmental parameters of the low-sodium MP system include ionic strength, NaCl / KCl dosage and ratio, protein concentration, KC addition amount and aldehyde concentration; Protein conformation data, rheological data; Flavor-related data, including the intensity of saltiness, umami, and bitterness, as well as the adsorption or binding rate of representative aldehyde compounds.

[0019] The above data serve as input features and supervision signals for a multi-objective flavor prediction model, used to characterize the flavor retention capability of the low-sodium MP system.

[0020] According to another aspect of this application, a random forest flavor prediction model apparatus is disclosed, comprising: The data acquisition module is used to collect and store data on KC addition amount, salt ratio, and physicochemical, conformational, rheological and aldehyde indicators. The data processing module is used to preprocess, standardize, and construct features from the data to generate the model input feature matrix; The prediction module is used to call the multi-output random forest model and output prediction results for various taste intensities and aldehyde adsorption rates. The model interpretation module is used to quantify the contribution of each input feature to the prediction result and generate visual analysis results; The user interaction module is used to receive user-defined KC addition amount, salt ratio constraints and target flavor indicators, and to perform prediction and optimization.

[0021] According to another aspect of this application, an electronic device is also disclosed, comprising: At least one processor and a memory, wherein the memory stores instructions that, when executed by the processor, implement the low-sodium myofibrillar flavor prediction method as described in any of the foregoing embodiments.

[0022] According to another aspect of this application, a computer-readable storage medium is also disclosed, on which instructions are stored, which, when executed by a processor, implement the low-sodium myofibrillar flavor prediction method as described in any of the foregoing embodiments.

[0023] The present invention includes, but is not limited to, the following beneficial effects: (1) A "quaternary system" research framework was constructed with a low-sodium system as the background, coupled with the interaction between κ-carrageenan (KC), salt ions, myofibril protein conformational changes and aldehyde flavor molecules. By setting up 11 controllable typical systems, a systematic characterization of protein conformation, aldehyde adsorption and taste response was achieved, breaking through the research limitations of existing technologies that are mainly based on binary systems or macroscopic parameters, and improving the completeness and reliability of conformation-flavor correlation analysis. (2) In terms of data acquisition, multimodal methods such as synchronous fluorescence, Trp fluorescence, UV-Vis spectroscopy, dynamic rheology and electronic tongue were used to form a multi-source heterogeneous dataset covering 8 types of high-dimensional spectroscopic / physical property data and 15 key physicochemical indicators, which significantly improved the characterizability and model recognizability of conformational changes and flavor responses under KC intervention. (3) To address the high-dimensionality, nonlinearity, and collinearity issues of spectral and rheological data, this invention introduces the Random Fourier Feature (RFF) method to map nonlinear relationships to a random feature space. This reduces computational complexity while maintaining sensitivity to spectral details, achieving a balance between expressive power and computational efficiency. (4) In terms of feature engineering, this invention performs intra-modal feature screening before cross-modal fusion and constructs a 95-dimensional comprehensive feature matrix. Simultaneously, it performs data augmentation by injecting Gaussian noise with adjustable amplitude to simulate actual production fluctuations, significantly improving model stability and generalization ability, and alleviating the problems of insufficient data and overfitting. (5) In terms of modeling strategy, this invention adopts the Multi-Output Random Forest Regressor framework. Independent random forest sub-models (e.g., 200 trees, 20 layers deep) are constructed for targets such as saltiness, bitterness, umami, and aldehyde binding rate, respectively, to achieve "target-specific modeling and overall integration," significantly improving prediction accuracy compared to traditional single-model schemes. (6) This invention is the first to apply SHAP-based interpretability analysis to the flavor prediction of low-sodium meat products. The TreeExplainer is used to quantitatively analyze the contribution and interaction effects of features such as KC concentration, α-helix content, surface hydrophobicity and gel strength on multiple flavor targets, and transform the model prediction results into mechanistic information that can support process decision-making. (7) Leave-One-System-Out Cross-Validation (LOSO-CV) combined with permutation test is used to ensure the applicability and statistical significance of the model in unseen formulation systems, and the stability of the model under perturbation conditions is evaluated by the prediction interval. (8) Based on the global integration of the importance of multi-target features, this invention can screen the top 10 core physicochemical variables with the highest contribution from 95 features, providing a clear basis for model lightweighting, industrial deployment and mechanism research.(9) This invention constructs an intelligent device integrating data acquisition, feature processing, model prediction, interpretable analysis and parameter optimization. It can output multi-flavor index prediction results and SHAP analysis results in real time according to the user input of KC addition amount and salt ratio, realize virtual screening of formula and process auxiliary decision-making, and significantly reduce experimental costs and R&D cycle. (10) Based on the above, this invention proposes a "system-level flavor response prediction-feedback linkage mechanism" suitable for actual low-sodium meat product flavor regulation scenarios, realizing bidirectional mapping from high-dimensional physicochemical input to multi-index flavor output. (11) Through systematic innovation in system construction, data acquisition, feature expression, modeling strategy and interpretability, this invention solves the problems of unclear mechanism, unstable prediction and uninterpretable model in low-sodium meat product flavor regulation, quantitatively reveals the key path of KC regulating protein conformation and aldehyde binding behavior, and forms a complete technical solution that can directly serve industrial formula optimization and product development. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0025] Figure 1 This is a flowchart illustrating the principle of the low-sodium MP flavor prediction method according to an embodiment of the present invention. Figure 2 This is a comparative graph showing the effect of low-sodium treatment with different sodium substitution ranges on the ability of MP to bind 3-methylbutyraldehyde, pentanal, hexanal, and heptanal. Figure 3 This invention relates to the effect of low-sodium treatment on the taste characteristics (umami, saltiness, bitterness) of the MP-aldehyde compound system. Figure 4 This is a diagram showing the proportion of secondary structures of MP-aldehyde complexes in the low-sodium treatment of this invention. Figure 5 This invention relates to the effect of low-sodium treatment on the particle size and zeta potential of MP-aldehyde complexes in an embodiment of the present invention. Figure 6 This describes the distribution of protein solubility and turbidity of MP-aldehyde complexes during the low-sodium treatment in this embodiment of the invention. Figure 7 This invention relates to the effect of low-sodium treatment on the surface hydrophobicity and thiol content of MP-aldehyde complexes. Figure 8 This invention illustrates the effect of aldehydes on the rheological properties of MP under different low-sodium systems in embodiments of the present invention. Figure 9 This invention illustrates the effect of aldehydes on MP synchronous fluorescence spectra at Δλ=15nm and Δλ=60nm under different low-sodium systems. Figure 10 This invention relates to the effects of aldehydes on MP tryptophan fluorescence (A) and ultraviolet spectrum (B) under different low-sodium systems in embodiments of the invention. Figure 11 This is a diagram of the protein taste and binding characteristics prediction model based on multi-source data fusion and multi-output random forest according to an embodiment of the present invention. Figure 12 This is a diagram illustrating the multi-index performance prediction (A) and error evaluation (B+C) method of random forest regression according to an embodiment of the present invention; Figure 13 This is a diagram illustrating the high-precision adsorption rate prediction method based on random forest learning curves and residual analysis, as described in this invention. Figure 14 This is a prediction graph of the multi-sensory index prediction model for low-sodium MP based on SHAP feature contribution analysis in an embodiment of this application; Figure 15 This is a schematic diagram of the indicator prediction device and system provided in the embodiments of the present invention; Figure 16 This is a flowchart of one step of the low-sodium MP flavor prediction method according to an embodiment of the present invention; Figure 17 This is a structural block diagram of the low-sodium MP flavor prediction device according to an embodiment of the present invention. Detailed Implementation

[0026] This invention provides a random forest flavor prediction model and device based on KC-regulated low-sodium myofibrillar protein to enhance aldehyde retention. The implementation is as follows: First, a low-sodium meat product model system is prepared according to the experimental design. Eleven representative formulation systems are constructed by adjusting the salt content and κ-carrageenan dosage. Then, multimodal data of each system are collected, including synchronous fluorescence, Trp fluorescence, UV-Vis spectroscopy, rheological parameters, and electronic tongue taste response, to obtain data on protein conformation, surface hydrophobicity, gel network characteristics, aldehyde binding rate, and salty, bitter, and umami flavors. Next, the high-dimensional spectral and physicochemical data are preprocessed, and nonlinear features are extracted using random Fourier transform (RFF) mapping. A comprehensive feature matrix is ​​constructed by combining intramodal feature screening and data augmentation. Then, an independent sub-model is trained using a multi-output random forest regression framework for salty, bitter, umami, and aldehyde binding rates, and the SHAP is calculated based on TreeExplainer. The model is evaluated using leave-one-out cross-validation and permutation tests to assess its predictive accuracy, stability, and generalization ability. Finally, the trained model is deployed in an intelligent device to instantly output multidimensional flavor prediction results, key feature contributions, and optimization suggestions based on user-inputted new formula parameters, enabling virtual formula screening and process optimization. Through these implementation methods, efficient flavor prediction and control can be achieved in the development of low-sodium meat products, significantly shortening the experimental cycle, reducing R&D costs, and providing quantitative decision support for formula optimization.

[0027] The terms "first," "second," etc., used in this specification and claims are only used to distinguish technical features and do not indicate any order limitation; the terms "comprising" and "having" should be understood as non-exclusive limitations. For ease of understanding, the specific process of the embodiments of this invention is as follows: Figure 16 As shown, Figure 16 This is a flowchart of a method for predicting the properties of protein gels and salty release, combined with... Figure 1 The schematic diagram shown illustrates the steps of the method.

[0028] S100, Obtain multiple sample data: The sample data were obtained from a low-sodium system with a number of targets. The sample data includes high-dimensional data of the first target type and conformational physicochemical data of the second target type. In this example, the first target type can be 8 types of spectral and rheological data, and the second target type can be 15 types of conformational physicochemical numerical data.

[0029] S102. Determine the aldehyde molecule adsorption simulation system based on multiple sample data: An aldehyde adsorption simulation system was constructed by combining low-sodium myofibrillar protein, potassium salt as a sodium salt substitute, κ-carrageenan as a texture modifier, and one or more representative aldehyde flavor molecules. Under controlled conditions such as ionic strength and temperature, the protein and target aldehydes were fully interacted, and a stable protein-aldehyde-polysaccharide interaction system was formed by adjusting the amount or ratio of κ-carrageenan. The system was characterized in multiple dimensions to obtain protein conformation data (including secondary structure ratio and surface hydrophobicity), binding rate data of aldehyde flavor molecules to proteins, and taste characteristic data.

[0030] S104. Preprocess and extract features from protein conformation data, construct input feature variables, and define output target variables: Protein structure parameters, flavor substance binding or residue data, and sensory or simulated sensory indicators are standardized, denoised, and feature-selected to construct input feature variables for modeling; aldehyde binding rate and taste characteristics such as saltiness and bitterness are defined as output target variables.

[0031] S106. Construct and train a flavor prediction model based on input feature variables and output target variables: A random forest flavor prediction model is constructed based on the aforementioned feature variables to establish a quantitative mapping relationship between the input feature variables and the output target variables, and model training and parameter optimization are completed through methods such as cross-validation.

[0032] S108. Input the target low-sodium formula to be predicted into the flavor prediction model for prediction: The characteristic data of the low-sodium formula to be predicted are input into the trained random forest model, and the corresponding flavor prediction results are output.

[0033] Furthermore, the mean squared error (MSE) and other indicators were used to evaluate the model fit and prediction performance; based on the SHAP theory, the importance of features was analyzed, and the influence weights of factors such as KC level, protein conformation changes and surface hydrophobicity on flavor retention and bitterness formation were quantified. Specifically, a computing device with machine learning capabilities is used as a modeling platform to acquire multimodal test data of myofibril proteins. The high-dimensional features include at least one or more of synchronous fluorescence spectroscopy, Trp fluorescence spectroscopy, UV-Vis spectroscopy, G′, G″, tan δ, and viscosity parameters. The original measurements of each modality are organized into high-dimensional feature vectors, and after standardization, they are input into a kernel approximation module based on random Fourier features (RFF) to enhance the model's ability to express nonlinear structure-flavor relationships.

[0034] The radial basis function kernel (RBF kernel) is selected as the target kernel function, as shown in the following formula: ; in: These are two standardized feature vectors; The square of the Euclidean distance; For the kernel parameter, in one embodiment γ = 1.0.

[0035] The RBF sampling unit (RBF Sampler) generates a random projection matrix for approximating the RBF kernel under the conditions of a preset kernel parameter γ (e.g., γ = 1.0) and a fixed random seed. By using a random projection matrix, the original high-dimensional features are mapped to random Fourier feature vectors of a predetermined dimension (e.g., 256-dimensional), thereby capturing the potential nonlinear relationship between the features and the target variable while maintaining computational efficiency. The form of the random projection matrix is ​​as follows: If all ωl are combined into a random projection matrix All offsets combined into For any physicochemical or high-dimensional spectral rheological data x obtained from an experiment, the above mapping can be written in matrix form: ; in, This is the element-wise cosine function of the vector.

[0036] For the random Fourier features obtained in the above steps, select the appropriate feature importance evaluation strategy based on the available sample size: When the number of modal samples is greater than or equal to a threshold (e.g., 5), mutual information (MI) is used to calculate the correlation between each random Fourier feature component and the target variable to obtain a supervised importance score; when the number of samples is less than the threshold, component variance is used for unsupervised scoring. For each modal feature, sort them by score, select the top 10 components as representative high-dimensional features, and discard low-contribution or redundant features. The modal features are concatenated in the sample dimension and combined with low-dimensional physicochemical indicators (such as ionic strength, salt type ratio, protein solubility, surface hydrophobicity, etc., totaling about 15 items) to form a unified input feature matrix X (approximately 23 dimensions).

[0037] To improve robustness to small samples, a small Gaussian noise (σ≈0.03) was superimposed on X and the corresponding target Y (taste score, electronic tongue response, aldehyde binding rate), and the samples were replicated multiple times to form an amplified training set (e.g., 10-fold). A random forest model is constructed and trained based on the amplified feature matrix X and the target Y. The number of trees, maximum depth, minimum number of leaf node samples, and feature random sampling ratio are set, and the parameters are optimized through cross-validation or training-validation split.

[0038] After training, the model can take in multimodal features and output flavor retention, bitterness intensity and aldehyde binding rate, and evaluate the model performance using untrained samples or hold-out data. Based on the model evaluation results, determine whether the preset performance threshold is met. If necessary, return to adjust the noise intensity, feature selection threshold, or random forest hyperparameters and retrain. A multi-output regression framework is adopted, with a tree model regressor (such as random forest) as the base learner. Independent sub-models are trained for the four target variables respectively, and multi-objective prediction results are obtained synchronously under a unified input, which facilitates formulation and process optimization. During the training phase, five-fold cross-validation is performed on the feature matrix and the target matrix. The samples are divided into five subsets, with one subset used as the validation set and the rest as the training set in turn. The stability and generalization ability of the model are evaluated. The results of each fold are calculated for the four targets and then averaged to obtain the overall prediction level.

[0039] Furthermore, under this method, the predicted value of each sample x, f(x), can be represented as the cumulative contribution of the feature and the baseline value, that is, measured by SHAP, generating a contribution matrix of "sample × feature", which quantitatively reflects the positive or negative influence of each feature on each sample. The mean contribution of the matrix's statistical features at the sample level can be used to verify its global importance, providing a quantitative basis for process control and formulation design. After preprocessing, feature construction, and standardization, the new formulation is input into a multi-output regression model. The four sub-models simultaneously output flavor and aldehyde adsorption predictions, and combine global features to interpret the matrix identification of key factors.

[0040] Experimental data were obtained from 11 groups of low-protein porcine systems with different salt forms and four aldehydes. Fifteen conformational physicochemical data and eight types of high-dimensional spectral / rheological data were collected. The 15 conformational physicochemical data characterized protein binding capacity and taste properties; aldehyde binding data were obtained quantitatively using GC-MS. Steps for extraction and low-sodium treatment of myofibril MP protein: 1) Take the longissimus dorsi muscle of pigs, remove fat and fascia, cut into about 1 small piece, and treat at 4°C; 2) Add 400mL of low-ion buffer A (20mM PIPES, 2mM MgCl2, 1mM EGTA, pH6.5) to 100g of meat piece, homogenize at 12,000rpm for 3×30s, with 1min intervals; 3) Centrifuge at 4°C, 2,000×g for 15min, discard the supernatant, resuspend the precipitate in buffer A and wash 3 times; 4) Wash twice with high-ion buffer B (0.1M NaCl), and resuspend in 0.1M NaCl; 5) Adjust the pH to 6.0±0.05, centrifuge at 4°C, 10,000×g for 20min to obtain high-purity milky white myofibril protein; 6) Dilute with PIPES buffer to 6.0mg / mL as the base sol.

[0041] Preparation of low-sodium system: As shown in Table 1, the total ionic strength is 0.6 M; KC-free: KCl replaces 10%-50% of NaCl to form KCl-10% to KCl-50%; KC-containing: KCl and KC composite salt replaces 10%-50% of NaCl to form Kcar-10% to Kcar-50%, KCl:KC=10:3; Homogenize at 8000 rpm for 1 min to ensure uniform salt dissolution; dispense and seal, and incubate at 4°C for 12–16 h.

[0042] Table 1. Different formulations for replacing NaCl in 130 mL MP suspension.

[0043]

[0044] Aldehyde flavor compounds and stock solutions: Four aldehydes: 3-methylbutanal (99%), pentanal (97%), hexanal (97%), and heptanal (97%); Stock solution preparation: Dissolve in chromatographic grade methanol, 500 mg / L, and sonicate for 20 min; Add myofibrillar protein system to a final concentration of 5 mg / L, protein concentration of 6 mg / mL, and total ionic strength of 0.6 M.

[0045] Sample grouping and processing: Proteome: 11 different salt systems + 4 aldehydes, 3 biological replicates; Blank group: replaced the proteolytic sol with an equal volume of PIPES buffer; stand at 25°C for equilibration for 16 hours.

[0046] Aldehyde binding assay: headspace solid-phase microextraction + GC-MS detection; binding rate (%) = [1 − (peak area of ​​proteome / peak area of ​​blank group)] × 100%.

[0047] Physicochemical, conformational, and high-dimensional characterization: determination of NaCl, KCl, KC mass, myofibrillar protein concentration, aldehyde concentration; particle size, zeta potential, protein solubility, turbidity, thiol content, and surface hydrophobicity; protein secondary structure: α-helix, β-sheet, β-turn, random coil; high-dimensional spectroscopy and rheology: synchronous fluorescence (Δλ=15 / 60nm), Trp fluorescence, UV–Vis; G′, G″, tanδ, and viscosity; used to evaluate the effects of different salt systems on protein conformation, rheological behavior, and aldehyde binding capacity, providing multimodal input for flavor prediction.

[0048] Furthermore, the effective taste properties, physicochemical and rheological characteristics, and fluorescence quenching of each low-sodium protein were measured to obtain the corresponding bitterness, saltiness, and umami intensity of the sample. The specific steps are as follows: Taste characteristics determination: Sample preparation: Mix 5g of protein with deionized water at a ratio of 1:5, disperse at 10,000×g for 1 min; centrifuge at 4°C and 3,000×g for 10 min, and take the clear supernatant; Electronic tongue test: A sample cup is placed in the supernatant. Following the equipment's standard procedure, multi-channel sensors collect electrical signals for saltiness, bitterness, and umami. Pattern recognition analysis is performed to obtain taste intensity parameters. Six parallel samples are used for each condition to ensure data reliability and provide quantitative basis for salt formulation optimization (e.g., ...). Figure 3 (and Table 2).

[0049] Table 2

[0050] Secondary structure content: Circular dichroism (CD) spectroscopy was used to detect aldehyde-treated proteins. Proteins were diluted to 0.6 mg / mL, with a path length of 1 mm, and scanned at 195-260 nm under nitrogen protection. Spectral data were imported into analysis software, and the contents of α-helices, β-sheets, β-turns, and random coils were calculated to characterize the secondary structure changes induced by aldehydes. Figure 4 ).

[0051] Other physicochemical values: A homogeneous suspension was prepared from the sample; particle size and zeta potential were measured to assess dispersibility and stability. Figure 5 Protein solubility: The ratio of supernatant protein content to initial concentration was determined by colorimetry; Turbidity: Visible absorbance reflects the aggregation state. Figure 6 ); Thiol content: Protein free thiols were determined by DTNB method, and absorbance was calculated ( Figure 7 ); Surface hydrophobicity: the protein reacts with the fluorescent probe, and the fluorescence intensity is measured ( Figure 7 ).

[0052] Rheological data: Storage modulus G′, loss modulus G″, loss factor tanδ, and viscosity were measured using a parallel plate rheometer. Figure 8 (Frequency scanning of the small-amplitude linear viscoelastic region, recording changes in G′ and G″ and calculating tanδ; temperature-controlled scanning to monitor changes in G′ and G″ with temperature, evaluating protein network formation and gelation behavior.)

[0053] Spectroscopic and Microenvironmental Analysis: Synchronous Fluorescence Spectroscopy: Δλ = 15 nm for analyzing the tyrosine microenvironment, Δλ = 60 nm for analyzing the tryptophan microenvironment, scanning 300–450 nm, and comparing changes in fluorescence intensity. Figure 9 Tryptophan intrinsic fluorescence: excitation at 280 nm, emission spectra recorded at 300–425 nm, analysis of maximum emission wavelength and fluorescence intensity, assessment of hydrophobic microenvironment and folding state ( Figure 10 A); UV-Vis spectroscopy: record absorption at 200–400 nm, focusing on the amide band and aromatic side chains to reflect the interaction between aldehydes and myofibrillar proteins ( Figure 10 B).

[0054] In one example, Figure 11The flowchart illustrates the data processing, augmentation, and construction of a machine learning model, as shown below. Figure 11 As shown, the procedure is as follows: The multi-output regression model uses a random forest regressor as the basic learner and encapsulates it within a multi-output regression framework to simultaneously predict multiple target indicators, such as changes in protein solubility, gel strength, viscosity, and taste intensity represented by the electronic tongue. The integration of multiple decision trees and multi-objective joint modeling helps to capture the nonlinear coupling effects between aldehyde introduction, salt ratio, and structural characterization parameters. To improve the model's ability to fit high-dimensional, multi-source features, the number of decision trees in the random forest is preferably set to 200, the maximum depth is limited to 20, and the minimum number of samples for each split and leaf node is appropriately constrained. This configuration ensures that complex inputs such as particle size, zeta potential, rheological parameters and spectral characteristics are fully learned, while suppressing overfitting to limited experimental data, thereby improving the model's generalization ability under unknown formulations or treatment conditions. To improve model training efficiency, a parallel computing strategy is enabled in the random forest model to parallelize the training process of different decision trees. By making full use of computing resources, it is possible to maintain a short training time even after introducing multimodal features such as particle size, fluorescence spectroscopy, UV-Vis, secondary structure ratio and rheological data, thus meeting the need for rapid model iteration in process optimization and formulation screening. To meet the needs of multi-output modeling, the random forest regressor is combined with the multi-output regression framework, enabling the model to simultaneously output multiple target variables, such as protein solubility, turbidity, thiol content, G′ / G″, tanδ, viscosity, and the intensity of saltiness, umami, and bitterness as assessed by the electronic tongue. A unified model structure enables collaborative prediction of multiple indicators, facilitating a comprehensive evaluation of the effects of formula adjustments in practical applications. To eliminate scale differences between features of different dimensions and magnitudes, all input features (including the characterization results of physicochemical indices, structural parameters, and spectral data) are first standardized to ensure that their mean and variance are on a uniform scale. By using stochastic Fourier feature mapping to perform nonlinear mapping on the standardized features, the complex relationship between particle size, Zeta potential, rheological parameters and spectral features coupled together is projected into a high-dimensional space to improve the model’s ability to learn nonlinear structure-function relationships. Furthermore, to reduce the interference of redundant features on model training, feature selection is performed from the high-dimensional features generated by random Fourier mapping. Preferably, the top 10 most important random Fourier features are selected for different data modes (such as physicochemical property modes, spectral modes, rheological modes, etc.) as the final input of the model; To enhance the model's robustness to experimental errors and measurement fluctuations, a data augmentation strategy involving Gaussian noise perturbation is implemented on the input data during the training phase. By superimposing small-amplitude random noise onto raw or characteristic data such as particle size, Zeta potential, fluorescence intensity, absorbance, G′ / G″, and viscosity, systematic errors and batch variations in actual measurements can be effectively simulated, thereby improving the model's applicability under real-world conditions. A 5-fold cross-validation strategy was introduced during the modeling process. The dataset, which contains multi-source features such as physicochemical, structural, spectral, and rheological characteristics, was divided into five subsets. A subset of these subsets was used as the validation set in turn. The mean squared error (MSE) and coefficient of determination of the model obtained from different training folds were compared. Model performance was quantitatively evaluated using mean squared error and coefficient of determination. Mean squared error measures the deviation of model predictions from actual experimental data (such as protein solubility, turbidity, thiol content, G′ / G″, viscosity, and taste intensity); coefficient of determination characterizes the proportion of the overall variance explained by the model. By combining the above indicators, we can intuitively compare the degree of fit of different model configurations or feature combinations on the prediction of particle size changes, structural rearrangement and flavor response. To reveal the degree of influence of different input features on the prediction results and improve the interpretability of the model, a tree-structure-based interpretive analysis is performed on the trained random forest multi-output regression model. By calculating the contribution and importance of particle size, Zeta potential, protein solubility, turbidity, thiol content, surface hydrophobicity, secondary structure ratio, rheological parameters and spectral characteristics in the prediction of various output indicators, key factors that significantly affect taste intensity, rheological behavior and structural stability can be identified. Understandably, a feature selection strategy is employed for the RFF features of each high-dimensional modality: When the number of modal samples is greater than or equal to a preset threshold, mutual information is used to evaluate the correlation between each RFF component and the target variable; when the number of samples is insufficient, component variance is used for unsupervised scoring. Based on the above scoring results, the Top 10 RFF components of each modality were selected and concatenated with 15 physicochemical indicators in the sample dimension to construct the final input feature matrix; Small Gaussian noise is superimposed on the feature matrix X and the target variable Y, and the data is augmented tenfold. The augmented samples are then vertically concatenated to form an amplified training set. The training set was expanded and input into the random forest multi-output regression model for training. As shown in Table 3, the model hyperparameters were preferably set to 200 trees, maximum depth 20, minimum number of split samples 5, and minimum number of leaf node samples 2. Parallel computing was enabled to improve training efficiency.

[0055] Table 3

[0056] In one example, Figure 12 The diagram illustrating the relationship between real and predicted data serves the purpose of: Verify the reliability and accuracy of the constructed model in predicting multidimensional physicochemical indicators and taste responses.

[0057] Figure 12 In this study, the measured values ​​of each target variable (such as protein solubility, viscosity, storage modulus G′, loss modulus G″, and taste intensity) are plotted on the x-axis and the model prediction values ​​are plotted on the y-axis to obtain a scatter plot. The theoretical consistency diagonal and linear fitting curve are also given. Depend on Figure 12 As can be seen, all scatter points are highly distributed near the diagonal, the linear fitting curve basically coincides with the diagonal, and the coefficient of determination R of each objective variable is high. 2 The values ​​are all close to 1, indicating that the model constructed in this invention has a good fitting ability and high prediction accuracy for multidimensional physicochemical indicators and taste response, and can be effectively used for rapid prediction and evaluation of system performance under different formulations and process conditions.

[0058] In one example, Figure 13 This diagram illustrates the learning curves and residual analysis of a random forest multi-output regression model. It is used to systematically evaluate the model's convergence behavior and generalization performance under different training sample sizes, including: The learning curves of the scores of each target variable in the training set and cross-validation as a function of the number of training samples show that as the number of training samples gradually increases from low to high, both the training score and the cross-validation score rise rapidly and tend to stabilize. The gap between the two curves is small and eventually approaches saturation, indicating that a relatively sufficient learning effect has been achieved with the current sample size. The residual analysis results of the model on the training and test sets. In each figure, the predicted value is plotted on the x-axis and the residual (measured value minus predicted value) is plotted on the y-axis. A histogram distribution of the residuals is also shown on the right to observe the concentration and symmetry of the error. The residuals of both the training and test sets are concentrated near zero and do not show a systematic deviation as the predicted value increases. The residual histogram shows a narrow peak with a zero center, indicating that the error is mainly a small random fluctuation and no obvious structural bias is observed. Figure 14This paper presents a feature importance analysis based on a random forest multi-output regression model to quantitatively assess the influence of various physicochemical and structural parameters on adsorption rate and taste attributes. The figure shows the SHAP colony distribution of the top ten key features when adsorption rate, bitterness, saltiness, and umami are the target variables. The horizontal axis represents the SHAP value, indicating the direction and magnitude of the feature's contribution to the model output, while the vertical axis represents the corresponding physicochemical or structural index. The color of the points reflects the intensity of the feature value. The model prediction results can be expressed as a linear sum of the input feature contributions.

[0059] from Figure 14 The subplots corresponding to the adsorption rate show that the amount of aldehyde added (C_Aldehyde), protein solubility, the content of β-sheets and α-helices in the secondary structure, turbidity, and β-turn angle have a significant impact on the adsorption rate prediction. Among them, high aldehyde addition and high protein solubility have a significant positive contribution to the adsorption rate, while turbidity and some conformational changes show a negative correlation. In the subplot with bitterness as the target variable, C_Aldehyde, thiol content, protein solubility, and surface hydrophobicity were the main contributing indicators. High aldehyde levels and high hydrophobic exposure were usually associated with increased bitterness, while moderate protein solubility was beneficial in mitigating bitterness. In the subplot corresponding to saltiness, protein solubility, surface hydrophobicity, zeta potential, and particle size rank highest. Furthermore, in the umami model, surface hydrophobicity and protein solubility are the most important characteristics, followed by parameters such as turbidity, C-Aldehyde, thiol content and particle size. High umami samples are mostly associated with higher surface hydrophobicity and suitable solubility.

[0060] In one specific embodiment, Figure 15 A schematic diagram of the electronic device structure used to implement the training and application of the random forest multi-output regression model is shown. like Figure 15 As shown, the electronic device may include a processor, a memory, an input unit, a display unit, and a communication unit. The processor is used to execute program instructions in the memory to preprocess, construct features, train models, cross-validate, and predict online data of collected physicochemical indicators, high-dimensional spectra, and rheological data. The memory is used to store raw experimental data, feature matrices, training model parameters, and generated graphical results; the input unit receives sample information, model parameters, and formulation optimization conditions. The display unit presents a comparison chart of actual and predicted values, a learning curve, a residual distribution chart, and a feature importance analysis chart; the communication unit enables data exchange and remote updates with external databases, experimental instruments, or servers. Working collaboratively, these units allow for model building, validation, and application on this device, providing reliable decision support for formulation design.

[0061] This application further discloses a low-sodium myofibrillar protein flavor prediction device, the device comprising the following modules, in the following order: Figure 17 As shown: Sample data acquisition module: used to acquire experimental data of multiple groups of low-sodium protein gels under different compound salt ratios, including protein conformation, physicochemical properties and flavor indicators; Aldehyde adsorption simulation system determination module: By combining low-sodium myofibrillar protein, potassium salt, κ-carrageenan and representative aldehyde flavor molecules, a protein-aldehyde-polysaccharide interaction system is constructed to characterize the sample in multiple dimensions and obtain data on protein secondary / tertiary structure, surface hydrophobicity, aldehyde binding rate and taste characteristics. Data processing module: preprocesses the collected high-dimensional spectral, physicochemical, and taste data, including standardization, feature extraction, random Fourier mapping, and feature selection, and combines data augmentation to form a comprehensive input feature matrix; Flavor prediction model construction module: Based on the processed input features, construct a multi-output random forest regression model or other multi-objective regression model, jointly train indicators such as protein solubility, turbidity, thiol content, rheological parameters and taste intensity, and improve the model's generalization ability through cross-validation and hyperparameter optimization; Flavor Prediction Module: Input the new formulation of the low-sodium protein gel to be predicted into the trained flavor prediction model, output the target flavor indicators and aldehyde binding rate, and provide suggestions for formulation optimization and process adjustment in combination with feature importance analysis.

[0062] The specific workflow of each module of the above-mentioned device can be implemented by referring to the corresponding steps in the method embodiments of this application.

[0063] This application further discloses an electronic device, including a memory and at least one processor, wherein the memory stores instructions, and the processor invokes the instructions to execute the steps of the low-sodium MP flavor prediction method described in this application.

[0064] This application also provides a computer-readable storage medium storing instructions that, when executed, enable a computer to perform all or part of the steps of the flavor prediction method of this application. The storage medium may be a USB flash drive, portable hard drive, ROM, RAM, magnetic disk, or optical disk, etc.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Those skilled in the art can modify or make equivalent substitutions to the embodiments without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A low-sodium myofibrillar protein flavor prediction method and apparatus, characterized by, The method includes: Multiple sample data are acquired, wherein the sample data is obtained based on a low-sodium system with a number of targets, and the sample data includes high-dimensional data of the first target type and conformational physicochemical data of the second target type; The aldehyde molecule adsorption simulation system corresponding to the sample data is determined. The aldehyde molecule adsorption simulation system is obtained by combining low sodium myofibrillar protein, potassium salt as a sodium salt substitute, κ-carrageenan as a texture modifier, and one or more representative aldehyde flavor molecules. The sample data is characterized in multiple dimensions to obtain the protein conformation data of the aldehyde molecule adsorption simulation system. The protein conformation data is preprocessed and features are extracted to construct input feature variables, and the binding rate data and the taste feature data are defined as output target variables. Based on the input feature variables and the output target variables, a flavor prediction model is constructed and trained to establish a quantitative relationship between the input feature variables and the output target variables; The target low-sodium formulation to be predicted is input into the flavor prediction model for flavor prediction, and the flavor prediction result of the target low-sodium formulation to be predicted is obtained.

2. The low-sodium myofibrillar protein flavor prediction method of claim 1, wherein, The preprocessing and feature extraction of the protein conformation data includes: To address the dimensional differences in multi-source heterogeneous data such as spectral data and rheological properties, standard deviation scaling is employed to convert each feature into a unified scale with a mean of 0 and a variance of 1, thereby obtaining spectral modal data and rheological modal data. The random Fourier features (RFF) are used to map each modal data to a high-dimensional space, and the nonlinear structure is approximated by random projection and sine and cosine transform. RFF high-dimensional mapping is performed on each modality data separately, and evaluation and ranking are performed within the modality based on feature variance or model importance index to achieve intra-modal feature dimensionality reduction; In the high-dimensional feature set of each modality, the top 10 features are selected as the key features of the modality based on their feature contribution. The key features of the number of targets obtained from each modality screening are combined according to a predetermined rule to form a subset of target features, which are used as input features for the prediction model.

3. The low-sodium myofibrillar protein flavor prediction method of claim 1, wherein, The process of constructing and training the flavor prediction model based on the input feature variables and the output target variable includes: Before model training, Gaussian noise is injected into the input feature variables. Gaussian noise with zero mean and a preset variance is added to the input features to generate multiple sets of training samples with random perturbation. Adjustable noise variance parameters can be set based on the statistical characteristics of different feature dimensions or different modal data. The training samples, which have been treated with Gaussian noise injection, are used together with the original training samples to train the random forest model. This allows each decision tree to learn from both the perturbation data and the original data during the modeling process, resulting in a random forest flavor prediction model.

4. The method for predicting the flavor of low-sodium myofibrillar proteins according to claim 1, characterized in that, The process of constructing and training the flavor prediction model based on the input feature variables and the output target variable includes: Kernel approximation was applied to κ-carrageenan-related indicators, aldehyde indicators and other physicochemical characteristics. RFF was introduced to extract features from the high-dimensional data of the first target species. The RFF and the conformational physicochemical data of the second target species were used as input to the tree model. In the subsequent interpretability analysis, all input features were included in the contribution calculation. An interpretability analysis module based on SHAP is set up to represent the tree model's predictions of taste attributes or aldehyde binding capacity as a linear summation of the input feature contributions: ； in, It is the true average value of all data from the experiment. The SHAP value of the i-th input feature is used to quantify the marginal contribution effect of each input feature to the prediction result. By statistically analyzing the SHAP values ​​and absolute values ​​of each input feature, the global importance ranking and main interaction features of κ-carrageenan-related features, aldehyde features, and RFF components on taste attributes and aldehyde binding capacity are obtained, which can be used to identify the key influencing factors and action patterns of the interaction between κ-carrageenan and aldehydes. Based on the marginal contribution and importance bias, the flavor prediction model is constructed and determined.

5. The method for predicting the flavor of low-sodium myofibrillar proteins according to claim 1, characterized in that, The method further includes: The input features of low-sodium myofibrillar protein conformation are mapped to the corresponding output prediction results to generate a baseline value. A histogram or stacked plot of SHAP values ​​is provided for reference, which is used to present the positive or negative contribution of each input feature to flavor attributes or aldehyde binding capacity. Two-dimensional or three-dimensional scatter plots or heatmaps were used to present the composition and aldehyde binding characteristics of low-sodium formulations to determine the combined effects of different protein conformational features on taste intensity and flavor retention, including feature interaction effects. The calculation formula is: ； in, Considered as a characteristic of collaborative consideration The model output, and The contribution of each feature, This is the baseline output.

6. The method for predicting the flavor of low-sodium myofibrillar proteins according to claim 1, characterized in that, The method further includes: Leave-one-out cross-validation is used for external validation. In each iteration, one system is selected from the low sodium systems of the target number as an independent external test set, and the remaining systems are used as the training set. The model hyperparameters are determined within the training set through 5-fold cross-validation. Each predicted value is calculated by resampling through quantile regression forest; Perform permutation tests to evaluate the statistical significance of the association between the obtained features and the target variable, combine Monte Carlo methods to obtain sensitivity data, and evaluate the stability data of the model output by introducing random perturbations into the input features. Construct an ensemble of n random forest sub-models, each trained on a different feature subset or random seed, and integrate the prediction results of the sub-models using a weighted voting method. , to obtain the predicted value : ； Weight The performance metrics are determined based on the performance metrics of each sub-model on the validation set, and these metrics are calculated as follows: ； in, This represents the physicochemical true value of the i-th sample in the dataset; This represents the model's prediction for the i-th sample; The represents the average of all true values ​​in the dataset; m represents the total number of samples. During the training phase of the RF model, the mean squared error (MSE) is used as the loss function for the regression task to determine the deviation between the predicted and true values. The calculation method is as follows: ； in, The true value is obtained from physical and chemical experiments. is the target prediction result, and m is the number of training data samples.

7. The method for predicting the flavor of low-sodium myofibrillar proteins according to claim 1, characterized in that, The method further includes: After standardizing the high-dimensional modal data of the first target category, the RFF method is used to map each high-dimensional modal data to a 256-dimensional random feature space by configuring a radial basis function sampler with a γ value of 1.

0. The importance of the 256-dimensional RFF features is evaluated and screened. When the number of modal samples is ≥5, mutual information is used to calculate the contribution of each feature to the output target variable. When the number of modal samples is <5, unsupervised scoring is performed based on feature variance to screen the RFF feature components with the number of target thresholds. The conformational physicochemical data of the second target category are horizontally concatenated with the selected RFF features in the sample dimension to form a fusion feature matrix; In the data augmentation process based on Gaussian noise, a small amplitude Gaussian noise is added to the fusion feature matrix and the corresponding output target variable, and the augmented samples of the target multiple are repeatedly generated and vertically concatenated with the original samples to construct the augmented training set; A flavor prediction model is trained based on the expanded training set.

8. The method for predicting the flavor of low-sodium myofibrillar proteins according to claim 7, characterized in that, The method also includes training and evaluating a multi-objective regression model, including: A multi-output regression framework is adopted, and a corresponding random forest regression sub-model is constructed and trained for each output target variable. Five-fold cross-validation was used to evaluate the generalization performance of each random forest regression sub-model; The coefficient of determination and root mean square error are used as performance indicators to quantitatively evaluate the prediction effect of each output target variable.

9. The method for predicting the flavor of low-sodium myofibrillar proteins according to claim 3, characterized in that, The method further includes: The process parameter combination of the target formulation is received, and the random forest model is called to output the saltiness intensity, umami intensity, bitterness intensity and aldehyde binding rate to obtain the virtual screening results of the target formulation. The process parameter combination includes the amount of κ-carrageenan added, the amount of NaCl, the amount of KCl and the ionic strength. Summarize the node splitting indices of each decision tree in the random forest model, quantify and rank the relative contributions of each input variable to the prediction results; The virtual screening results and the relative contribution ranking are comprehensively analyzed to generate a visual analysis result that shows the relationship between formula parameters, protein conformation and flavor index. Based on the user-defined multi-objective flavor optimization function, a search is performed within the preset process parameter space to output the target process parameter combination that meets the optimization conditions.

10. A low-sodium myofibrillar protein flavor prediction device, characterized in that, The device includes: The sample data acquisition module is used to acquire multiple sample data, which are obtained based on a low-sodium system with a target number of targets, including high-dimensional data of the first target type and conformational physicochemical data of the second target type. The aldehyde molecule adsorption simulation system determination module is used to construct the aldehyde molecule adsorption simulation system corresponding to the sample data. It obtains protein conformation data by simulating the binding of low sodium myofibrillar protein, potassium salt, κ-carrageenan and one or more aldehyde flavor molecules. The protein conformation data includes secondary structure ratio, surface hydrophobicity, aldehyde binding rate data and taste characteristic data. The data processing module is used to preprocess and extract features from the protein conformation data, construct input feature variables, and use the binding rate data and taste feature data as output target variables. The flavor prediction model building module is used to build and train a flavor prediction model based on the input feature variables and the output target variables. The flavor prediction module is used to input the target low-sodium formula to be predicted into the flavor prediction model and output the corresponding flavor prediction results.

Images