Perfume sensory characteristic value prediction method and perfume formula substitution method and system

Abstract

The invention provides a perfume sensory characteristic value prediction method and a perfume formula substitution method and system, which are applied to the technical field of perfume essences, and the method comprises the following steps: collecting perfume molecules and corresponding sensory characteristic scores; the perfume molecules are converted into a graph structure, nodes represent atoms, and edges represent chemical bonds; constructing a perfume sensory evaluation database for storing graph structure information and sensory feature scores of each perfume molecule; constructing a perfume sensory feature regression model based on the sensory evaluation database; the sensory characteristic value of each perfume is predicted by using the perfume sensory characteristic regression model, so that the accuracy and efficiency of perfume sensory characteristic prediction are remarkably improved, an effective tool is provided for creating a new perfume combination, the innovation ability and market competitiveness of the perfume industry are enhanced, and the cost is reduced.

Description

Technical Field

[0001] The present application relates to the technical field of flavors and fragrances, and in particular to a method for predicting flavor sensory characteristic values ​​and a flavor formula replacement method and system. Background Art

[0002] Dairy products (such as milk, butter, cheese and yogurt) are an important part of the modern diet. They not only provide high-quality protein and key nutrients, but are also widely welcomed by consumers for their unique flavor. In order to improve the sensory quality of dairy products, milk flavoring agents are widely used as key ingredients in the dairy industry. These flavoring agents increase the attractiveness of products and enhance consumer satisfaction by giving foods a unique creamy aroma. As people become more and more aware of food safety, the potential health risks of some traditional flavoring ingredients have begun to attract attention, which has brought new challenges to the food industry in selecting flavoring agents.

[0003] Diacetyl (DA) is a core ingredient in dairy flavors and is widely used in dairy products, baking, and snacks to provide a creamy aroma. Although DA is effective in flavor enhancement, recent studies have shown that long-term exposure to DA vapor may lead to serious respiratory diseases such as bronchiolitis obliterans ("popcorn lung"). Several countries are considering stricter restrictions on the use of DA, and may even impose a total ban on it.

[0004] Therefore, how to find safer alternatives while maintaining the sensory characteristics of milk flavoring agents has become an urgent problem to be solved in the food industry. Summary of the invention

[0005] In view of this, the embodiments of this specification provide a method for predicting the sensory characteristic values ​​of spices and a method and system for replacing spices formula, which are applied to the field of spices and flavors technology, and predict the sensory characteristic values ​​of spices by analyzing the relationship between the structural characteristics and sensory characteristics of spice molecules.

[0006] The embodiments of this specification provide the following technical solutions:

[0007] The present specification provides a method for predicting sensory characteristic values ​​of spices based on machine learning, comprising:

[0008] Collect flavor molecules and corresponding sensory characteristic scores;

[0009] Converting the fragrance molecule into a graph structure, wherein nodes represent atoms and edges represent chemical bonds;

[0010] Construct a flavor sensory evaluation database to store the graph structure information and sensory characteristic scores of each flavor molecule;

[0011] Based on the sensory evaluation database, a regression model of flavor sensory characteristics is constructed;

[0012] The sensory characteristic regression model of spices is used to predict the sensory characteristic value of each spice.

[0013] The present specification also provides a method for replacing a fragrance formula, comprising:

[0014] Predicting the sensory characteristic value of each flavor by using any of the above-mentioned flavor sensory characteristic value prediction methods based on machine learning;

[0015] Generate an initial population of spice formulas according to the sensory characteristic values ​​of the spices, wherein one chromosome in the initial population represents one spice formula; one gene represents one spice;

[0016] Based on the similarity between the sensory characteristic values ​​of the spices and the target aroma characteristics, the fitness function is set, and the formula is:

[0017] Where x and y are the molecular fingerprint vectors of two molecules; |x∩y| represents the intersection of the fingerprint vectors; |x| and |y| represent the number of non-zero elements of the two vectors respectively;

[0018] According to the score of the fitness function, the spice formula with high fitness is selected;

[0019] Introducing new flavor ingredients or changing the proportion of existing ingredients in the highly adaptable flavor formula;

[0020] The selection, crossover and mutation operations are iterated until the score value of the fitness function reaches a preset threshold, and the final spice formula is obtained.

[0021] The embodiment of this specification also provides a fragrance formula replacement system, the fragrance formula replacement system comprising: a fragrance sensory characteristic value prediction module and a formula generation module;

[0022] The flavor sensory characteristic value prediction module predicts the sensory characteristic value of each flavor using any of the flavor sensory characteristic value prediction methods based on machine learning described above;

[0023] The formula generation module is used to generate an initial population of a spice formula according to the sensory characteristic values ​​of the spice, wherein one chromosome in the initial population represents one spice formula; one gene represents one spice;

[0024] Based on the similarity between the sensory characteristic value of the spice and the target aroma characteristic, the fitness function is set, and the formula is:

[0025] Where x and y are the molecular fingerprint vectors of two molecules; |x∩y| represents the intersection of the fingerprint vectors; |x| and |y| represent the number of non-zero elements of the two vectors respectively;

[0026] According to the score of the fitness function, the spice formula with high fitness is selected;

[0027] Introducing new flavor ingredients or changing the proportion of existing ingredients in the highly adaptable flavor formula;

[0028] The selection, crossover and mutation operations are iterated until the score value of the fitness function reaches a preset threshold, and the final spice formula is obtained.

[0029] Compared with the prior art, the beneficial effects that can be achieved by at least one of the above technical solutions adopted in the embodiments of this specification include at least:

[0030] By collecting spice molecules and their corresponding sensory characteristic scores and converting these molecules into graph structures, a spice sensory evaluation database is established. Based on this database, a spice sensory characteristic regression model is constructed to predict the sensory characteristic values ​​of spices. This can efficiently predict the sensory characteristics of spices, which not only significantly improves the accuracy and efficiency of predicting the sensory characteristics of spices, but also provides an effective tool for creating new spice combinations, enhancing the innovation ability and market competitiveness of the spice industry and reducing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] Figure 1 It is the forward prediction and reverse optimization process of the SensoryGAN model;

[0033] Figure 2 is a data graph of stimulus prompts and molecular characteristics used for analysis in the experiment of this application; wherein, Figure 2 (A) Diagram of the experimental steps for sensory evaluation for each olfactory stimulus, N represents the total number of responses of all stimuli and subjects; Figure 2 (B) and Figure 2 (C) Molecular weight and molecular complexity diagram of the molecules used in this application, respectively; Figure 2 (D) is a graph of molecular weight and molecular complexity analyzed by chemical function;

[0034] Figure 3 This is a diagram of the construction process of the GCN model in this application;

[0035] Figure 4 It is the graph of the sensory evaluation data analysis results in this application; among them, Figure 4 (A) is the hierarchical clustering analysis graph; Figure 4 (B) is the multidimensional scaling analysis (MDS) graph, and different ellipses represent different odor feature categories; Figure 4 (C) is the box plot of sensory scores at different concentrations, comparing the score distributions and significant differences of odor categories (* represents the significance level); Figure 4 (D) is the graph of the connection strength of sensory characteristics, and the dark and light edges represent positive and negative correlations respectively; Figure 4 (E) is the heat map of the correlation of sensory characteristics at 10% concentration; Figure 4 (F) is the heat map of the correlation of sensory characteristics at 1% concentration;

[0036] Figure 5 It is the comparison result graph of MSE, RMSE and MAE of three models at different concentrations;

[0037] Figure 6 It is the radar chart comparison graph of five optimized formulas and target sensory characteristics;

[0038] Figure 7 It is the EEG analysis result graph of the optimal formula and DA at different concentrations; among them: Figure 7 (A) The topographic maps of electroencephalogram activities of four brain waves (Theta, Alpha, Beta, Gamma) under different concentrations of essence (20%-100%) and DA conditions, the dark areas indicate weaker activities, and the light areas indicate stronger activities; Figure 7 (B) The energy differences of different concentrations of essence (20%-100%) and DA in different frequency bands, * indicates significance (0.01 < p ≤ 0.05), ** indicates highly significant (0.001 < p ≤ 0.01), *** indicates extremely significant (p ≤ 0.001), and no label indicates no difference (p > 0.05); Figure 7 (C) Comparing the CSD of the 80% concentration essence formula and DA in the Alpha band; Figure 7 (D) Comparing the CSD of the 100% concentration essence formula and DA in the Alpha band. Detailed implementation manners

[0039] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0040] The following describes the implementation methods of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the following embodiments and the features in the embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work belong to the scope of protection of the present application.

[0041] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present application, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspect described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0042] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. The drawings only show components related to the present application rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0043] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the examples can be practiced without these specific details.

[0044] In order to ensure food safety while meeting sensory quality requirements, the industry is exploring a variety of flavor alternatives, including synthetic flavors, plant extracts, and nanotechnology (such as patents CN106565640A, CN118844594A, or CN118725981A), in order to achieve a balance between safety and sensory characteristics.

[0045] However, these alternatives have some shortcomings, such as poor stability, high cost, lack of environmental friendliness, and difficulty in accurately replicating specific aroma characteristics, such as the creamy aroma of diacetyl (DA). Therefore, the existing technology cannot simultaneously meet the requirements of food safety, economy and sensory characteristics, which restricts the development and application of safe flavor substitutes in the food industry.

[0046] In view of this, the inventors have found through research and improvement exploration that although existing machine learning, such as Graph Convolutional Network (GCN), Decision Tree Regression (DTR) and K-Nearest Neighbors (KNN), can capture the complex relationship between molecular structure and odor characteristics through big data analysis, these traditional machine learning methods still have significant limitations in dealing with the complex nonlinear relationship between molecular structure and sensory characteristics, which are specifically manifested as follows:

[0047] Compounds with similar structures may have very different odor characteristics, while compounds with large structural differences may exhibit similar aroma characteristics. For example, steroidal musk and indole musk have different structures, but both have the characteristic aroma of musk, which makes it difficult for current schemes to accurately capture the complex relationship between molecular structural diversity and sensory characteristics, thus affecting the accuracy of fragrance replacement design.

[0048] In addition, chiral isomers have significant differences in aroma characteristics, which further limits the performance of traditional methods in predicting the sensory characteristics of flavors.

[0049] Based on this, the present specification proposes a method for predicting the sensory characteristic value of a flavor based on machine learning, such as Figure 1 As shown in the figure, the overall idea is: collect the chemical structures of spice molecules and the corresponding sensory characteristic scores, convert the chemical structures of spice molecules into graph structures and store them in a database, construct a spice sensory characteristic regression model based on the graph structure information and sensory characteristic scores of the spice molecules in the database, and then use the constructed spice sensory characteristic regression model to predict the sensory characteristic values ​​of the spice, thereby achieving efficient prediction of the sensory characteristics of spice molecules, which effectively solves the limitations of the existing solutions in predicting the sensory characteristics of spices and designing formulas through molecular fingerprints or physicochemical parameters.

[0050] The technical solutions provided by various embodiments of the present application are described below in conjunction with the accompanying drawings.

[0051] The present specification provides a method for predicting sensory characteristic values ​​of spices based on machine learning, comprising:

[0052] Collect flavor molecules and corresponding sensory characteristic scores;

[0053] Converting the fragrance molecule into a graph structure, wherein nodes represent atoms and edges represent chemical bonds;

[0054] Construct a flavor sensory evaluation database to store the graph structure information and sensory characteristic scores of each flavor molecule;

[0055] Based on the sensory evaluation database, a regression model of flavor sensory characteristics is constructed;

[0056] The sensory characteristic value of each flavor is predicted using the flavor sensory characteristic regression model.

[0057] Specifically, the chemical structure information of the fragrance molecules is collected, which includes details such as the types of atoms in the molecules, the chemical bonds between atoms, and the spatial arrangement of the molecules. At the same time, the sensory characteristic scores of the fragrance molecules are collected. These scores are the results of quantitative evaluation of the sensory attributes of the fragrance molecules, such as odor intensity, pleasantness, familiarity, etc.

[0058] The spice molecules are converted into a graph structure, where the nodes (vertices) represent atoms and the edges represent the chemical bonds between atoms, thus capturing the structural information of the molecules, including the types of atoms and how they are connected.

[0059] Then, a flavor sensory evaluation database is constructed to store the graph structure information and sensory characteristic score of each flavor molecule, and the sensory characteristic score in the flavor sensory evaluation database is associated with the corresponding flavor molecule graph structure data. In this way, the graph structure of each flavor molecule matches its sensory characteristic score, so that the relationship between the structural characteristics of the flavor molecule and its sensory characteristics can be analyzed.

[0060] Use the associated data to train machine learning models, such as graph convolutional networks (GCNs), to learn the complex relationship between the molecular structure of spices and their sensory characteristics, thereby predicting the sensory characteristic values ​​of each spice.

[0061] Through the above method, the chemical structure information of spice molecules can be effectively associated with their sensory characteristic scores, and analyzed and predicted through machine learning models, providing a scientific basis and innovative tools for the spice industry.

[0062] In some embodiments, the step of constructing a regression model for flavor sensory characteristics comprises:

[0063] Use graph convolutional network layers to perform convolution operations on molecular graphs and aggregate node neighbor information layer by layer;

[0064] The node feature update formula of each layer of graph convolutional network is:

[0065]

[0066] in, is the node feature of the l+1th layer, σ is the activation function, N(i) is the neighbor set of the node, d i and d j is the degree of the node, W (l) is the weight matrix.

[0067] In some embodiments, the fragrance molecules are collected by the following steps:

[0068] According to the preset aroma characteristic descriptors of the target aroma, fragrances matching the aroma characteristic descriptors are screened out to obtain a first fragrance set;

[0069] Select spices that can represent different aroma characteristics from the first spice set to obtain a second spice set.

[0070] Specifically, according to the requirements of the target aroma, a series of aroma characteristic descriptors are determined.

[0071] For example: for DA aroma, 23 evaluation criteria for aroma characteristics are defined, covering odor intensity, pleasantness, familiarity, edibility, baking, sweetness, fruitiness, sourness, woodiness, floral, grassy, ​​garlic, spicy, fishy, ​​burnt, warm, cold, musky, sweaty, ammonia / urine, rotten, chemical and rancid, etc., describing the characteristics of the expected aroma in detail.

[0072] Then, the aroma characteristic descriptors are used to conduct a preliminary screening of the spices, for example, using the keywords containing the spices molecules of the target aroma characteristics in the chemical database to obtain a collection of spices with multiple similar aroma characteristics.

[0073] According to the diversity and representativeness of aroma, spices that can represent different aroma characteristics are selected. The aroma intensity and uniqueness of spices and their performance in sensory evaluation can be considered to obtain the second spice set.

[0074] In some embodiments, the fragrance molecules are collected by the following steps:

[0075] According to the preset aroma characteristic descriptors of the target aroma, fragrances matching the aroma characteristic descriptors are screened out to obtain a first fragrance set;

[0076] Select spices that can represent different aroma characteristics from the first spice set to obtain a second spice set.

[0077] Specifically, words can be collected from the documents. This application takes DA as an example, collects literature on DA aroma, and extracts high-frequency keywords.

[0078] Then, through multiple rounds of screening, we removed vague words or words with low correlation with DA, and finally determined 11 descriptive words, including: milk, cream, butter, cheese, cottage cheese, sweet, caramel, toasted, sour, butterscotch, and vanilla;

[0079] The above vocabulary is used to screen out an initial 1,395 flavor molecules in the flavor database (FlavorDB database), thereby obtaining the first flavor set.

[0080] Then, a toxicity screening was performed in the public compound database (PubChem database) to exclude potentially toxic molecules and screen out 125 spices with higher safety to obtain the second spice set.

[0081] The assessment contents include: pulmonary toxicity, hepatotoxicity, skin toxicity, inhalation toxicity, food toxicity and neurotoxicity.

[0082] In some embodiments, after the initial 1395 flavor molecules are screened, duplicate data are cleaned up, for example, in this embodiment, 674 duplicate names and 35 duplicate SMILES format records are deleted to ensure data uniqueness.

[0083] In some embodiments, the sensory characteristic scores corresponding to the flavor molecules are collected by the following steps:

[0084] Conduct sensory tests on spices of different concentrations and establish a spice sensory evaluation database, including:

[0085] dividing the flavor sensory evaluation database into different subsets according to flavor concentration;

[0086] Using the stratified random sampling method, each concentration subset is divided into a training set and a test set, the formula is:

[0087] Among them, n i is the number of samples after stratification, N i is the total number of samples in the stratum, N is the total number of samples, and n is the total number of samples.

[0088] Specifically, in combination with the descriptors determined in the above examples, compounds that can represent different aroma characteristics are selected from 125 safe compounds (for compound attribute analysis, see Figure 2B, C, D), such as vanillyl acetone, benzyl lactate, ethyl vanillate, (S)-(+)-2-heptanol, milk lactone, methyl dantilis, cocaaldehyde, lactic acid, 3-hexenoic acid, cis-4-hexenoic acid, methyl lactone, 2-ethyl-4-hydroxy-5-methyl-3(2H)-furanone (soy sauce ketone), 4-methyl-5-thiazoleethanol, acetoin, butyl butyrolactate, 2-methylpyrazine, 3-mercapto-2-pentanone, 2-acetyl-1-pyrroline, butyl dodecanol, 3-ethyl-2,5-dimethylpyrazine, butyrophenone, methyl cyclohexanecarboxylate, 2-butenoic acid hexyl ester, 4-ethylguaiacol, ethyl Vanillin acetate, 3-ethyl-2,5-dimethylpyrazine, 2,4-dimethylthiazole, 2-methoxy-3-methylpyrazine, myrtenol, propyl dodecalactone, 5-hydroxymethylfurfural, 4-(4-methoxyphenyl)-2-butanone, cis-3-hexenyl butyl ester, 2-methylthiophene, acetoin dimer, S-allyl-L-cysteine, star apple ester (strawberry ester), 2-methylquinoxaline, 2-methoxythiazole, 2-ethyl-5-methylpyrazine, 2,3-butanediol, tributyrin, methyl eugenol, L-butyl lactate, methylcyclopentenolone, 2-acetyl-3-5-dimethylpyrazine.

[0089] Sample dilution: dilute the selected flavor (compound) to 10% and 1% concentrations for subsequent sensory testing. The dilution calculation formula is as follows:

[0090] Among them, C f is the final concentration, C i is the initial concentration, V i and V f are the initial and final volumes;

[0091] Standardization of sensory descriptions: Describing standards for 23 aroma characteristics, including intensity, pleasantness, familiarity, edible, toasty, sweet, fruity, sour, woody, floral, grassy, ​​garlicky, spicy, fishy, ​​burnt, warm, cold, musky, sweaty, ammonia / urine, rotten, chemical, and rancid, to ensure consistency in evaluation;

[0092] Sensory test: 40 subjects were organized to conduct sensory evaluation on 47 flavor samples (at 10% and 1% concentrations) ( Figure 2 A), the 23 aroma characteristics of each sample were scored in the range of 0-100, and a total of 86,480 data were collected (see Figure 4 );

[0093] Database establishment: Integrate the collected data into a "milk flavor sensory evaluation database" as the basis for subsequent machine learning modeling.

[0094] Dataset concentration division: divided into 0.1 and 0.01 data sets according to the fragrance concentration, where:

[0095] In the 0.1 concentration dataset, the training set accounts for 60% and the test set accounts for 40%;

[0096] In the 0.01 concentration dataset, the training set accounts for 70% and the test set accounts for 30%;

[0097] Group randomization: A stratified random sampling method was used to ensure that each group of data was representative of the 23 aroma characteristics. The stratified sampling formula was:

[0098]

[0099] Among them, n i is the number of samples after stratification, N i is the total number of samples in the stratification, N is the total number of samples, and n is the total number of samples;

[0100] Verification and adjustment: Perform frequency distribution test on the divided data to ensure the consistency of the distribution of each feature between the training set and the test set to improve the generalization ability of the model.

[0101] In some embodiments, Min-Max normalization is used on the training data to scale the sensory feature scores to the range of 0-1, as follows:

[0102] Among them, x is the original data, x ′ To standardize data.

[0103] In some embodiments, Figure 3 As shown, the flavor sensory feature regression model is trained, including:

[0104] The graph convolutional network model is trained using the mean square error as the loss function, where the loss function formula is:

[0105] Where n represents the number of samples; y i Indicates the actual sensory characteristic value; represents the sensory feature value predicted by the model;

[0106] The Adam optimizer is used to adjust the learning rate of the graph convolutional network model parameters;

[0107] After each training cycle, the weight matrix is ​​updated according to the gradient of the loss function until the goal of minimizing the MSE loss is achieved.

[0108] Specifically, the weights of the model are continuously updated through multiple rounds of training (epochs) to improve the prediction accuracy.

[0109] In some embodiments, when training the flavor sensory feature regression model, the model performance is quantified by the mean square error (MSE) and R2 indicators in the test set, and the number of layers or learning rate hyperparameters of the GCN are adjusted by 5-fold cross validation;

[0110] Where R2 is defined as follows:

[0111] in, Represents the mean of the sensory characteristics.

[0112] In some embodiments, model parameters are adjusted based on the test set to optimize model performance.

[0113] Hyperparameter adjustment: Use 5-fold cross validation to test different combinations of number of layers, learning rate, batch size, etc. Verification formula:

[0114]

[0115] Error analysis: MSE and mean absolute error (MAE) are calculated on different concentration data sets to ensure the stability of the model under various conditions. The definition of MAE is as follows:

[0116]

[0117] Among them, y i is the actual sensory characteristic value, sensory feature values ​​predicted by the model;

[0118] Optimizer and activation function selection: Select the Adam optimizer and ReLU activation function to accelerate convergence and improve nonlinear expression capabilities, respectively.

[0119] This application evaluates the performance of three different machine learning models, including graph convolutional networks (GCN), K-nearest neighbor algorithms (KNN), and decision tree regression (DTR), to predict the sensory characteristics of spices at different concentrations. In order to measure the prediction accuracy of these models, the error generated by each model during the prediction process was calculated, and the three statistical indicators of mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) were used for comparison. The comparison results of these error indicators are shown in Figure 5 The results are presented in detail in , which provides us with an intuitive understanding of which model performs best under specific conditions. Through this comparison, we can better understand the advantages and disadvantages of different models in dealing with the task of predicting the sensory characteristics of spices, thus providing a basis for selecting the most appropriate prediction tool.

[0120] After careful comparative analysis, it was found that graph convolutional networks (GCN) performed best in predicting the sensory feature values ​​of spices.

[0121] GCN (graph convolutional network) is applied to the present invention, which has the following advantages:

[0122] Effectively capture the complex relationship between molecular structure and sensory characteristics: Traditional molecular fingerprint or chemical parameter methods are difficult to accurately describe the overall picture of the molecular structure, while GCN can fully represent the local and global topological characteristics of molecules through graph structures, thereby more accurately modeling the nonlinear relationship between molecular structure and sensory characteristics;

[0123] Enhance the model's ability to distinguish compounds with similar structures: GCN can distinguish molecules with similar structures but different smells by weighting the features of atoms and chemical bonds in the graph structure. This ability allows the model to maintain high accuracy and resolution when facing fragrances with similar structures.

[0124] Improve the recognition of similar odors of compounds with different structures: GCN's multi-layer convolution processing can still find potential structural feature associations even when the molecular structures are quite different, thereby identifying molecules with similar odor characteristics. This ability to recognize similar odors of heterogeneous molecules enhances the generalization of the model.

[0125] Supporting parallel prediction of multiple sensory features: GCN has the ability to process multiple sensory features in parallel. In the application of the present invention, aroma features of multiple dimensions (such as sweet aroma, fruity aroma, etc.) can be predicted at one time, which improves the prediction efficiency and functionality of the model.

[0126] Optimize computing resources and support efficient training: GCN can effectively utilize the sparsity of graph data, reduce computing overhead, and improve the training speed of the model. Compared with other deep learning models, GCN is more efficient when processing large-scale data, which helps the present invention to achieve real-time prediction and alternative formulation optimization under diversified data input.

[0127] Based on the same inventive concept, Figure 6 As shown, the present application also provides a fragrance formulation alternative method, including:

[0128] Predicting the sensory characteristic value of each flavor by using the flavor sensory characteristic value prediction method based on machine learning described in any one of the above embodiments;

[0129] Generate an initial population of spice formulas according to the sensory characteristic values ​​of the spices, wherein one chromosome in the initial population represents one spice formula; one gene represents one spice;

[0130] Based on the similarity between the sensory characteristic value of the spice and the target aroma characteristic, the fitness function is set, and the formula is:

[0131] Where x and y are the molecular fingerprint vectors of two molecules; |x∩y| represents the intersection of the fingerprint vectors (i.e., the number of common features); |x| and |y| represent the number of non-zero elements of the two vectors (i.e., the total number of features);

[0132] According to the score of the fitness function, the spice formula with high fitness is selected;

[0133] Introducing new flavor ingredients or changing the proportion of existing ingredients in the highly adaptable flavor formula;

[0134] The selection, crossover and mutation operations are iterated until the score value of the fitness function reaches a preset threshold, and the final spice formula is obtained.

[0135] Specifically, an initial population of spice combinations (1000) was generated, and molecular fingerprints were used to uniquely identify the population and capture molecular characteristics;

[0136] Based on the similarity between the aroma feature scores predicted by the GCN model and the target features, the fitness is calculated using the Tanimoto coefficient.

[0137] The tournament selection method is used to retain the combinations with high fitness. The mutation mechanism (randomly changing the molecular fingerprint) is introduced to improve the diversity of the population;

[0138] like Figure 6 As shown, after multiple rounds of iterations, until there is no significant change in fitness improvement, the spice formula with the best adaptability is finally screened out.

[0139] In some embodiments, the fragrance formula replacement method further comprises:

[0140] The three-point test method and electroencephalogram were used to verify the odor similarity between the alternative formula and the target aroma.

[0141] Specifically, the three-point test method was used to evaluate the similarity between the generated formula and the target aroma. Beta risk (Type II error) was used to evaluate the discrimination ability, and Pd (sensory discrimination ratio) was set to 40%.

[0142] like Figure 7 As shown, EEG data analysis is used to convert the EEG signal into frequency domain data by fast Fourier transform (FFT) to obtain the power spectral density (PSD) of each electrode. i EEG signal x i (t), and its frequency domain representation X is obtained through Fourier transform i(f). Then, the power value of a specific frequency band is extracted, and the average power spectral density (PSD) of each electrode in this frequency band is calculated. Next, the power value of each electrode is extended to the entire scalp surface by interpolation to generate a brain topography map of a specific frequency band to show the activity intensity of each area of ​​the scalp at a specific frequency, which is convenient for observing the frequency response characteristics of the brain under different olfactory conditions. The formula is as follows:

[0143]

[0144] Frequency band response comparison: Analyze the neurophysiological response of the formula in the Theta, Alpha, Beta, Gamma and other frequency bands to quantify the perceptual similarity;

[0145] CSD calculation: Calculate the CSD value of the formula and the target aroma in the key frequency band, evaluate the consistency of neural response, and finally verify the effect of formula substitution. Given two time series x[n] and y[n], their CSD is defined as:

[0146] P xy (e iω )=X(e iω )Y * (e iω )

[0147] Among them, X(e iω ) and Y(e iω ) are the Fourier transforms of x[n] and y[n], respectively, and Y * (e iω ) is Y(e iω ). In order to reduce the influence of noise, the segmented average method (Welch method) is used to segment the signal and take the CSD average of each segment.

[0148] Based on the same inventive concept, the present application also provides a fragrance formula replacement system, the fragrance formula replacement system comprising: a fragrance sensory characteristic value prediction module and a formula generation module;

[0149] The flavor sensory characteristic value prediction module uses the flavor sensory characteristic value prediction method based on machine learning described in any one of the above embodiments to predict the sensory characteristic value of each flavor;

[0150] The formula generation module is used to generate an initial population of a spice formula according to the sensory characteristic values ​​of the spice, wherein one chromosome in the initial population represents one spice formula; one gene represents one spice;

[0151] Based on the similarity between the sensory characteristic value of the spice and the target aroma characteristic, the fitness function is set, and the formula is:

[0152] Where x and y are the molecular fingerprint vectors of two molecules; |x∩y| represents the intersection of the fingerprint vectors; |x| and |y| represent the number of non-zero elements of the two vectors respectively;

[0153] According to the score of the fitness function, the spice formula with high fitness is selected;

[0154] Introducing new flavor ingredients or changing the proportion of existing ingredients in the highly adaptable flavor formula;

[0155] The selection, crossover and mutation operations are iterated until the score value of the fitness function reaches a preset threshold, and the final spice formula is obtained.

[0156] This application effectively solves the following problems:

[0157] It is difficult to fully capture the complex nonlinear relationship between molecular structure and sensory experience;

[0158] Compounds with similar structures may exhibit different odor characteristics, affecting the prediction accuracy;

[0159] Compounds with large structural differences sometimes have similar odor characteristics, which makes it difficult to accurately identify them using traditional methods;

[0160] There is a lack of effective ways to apply sensory profile prediction results to flavor formulation design.

[0161] In this specification, the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the embodiments described later, the description is relatively simple, and the relevant parts can be referred to the partial description of the previous embodiments.

[0162] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A method for predicting sensory characteristic values ​​of spices based on machine learning, characterized in that: include: Collect flavor molecules and corresponding sensory characteristic scores; Converting the fragrance molecule into a graph structure, wherein nodes represent atoms and edges represent chemical bonds; Construct a flavor sensory evaluation database to store the graph structure information and sensory characteristic scores of each flavor molecule; Based on the sensory evaluation database, a regression model of flavor sensory characteristics is constructed; The sensory characteristic regression model of spices is used to predict the sensory characteristic value of each spice.

2. The method for predicting flavor sensory characteristic values ​​based on machine learning according to claim 1, characterized in that: The method of constructing a regression model for flavor sensory characteristics comprises: Use graph convolutional network layers to perform convolution operations on molecular graphs and aggregate node neighbor information layer by layer; The node feature update formula of each layer of graph convolutional network is: in, is the node feature of the l+1th layer, σ is the activation function, N(i) is the neighbor set of the node, d i and d j is the degree of the node, W (l) is the weight matrix.

3. The method for predicting flavor sensory characteristic values ​​based on machine learning according to claim 1, characterized in that: The fragrance molecules were collected by the following steps: According to the preset aroma characteristic descriptors of the target aroma, fragrances matching the aroma characteristic descriptors are screened out to obtain a first fragrance set; Select spices that can represent different aroma characteristics from the first spice set to obtain a second spice set.

4. The method for predicting flavor sensory characteristic values ​​based on machine learning according to claim 3, characterized in that: When you get the first spice set, it includes: Collect information about aromas described in literature, extract keywords and perform multiple rounds of screening, and preset aroma feature descriptors; Using the aroma characteristic descriptors, screening out initial flavor molecules in a flavor database; The initial fragrance molecules are screened for toxicity in a public compound database to exclude molecules with potential toxicity, thereby obtaining a first fragrance set.

5. The method for predicting flavor sensory characteristic values ​​based on machine learning according to claim 1, characterized in that: Collecting sensory characteristic scores corresponding to flavor molecules: performing sensory tests on flavors of different concentrations to establish a flavor sensory evaluation database, including: dividing the flavor sensory evaluation database into different subsets according to flavor concentration; using a stratified random sampling method to divide each concentration subset into a training set and a test set, the formula is: Among them, n i is the number of samples after stratification, N i is the total number of samples in the stratification, N is the total number of samples, and n is the total number of samples; And / or, scale the sensory characteristic scores to the 0-1 interval using the formula: Among them, x is the original data, x ′ To standardize data.

6. The method for predicting flavor sensory characteristic values ​​based on machine learning according to claim 1, characterized in that: Training the flavor sensory feature regression model includes: The graph convolutional network model is trained using the mean square error as the loss function, where the loss function formula is: Where n represents the number of samples; y i Indicates the actual sensory characteristic value; represents the sensory feature value predicted by the model; The Adam optimizer is used to adjust the learning rate of the graph convolutional network model parameters; After each training cycle, the weight matrix is ​​updated according to the gradient of the loss function until the goal of minimizing the MSE loss is achieved.

7. The method for predicting flavor sensory characteristic values ​​based on machine learning according to claim 6, characterized in that: When training the regression model of flavor sensory features, the mean square error and R 2 The indicators quantify the model performance, and 5-fold cross-validation is used to adjust the number of layers or learning rate hyperparameters of GCN; Where R2 is defined as follows: in, Represents the mean of the sensory characteristics.

8. A method for replacing a spice formula, characterized in that: include: Predicting the sensory characteristic value of each flavor by using the flavor sensory characteristic value prediction method based on machine learning as described in any one of claims 1 to 7; Generate an initial population of spice formulas according to the sensory characteristic values ​​of the spices, wherein one chromosome in the initial population represents one spice formula; one gene represents one spice; Based on the similarity between the sensory characteristic value of the spice and the target aroma characteristic, the fitness function is set, and the formula is: Where x and y are the molecular fingerprint vectors of two molecules; |x∩y| represents the intersection of the fingerprint vectors; |x| and |y| represent the number of non-zero elements of the two vectors respectively; According to the score of the fitness function, the spice formula with high fitness is selected; Introducing new flavor ingredients or changing the proportions of existing ingredients in the highly adaptable flavor formula; The selection, crossover and mutation operations are iterated until the score value of the fitness function reaches a preset threshold, and the final spice formula is obtained.

9. The method for replacing a flavor formula according to claim 8, characterized in that: The spice formula replacement method also includes: The three-point test method and electroencephalogram were used to verify the odor similarity between the alternative formula and the target aroma.

10. A fragrance formula replacement system, characterized in that: The fragrance formula replacement system comprises: a fragrance sensory characteristic value prediction module and a formula generation module; The flavor sensory characteristic value prediction module predicts the sensory characteristic value of each flavor using the flavor sensory characteristic value prediction method based on machine learning as described in any one of claims 1 to 7; The formula generation module is used to generate an initial population of a spice formula according to the sensory characteristic values ​​of the spice, wherein one chromosome in the initial population represents one spice formula; one gene represents one spice; Based on the similarity between the sensory characteristic value of the spice and the target aroma characteristic, the fitness function is set, and the formula is: Where x and y are the molecular fingerprint vectors of two molecules; |x∩y| represents the intersection of the fingerprint vectors; |x| and |y| represent the number of non-zero elements of the two vectors respectively; According to the score of the fitness function, the spice formula with high fitness is selected; Introducing new flavor ingredients or changing the proportion of existing ingredients in the highly adaptable flavor formula; The selection, crossover and mutation operations are iterated until the score value of the fitness function reaches a preset threshold, and the final spice formula is obtained.

Images