A method for constructing a model for identifying a focal sweet flavor molecule odor threshold

Abstract

The application discloses a kind of for identifying the construction method of caramelly sweet spice molecule smell threshold model, by combining quantum chemistry calculation and machine learning, grasp the mechanism of the smell of smell molecule on atomic scale, construct the smell intensity fitting model between molecular structure and smell intensity, utilize the smell intensity fitting model to predict the threshold of different caramelly sweet spice molecules.The application creatively realizes the accurate prediction of molecular smell threshold based on the characteristics of functional groups in molecular structure, provides a reliable tool for odor optimization in cigarette production, effectively improves the sensory quality of caramelly sweet characteristics in Chinese-style cigarettes.

Description

Technical Field

[0001] The present invention relates to a method for constructing an aroma threshold model for identifying burnt-sweet aroma spice molecules. Background Art

[0002] Sweet, burnt aroma is one of the primary notes in Chinese cigarettes. Improving this characteristic smoke quality can contribute to the sensory quality of Chinese cigarettes. Smoke contains a variety of compounds, which interact with each other to influence the flavor of the cigarette. The aroma of a mixture is not simply a summation of the original aromas, but rather a "combination of forces" effect. Therefore, understanding the aroma-producing mechanisms of these characteristic compounds and further studying the impact of changes in aroma groups and molecular charge distribution caused by interactions between aroma molecules in solvents on aroma production are crucial for regulating cigarette aroma.

[0003] With the improvement in the accuracy of quantum chemistry and density functional theory methods and the improvement in computing power, people can use theoretical calculations to obtain accurate molecular structure information, charge density distribution, electronic structure information, etc. They can also obtain changes in geometric structure, electron density distribution, and energy level structure under interactions with other molecules, solvents, etc., and can also obtain some spectroscopic information that can be compared with experimental observations through simulation, such as infrared, Raman and other vibrations. Artificial intelligence machine learning models can analyze and process new data much faster than humans, allowing people to connect information at the atomic scale with observable and perceptible information. They help people extract the key factors that determine a certain property from the complex information, and then predict the properties of molecules or materials. At present, there are no reports on the use of quantitative calculations combined with machine learning to build a model to predict the aroma threshold of burnt sweet-scented molecules. Summary of the Invention

[0004] In order to avoid the deficiencies of the above-mentioned prior art, the present invention provides a method for constructing an aroma threshold model for identifying the molecules of caramel-sweet fragrance, in order to reveal the relationship between changes in molecular structure and changes in aroma intensity, and provide a theoretical basis for the accurate prediction of molecular aroma thresholds.

[0005] The present invention adopts the following technical solutions to solve the technical problems:

[0006] The method for constructing a fragrance threshold model for identifying burnt sweet fragrance molecules of the present invention is to grasp the fragrance-producing mechanism of fragrance molecules at the atomic scale by combining quantum chemical calculations and machine learning, construct a fragrance intensity fitting model between molecular structure and fragrance intensity, and use the fragrance intensity fitting model to predict the thresholds of different burnt sweet fragrance molecules.

[0007] The method for constructing a fragrance threshold model for identifying burnt sweet fragrance molecules of the present invention comprises the following steps: first, obtaining structural information of the burnt sweet fragrance molecules by using a CAS number or a SMILES character string through a CASSciFinder database or an RDKit software package; optimizing the molecular structure by using quantum computing software Gaussian based on the structural information to obtain an optimized molecular structure, and obtaining an infrared spectrum of the optimized molecular structure; then, obtaining structure-related descriptors by performing piecewise integration processing on the infrared spectrum; then, extracting key feature descriptors for the structure-related descriptors by using a machine learning tool SISSO, constructing a model of the relationship between functional groups and thresholds by using the key feature descriptors as a fragrance intensity fitting model; and predicting the thresholds of different burnt sweet fragrance molecules by using the fragrance intensity fitting model.

[0008] The method for constructing an aroma threshold model for identifying burnt sweet aroma spice molecules of the present invention comprises the following steps:

[0009] Step 1: Collecting information on caramel-sweet flavor molecules with a threshold value; including: obtaining the CAS number or SMIELS string of the molecule by searching; obtaining a structure file by downloading it from the CAS-SciFinder database based on the CAS number, or converting the SMILES string to a structure file using the RDKit software package in Python to obtain the structure file; completing the collection of information on caramel-sweet flavor molecules with a threshold value;

[0010] Step 2: Optimize the molecular structure and predict its infrared spectrum using the quantum computing software Gaussian. This includes obtaining the experimental infrared spectrum of the caramel flavoring molecule from the CAS-SciFinder database. In Gaussian, optimize the structure, calculate the infrared spectrum, and perform frequency correction using different functionals and basis sets to obtain the infrared spectrum of the optimized molecular structure. Compare the infrared spectrum of the optimized molecular structure with the experimental infrared spectrum, and select the functional and basis set with the smallest error with the characteristic peaks for subsequent calculation of the infrared spectrum information.

[0011] Step 3: Use Python to analyze the infrared spectrum from 500-4000cm -1 The frequency is processed by 50 segment integration, each segment of the integration corresponds to a descriptor, and a total of 50 descriptor features are obtained, which are recorded as Feature (N) (N=1-50);

[0012] Step 4: Screening and extracting key Features from 50 descriptors Features by machine learning tool SISSO, and constructing a model of functional group and threshold value relationship as a fragrance intensity fitting model by the key Features;

[0013] Step 5: predicting the threshold value of different caramelic flavor molecules by using the fragrance intensity fitting model.

[0014] The step 4 in the method for constructing a model for identifying the fragrance threshold value of caramelic flavor molecules of the present application is carried out according to the following process:

[0015] Step 4.1: The collected threshold values of caramelic flavor molecules are subjected to negative logarithm processing to obtain threshold negative logarithms, and the threshold negative logarithms and the 50 descriptors Features of the infrared spectrum of each molecule obtained in step 3 are input into SISSO; the allowed operator selection in the model is: (+) (-) (*) ( / ) (exp) (exp-) (^-1) (^2) (^3) (sqrt) (cbrt) (log); the size selection of descriptors / model is 2; and finally the key Features are Feature {7,8,10,19,20,38,46} ;

[0016] The Feature {7,8,10,19,20,38,46} corresponds to the frequency and characteristic functional group as follows:

[0017] Feature 7,8 -904cm-1: C-H bending vibration on olefin or aromatic ring;

[0018] Feature 10 -1084cm-1: C-O stretching vibration on alcohol, ether, ester group;

[0019] Feature 19,20 -1768cm-1: C=O stretching vibration on ester, ketone, anhydride group;

[0020] Feature 38 -3100cm-1: sp 2 hybrid C-H stretching vibration on olefin, aromatic ring;

[0021] Feature 46 -3676cm-1: O-H stretching vibration on hydroxyl group;

[0022] Then, two new descriptors obtained by operator operation are respectively a descriptor D1 represented by formula (1) and a descriptor D2 represented by formula (2):

[0023]

[0024] D2= [(feature 36 -feature7)+(feature 38 +feature 46 )] (2)

[0025] Step 4.2: Establish a model for the relationship between the frequency characteristics and threshold value of the burnt sweet aroma spice molecules represented by formula (3):

[0026] P=0.00035×D1+0.00082×D2-0.91087 (3)

[0027] Where P is the negative natural logarithm of the predicted value.

[0028] Step 5 of the method for constructing a threshold value model for identifying burnt sweet aroma spices of the present invention is performed as follows:

[0029] Step 5.1: Obtain the molecular structure file through the database;

[0030] Step 5.2: Optimize the molecular structure using Gaussian and obtain the infrared spectrum; -1 The frequency is processed by 50 segments of segment integration, using the Feature {7,8,10,19,20,38,46} Descriptors D1 and D2 are calculated by equations (1) and (2), and then the negative natural logarithm prediction value P of the threshold is calculated by equation (3) using the descriptors D1 and D2. The threshold prediction value T is obtained by performing exponential calculation on P.

[0031] Compared with the existing technology, the beneficial effects of the present invention are embodied in:

[0032] This study combines quantum chemical calculations with machine learning to gain a deeper understanding of the aroma-producing mechanisms of aroma molecules at the atomic and molecular scale, establishing a structure-activity relationship between molecular structure and aroma intensity. Quantum chemical calculations are used to obtain information such as the structure and vibrational frequency of compounds associated with burnt-sweet aroma. Using aroma thresholds as a criterion, a machine learning approach is used to construct a fitting model for compound aroma intensity. This model accurately predicts the molecular aroma threshold based on the characteristics of functional groups in the molecular structure, providing a theoretical basis for odor optimization and flavoring in cigarette production. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the modeling scheme of the present invention;

[0034] Figure 2 This is the optimized structure file of hydroxyacetone in the model test;

[0035] Figure 3This is the optimized structure file of ethyl vanillin in the model test;

[0036] Figure 4 The calculated infrared spectrum of hydroxyacetone in the model test;

[0037] Figure 5 Calculated infrared spectrum of ethyl vanillin in model tests. DETAILED DESCRIPTION

[0038] The method for constructing the aroma threshold model for identifying burnt sweet aroma flavor molecules in this embodiment is to grasp the aroma-producing mechanism of aroma molecules at the atomic scale by combining quantum chemical calculations and machine learning, construct an aroma intensity fitting model between molecular structure and aroma intensity, and use the aroma intensity fitting model to predict the thresholds of different burnt sweet aroma flavor molecules.

[0039] In this embodiment, the construction of the aroma threshold model for identifying burnt sweet flavor molecules is carried out according to the following process:

[0040] First, the structural information of the sweet-smelling aroma molecule was obtained using the CAS number or SMILES string through the CAS SciFinder database or RDKit software package. The molecular structure was then optimized using the quantum computing software Gaussian to obtain the optimized molecular structure and infrared spectrum of the optimized molecular structure.

[0041] Then, the structure-related descriptors are obtained by performing segmented integration processing on the infrared spectrum;

[0042] Then, for the structure-related descriptors, key feature descriptors were extracted using the machine learning tool SISSO, and a model of the relationship between functional groups and thresholds was constructed using the key feature descriptors as a fragrance intensity fitting model; the fragrance intensity fitting model was used to predict the thresholds of different caramel-sweet fragrance molecules.

[0043] In this embodiment, the steps for constructing the aroma threshold model for identifying burnt sweet flavor molecules are as follows:

[0044] Step 1: Collect information on caramel-sweet aroma molecules with existing thresholds;

[0045] The method includes: obtaining the names and thresholds of 32 molecules with existing thresholds and burnt-sweet aroma through literature and databases, obtaining the CAS number or SMIELS string of the molecule through the Chinese and English names of the molecule through various database websites such as PubChem, CAS-SciFinder, Good Scents and Leffingwell & Associates; obtaining the structure file by downloading from the CAS-SciFinder database based on the CAS number, or converting SMILES to structure file (.mol) through the RDKit software package in Python based on the SMILES string to obtain the structure file; at the same time, obtaining the experimental infrared spectrum corresponding to the molecule in the CAS-SciFinder database; completing the collection of burnt-sweet aroma spice molecule information with existing thresholds.

[0046] Step 2: Optimize the molecular structure and predict its infrared spectrum using the quantum computing software Gaussian;

[0047] This includes obtaining the experimental infrared spectrum of the caramel-sweet flavor molecule from the CAS-SciFinder database; optimizing the structure and calculating the infrared spectrum in Gaussian using different functionals and basis sets: B3LYP / 6-31g**, B3LYP / 6-311++g**, B3LYP / aug-cc-pvtz, WB97XD / 6-311++g**, M06 / 6-31+g**, M06 / 6-311++g**, and M06-2X / 6-311++g**. The infrared spectrum of the optimized molecular structure is then compared with the experimental infrared spectrum, and the functional and basis set B3LYP / 6-311++g** with the smallest error with the characteristic peaks is selected for subsequent calculations of infrared spectrum information.

[0048] Step 3: Use Python to analyze the infrared spectrum from 500-4000cm -1 The frequency is processed by 50 segment integration, each segment of the integration corresponds to a descriptor, and a total of 50 descriptor features are obtained, which are recorded as Feature (N) (N=1-50);

[0049] Step 4: Filter the 50 descriptor features using the machine learning tool SISSO and extract key features. Use the key features to build a model of the relationship between functional groups and thresholds as a fragrance intensity fitting model.

[0050] Step 4.1: Due to the large difference in threshold values, the gap between the larger and smaller values ​​is usually 4-5 orders of magnitude, and the molecules with smaller thresholds (i.e., low concentration and strong odor) are the desired ones. Therefore, the thresholds of the collected sweet and fragrant spice molecules are negatively logarithmized to obtain the negative logarithm of the thresholds. The negative logarithm of the thresholds and the 50 descriptor features of the infrared spectra of each molecule obtained in step 3 are input into SISSO. The operators allowed in the model are: (+)(-)(*)( / )(exp)(exp-)(^-1)(^2)(^3)(sqrt)(cbrt)(log); the size of the descriptor / model is 2; the key features finally obtained are Feature {7,8,10,19,20,38,46} ;

[0051] Then, two new descriptors are obtained through operator operation, namely, descriptor D1 represented by formula (1) and descriptor D2 represented by formula (2):

[0052]

[0053] D2=[(feature 36 -feature7)+(feature 38 +feature 46 )] (2)

[0054] Step 4.2: Calculate the model of the relationship between the frequency characteristics and threshold value of the burnt sweet aroma molecule represented by formula (3) through two new descriptors:

[0055] P=0.00035×D1+0.00082×D2-0.91087 (3)

[0056] Among them, P is the negative natural logarithm predicted value of the threshold, and the determination coefficient R of the model is obtained 2 is 84.16%, R 2 The closer the value is to 1, the better the fitting effect is.

[0057] Feature {7,8,10,19,20,38,46} The corresponding frequencies and characteristic functional groups are:

[0058] Feature 7,8 -904cm-1: CH bending vibration on olefins or aromatic rings;

[0059] Feature 10 -1084cm-1: CO stretching vibration on alcohol, ether and ester groups;

[0060] Feature 19,20- 1768 cm"1: C=0 stretch on ester, ketone, anhydride groups;

[0061] Feature 38 - 3100 cm"1: sp 2 hybridized C-H stretch on alkenes, aromatic rings;

[0062] Feature 46 - 3676 cm"1: O-H stretch on hydroxyl groups;

[0063] Thus, the functional groups that mainly affect the threshold of caramelic flavor molecules are analyzed as follows: alkenyl, ketone, hydroxyl, ether and aromatic ring. Among them, furan ketone compounds with caramelic flavor generally have alkenyl, ketone and ether groups at the same time; and ethyl maltol compounds commonly used in tobacco for caramelic flavor have aromatic ring, ketone and hydroxyl groups at the same time, which matches the prediction results.

[0064] Step 5: predicting the threshold of different caramelic flavor molecules by using the flavor intensity fitting model.

[0065] Step 5.1: obtaining the structure file of the molecule from the database;

[0066] Step 5.2: optimizing the molecular structure by Gaussian to obtain the infrared spectrum; performing 50-segment segmented integral processing on the frequencies of the spectrum 500-4000 cm -1 - 3100 cm"1: sp {7,8,10,19,20,38,46} hybridized C-H stretch on alkenes, aromatic rings;

[0067] Model test: taking hydroxyacetone and ethyl vanillin, which are known to have caramelic flavor but have not yet obtained the threshold, as examples to test the model.

[0068] Hydroxyacetone is an organic compound with the molecular formula C3H6O2, which is composed of one acetone molecule (C3H6O) and one hydroxyl group (-OH) in structure. It is an α-hydroxy ketone compound and is usually considered as an organic synthesis intermediate or byproduct of chemical reaction. In aroma chemistry, it is usually considered to have caramel, sweet and slight smoke aroma, so it is often used in flavor related to caramel, baking and smoking flavor. Because it can produce caramel-like sweet aroma when heated, this feature makes it have potential application in flavor blending. Ethyl vanillin is a flavor compound with the molecular formula C9H 10O3 is an ethylated derivative of vanillin. Its chemical structure is similar to vanillin, but it has an ethoxy group (-OCH2CH3) on the aromatic ring replacing the methoxy group (-OCH3) of vanillin. Compared to regular vanillin, ethyl vanillin has a more intense aroma. Its aroma is rich and sweet, similar to vanilla, but more intense and long-lasting. Due to its caramel-like sweetness and warm notes, it is often used in the food and flavoring industries, particularly in chocolate, desserts, and beverages, imparting a caramelized, sweet flavor.

[0069] (1) Obtain molecular structure files through CAS-SciFinde search, optimize the structures using Gaussian's B3LYP / 6-311++g** basis set and functionals, and obtain their infrared spectra. Figure 2 The optimized structure file of hydroxyacetone is shown; Figure 3 The optimized structure file of ethyl vanillin is shown; Figure 4 Shown is the calculated infrared spectrum of hydroxyacetone; Figure 5 Shown is the calculated infrared spectrum of ethyl vanillin.

[0070] (2) Perform 50-segment integration on the infrared spectrum and obtain the integral value of each segment as the basic descriptor Feature 1-50 . Select the Feature {7,8,10,19,20,38,46} Substitute (1) and (2) to calculate and obtain D1 and D2 respectively

[0071]

[0072] D2=[(feature 36 -feature7)+(feature 38 +feature 46 )] (2)

[0073] P=0.00035×D1+0.00082×D2-0.91087 (3)

[0074] Table 1: Feature display of some descriptors of two molecules:

[0075] Feature 7 8 10 19 20 38 46 Hydroxyacetone 127.51 2119.93 3224.51 7598.86 309.38 8.35 624.02 Ethyl vanillin 2347.93 897.55 5904.01 5926.12 117.12 117.79 3245.22

[0076] Substituting D1 and D2 into formula (3), we can obtain the negative natural logarithm prediction value P of the threshold of the two molecules:

[0077] P Hydroxyacetone =0.00035×2128.37+0.00082×1036.67=6.184;

[0078] PEthyl vanillin =0.00035×138.63+0.00082×2346.29=1.9725;

[0079] Then perform exponential processing to obtain the threshold prediction value:

[0080] T Hydroxyacetone =EXP(-6.184)=0.00206ppm;

[0081] T Ethyl vanillin =EXP(-1.9725)=0.1391ppm.

[0082] In summary, molecules with lower odor thresholds can release stronger aromas at the same concentration. Hydroxyacetone clearly performs better in this regard, exhibiting a higher threshold effect. Therefore, hydroxyacetone is more suitable for experimental verification to more accurately assess its aroma intensity and potential for application in flavoring and tobacco products.

Claims

1. A method for constructing a threshold model for identifying aroma molecules of burnt sweet aroma, characterized by By combining quantum chemical calculations with machine learning, the aroma-generating mechanism of aroma molecules at the atomic scale is grasped, and an aroma intensity fitting model between molecular structure and aroma intensity is constructed. The threshold values ​​of different burnt-sweet aroma molecules are predicted using the aroma intensity fitting model. First, the structural information of the burnt-sweet aroma molecules is obtained using the CAS number or SMILES string through the CAS SciFinder database or the RDKit software package. The molecular structure is optimized based on the structural information using the quantum computing software Gaussian to obtain an optimized molecular structure and an infrared spectrum of the optimized molecular structure. Then, structure-related descriptors are obtained by performing piecewise integration processing on the infrared spectrum. Finally, key feature descriptors are extracted from the structure-related descriptors using the machine learning tool SISSO. A model of the relationship between functional groups and threshold values ​​is constructed using the key feature descriptors as a fragrance intensity fitting model. The threshold values ​​of different burnt-sweet aroma molecules are predicted using the aroma intensity fitting model. The method for constructing a flavor threshold model for identifying burnt sweet flavor molecules comprises the following steps: Step 1: Collecting information on caramel-sweet flavor molecules with a threshold value; including: obtaining the CAS number or SMIELS string of the molecule by searching; obtaining a structure file by downloading it from the CAS-SciFinder database based on the CAS number, or converting the SMILES string to a structure file using the RDKit software package in Python to obtain the structure file; completing the collection of information on caramel-sweet flavor molecules with a threshold value; Step 2: Optimize the molecular structure and predict its infrared spectrum using the quantum computing software Gaussian. This includes obtaining the experimental infrared spectrum of the caramel flavoring molecule from the CAS-SciFinder database. In Gaussian, optimize the structure, calculate the infrared spectrum, and perform frequency correction using different functionals and basis sets to obtain the infrared spectrum of the optimized molecular structure. Compare the infrared spectrum of the optimized molecular structure with the experimental infrared spectrum, and select the functional and basis set with the smallest error with the characteristic peaks for subsequent calculation of the infrared spectrum information. Step 3: Use Python to analyze the infrared spectrum from 500-4000cm -1 The frequency is processed by 50 segment integration, each segment of the integration corresponds to a descriptor, and a total of 50 descriptor features are obtained, which are recorded as Feature {N} , N=1-50; Step 4: Filter the 50 descriptor features using the machine learning tool SISSO and extract key features. Use the key features to construct a model of the relationship between functional groups and thresholds as a fragrance intensity fitting model. Step 5: Use the aroma intensity fitting model to predict the thresholds of different caramel-sweet aroma molecules.

2. The method for constructing a flavor threshold model for identifying burnt sweet fragrance molecules according to claim 1, characterized in that Described step 4 is carried out as follows: Step 4.1: Perform negative logarithm processing on the thresholds of the collected burnt sweet flavor molecules to obtain the negative logarithm of the threshold, and input the negative logarithm of the threshold and the 50 descriptor features of the infrared spectrum of each molecule obtained in step 3 into SISSO; the operators allowed in the model are: (+), (-), (*), ( / ), (exp), (exp-), (^-1), (^2), (^3), (sqrt), (cbrt), (log); the size of the descriptor / model is 2; the key features finally obtained are Feature {7,8,10,19,20,38,46} ; Feature {7,8,10,19,20,38,46} The corresponding frequencies and characteristic functional groups are: Feature {7,8} -904cm -1 : CH bending vibration on olefins or aromatic rings; Feature {10} -1084 cm -1 : CO stretching vibration on alcohol, ether and ester groups; Feature {19,20} -1768 cm -1 : C=O stretching vibration on ester, ketone and anhydride groups; Feature {38} -3100 cm -1 : sp² hybridized CH stretching vibration on olefins and aromatic rings; Feature {46} -3676 cm -1 : OH stretching vibration on hydroxyl group; Then, two new descriptors are obtained through operator operation, namely the descriptor D1 represented by formula (1) and the descriptor D2 represented by formula (2): ； ； Step 4.2: Establish a model for the relationship between the frequency characteristics of the burnt sweet flavor molecule and the threshold value represented by formula (3): ； Where P is the negative natural logarithm of the predicted value.

3. The method for constructing a molecular aroma threshold model for identifying burnt sweet fragrance according to claim 2, characterized in that Step 5 is performed as follows Step 5.1: Obtain the molecular structure file through the database; Step 5.2: Optimize the molecular structure using Gaussian and obtain the infrared spectrum; -1 The frequency is processed by 50 segments of segment integration, using the Feature {7,8,10,19,20,38,46} Descriptors D1 and D2 are calculated by equations (1) and (2), and then the negative natural logarithm prediction value P of the threshold is calculated by equation (3) using the descriptors D1 and D2. The threshold prediction value T is obtained by performing exponential calculation on P.

Images