Smell knowledge base construction method and system based on molecular modeling and deep learning

Abstract

The invention discloses an odor knowledge base construction method and system based on molecular modeling and deep learning, and belongs to the technical field of artificial intelligence and digital odor. According to the method, firstly, smell molecules are subjected to standardization treatment, and smell perception labels of the smell molecules are extracted to construct a smell knowledge base; expressing the odor molecules as a molecular graph structure, introducing molecular position coding to model a three-dimensional geometric configuration, and learning high-dimensional feature expression through a first graph neural network; then selecting odor molecules from the knowledge base and generating mixture candidates based on the similarity so as to form a mixture molecular map; and finally, processing the molecular map of the mixture by using a label prediction model, predicting an olfactory perception label of the mixture, and expanding a result to a knowledge base. According to the method, unified characterization of the single molecule and the mixture is realized, the technical vacancy of mixture olfaction prediction is filled, the dimensionality and expansibility of the knowledge base are improved, and the method can be widely applied to the fields of digital blending, smell recommendation and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and digital odor, and particularly relates to an odor knowledge base construction method and system based on molecular modeling and deep learning. BACKGROUND

[0002] In recent years, researchers have proposed digital odor technology, that is, through digital representation of odor molecules, combined with sensor detection and release devices, programmable description and control of odor are realized. This kind of technology provides new possibilities for virtual reality, personalized recommendation and intelligent interaction. However, in the development process of artificial intelligence and digital odor technology, how to structurally and standardize the representation and storage of odor molecules is still the core prerequisite for realizing digital perfumery and olfactory computing. The generation and perception process of odor essentially depends on the chemical structure characteristics of odor molecules, the interaction between molecules and the binding rules of odor receptors. Specifically, the atomic type, chemical bond type, hybrid state, charge distribution, spatial geometry of a single odor molecule and the interaction between odor molecules will have a significant impact on olfactory perception. At the same time, the subjective perception of odor will also be affected by individual differences, environmental factors and application scenarios, resulting in different olfactory experiences of the same odor molecule combination in different groups of people.

[0003] In the prior art, odor molecule knowledge bases often rely on sensory experiments or single data dimensions, and only record odor molecule names, CAS numbers or simple olfactory descriptions, lack systematic association of odor molecule chemical structure and olfactory perception labels, and the data structure is scattered, the standardization degree is low, and it is difficult to support large-scale algorithm training and mixed odor prediction. For example, for single odor molecule information, the existing knowledge base cannot effectively utilize the odor molecule graph structure for atomic and bond feature modeling, nor can it introduce three-dimensional spatial information of odor molecules; for mixed odor prediction, there is a lack of unified multi-dimensional labels and auxiliary features such as concentration and volatility rate, which limits the model training and prediction effect.

[0004] Therefore, how to design a novel odor molecule representation and management method, and construct an odor molecule knowledge base that can support single odor molecule and mixed odor molecule label prediction, high-precision odor prediction and perfumery application, is a technical problem to be solved in the field. SUMMARY

[0005] The technical problem to be solved by the present application is to overcome the deficiencies in the construction and application of the existing odor molecule knowledge base: the existing technology only records the physical and chemical properties of odor molecules, lacks systematic correlation between the chemical structure, spatial configuration and olfactory perception tags (such as floral, fruity, woody, etc.) of odor molecules, and the feature representation dimension of odor molecules is single, making it difficult to fully capture the atomic topological structure, chemical bond information and three-dimensional spatial geometric characteristics of odor molecules; the knowledge storage lacks unified structured specification, making it difficult to support representation learning based on odor molecule graphs, single odor molecule and mixed odor molecule tag prediction, and efficient retrieval in large-scale algorithm training and complex application scenarios. Therefore, the present application proposes a method and system for constructing an odor knowledge base based on molecular modeling and deep learning, which can finely characterize single odor molecules and odor molecule combinations and achieve standardized and systematic storage, thereby providing reliable data support for subsequent odor prediction and intelligent perfumery applications.

[0006] In order to achieve the above-mentioned application purposes, the present application specifically adopts the following technical solutions:

[0007] In the first aspect, the present application provides a method for constructing an odor knowledge base based on molecular modeling and deep learning, which includes the following steps:

[0008] S1, after obtaining odor molecule data and performing standardization processing, generating the olfactory perception tags and quantitative characteristics corresponding to each odor molecule, and using the odor molecules and their corresponding olfactory perception tags and quantitative characteristics to construct an odor knowledge base; the quantitative characteristics include the perception intensity score and the evaporation rate of each odor molecule;

[0009] S2, the atoms in each odor molecule are taken as nodes in the molecular graph structure and encoded, the chemical bonds in the odor molecule are taken as edges in the molecular graph structure and encoded, and a molecular position encoding method is introduced to model the three-dimensional geometric configuration of the odor molecule, so that the chemical structure information of the odor molecule is represented as a molecular graph structure, and then the molecular graph is input into the trained first graph neural network to obtain a high-dimensional feature representation containing the chemical structure and spatial information of each odor molecule; wherein the molecular graph is composed of a node feature matrix and an edge feature matrix;

[0010] S3, selecting several odor molecules from the odor knowledge base to generate a mixture candidate, generating a final mixture according to the similarity between the two odor molecules in the mixture candidate, taking the odor molecules in the final mixture as candidate molecules, calculating the concentration of the candidate molecules according to the perception intensity score and the evaporation rate, weighting the molecular graphs of all candidate molecules with the concentration as the weight, and splicing all the weighted molecular graphs to form a mixture molecular graph;

[0011] S4, input the constructed mixture molecular graph into a label prediction model for supervised learning. In the supervised learning, the olfactory perception label of the candidate molecule is used to generate a corresponding pseudo-real olfactory perception label for the final mixture. The pseudo-real olfactory perception label is used as a supervision signal. The second graph neural network in the model extracts local interaction features between atoms, aggregates the local interaction features through global pooling operation, and forms a global vector representation of the final mixture. Finally, the feedforward neural network in the model is used to regress and map the global vector representation, so as to predict the olfactory perception label of the final mixture. The final mixture and the corresponding olfactory perception label are added to the odor knowledge base.

[0012] On the basis of the above scheme, each step can be implemented in the following preferred specific manner.

[0013] As a preferred embodiment of the above first aspect, in the S1 olfactory perception label collection process, each odor molecule is independently evaluated by at least three perfumers to obtain the olfactory perception label, and the perception intensity score and the evaporation rate of the gas molecule are obtained through multiple experimental measurements; in the standardization process, the evaporation rate is normalized to the [0, 1] interval according to the maximum value, the perception intensity score is converted to a percentage value according to a preset standard score, and the olfactory perception label is corrected for consistency in combination with expert voting results.

[0014] As a preferred embodiment of the above first aspect, in S2, the node encoding needs to encode the atom type, charge, hybrid state and topological feature; the edge encoding needs to encode the bond type, bond length and bond order.

[0015] As a preferred embodiment of the above first aspect, in S2, the specific process of obtaining the molecular graph is as follows: first, the original three-dimensional coordinates of each atom in the odor molecule are translated according to the geometric center of the odor molecule, so that the spatial position of the odor molecule is aligned relative to the geometric center; then the original three-dimensional coordinates are scaled and normalized, so that all odor molecules are represented in a unified scale range, thereby obtaining the processed three-dimensional coordinates of each atom; after position encoding the processed three-dimensional coordinates, the position encoding results are fused with the node feature vectors in a splicing manner to obtain the fusion features of each atom; then the fusion vectors of all atoms belonging to the same odor molecule are constructed, thereby constructing the node feature matrix from the fusion vectors of all odor molecules, and constructing the edge feature matrix from the edge feature vectors of all atoms, and completing the construction of the molecular graph; wherein the node feature vector is the feature vector obtained after encoding each node, and the edge feature vector is the feature vector obtained after encoding each edge.

[0016] As a preferred embodiment of the above first aspect, in S3, the specific process of generating the final mixture is as follows:

[0017] S31, selecting a plurality of odor molecules from the odor knowledge base to generate a mixture candidate, wherein a high-dimensional feature representation of the odor molecules in the mixture candidate constitutes a mixture vector;

[0018] S32, calculating the similarity between the high-dimensional feature representations in the mixture vector, and when there is a calculated similarity result greater than a preset similarity threshold, the two odor molecules corresponding to the similarity result are regarded as similar molecules, and a weighted average of the similarity between each similar molecule and the remaining odor molecules in the mixture candidate is calculated, and the similar molecules with smaller weighted average results are retained to form a final mixture; when the calculated similarity results are all less than the similarity threshold, the mixture candidate is regarded as the final mixture.

[0019] As a preferred embodiment of the first aspect, in S32, the similarity between two odor molecules is measured using cosine similarity.

[0020] As a preferred embodiment of the first aspect, in S3, the concentration of a candidate molecule is calculated as follows: the product of the perceptual intensity score and the evaporation rate of the candidate molecule is taken as the comprehensive score of the candidate molecule, and the ratio of the comprehensive score of the candidate molecule to the comprehensive scores of all candidate molecules is taken as the concentration of the candidate molecule.

[0021] As a preferred embodiment of the first aspect, in the supervised learning of S4, an adaptive focal loss function is calculated based on the error between the pseudo-real olfactory perception label and the predicted olfactory perception label, and the label prediction model parameters are updated based on minimizing the adaptive focal loss function.

[0022] Secondly, the present application provides an odor knowledge base construction system based on molecular modeling and deep learning, which comprises:

[0023] A data acquisition module is configured to acquire odor molecule data, perform standardization processing, generate an olfactory perception label and a quantitative feature corresponding to each odor molecule, and construct an odor knowledge base using the odor molecules and the corresponding olfactory perception labels and quantitative features; the quantitative features include the perceptual intensity score and the evaporation rate of each odor molecule.

[0024] A molecular graph generation module is configured to encode the atoms in each odor molecule as nodes in a molecular graph structure, encode the chemical bonds in the odor molecules as edges in the molecular graph structure, and introduce a molecular position encoding method to model the three-dimensional geometric configuration of the odor molecules, so as to represent the chemical structure information of the odor molecules as a molecular graph structure, and then input the molecular graph into a trained first graph neural network to obtain a high-dimensional feature representation containing the chemical structure and spatial information of each odor molecule; wherein the molecular graph is composed of a node feature matrix and an edge feature matrix.

[0025] The mixture molecule graph generation module is configured to select a plurality of odor molecules from the odor knowledge base to generate a mixture candidate, generate a final mixture according to the similarity between two odor molecules in the mixture candidate, take the odor molecules in the final mixture as candidate molecules, calculate the concentration of the candidate molecules according to the perception intensity score and the volatilization rate of the candidate molecules, weight the molecular graphs of all the candidate molecules by taking the concentration as a weight, and splice all the weighted molecular graphs to form a mixture molecular graph.

[0026] The prediction module is configured to input the constructed mixture molecular graph into a label prediction model for supervised learning. In the supervised learning, a pseudo-real olfactory perception label corresponding to the final mixture is generated according to the olfactory perception label of the candidate molecules, the pseudo-real olfactory perception label is taken as a supervision signal, the local interaction features between atoms are extracted by a second graph neural network in the model, the local interaction features are aggregated by a global pooling operation to form a global vector representation of the final mixture, and finally, a feedforward neural network in the model is used to regress and map the global vector representation, so as to predict the olfactory perception label of the final mixture. The final mixture and the corresponding olfactory perception label are added to the odor knowledge base.

[0027] In a third aspect, the present application provides a computer electronic device, comprising a memory and a processor;

[0028] The memory is configured to store a computer program.

[0029] The processor is configured to implement the odor knowledge base construction method based on molecular modeling and deep learning according to any one of the above first aspect when executing the computer program.

[0030] Compared with the prior art, the present application has the following beneficial effects:

[0031] The application realizes high-dimensional and refined representation of the chemical structure, spatial configuration and olfactory perception label of the odor molecule by representing the odor molecule as an odor molecule graph structure and combining odor molecule position coding and graph neural network feature learning, and overcomes the deficiency of the existing knowledge base that only relies on physical and chemical properties and has single feature representation dimension; the olfactory perception label of the mixture is predicted through a label prediction model, unified representation and storage of single odor molecules and mixed odor molecules are realized, and the technical gap in mixture olfactory prediction is filled; meanwhile, the multi-dimensional information such as odor molecule feature vector, olfactory perception label, perception intensity, volatile rate and application scenario is standardized and uniformly managed, an extensible and searchable knowledge base structure is formed, and the operation efficiency in large-scale algorithm training and complex application scenarios is improved; in the prediction process, a multi-layer graph attention network and an adaptive focal loss function are used, which effectively improves the prediction accuracy and model generalization ability; the knowledge base can be widely applied to odor prediction, mixed label generation, digital perfumery, personalized odor recommendation and multi-modal interaction, and provides data support for the popularization of digital odor technology in the field of artificial intelligence. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 It is a step flow diagram of the method of the application;

[0033] Figure 2 It is a prediction effect diagram provided by the embodiment of the application;

[0034] Figure 3 It is a mixed odor molecule olfactory perception label prediction schematic diagram provided by the embodiment of the application;

[0035] Figure 4 It is a system block diagram of the application;

[0036] Figure 5 It is a computer electronic equipment composition schematic diagram provided by the embodiment of the application. DETAILED DESCRIPTION

[0037] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings. In the following description, a large number of specific details are set forth in order to facilitate a full understanding of the application. However, the application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the application, so the application is not limited by the specific embodiments disclosed below. The technical features in each embodiment of the application can be combined accordingly without conflict.

[0038] In the description of the present application, it should be understood that the terms "first", "second" are only used for distinguishing description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features.

[0039] The present application provides a method for constructing an odor knowledge base based on molecular modeling and deep learning, which realizes the unified representation and knowledge base construction of single odor molecules and mixed odor molecules by systematically collecting single odor molecule odor information, converting odor molecule chemical structure into molecular graph and combining position coding, generating mixture molecular graph with diversity constraint, and predicting the olfactory perception label of the mixture based on the label prediction model, and provides expandable and searchable technical support for digital perfumery and intelligent odor design. Figure 1 As shown in the preferred embodiment of the present application, the method comprises the following steps:

[0040] S1, single molecule odor label collection

[0041] After obtaining the odor molecule data and performing standardization processing, the corresponding olfactory perception label and quantitative characteristics of each odor molecule are generated, and the odor knowledge base is constructed using the odor molecule and its corresponding olfactory perception label and quantitative characteristics; the quantitative characteristics include the perception intensity score and the volatilization rate of each odor molecule.

[0042] It should be noted that in the process of collecting the olfactory perception label, each odor molecule is independently evaluated by at least three perfumers to obtain the olfactory perception label, and the perception intensity score and the volatilization rate of the gas molecule are obtained through multiple experimental measurements; in the standardization processing, the volatilization rate is normalized to the [0, 1] interval according to the maximum value, the perception intensity score is converted to a percentage value according to the preset standard score, and the olfactory perception label is corrected for consistency in combination with the expert voting results.

[0043] It should be noted that in the expression of the present application, the odor molecule is a specific, objectively existing chemical substance, which can be regarded as an entity in the physical world, which has a specific chemical structure (such as molecular formula, functional group, stereo configuration, etc.), and itself is the physical and chemical basis of causing olfactory perception. For example, citronellol molecules in rose fragrance, limonene molecules in lemon fragrance, etc. The odor molecule data is a set of digitized information describing the characteristics and perception effect of the odor molecule, which is equivalent to the multidimensional description of the entity of the "odor molecule", which is stored in the computer, and is used to quantify, analyze and predict the properties and effects of the odor molecule. For example, benzyl acetate in jasmine flower fragrance is an odor molecule, and in the database, the benzyl acetate record contains: olfactory perception label: jasmine flower, sweet fragrance; perception intensity score: 8.5 / 10; volatile rate: moderate; applicable scene: perfume, skin care product, air freshener.

[0044] It should be noted that the output of S1 of the present application is the olfactory perception label and quantitative characteristics of each odor molecule, which provides reliable data support for subsequent molecular modeling and knowledge base construction. Specifically, in S1, the present embodiment first systematically collects and encodes individual odor molecules. The obtained odor molecule data includes the olfactory perception label (such as floral, fruity, woody, spicy, etc.), perception intensity score, volatile rate, and applicable scene of each odor molecule, and other multi-dimensional information.

[0045] The above multi-dimensional information sources include experimental measurement data, literature data and perfumer experience. In the data collection process, in order to ensure data consistency and reliability, the present embodiment adopts a combination of multiple experimental measurements and expert voting for collection, and is processed by unified standardization. Specifically, each odor molecule needs to be independently evaluated by at least three perfumers, and the corresponding olfactory perception label (such as floral, fruity, woody, spicy, etc.) is recorded, which is subjective data, at the same time, the volatile rate and the perception intensity score and the applicable scene of the gas molecule are obtained through repeated experimental measurement. In the present embodiment, the volatile rate is specifically measured by professional instruments. Considering that there may be missing values or abnormal values in the experimental measurement results, therefore, before processing, the missing data needs to be completed, and the abnormal values beyond the physical reasonable range are removed or smoothed, in order to ensure the accuracy of the data.

[0046] The interpolated numerical data is normalized (such as normalizing the volatile rate to the maximum value to the 0-1 interval, and converting the perception intensity to a percentage numerical value according to the standard score), and the olfactory perception label is corrected for consistency in combination with the expert voting results, forming a unified standardized olfactory perception label set, which provides training data that can be directly input into the model for subsequent molecular graph modeling, vectorization representation and knowledge base construction.

[0047] It should be noted that the entire process of data collection and processing described above can be realized automatically by a program. For example, the paper score sheet or experimental measurement file of the perfumer is read into an electronic table through OCR or data import script, and automatic normalization and missing value completion processing are performed, while abnormal value is detected and corrected.

[0048] The specific normalization formula for the volatilization rate of the th odor molecule is:

[0049]

[0050] wherein, represents the normalized volatilization rate of the th odor molecule; represents the minimum value in the volatilization rate; represents the maximum value in the volatilization rate.

[0051] The perception intensity of the th odor molecule is processed according to the above formula. In this way, standardized data can be formed for model training.

[0052] S2, Modeling and Feature Extraction of Single Molecule Graph

[0053] The atoms in each odor molecule are taken as nodes in the molecular graph structure and the atomic types, charges, hybrid states and topological features are encoded, the chemical bonds in the odor molecule are taken as edges in the molecular graph structure and the chemical bond types, bond lengths and bond orders are encoded, and the molecular position encoding (Mol-PECO) method is introduced to model the three-dimensional geometric configuration of the odor molecule, so as to represent the chemical structure information of the odor molecule as a molecular graph structure. The molecular graph is input into the trained first graph neural network to obtain a high-dimensional feature representation containing the chemical structure and spatial information of each odor molecule; wherein the molecular graph is composed of a node feature matrix and an edge feature matrix.

[0054] ​It should be noted that the specific process of obtaining the molecular graph in S2 of this invention is as follows: First, the original three-dimensional coordinates of each atom in the odor molecule are translated according to the geometric center of the odor molecule, so that the spatial position of the odor molecule is aligned with the geometric center; then, the original three-dimensional coordinates are scaled to ensure that all odor molecules are represented within a uniform scale range, avoiding the influence of size differences on feature learning, thereby obtaining the processed three-dimensional coordinates of each atom. After position encoding of the processed three-dimensional coordinates, the position encoding result is fused with the node feature vector by concatenation to obtain the fused feature of each atom. Then, the fused features of all atoms belonging to the same odor molecule are used to construct a fused vector, thereby constructing a node feature matrix from the fused vectors of all odor molecules, and constructing an edge feature matrix from the edge feature vectors of all atoms, thus completing the construction of the molecular graph; wherein, the node feature vector is the feature vector obtained after encoding each node, and the edge feature vector is the feature vector obtained after encoding each edge.

[0055] It should be noted that in embodiment S2 of the present invention, after completing the single-molecule tag acquisition, this embodiment first parses the chemical structure information of each odor molecule (such as the SMILES string or Mol file) into a computer-processable molecular diagram structure. (The rest of the text appears to be a separate, unrelated sentence fragment.) Taking an odor molecule as an example, its corresponding molecular diagram structure can be generally represented as follows: ,in Indicates the first A set of nodes for each odor molecule. Indicates the first An edge set of odor molecules This indicates the node in the above set of nodes. 1 node Indicates the number of nodes.

[0056] In this embodiment, the specific process for constructing the above molecular map is as follows:

[0057] 1) Atom Node Encoding: Parsing the chemical structure file of odor molecules to extract the atom type (carbon C, hydrogen H, oxygen O, nitrogen N, etc.), charge, hybridization state, and topological characteristics (including the number of covalent bonds and the number of neighboring atoms) of each atom. For the first... The odor molecule, and the first one of them Nodes Represented as node feature vectors .

[0058] Furthermore, this embodiment employs one-hot encoding for discrete features (hybrid state, atom type) and normalizes continuous features (such as charge, number of covalent bonds, and number of neighboring atoms) to ensure that the node feature vectors are within the same scale range.

[0059] 2) Chemical bond edge coding: Extract the bond type (single bond, double bond, triple bond, aromatic bond), bond length, bond order, etc. information for each chemical bond, and form an edge feature vector. For the first odor molecule, the first node and the second node connected by the first edge form an edge, and the edge feature vector corresponding to the edge is denoted as

[0060] 3) Molecular position coding (Mol-PECO): The original three-dimensional coordinates of the first atom in the first odor molecule are centered and scaled as follows:

[0061]

[0062]

[0063]

[0064] wherein the atom and the node are one-to-one corresponding, and thus the first atom is actually the first node; is the three-dimensional coordinates of the geometric center of the odor molecule; is the three-dimensional coordinates of the first atom after processing; is the scale normalization coefficient. Fusion of node feature vector and position coding result to obtain the fusion feature of the first node in the first odor molecule :

[0065] wherein is the position coding result obtained after the three-dimensional coordinates after processing are position coded;

[0066] is the weight coefficient of position coding, used to adjust the influence strength of the position coding in the fusion feature; is the vector splicing.

[0067] The fusion vector is constructed by the fusion features of all atoms belonging to the same odor molecule. Taking the first odor molecule as an example, the fusion vector corresponding to the first odor molecule is denoted as :

[0068] In the formula,

[0069]

[0070] In the formula, ​​​​​​​​​​​These represent the first odor molecule in the first odor molecule. The fusion characteristics of each node.

[0071] Finally, based on the fusion vector of all odor molecules, the molecular diagram is constructed using the following formula:

[0072]

[0073]

[0074] in, Represents the node feature matrix; They represent the first The fusion vector of odor molecules; superscript Indicates matrix transpose; Represents the real number field; Indicates the number of odor molecules; Indicates the dimension of the node feature vector; Represents the edge feature matrix; This represents the dimension of the edge feature vector.

[0075] In this embodiment, the molecular graph is input into a first graph neural network (GNN) for iterative feature learning. The specific implementation of the first graph neural network is prior art. To facilitate a better understanding of the present invention by those skilled in the art, the following describes the first graph neural network's... Taking the layer as an example, describe the first layer. The first odor molecule The feature update process of a node in this layer is shown in the following equation:

[0076]

[0077]

[0078] in, Indicates the first Layer The node representation of the fused edge feature vector corresponding to each node; Represents the ReLU activation function; This represents the first trainable weight matrix; Indicates the first Layer The node representation of the fused edge feature vector corresponding to each node; Indicates the node index; Indicates the first The set of neighboring nodes of a node; Indicates the attention coefficient; This represents the trainable second weight matrix; Indicates the first Layer The node representation of the fused edge feature vector corresponding to each node; This represents the softmax activation function; Represents a learnable attention weight vector; Represents the L1 norm; Indicates the first Layer The initial node representation corresponding to each node.

[0079] Following the above procedure, after After processing by the first layer of the neural network, the 1st layer is obtained. The first odor molecule The final node representation of each node Then, the final node representation of all nodes in the odor molecule is processed using an average pooling function to obtain the high-dimensional feature representation of the odor molecule. , used for vector representation of mixtures and prediction of final mixture labels.

[0080] As can be seen, the method of this invention uses graph neural networks for iterative feature learning to achieve vectorized representation of single molecules, providing a foundational feature base for mixture prediction and knowledge base construction. During the iterative feature learning process of the graph neural network, a multi-layered stacked message-passing layer is preferred. Each layer updates the node representation through interactions between nodes, neighboring nodes, and edge features. Skip connections are introduced to alleviate the gradient vanishing problem in deep networks and preserve underlying feature information. To highlight the representation capability of key atoms or functional groups, an attention mechanism is preferably added to each layer, assigning different weights to the features of neighboring nodes, thereby enhancing the model's sensitivity to important regions of the molecular structure. For different types of node and edge features, normalization processing is preferred to ensure stable feature value distribution and prevent information loss.

[0081] Additionally, it's worth noting that all steps in S2 can be automated via scripts: SMILES / Mol file parsing, node / edge vectorization, coordinate normalization, tensor construction, and GNN feature learning outputs are all data that the program can directly manipulate. The molecular graph of an odor molecule includes: a node feature matrix (containing numerical information such as atom type, charge, hybridization state, and number of neighbors); an edge feature matrix (containing numerical information such as bond type, bond length, and aromaticity); and a position encoding matrix (containing the normalized position vectors of the atoms' three-dimensional coordinates). These molecular graphs are stored in standard file formats (such as JSON, HDF5, or PyTorch Geometric Data objects) to ensure direct access by the computer. Through a unified data structure and standardized storage, each odor molecule can form a computable and trainable high-dimensional feature representation, facilitating subsequent algorithm calls.

[0082] S3, mixture generation and its molecular graph construction

[0083] select several odor molecules from the odor knowledge base to generate mixture candidates, generate the final mixture according to the similarity between two odor molecules in the mixture candidate, take the odor molecules in the final mixture as candidate molecules, calculate the concentration of the candidate molecules according to the perception intensity score and the volatilization rate, weight all the molecular graphs of the candidate molecules according to the concentration as the weight, and splice all the weighted molecular graphs to form the molecular graph of the mixture.

[0084] It should be noted that in S3 of the present application, after the construction of the odor knowledge base and the molecular graph characterization and position coding of single odor molecules are completed, the mixture molecular graph construction phase is entered, and the core task is to unify the structure and proportion information of multiple single molecular odors into a computer processable data structure, so as to be input into the subsequent second graph neural network. Specifically, the present embodiment first selects several single odor molecules from the knowledge base according to the expert perfuming rules to generate mixture candidates that meet the harmony rules. In the above mixture candidate generation process, first, the mixture candidate is obtained according to the odor knowledge base, and the chemical similarity between molecules is calculated as a similarity constraint to ensure that high-similarity molecules are not excessively concentrated in the same mixture candidate, avoiding monotony or conflict of the fragrance type; at the same time, the perception intensity score and the volatilization rate of the single odor molecules are considered as diversity constraints in the combination process to ensure that the mixture candidate contains odor molecules of different categories of fragrance types, thereby improving the balance of the overall harmony, and finally forming a concentration distribution vector through proportion allocation. Therefore, the final output of this step is the combination information of the odor molecules of the mixture (i.e. the above final mixture) and the concentration distribution vector, which provides an input condition for the subsequent olfactory label prediction.

[0085] It should be noted that in S3 of the present application, the specific process of generating the final mixture is as follows:

[0086] S31, select several odor molecules from the odor knowledge base to generate mixture candidates, and the mixture vector is composed of the high-dimensional feature representation of the odor molecules in the mixture candidate.

[0087] It should be noted that in step S31 of the present embodiment, the above mixture vector can be represented as , respectively represent the high-dimensional feature representation of the odor molecules in the mixture candidate; is the number of odor molecules in the mixture candidate.

[0088] ​S32, in the mixture vector, similarity between high-dimensional feature representations of two-by-two is calculated: when there is a calculated similarity result greater than a preset similarity threshold, then the two odor molecules corresponding to the similarity result are considered as similar molecules, and then a weighted average of similarity between each similar molecule and the remaining odor molecules in the mixture candidate is calculated, and the similar molecule with a smaller weighted average result is retained to form a final mixture; when the calculated similarity result is less than the similarity threshold, then the mixture candidate is considered as the final mixture.

[0089] It should be noted that, in step S32 of the present embodiment, in order to ensure the diversity of the fragrance type of the mixture candidate, the similarity between any two odor molecules is calculated as a similarity constraint to avoid excessive concentration of highly similar odor molecules in the mixture candidate. Specifically, assuming that there are 5 odor molecules in the mixture candidate, denoted as a first odor molecule, a second odor molecule, a third odor molecule, a fourth odor molecule, and a fifth odor molecule, then the similarity between the high-dimensional feature representation of the first odor molecule and the high-dimensional feature representation of the other four odor molecules (i.e., the second odor molecule, the third odor molecule, the fourth odor molecule, and the fifth odor molecule) is calculated, the similarity between the high-dimensional feature representation of the second odor molecule and the high-dimensional feature representation of the other three odor molecules (i.e., the third odor molecule, the fourth odor molecule, and the fifth odor molecule) is calculated, the similarity between the high-dimensional feature representation of the third odor molecule and the high-dimensional feature representation of the other two odor molecules (i.e., the fourth odor molecule and the fifth odor molecule) is calculated, and the similarity between the high-dimensional feature representation of the fourth odor molecule and the high-dimensional feature representation of the fifth odor molecule is calculated. After the similarity between two-by-two is calculated, if it is found that the similarity between the first odor molecule and the fifth odor molecule exceeds the above-mentioned similarity threshold, then the first odor molecule and the fifth odor molecule are considered as similar molecules, and then a weighted average of the similarity between the first odor molecule and the other second, third, and fourth odor molecules is calculated, a weighted average of the similarity between the fifth odor molecule and the other second, third, and fourth odor molecules is calculated, and then the two weighted average results are compared, the similar molecule with a smaller weighted average result is retained in the mixture candidate, and the similar molecule with a larger weighted average result is removed, thereby obtaining a final mixture. Further, when the weighted average results are equal, then one of the two similar molecules is randomly selected and retained in the mixture candidate. Conversely, if the similarity between any two odor molecules does not exceed the above-mentioned similarity threshold, then it is indicated that there is no close odor in the mixture candidate, and thus the mixture candidate does not need to be processed and is directly considered as the final mixture.

[0090] It should be noted that in step S32 of the embodiment, the cosine similarity is used to measure the similarity of two odor molecules, but in addition to the cosine similarity, other indicators can also be selected by those skilled in the art as long as they can measure the similarity. In order to facilitate those skilled in the art to understand, in the embodiment, the cosine similarity between two odor molecules is calculated as follows:

[0091]

[0092] wherein, and represent any two odor molecules in the mixture candidate; represents the odor molecule and the cosine similarity between them; represents the dot product; is a preset similarity threshold.

[0093] It should be noted that in S3 of the present application, the specific process of calculating the concentration of a candidate molecule is as follows: the perceived intensity score and the volatilization rate of the candidate molecule are multiplied to obtain the comprehensive score of the candidate molecule, and the ratio of the comprehensive score of the candidate molecule to the comprehensive scores of all candidate molecules is taken as the concentration of the candidate molecule.

[0094] In the embodiment, the concentration of the first candidate molecule is calculated as follows:

[0095]

[0096] wherein, and represent the perceived intensity score and the volatilization rate of the first candidate molecule; and represent the perceived intensity score and the volatilization rate of the first candidate molecule. After calculating the above concentration, the molecular graph of all candidate molecules is spliced according to the concentration weight to generate a mixture molecular graph for odor prediction. It should be noted that in the above mixture construction, the present application combines the node set and edge set of multiple monomer odor molecules by molecular splicing and index expansion. In order to ensure the efficiency of computer processing, each odor molecule is assigned a unique molecular ID when input, and the odor molecule ID is marked in the node index, so that subsequent calculations can distinguish nodes from different sources.

[0097] In addition, it should be noted that in the above mixture construction, the present application combines the node set and edge set of multiple monomer odor molecules by molecular splicing and index expansion. In order to ensure the efficiency of computer processing, each odor molecule is assigned a unique molecular ID when input, and the odor molecule ID is marked in the node index, so that subsequent calculations can distinguish nodes from different sources.

[0098] S4, mixture olfactory label prediction and knowledge base update ​​​

[0099] The constructed mixture molecular graph is input into a label prediction model for supervised learning. In the supervised learning, a pseudo-real olfactory perception label of the final mixture is generated according to the olfactory perception label of the candidate molecule, the pseudo-real olfactory perception label is used as a supervision signal, the local interaction features between atoms are extracted by a second graph neural network in the model, the local interaction features are aggregated by a global pooling operation to form a global vector representation of the final mixture, and finally a feedforward neural network in the model is used for regression mapping of the global vector representation, so as to predict the olfactory perception label of the final mixture, and the final mixture and the corresponding olfactory perception label are added to the odor knowledge base.

[0100] It should be noted that in S4 of the present application, after the construction of the mixture molecular graph, the embodiment enters the graph neural network modeling phase. The core task of this step is to input the mixture molecular graph into a computer executable deep learning framework, aggregate the features of atoms and bonds by using a graph neural network, and output the global vector representation of the mixture, and finally regress to the olfactory perception label.

[0101] Specifically, in the mixture olfactory perception label prediction process, first, the graph loading phase is performed. In this phase, multiple mixture molecular graphs are batched (batching), packaged into a unified tensor input to GPU for parallel computing, and then the mixture molecular graph is input into the label prediction model. In the feature updating phase, the model iteratively updates the features of each atom node and its neighbor nodes by using a multi-layer graph attention network (GAT), calculates the important weights of neighbor nodes to the center node by using a self-attention mechanism, so as to capture the key interaction relationship between molecules and between molecules, and obtains the global vector representation of the mixture by using a global pooling operation. In the prediction phase, a feedforward neural network (MLP) is used to regress map the global vector representation to obtain the predicted olfactory perception label.

[0102] It should be noted that in S4 of the present application, the second graph neural network described above adopts a graph attention network (Graph Attention Network, GAT), and the implementation of the graph attention network belongs to the prior art and will not be described again.

[0103] In addition, it should be noted that in the supervised learning of S4 of the present application, an adaptive focal loss function is calculated between the pseudo-real olfactory perception label and the predicted olfactory perception label, and the label prediction model parameters are updated based on minimizing the adaptive focal loss function.

[0104] In this embodiment S4, the error between the predicted label (the predicted olfactory perception label mentioned above) and the true label (the pseudo-true olfactory perception label mentioned above) is calculated by defining an adaptive focal loss function. The model parameters are iteratively updated using gradient descent, allowing the predicted label to gradually approach the true label. Furthermore, to enhance training performance, rare fragrance labels or difficult-to-predict combinations are weighted, making the model pay more attention to complex mixtures that are difficult to classify, thus improving prediction accuracy and generalization ability. The adaptive focal loss function... The calculation method is existing technology. In this embodiment, the following formula can be used for calculation:

[0105]

[0106] in, Indicates the total number of tags. This indicates the weight of the label. Indicates the predicted probability of the label; This indicates the actual label.

[0107] To better demonstrate the specific implementation and technical effects of the present invention, the method for constructing an odor knowledge base based on molecular modeling and deep learning, as shown in steps S1 to S4 of the above preferred implementation, is applied to a specific example below.

[0108] Example

[0109] The specific implementation process of the odor knowledge base construction method based on molecular modeling and deep learning used in this embodiment is as described above and will not be repeated here.

[0110] like Figure 2 As shown in the figure, this invention demonstrates its predictive performance on olfactory tags in a mixed odor molecule olfactory tag prediction task, where the vertical axis represents the prediction probability of each olfactory tag (range 0-1), and the horizontal axis represents the olfactory tag number (in...). Figure 2 (These labels are hidden to highlight the label distribution trend). The bar colors are used to distinguish high-prediction-probability labels from the remaining labels: dark gray bars represent the 10 olfactory perception labels with the highest predicted probabilities and are labeled with specific olfactory descriptions, while light gray bars represent the predicted probabilities of the remaining labels. High-prediction-probability labels are mainly concentrated in odors with obvious sensory characteristics, such as fresh, fruity, woody, and creamy scents, while low-prediction-probability labels are distributed between 0 and 0.3, including herbal and earthy odor types. In this embodiment, the Pearson correlation coefficient R is 0.63, indicating a moderate linear correlation between the model's predicted probabilities and the actual labels.

[0111] As shown in Table 1, the table shows the structure and characteristic code of part of the odor molecules in the odor molecule knowledge base of the application. Each odor molecule contains atomic node features, chemical bond edge features, and three-dimensional position codes generated by the Mol-PECO method to support the prediction and analysis of odor attributes by a deep learning model. The above atomic node features are represented by the vector [atomic number, hybridization type, valence, charge]; the Mol-PECO position code is the three-dimensional coordinate obtained after the space normalization of the odor molecule; the perception intensity score and the volatilization rate are standardized values between 0 and 1, which are used to represent the strength of the odor characteristics and the volatilization trend of the odor molecules.

[0112] Table 1. Structure and characteristic code of part of the odor molecules in the odor knowledge base of the application

[0113]

[0114] Table 2 lists some examples of odor molecule formulations in the odor knowledge base of the application, including the composition of each formulation, the chemical formula, the proportion, and the corresponding olfactory perception label. Through Table 2, the correspondence between each odor molecule in the odor formulation and its olfactory characteristics can be clearly understood, providing a reference basis for the odor prediction, generation and combination of the application. The specific description is as follows:

[0115] Formulation number: used to identify different odor formulations, such as Mix01, Mix02, Mix03, etc.

[0116] Odor molecule name: the odor molecules contained in each formulation, such as isoamyl alcohol, cinnamyl alcohol, benzyl alcohol, etc.

[0117] Odor molecule chemical formula: the chemical structure representation corresponding to each odor molecule.

[0118] Proportion: the proportion of each odor molecule in the formulation, used to describe the relative contribution of each odor molecule in the formulation.

[0119] Olfactory label: the olfactory perception label corresponding to each formulation, such as "floral", "citrus", "woody", "soap", "mint", etc., used to describe the overall olfactory characteristics of the odor molecules.

[0120] Table 2. Composition and olfactory label correspondence table of part of the odor molecule formulations in the odor knowledge base of the application

[0121]

[0122] In summary, as Figure 3As shown, the present application realizes unified representation from single odor molecule to mixture and high-precision olfactory label prediction by constructing an odor knowledge base containing multi-dimensional information such as odor molecule graph structure, odor molecule position encoding, olfactory perception label, perception intensity and volatile characteristics, and combining a prediction model based on graph attention mechanism. The knowledge base not only breaks through the limitations of the prior art on single physical and chemical properties, but also supports large-scale retrieval, expansion and training, significantly improving the intelligent level of application scenarios such as odor prediction, mixture blending and personalized recommendation.

[0123] In addition, it should be noted that the odor knowledge base construction method based on molecular modeling and deep learning in the above embodiment can be essentially executed by a computer program or module. Therefore, based on the same inventive concept, another preferred embodiment of the present application also provides an odor knowledge base construction system based on molecular modeling and deep learning corresponding to the odor knowledge base construction method based on molecular modeling and deep learning provided by the above embodiment, as shown in Figure 4 As shown, it comprises:

[0124] A data acquisition module is configured to acquire odor molecule data, perform standardization processing, generate an olfactory perception label and quantitative characteristics corresponding to each odor molecule, and use the odor molecule and the olfactory perception label and quantitative characteristics corresponding thereto to construct an odor knowledge base. The quantitative characteristics include the perception intensity score and the volatile rate of each odor molecule.

[0125] A molecular graph generation module is configured to take atoms in each odor molecule as nodes in a molecular graph structure and encode, take chemical bonds in the odor molecule as edges in the molecular graph structure and encode, and introduce a molecular position encoding method to model the three-dimensional geometric configuration of the odor molecule, so as to represent the chemical structure information of the odor molecule as a molecular graph structure, and then input the molecular graph into a trained first graph neural network to obtain a high-dimensional feature representation containing the chemical structure and spatial information of each odor molecule. The molecular graph is composed of a node feature matrix and an edge feature matrix.

[0126] A mixture molecular graph generation module is configured to select several odor molecules from the odor knowledge base to generate a mixture candidate, generate a final mixture according to the similarity between two odor molecules in the mixture candidate, take the odor molecules in the final mixture as candidate molecules, calculate the concentration of the candidate molecules according to the perception intensity score and the volatile rate, weight the molecular graphs of all candidate molecules according to the concentration as a weight, and splice all weighted molecular graphs to form a mixture molecular graph.

[0127] The prediction module is configured to input the constructed mixture molecule graph into a label prediction model for supervised learning, in which a pseudo-real olfactory perception label of the candidate molecule is generated as a final mixture according to an olfactory perception label of the candidate molecule, the pseudo-real olfactory perception label is used as a supervision signal, local interaction features between atoms are extracted by a second graph neural network in the model, the local interaction features are aggregated by a global pooling operation to form a global vector representation of the final mixture, and finally, a feedforward neural network in the model is used to regress and map the global vector representation, so as to predict the olfactory perception label of the final mixture, and the final mixture and the corresponding olfactory perception label are added to the odor knowledge base.

[0128] Similarly, based on the same inventive concept, another preferred embodiment of the present application also provides a computer electronic device corresponding to the odor knowledge base construction method based on molecular modeling and deep learning provided by the above-mentioned embodiment, as shown in the figure, which comprises a memory and a processor. Figure 5

[0129] The memory is configured to store a computer program.

[0130] The processor is configured to implement the odor knowledge base construction method based on molecular modeling and deep learning in the above-mentioned embodiment when executing the computer program.

[0131] In addition, the logical instructions in the above-mentioned memory can be realized in the form of a software functional unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application.

[0132] It can be understood that the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. ​

[0133] In addition, it should be noted that, for the convenience and brevity of the description, the specific working process of the system described above can refer to the corresponding process in the foregoing method embodiments, and will not be described here. In the embodiments provided in the present application, the division of steps or modules in the system and method described is only a logical functional division, and there can be another division manner in actual implementation, for example, multiple modules or steps can be combined or integrated together, or a module or step can be split.

[0134] The above-described embodiments are only a preferred scheme of the present application, and are not intended to limit the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application. Therefore, any technical scheme obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present application.

Claims

1. A method for constructing an odor knowledge base based on molecular modeling and deep learning, characterized in that, Includes the following steps: S1. After acquiring odor molecule data and performing standardization processing, generate olfactory perception labels and quantitative features corresponding to each odor molecule, and use odor molecules and their corresponding olfactory perception labels and quantitative features to construct an odor knowledge base; The quantitative features include a perceived intensity score and evaporation rate for each odor molecule; S2. The atoms in each odor molecule are treated as nodes in the molecular graph structure and encoded, and the chemical bonds in the odor molecule are treated as edges in the molecular graph structure and encoded. At the same time, a molecular position encoding method is introduced to model the three-dimensional geometric configuration of the odor molecule, thereby representing the chemical structure information of the odor molecule as a molecular graph structure. The molecular graph is then input into a trained first graph neural network to obtain a high-dimensional feature representation containing the chemical structure and spatial information of each odor molecule. The molecular graph consists of a node feature matrix and an edge feature matrix. S3. Select several odor molecules from the odor knowledge base to generate mixture candidates. Generate the final mixture based on the similarity between two odor molecules in the mixture candidates. Use the odor molecules in the final mixture as candidate molecules. Calculate the concentration of the candidate molecules based on their perceived intensity score and evaporation rate. Use the concentration as a weight to weight the molecular graphs of all candidate molecules. Then, stitch together all the weighted molecular graphs to form the mixture molecular graph. S4. Input the constructed mixture molecular graph into the label prediction model for supervised learning. In supervised learning, generate corresponding pseudo-real olfactory perception labels for the final mixture based on the olfactory perception labels of the candidate molecules. Use the pseudo-real olfactory perception labels as supervision signals. Extract the local interaction features between atoms by the second graph neural network in the model, and aggregate the local interaction features through global pooling to form the global vector representation of the final mixture. Finally, use the feedforward neural network in the model to perform regression mapping on the global vector representation to predict the olfactory perception label of the final mixture. Add the final mixture and its corresponding olfactory perception label to the odor knowledge base.

2. The method for constructing an odor knowledge base based on molecular modeling and deep learning as described in claim 1, characterized in that, During the S1 olfactory perception tag collection process, each odor molecule was independently sensorily evaluated by at least three perfumers to obtain an olfactory perception tag. The perceived intensity score and evaporation rate of the gas molecules were obtained through multiple rounds of experimental measurements. In the standardization process, the evaporation rate was normalized to the [0,1] interval according to the maximum value, the perceived intensity score was converted into a percentage value according to the preset standard score, and the olfactory perception tag was consistently corrected in combination with the expert voting results.

3. The method for constructing an odor knowledge base based on molecular modeling and deep learning as described in claim 1, characterized in that, In S2, node encoding requires encoding the atom type, charge, hybridization state, and topological characteristics; edge encoding requires encoding the chemical bond type, bond length, and bond sequence.

4. The method for constructing an odor knowledge base based on molecular modeling and deep learning as described in claim 3, characterized in that, In S2, the specific process of obtaining the molecular map is as follows: First, the original three-dimensional coordinates of each atom in the odor molecule are translated according to the geometric center of the odor molecule, so that the spatial position of the odor molecule is aligned with the geometric center; then, the original three-dimensional coordinates are scaled to normalize the scale so that all odor molecules are represented within a uniform scale range, thereby obtaining the processed three-dimensional coordinates of each atom; after the processed three-dimensional coordinates are position-encoded, the position-encoded results are fused with the node feature vectors by concatenation to obtain the fused features of each atom; Then, a fusion vector is constructed from the fusion features of all atoms belonging to the same odor molecule. A node feature matrix is ​​constructed from the fusion vectors of all odor molecules, and an edge feature matrix is ​​constructed from the edge feature vectors of all atoms, thus completing the construction of the molecular graph. Among them, the node feature vector is the feature vector obtained by encoding each node, and the edge feature vector is the feature vector obtained by encoding each edge.

5. The method for constructing an odor knowledge base based on molecular modeling and deep learning as described in claim 1, characterized in that, In S3, the specific process for generating the final mixture is as follows: S31. Select several odor molecules from the odor knowledge base to generate mixture candidates, and form a mixture vector from the high-dimensional feature representations of the odor molecules in the mixture candidates. S32. In the mixture vector, calculate the similarity between the high-dimensional feature representations pairwise: when there is a similarity result that is greater than the preset similarity threshold, the two odor molecules corresponding to the similarity result are regarded as similar molecules, and the similarity of each similar molecule and the remaining odor molecules in the mixture candidate is weighted and averaged. The similar molecules with the smaller weighted average result are retained to form the final mixture; when the similarity results are all less than the similarity threshold, the mixture candidate is regarded as the final mixture.

6. The method for constructing an odor knowledge base based on molecular modeling and deep learning as described in claim 5, characterized in that, In S32, cosine similarity is used to measure the similarity between two odor molecules.

7. The method for constructing an odor knowledge base based on molecular modeling and deep learning as described in claim 1, characterized in that, In S3, the specific process for calculating the concentration of a candidate molecule is as follows: multiply the perceived intensity score and the evaporation rate of the candidate molecule to obtain the comprehensive score of the candidate molecule, and then take the ratio of the comprehensive score of the candidate molecule to the comprehensive score of all candidate molecules as the concentration of the candidate molecule.

8. The method for constructing an odor knowledge base based on molecular modeling and deep learning as described in claim 1, characterized in that, In the supervised learning of S4, an adaptive focus loss function is calculated based on the error between the pseudo-real olfactory perception label and the predicted olfactory perception label, and the label prediction model parameters are updated based on minimizing the adaptive focus loss function.

9. A system for constructing an odor knowledge base based on molecular modeling and deep learning, characterized in that, include: The data acquisition module is used to acquire odor molecule data, perform standardization processing, generate olfactory perception labels and quantitative features corresponding to each odor molecule, and construct an odor knowledge base using odor molecules and their corresponding olfactory perception labels and quantitative features. The quantitative features include a perceived intensity score and evaporation rate for each odor molecule; The molecular graph generation module encodes the atoms in each odor molecule as nodes and the chemical bonds as edges in the molecular graph structure. It also introduces a molecular position encoding method to model the three-dimensional geometry of the odor molecules, thus representing the chemical structure information of the odor molecules as a molecular graph structure. This molecular graph is then input into a pre-trained first graph neural network to obtain a high-dimensional feature representation containing the chemical structure and spatial information of each odor molecule. The molecular graph consists of a node feature matrix and an edge feature matrix. The mixture molecular graph generation module is used to select several odor molecules from the odor knowledge base to generate mixture candidates, generate the final mixture based on the similarity between two odor molecules in the mixture candidates, take the odor molecules in the final mixture as candidate molecules, calculate their concentration based on the perceived intensity score and evaporation rate of the candidate molecules, use the concentration as a weight to weight the molecular graphs of all candidate molecules, and stitch together all weighted molecular graphs to form the mixture molecular graph. The prediction module is used to input the constructed molecular graph of the mixture into the label prediction model for supervised learning. In supervised learning, based on the olfactory perception labels of the candidate molecules, corresponding pseudo-real olfactory perception labels are generated for the final mixture. The pseudo-real olfactory perception labels are used as supervision signals. The second graph neural network in the model extracts the local interaction features between atoms and aggregates the local interaction features through global pooling to form a global vector representation of the final mixture. Finally, the feedforward neural network in the model is used to perform regression mapping on the global vector representation to predict the olfactory perception label of the final mixture. The final mixture and its corresponding olfactory perception label are then added to the odor knowledge base.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement, when executing the computer program, the method for constructing an odor knowledge base based on molecular modeling and deep learning as described in any one of claims 1 to 8.

Images