An olfactory perception method and device based on two-dimensional and three-dimensional molecular structure fusion

Abstract

The application discloses an olfactory perception method and device based on two-dimensional and three-dimensional molecular structure fusion, and relates to the technical field of information perception and intelligent modeling. The method realizes effective description of a complex mapping relationship between a molecular structure and human olfactory perception by jointly modeling two-dimensional topological structure and three-dimensional spatial configuration characteristics of an odor molecule; connection relationship, functional group distribution and spatial geometric configuration information of the molecule are extracted by using multi-level molecular structure representation, and are fused and represented in a unified high-dimensional perception space, so that physical and chemical consistency of olfactory perception characteristics is enhanced; a continuous and smooth olfactory perception representation space is adaptively constructed, so that expression capability of the model for fine-grained odor difference and complex odor cases is improved. The method can realize more stable and accurate olfactory perception prediction in single-molecule odor prediction and mixed odor modeling tasks.

Description

Technical Field

[0001] This invention relates to the field of information perception and intelligent modeling technology, and more specifically to an olfactory perception method and device based on the fusion of two-dimensional and three-dimensional molecular structures. Background Technology

[0002] Olfactory perception is one of the most important sensory ways for humans to obtain information from the outside world, and it has wide applications in fields such as odor recognition, fragrance and flavor design, environmental monitoring, and digital olfaction. This task aims to establish a model that can characterize the mapping relationship between the structural features of odor molecules and human subjective olfactory perception, and to predict the olfactory properties of odor molecules computationally, thereby achieving objective modeling and analysis of the olfactory perception process.

[0003] Existing olfactory perception modeling methods typically predict odor tags corresponding to molecules based on molecular structure information. However, due to the high complexity of odor molecules in terms of functional group composition, topological structure, and spatial configuration, and the lack of a simple one-to-one correspondence between molecular structure and human olfactory perception, existing olfactory perception modeling methods still face many challenges. Some methods mainly rely on two-dimensional molecular structures or artificially constructed molecular descriptors, which are insufficient to fully characterize the geometric configuration and stereoscopic effects of molecules in three-dimensional space, resulting in insufficient ability of the models to express fine-grained olfactory differences and complex odor situations. In addition, there are often correlations and uneven distributions between different odor perception dimensions, further increasing the difficulty of olfactory perception modeling.

[0004] Therefore, how to construct an olfactory perception modeling method that can simultaneously reflect molecular topological and spatial configuration features based on fully utilizing molecular structural information, and improve the model's predictive consistency with human subjective olfactory perception, remains a technical problem that urgently needs to be solved in the field of olfactory information perception. Summary of the Invention

[0005] In view of the above problems, this invention is proposed to provide an olfactory perception method and device based on the fusion of two-dimensional and three-dimensional molecular structures to overcome or at least partially solve the above problems. This method jointly models the two-dimensional topological structure and three-dimensional spatial configuration of odor molecules, extracting multi-level structural features of molecules such as atomic connections, functional group distribution, and spatial geometry. The molecular structural information is then analyzed in a unified perceptual representation space, thereby effectively characterizing the complex mapping relationship between molecular structure and human olfactory perception. By fusing molecular features from different structural levels, this method enhances the model's ability to express complex odor situations, making the olfactory perception prediction results more consistent with the characteristics of human subjective olfactory perception.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide an olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures, comprising the following steps: S1. Obtain the SMILES string of the olfactory molecule to be modeled, construct a two-dimensional topological structure representation of the molecule based on the SMILES string, and obtain or generate a three-dimensional molecular geometric structure representation of the corresponding molecule. S2. Input the two-dimensional topological structure representation into a two-dimensional molecular graph neural network to extract the two-dimensional structural feature representation of the molecule; S3. Input the three-dimensional molecular geometric structure representation into a three-dimensional molecular neural network to extract the three-dimensional structural feature representation of the molecule; S4. Based on the two-dimensional and three-dimensional structural feature representations, a bidirectional pre-training process is performed through a diffusion model to generate two-dimensional and three-dimensional structures from each other, resulting in a pre-trained graph neural network model that integrates two-dimensional and three-dimensional molecular structural information. S5. Input the two-dimensional topological structure diagram of the molecule to be predicted into the pre-trained graph neural network model, and output the molecular depth feature representation containing the implicit three-dimensional structural information; S6. Input the molecular deep feature representation into the perception prediction module composed of a multi-layer fully connected neural network, and output the prediction results of the molecule in multiple olfactory dimensions. S7. Using the trained perception prediction module, for an odor mixture composed of multiple single molecules, the molecular depth feature representation of each single molecule is obtained and spliced ​​together to obtain the feature representation of the mixture; by calculating the cosine similarity between the feature representations of two mixtures, the odor similarity is predicted.

[0008] In one embodiment, in step S1, constructing a two-dimensional topological representation of the molecule based on the SMILES string includes: Based on the SMILES string, the olfactory molecule is parsed to construct a two-dimensional topological structure diagram of the molecule, where nodes are used to represent atoms in the molecule and edges are used to represent chemical bond relationships between atoms. The atom type and bond type are encoded into corresponding node features and edge features to form a two-dimensional topological structure representation.

[0009] In one embodiment, step S2 includes: The atom types and chemical bond types in the two-dimensional topological structure representation are encoded as node features and edge features; By using a two-dimensional molecular graph neural network through multi-layer graph convolution or transfer mechanisms, local and global structural information between atoms in a molecule is modeled, and a two-dimensional structural feature representation of the molecule is extracted.

[0010] In one embodiment, step S3 includes: The three-dimensional molecular geometry representation is input into a three-dimensional molecular neural network to model the spatial positions of atoms and their relative relationships, and to extract the three-dimensional structural features of the molecule, which are used to characterize the structural information of the molecule at the spatial structure level.

[0011] In one embodiment, the bidirectional pre-training in step S4 includes the following two processes: The process of generating three dimensions from two dimensions: using the two-dimensional structural feature representation as a condition, the three-dimensional atomic coordinates are denoised and denoised through a diffusion model, and the three-dimensional spatial configuration is learned from the two-dimensional topological structure. The process of generating a two-dimensional structure from a three-dimensional structure: using the three-dimensional structural feature representation as a condition, the atomic and bond properties are denoised and denoised through a diffusion model, and a two-dimensional topological structure is generated from the three-dimensional spatial configuration.

[0012] In one embodiment, during the two-dimensional to three-dimensional generation process, a scoring network is used to guide the denoising process. This scoring network calculates atomic scores based on the two-dimensional features and spatial direction vectors between atoms, which are used to gradually recover the three-dimensional coordinates.

[0013] In one embodiment, in step S6, the perception prediction module is a multi-layer fully connected neural network that outputs the probability of a molecule belonging to each olfactory dimension through a Softmax function.

[0014] Secondly, embodiments of the present invention provide an olfactory sensing device based on the fusion of two-dimensional and three-dimensional molecular structures, comprising: Molecular structure information acquisition module: acquires the SMILES string of the olfactory molecule to be modeled, constructs a two-dimensional topological structure representation of the molecule based on the SMILES string, and acquires or generates a three-dimensional molecular geometric structure representation of the corresponding molecule; Two-dimensional structure feature extraction module: Input the two-dimensional topological structure representation into a two-dimensional molecular graph neural network to extract the two-dimensional structural feature representation of the molecule; Three-dimensional structural feature extraction module: Input the three-dimensional molecular geometric structure representation into the three-dimensional molecular neural network to extract the three-dimensional structural feature representation of the molecule; Cross-modal diffusion pre-training module: Based on the two-dimensional and three-dimensional structural feature representations, a bidirectional pre-training process is performed through the diffusion model to generate two-dimensional and three-dimensional structures from each other, resulting in a pre-trained graph neural network model that integrates two-dimensional and three-dimensional molecular structural information; Deep feature extraction module: Input the two-dimensional topological structure map of the molecule to be predicted into the pre-trained graph neural network model, and output the molecular deep feature representation containing implicit three-dimensional structural information; Single-molecule olfactory prediction module: The molecular deep feature representation is input into the perception prediction module composed of a multi-layer fully connected neural network, and the prediction results of the molecule in multiple olfactory dimensions are output. Odor Similarity Prediction Module for Mixtures: Using the trained perception prediction module, for an odor mixture composed of multiple single molecules, the molecular depth feature representation of each single molecule is obtained and spliced ​​together to obtain the feature representation of the mixture; by calculating the cosine similarity between the feature representations of two mixtures, the odor similarity is predicted.

[0015] Thirdly, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures as described in any one of the first aspects.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures as described in any one of the first aspects.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages: First, this invention introduces both the two-dimensional topological structure and the three-dimensional spatial configuration information of molecules to jointly model the structural features of odor molecules. Compared with olfactory perception methods that rely solely on two-dimensional molecular descriptors or single structural information, this invention can more comprehensively characterize key structural features such as atomic connections, functional group distribution, and spatial geometric configuration in molecules, thereby improving the physical rationality and chemical consistency of molecular structure representation.

[0018] Secondly, this invention introduces a diffusion modeling mechanism based on the fusion features of two-dimensional and three-dimensional molecular structures. By performing random perturbation and reverse recovery modeling in the structural feature space, the model can learn the inherent laws of molecular structure distribution during noise perturbation. This effectively alleviates the problems of discontinuous perceptual representation and insufficient generalization ability in traditional supervised learning methods, and constructs a smoother and more stable olfactory perception representation space, which is conducive to improving the model's ability to model complex odor structures.

[0019] Furthermore, this invention extracts the structural depth features of molecules based on pre-trained two-dimensional and three-dimensional fused graph neural networks, and realizes single-molecule olfactory perception prediction through molecular-level embedding representation and multilayer perception prediction module. This enables the model to model multiple olfactory perception dimensions simultaneously, enhances the ability to distinguish fine-grained olfactory differences, and makes the prediction results more consistent with the characteristics of human subjective olfactory perception.

[0020] Furthermore, this invention addresses the characteristic that odor mixtures are composed of multiple single molecules. By splicing together the olfactory perception features of each single molecule and introducing a prediction mechanism based on similarity measurement, it achieves modeling and prediction of the odor similarity of mixtures. Compared with the traditional method of simplifying mixed odors into a single feature, this method can more effectively reflect the comprehensive perception relationship between different molecules in the mixture, and improve the accuracy and stability of mixed odor perception prediction.

[0021] In summary, the olfactory perception modeling method based on the fusion of two-dimensional and three-dimensional molecular structures proposed in this invention has significant advantages in terms of the completeness of molecular structure expression, consistency of olfactory perception, and applicability to single-molecule and mixed odors. It can provide an effective technical means for digital olfaction, odor recognition, and related intelligent sensing systems. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0023] Figure 1 This is a flowchart of an olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures provided in an embodiment of the present invention. Figure 2 This is an overall architecture diagram of olfactory perception modeling based on the fusion of two-dimensional and three-dimensional molecular structures provided in the embodiments of the present invention; Figure 3 This is a block diagram of an olfactory sensing device based on the fusion of two-dimensional and three-dimensional molecular structures provided in an embodiment of the present invention. Figure 4 This is a structural diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: This invention discloses an olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures, referring to... Figure 1As shown, the process includes the following steps S1 to S7. The entire process uses multimodal fusion and diffusion modeling to make olfactory prediction more in line with human perception.

[0026] S1. Molecular structure information acquisition steps: Obtain the SMILES string of the olfactory molecule to be modeled, construct a two-dimensional topological representation of the molecule based on the SMILES string, and obtain or generate a three-dimensional molecular geometric representation of the corresponding molecule. For example, use the RDKit toolkit to parse the topological information of the molecular diagram, including atom types and chemical bond types, to construct a two-dimensional topological representation of the molecule; simultaneously, generate a three-dimensional molecular geometric representation of the corresponding molecule, including the three-dimensional spatial coordinates of the atoms, through a conformation generation algorithm.

[0027] The purpose of this step is to obtain the basic structural data of odor molecules. Specifically, this includes: inputting the SMILES string of the molecule, which is a standard format for representing molecular structure in text; then parsing the olfactory molecule based on the SMILES to construct a two-dimensional topological structure, represented by a graph, where nodes represent atoms and edges represent chemical bonds; and encoding the atom type and bond type into corresponding node features and edge features to form a two-dimensional topological structure representation.

[0028] Simultaneously, the three-dimensional spatial configuration information of molecules is generated or acquired, that is, the coordinates of atoms in three-dimensional space, forming a three-dimensional molecular geometric structure representation.

[0029] S2, Two-dimensional structural feature extraction step: Input the two-dimensional topological structure representation into a two-dimensional molecular graph neural network to extract the two-dimensional structural feature representation of the molecule; Specifically, it includes: 1) Encode the atom types and chemical bond types in the two-dimensional topological structure representation as node features and edge features; 2) By using a two-dimensional molecular graph neural network through multi-layer graph convolution or transfer mechanism, the local and global structural information between atoms in the molecule is modeled, and the two-dimensional structural feature representation of the molecule is extracted.

[0030] The two-dimensional topological representation obtained from S1 is input into a two-dimensional molecular graph neural network, such as GCN, GIN, or GAT. This network encodes features such as atom types and chemical bond types through multi-layer graph convolution or message passing mechanisms, learning the local and global connectivity relationships between atoms in the molecule, and ultimately outputting a two-dimensional structural feature representation. This feature focuses on the topological information of the molecule, characterizing its structural information at the level of topological connectivity, such as the distribution of functional groups, ignoring spatial shape, but providing crucial input for subsequent fusion.

[0031] For example, the 2D molecular graph neural network (2D GNN) adopts the graph isomorphic network (GIN) architecture, which contains multiple stacked graph convolutional layers. Each layer aggregates and updates the features of neighboring nodes through a multilayer perceptron and the ReLU activation function, thereby extracting a two-dimensional structural feature representation containing local chemical structure and global topological information.

[0032] S3. Three-dimensional structural feature extraction step: Input the three-dimensional molecular geometric structure representation into the three-dimensional molecular neural network to extract the three-dimensional structural feature representation of the molecule; This step involves inputting a three-dimensional molecular geometric representation into a three-dimensional molecular neural network to model the spatial positions of atoms and their relative relationships, extracting the three-dimensional structural features of the molecule, and using this representation to characterize the molecule's structural information at the spatial structural level.

[0033] Among them, three-dimensional molecular neural networks such as SchNet or SE(3)-Transformer directly process the three-dimensional coordinates of atoms, model the spatial distance and relative orientation between atoms, such as bond angles and dihedral angles, and output a three-dimensional structural feature representation. That is, it models the interaction of atoms in three-dimensional space and extracts a three-dimensional structural feature representation that is invariant to rotation and translation. This step captures the stereoscopic effects and geometric configuration of molecules, makes up for the deficiencies of two-dimensional features, and enables the model to better simulate the sensitivity of the human nose to molecular shapes.

[0034] S4. Cross-modal diffusion pre-training step: Based on the two-dimensional structural feature representation and the three-dimensional structural feature representation, bidirectional pre-training is performed through the diffusion model to generate two-dimensional and three-dimensional structures from each other, resulting in a pre-trained graph neural network model that integrates two-dimensional and three-dimensional molecular structural information. This step introduces a diffusion modeling mechanism to perform multi-scale random perturbation and modeling on the fusion features. During the back diffusion process, conditional constraints are constructed based on two-dimensional and three-dimensional structural features, and a cross-modal scoring network is introduced to model the dependencies between different modal structural information.

[0035] The aforementioned bidirectional pre-training includes the following two processes: 1) Generation process from two-dimensional to three-dimensional: Based on the two-dimensional structural feature representation, the three-dimensional atomic coordinates are denoised and denoised through a diffusion model to learn how to generate a three-dimensional spatial configuration from a two-dimensional topological structure; for example, a scoring network is used to guide the denoising process. This scoring network calculates the atomic score based on the two-dimensional features and spatial direction vectors between atoms, which is used to gradually recover the three-dimensional coordinates.

[0036] The diffusion generation process from 2D to 3D first uses multi-layer graph convolution to aggregate local features of atoms and bonds, capturing chemical information such as the bonding environment and bond relationships. Then, noisy 3D molecular coordinates corresponding to the 2D molecule are added to generate noisy 3D coordinates. Next, a scoring network with physical constraints generates a denoising gradient to gradually recover the 3D coordinates. The formula is expressed as:

[0037] in, The index after multi-layer graph convolution aggregation is 2D feature vectors of atoms,

[0038]

[0039]

[0040]

[0041]

[0042] in, The direction between the two atoms. For vertical direction, For vertical and The directions that make up the plane. Indicates the first A vector of atoms The vector representing the j-th atom. This means that the noisy three-dimensional coordinates are projected into the above coordinate system to reflect the spatial attitude and ensure equivariance.

[0043]

[0044] in, For the first The score of each atom is used to guide the denoising process, ultimately generating clean 3D coordinates. For Hadama accumulation.

[0045] 2) Generation process from three-dimensional to two-dimensional: Based on the three-dimensional structural feature representation, the atomic and bond properties are denoised and denoised through a diffusion model, and two-dimensional topological structures are generated from three-dimensional spatial configurations.

[0046] The diffusion generation process from 3D to 2D first extracts 3D feature codes using a 3D molecular neural network, then simultaneously adds noise to atomic and bond properties. A two-branch fractional prediction method is used to generate denoised gradients for atoms and bonds, gradually recovering the true 2D topological structure. The formula is expressed as:

[0047] in, For noisy atomic properties, It is a three-dimensional feature encoding. It is aggregated through multiple dense GCNs. To obtain multi-scale features , ... Multi-scale features are spliced ​​together and then processed by an MLP to output atomic-level scores. .

[0048]

[0049] in, For atoms and The higher the feature similarity value between them, the higher the probability of bonding. Ultimately, this will... Multi-scale bonding features are concatenated and then processed by an MLP to output edge-level scores. .

[0050] S5. Deep feature extraction step: Input the two-dimensional topological structure diagram of the molecule to be predicted into the pre-trained graph neural network model, and output the molecular deep feature representation containing the implicit three-dimensional structural information; The two-dimensional topological structure map of the olfactory molecule to be predicted is input into a pre-trained graph neural network model, i.e., a trained 2D GNN encoder. Since the encoder has incorporated three-dimensional geometric knowledge through cross-modal pre-training in step S4, its output molecular depth feature representation can still implicitly contain the three-dimensional spatial configuration information and physicochemical properties of the molecule even when only a two-dimensional map is input.

[0051] For a graph neural network that has been pre-trained by mutual generation of two-dimensional and three-dimensional structures, the two-dimensional molecular graph structure data to be predicted is input into the graph neural network. Through a multi-layer message passing mechanism, the local and global structural information between atoms in the molecule is modeled, and the two-dimensional structural depth feature representation of the implicit three-dimensional information of the molecule is extracted.

[0052]

[0053] Among them, is A two-dimensional representation of the depth features of odor molecules that contain the extracted implicit three-dimensional information.

[0054] After the diffusion model pre-training is completed in step S4, a graph neural network that integrates two-dimensional and three-dimensional information is used to process the input two-dimensional structure data. Through a multi-layer message passing mechanism, the network can extract "two-dimensional depth features with implicit three-dimensional information". This means that even if only a two-dimensional structure is input, the model can indirectly utilize three-dimensional prior knowledge to output richer feature representations, preparing for the prediction stage.

[0055] S6. Single-molecule olfactory prediction step: Input the molecular deep feature representation into the perception prediction module composed of a multi-layer fully connected neural network, and output the prediction results of the molecule in multiple olfactory dimensions. The perception prediction module consists of a multi-layer fully connected neural network (MLP), with dropout layers between layers to prevent overfitting, and ReLU is used as the activation function. The output layer is connected to a softmax classifier, which outputs the probability prediction results of the molecule in multiple olfactory dimensions such as "floral", "fruity", and "woody".

[0056] Step S6 specifically includes: 1) Perform a global readout operation on the node-level features output by the graph neural network model, and aggregate the node features into molecular-level feature vectors to obtain molecular-level embedding representations that can characterize the overall structural properties of the molecule. 2) Input the molecular-level embedded representation into the perception prediction module composed of a multi-layer fully connected neural network, and further model the molecular structure features through nonlinear mapping to generate a high-dimensional feature representation related to olfactory perception; 3) Using the trained olfactory perception modeling network, the olfactory perception characteristics of unknown molecules are predicted, and the response results of molecules in multiple olfactory perception dimensions are output, thereby realizing the modeling and prediction of molecular olfactory perception characteristics.

[0057] The molecular deep feature representation from step S5 is input into a perception prediction module composed of a multi-layer fully connected neural network. The molecular structural features are further modeled through nonlinear mapping to generate high-dimensional feature representations related to olfactory perception. Finally, the odor is classified using a softmax function, thus achieving single-molecule olfactory perception prediction. For example, the final output shows the molecule's prediction results in multiple olfactory dimensions, such as pleasant fragrance or putrid odor. This step achieves the classification or regression prediction of single-molecule odors, which is the direct application goal of the method.

[0058] S7. Odor Similarity Prediction Steps for Mixtures: Using the trained perception prediction module, for an odor mixture composed of multiple single molecules, the molecular depth feature representation of each single molecule is obtained and spliced ​​together to obtain the feature representation of the mixture; by calculating the cosine similarity between the feature representations of two mixtures, the odor similarity is predicted.

[0059] Since the odor mixture is composed of multiple single molecules, the high-dimensional feature representations related to olfactory perception generated in step S6 are spliced ​​together to obtain the spliced ​​features corresponding to the odor of the mixture; the spliced ​​features are input into the cosine similarity prediction head to obtain the similarity scores of the two sets of odor mixtures through supervised learning.

[0060] The formula is expressed as:

[0061]

[0062] in, This refers to the splicing characteristics of molecules in a mixture. This represents the corresponding two-dimensional odor molecule depth characteristics. The similarity score between the odors of two mixtures is measured by a cosine similarity function. These represent the characteristics of the two mixtures, respectively.

[0063] For real-world mixtures such as perfumes, a pre-trained S6 network is first used to extract the olfactory features of each individual molecule, and then these features are concatenated into a long vector. Finally, cosine similarity is used to calculate the similarity score between the features of two mixtures. This method avoids simplifying mixtures into single entities, better reflects the synergistic effects between molecules, and improves the accuracy of mixed odor prediction.

[0064] Reference Figure 2 The diagram shown illustrates the overall architecture of the olfactory perception modeling based on the fusion of two-dimensional and three-dimensional molecular structures according to the present invention. Taking a specific molecule, vanillin, as an example, the process of steps S1 to S7 of the present invention is used to illustrate the prediction of the odor of vanillin: 1. Obtaining the SMILES string: First, obtain the SMILES string for vanillin: COc1cc(C=O)ccc1O; 2. Two-dimensional topology construction: The algorithm parses SMILES into a graph G=(V, E); 3. Generation of 3D geometry: Obtain its 3D coordinates using RDKit's built-in library functions; 4. Feature Extraction Stage: The above two-dimensional image is input into GIN to extract two-dimensional features, resulting in a latent variable vector representing two-dimensional information. The three-dimensional coordinates are then input into SchNet to obtain a latent variable vector representing three-dimensional information.

[0065] 5. Pre-training phase: Using a diffusion model to connect two-dimensional and three-dimensional models.

[0066] 2D to 3D: Input the two-dimensional features of vanillin and random noise, and the model restores the three-dimensional coordinates of the atoms to the correct positions from the noise, given the two-dimensional topology.

[0067] 3D to 2D: Input the three-dimensional features of vanillin and random noise. The model infers whether the atoms are connected by single or double bonds and the type of atoms based on the point cloud of the atoms' spatial arrangement.

[0068] Comparative learning: bringing the two-dimensional and three-dimensional features of the same molecule closer together in the feature space, while pushing the two-dimensional and three-dimensional features of different molecules further apart.

[0069] 6. Olfactory Prediction Stage: Now, a new molecule arrives, and we only have its SMILES string (two-dimensional data) as input into the trained GIN. Because it has been trained using S4, the model automatically "associates" its three-dimensional spatial structure when processing the two-dimensional image of a new molecule. The output feature vector is analyzed by a fully connected layer to output the final odor prediction probability: for example, sweetness: 98%, milkiness: 55%, woodiness: 10%.

[0070] This invention provides an olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures. By jointly modeling the two-dimensional topological structure and three-dimensional spatial configuration features of odor molecules, it effectively characterizes the complex mapping relationship between molecular structure and human olfactory perception. It utilizes multi-level molecular structure characterization to extract molecular connectivity, functional group distribution, and spatial geometric configuration information, and performs fusion representation in a unified high-dimensional perception space to enhance the physical and chemical consistency of olfactory perception features. By adaptively constructing a continuous and smooth olfactory perception representation space, it improves the model's ability to express fine-grained odor differences and complex odor situations. This method can achieve more stable and accurate olfactory perception prediction in single-molecule odor prediction and mixed odor modeling tasks.

[0071] Example 2: Based on the same inventive concept, this invention also provides an olfactory sensing device based on the fusion of two-dimensional and three-dimensional molecular structures. Since the principle of solving the problem by these devices is similar to that of the aforementioned olfactory sensing method based on the fusion of two-dimensional and three-dimensional molecular structures, the implementation of this device can refer to the implementation of the aforementioned method, and the repeated parts will not be described again.

[0072] like Figure 3 As shown, the device includes: Molecular structure information acquisition module: acquires the SMILES string of the olfactory molecule to be modeled, constructs a two-dimensional topological structure representation of the molecule based on the SMILES string, and acquires or generates a three-dimensional molecular geometric structure representation of the corresponding molecule; Two-dimensional structure feature extraction module: Input the two-dimensional topological structure representation into a two-dimensional molecular graph neural network to extract the two-dimensional structural feature representation of the molecule; Three-dimensional structural feature extraction module: Input the three-dimensional molecular geometric structure representation into the three-dimensional molecular neural network to extract the three-dimensional structural feature representation of the molecule; Cross-modal diffusion pre-training module: Based on the two-dimensional and three-dimensional structural feature representations, a bidirectional pre-training process is performed through the diffusion model to generate two-dimensional and three-dimensional structures from each other, resulting in a pre-trained graph neural network model that integrates two-dimensional and three-dimensional molecular structural information; Deep feature extraction module: Input the two-dimensional topological structure map of the molecule to be predicted into the pre-trained graph neural network model, and output the molecular deep feature representation containing implicit three-dimensional structural information; Single-molecule olfactory prediction module: The molecular deep feature representation is input into the perception prediction module composed of a multi-layer fully connected neural network, and the prediction results of the molecule in multiple olfactory dimensions are output. Odor Similarity Prediction Module for Mixtures: Using the trained perception prediction module, for an odor mixture composed of multiple single molecules, the molecular depth feature representation of each single molecule is obtained and spliced ​​together to obtain the feature representation of the mixture; by calculating the cosine similarity between the feature representations of two mixtures, the odor similarity is predicted.

[0073] Example 3: Based on the same inventive concept, the present invention also provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it is able to implement the olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures as described in any one of Embodiments 1.

[0074] like Figure 4 As shown, the electronic device may include: a processor 10, a communication interface 20, a memory 30, and a communication bus 40, wherein the processor 10, the communication interface 20, and the memory 30 communicate with each other via the communication bus 40. The processor 10 can call logical instructions in the memory 30 to execute an olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures, the method including: S1. Obtain the SMILES string of the olfactory molecule to be modeled, construct a two-dimensional topological structure representation of the molecule based on the SMILES string, and obtain or generate a three-dimensional molecular geometric structure representation of the corresponding molecule. S2. Input the two-dimensional topological structure representation into a two-dimensional molecular graph neural network to extract the two-dimensional structural feature representation of the molecule; S3. Input the three-dimensional molecular geometric structure representation into a three-dimensional molecular neural network to extract the three-dimensional structural feature representation of the molecule; S4. Based on the two-dimensional and three-dimensional structural feature representations, a bidirectional pre-training process is performed through a diffusion model to generate two-dimensional and three-dimensional structures from each other, resulting in a pre-trained graph neural network model that integrates two-dimensional and three-dimensional molecular structural information. S5. Input the two-dimensional topological structure diagram of the molecule to be predicted into the pre-trained graph neural network model, and output the molecular depth feature representation containing the implicit three-dimensional structural information; S6. Input the molecular deep feature representation into the perception prediction module composed of a multi-layer fully connected neural network, and output the prediction results of the molecule in multiple olfactory dimensions. S7. Using the trained perception prediction module, for an odor mixture composed of multiple single molecules, the molecular depth feature representation of each single molecule is obtained and spliced ​​together to obtain the feature representation of the mixture; by calculating the cosine similarity between the feature representations of two mixtures, the odor similarity is predicted.

[0075] Example 4: This invention also provides a computer-readable storage medium containing a program for executing an olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures described in Embodiment 1. The program can be executed on a processor.

[0076] Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] The program stored on this medium is loaded into the processor's memory and executed to perform various functions. This storage medium, connected to hardware devices, enables the computer to perform the steps of Embodiment 1 described above.

[0078] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0079] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures, characterized in that, Includes the following steps: S1. Obtain the SMILES string of the olfactory molecule to be modeled, construct a two-dimensional topological structure representation of the molecule based on the SMILES string, and obtain or generate a three-dimensional molecular geometric structure representation of the corresponding molecule. S2. Input the two-dimensional topological structure representation into a two-dimensional molecular graph neural network to extract the two-dimensional structural feature representation of the molecule; S3. Input the three-dimensional molecular geometric structure representation into a three-dimensional molecular neural network to extract the three-dimensional structural feature representation of the molecule; S4. Based on the two-dimensional and three-dimensional structural feature representations, a bidirectional pre-training process is performed through a diffusion model to generate two-dimensional and three-dimensional structures from each other, resulting in a pre-trained graph neural network model that integrates two-dimensional and three-dimensional molecular structural information. S5. Input the two-dimensional topological structure diagram of the molecule to be predicted into the pre-trained graph neural network model, and output the molecular depth feature representation containing the implicit three-dimensional structural information; S6. Input the molecular deep feature representation into the perception prediction module composed of a multi-layer fully connected neural network, and output the prediction results of the molecule in multiple olfactory dimensions. S7. Using the trained perception prediction module, for an odor mixture composed of multiple single molecules, the molecular depth feature representation of each single molecule is obtained and spliced ​​together to obtain the feature representation of the mixture; by calculating the cosine similarity between the feature representations of two mixtures, the odor similarity is predicted.

2. The method as described in claim 1, characterized in that, In step S1, constructing a two-dimensional topological representation of the molecule based on the SMILES string includes: Based on the SMILES string, the olfactory molecule is parsed to construct a two-dimensional topological structure diagram of the molecule, where nodes are used to represent atoms in the molecule and edges are used to represent chemical bond relationships between atoms. The atom type and bond type are encoded into corresponding node features and edge features to form a two-dimensional topological structure representation.

3. The method as described in claim 2, characterized in that, Step S2 includes: The atom types and chemical bond types in the two-dimensional topological structure representation are encoded as node features and edge features; By using a two-dimensional molecular graph neural network through multi-layer graph convolution or transfer mechanisms, local and global structural information between atoms in a molecule is modeled, and a two-dimensional structural feature representation of the molecule is extracted.

4. The method as described in claim 1, characterized in that, Step S3 includes: The three-dimensional molecular geometry representation is input into a three-dimensional molecular neural network to model the spatial positions of atoms and their relative relationships, and to extract the three-dimensional structural features of the molecule, which are used to characterize the structural information of the molecule at the spatial structure level.

5. The method as described in claim 1, characterized in that, The bidirectional pre-training in step S4 includes the following two processes: The process of generating three dimensions from two dimensions: using the two-dimensional structural feature representation as a condition, the three-dimensional atomic coordinates are denoised and denoised through a diffusion model, and the three-dimensional spatial configuration is learned from the two-dimensional topological structure. The process of generating a two-dimensional structure from a three-dimensional structure: using the three-dimensional structural feature representation as a condition, the atomic and bond properties are denoised and denoised through a diffusion model, and a two-dimensional topological structure is generated from the three-dimensional spatial configuration.

6. The method as described in claim 5, characterized in that, In the process of generating two-dimensional to three-dimensional data, a scoring network is used to guide the denoising process. This scoring network calculates the atomic score based on the two-dimensional features and spatial direction vectors between atoms, which is used to gradually recover the three-dimensional coordinates.

7. The method as described in claim 1, characterized in that, In step S6, the perception prediction module is a multi-layer fully connected neural network that outputs the probability of a molecule belonging to each olfactory dimension through the Softmax function.

8. An olfactory sensing device based on the fusion of two-dimensional and three-dimensional molecular structures, characterized in that, include: Molecular structure information acquisition module: acquires the SMILES string of the olfactory molecule to be modeled, constructs a two-dimensional topological structure representation of the molecule based on the SMILES string, and acquires or generates a three-dimensional molecular geometric structure representation of the corresponding molecule; Two-dimensional structure feature extraction module: Input the two-dimensional topological structure representation into a two-dimensional molecular graph neural network to extract the two-dimensional structural feature representation of the molecule; Three-dimensional structural feature extraction module: Input the three-dimensional molecular geometric structure representation into the three-dimensional molecular neural network to extract the three-dimensional structural feature representation of the molecule; Cross-modal diffusion pre-training module: Based on the two-dimensional and three-dimensional structural feature representations, a bidirectional pre-training process is performed through the diffusion model to generate two-dimensional and three-dimensional structures from each other, resulting in a pre-trained graph neural network model that integrates two-dimensional and three-dimensional molecular structural information; Deep feature extraction module: Input the two-dimensional topological structure map of the molecule to be predicted into the pre-trained graph neural network model, and output the molecular deep feature representation containing implicit three-dimensional structural information; Single-molecule olfactory prediction module: The molecular deep feature representation is input into the perception prediction module composed of a multi-layer fully connected neural network, and the prediction results of the molecule in multiple olfactory dimensions are output. Odor Similarity Prediction Module for Mixtures: Using the trained perception prediction module, for an odor mixture composed of multiple single molecules, the molecular depth feature representation of each single molecule is obtained and spliced ​​together to obtain the feature representation of the mixture; by calculating the cosine similarity between the feature representations of two mixtures, the odor similarity is predicted.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the olfactory perception method based on the fusion of two-dimensional and three-dimensional molecular structures as described in any one of claims 1 to 7.

Images