Olfactory perception prediction method based on path enhancement and olfactory relationship fusion
By introducing path enhancement and olfactory relation fusion methods into olfactory perception prediction, and utilizing molecular graph self-attention modeling and olfactory attribute relation graphs, the problem of characterizing the relationship between molecular structure and olfactory attributes is solved, thereby improving the accuracy and consistency of olfactory perception prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-29
AI Technical Summary
Existing olfactory perception prediction methods are unable to effectively characterize the relationship between pathway structure relationships in molecular structures and olfactory attributes, resulting in poor prediction performance.
We employ a method based on path enhancement and olfactory relationship fusion. By introducing topological distance, molecular fragment relationships and atom pair path structure features through molecular graph self-attention modeling, and combining olfactory attribute relationship graph and label conditional attention mechanism, we can achieve multi-layer contextual modeling and prediction of molecular olfactory perception attributes.
It improves the accuracy and consistency of molecular olfactory perception prediction, better characterizes the relationship between molecular structure and olfactory properties, and enhances the stability of multi-label prediction and the unified expression of molecular olfactory perception.
Smart Images
Figure CN122117153A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of molecular perception modeling and artificial intelligence technology, and more specifically to an olfactory perception prediction method based on the fusion of path enhancement and olfactory relationship. Background Technology
[0002] Olfactory perception prediction is a fundamental problem in the fields of computational olfaction and chemical information processing, and is widely used in scenarios such as flavor and fragrance research and development, food flavor molecule screening, daily chemical product scent design, and environmental volatile substance identification. This task establishes a computational model that can characterize the mapping relationship between molecular structure and olfactory properties, and predicts the olfactory perception characteristics of molecules in a numerical manner, thereby assisting the molecular screening and design process in related fields.
[0003] Currently, most existing olfactory perception prediction methods are based on molecular structural features to model molecules, including molecular descriptors, molecular fingerprints, and deep learning methods based on molecular graphs. With the development of deep learning technology, some methods have begun to comprehensively utilize local structural information and overall topological information of molecules to perform multi-label prediction of olfactory attributes.
[0004] However, how to effectively extract molecular structural features, fully characterize the path structure relationships between atoms in the molecule, and reasonably model the correlation between different olfactory attributes remains a challenging problem in the field of olfactory perception prediction. Summary of the Invention
[0005] In view of the above problems, this invention provides an olfactory perception prediction method based on the fusion of path enhancement and olfactory relationship, in order to overcome or at least partially solve the above problems. Specifically, by introducing topological distance, molecular fragment relationships, and atom pair path structure features into the molecular graph self-attention modeling process, multi-layer contextual modeling of molecular structure is performed; simultaneously, by constructing an olfactory attribute relationship graph and combining it with a label conditional attention mechanism, the correlation between molecular structure information and olfactory attributes is comprehensively characterized, thereby achieving effective prediction of molecular olfactory perception attributes.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] This invention provides an olfactory perception prediction method based on the fusion of path enhancement and olfactory relationship, the steps of which include: S1. Determine the molecular graph input representation based on the molecular graph of the target molecule. The molecular graph input representation includes atomic features, interatomic distance information, molecular segment relationships, and atom pair path structure features. S2. Perform multi-layer self-attention modeling on the molecular graph input representation to obtain node-level structural features; the modeling introduces inter-atomic distance bucket bias, molecular fragment bias, and fused atom pair path structural features; S3. Using olfactory attribute labels as conditions, perform label-conditional attention aggregation on node-level structural features to obtain label-specific molecular representations; S4. Calculate the prediction score based on the label-specific molecular representation and the olfactory attribute label, and output the olfactory perception prediction result of the target molecule.
[0008] Preferably, in S1, the atomic characteristics include at least one of atomic element type, atomic degree, number of hydrogen atoms, and valence state related characteristics.
[0009] Preferably, in S1, the interatomic distance information is obtained by calculating the topological shortest path distance between any pair of atoms in the molecular graph, and the distance is discretely bucketed and encoded.
[0010] Preferably, in S1, the molecular fragment relationship includes structural segmentation of the molecule based on molecular fragment partitioning rules to generate a molecular fragment relationship mask, which is used to indicate whether a pair of atoms belongs to the same molecular fragment.
[0011] Preferably, in S1, the atom pair path structure feature includes at least one or more of the following features: Discrete coding features of shortest path length; Statistical characteristics of different types of chemical bonds along the shortest path; Characteristics of the proportion of aromatic bonds, conjugated bonds, or intracyclic bonds along the shortest path; Indicative features of aromatic or non-aromatic conjugated bonds on the shortest path; Path directionality characteristics based on the properties of atomic donors and acceptors.
[0012] Preferably, in S2, the modeling incorporates interatomic distance binning bias and molecular fragment bias, including:
[0013] In the formula, Indicates atomic pairs and Attention weights between and Indicates the atomic node index. Indicates the first The query vector of atomic feature mappings Indicates the first The bond vector of each atom feature map. Indicates transpose. This represents the feature dimensions of the query vector and the key vector. Indicates atomic pairs and The bias term for the topological distance bucket mapping between nodes. Indicates atomic pairs and Bias term for the relationship between molecular fragments.
[0014] Preferably, in S2, the fused atom pair path structure features include: The atomic pair path structure features are mapped to path value vectors; The enhanced vector is obtained by fusing the standard attention value vector and the path value vector through a gating function; Based on the enhancement vector, node-level structural features are generated according to the following formula:
[0015] In the formula, Indicates the first The atomic node at the th th Layer feature representation, This indicates the atom pairs after introducing interatomic distance binning bias and molecular fragment bias. and Attention weights between This represents the augmentation vector.
[0016] Preferably, S3, label conditional attention aggregation includes:
[0017]
[0018] In the formula, For the first Individual atomic node-level structural features For the first An olfactory attribute label, Aggregate weights for attention. This represents a label-specific molecule.
[0019] Preferably, in S4, the calculation process includes:
[0020]
[0021] In the formula, For the first An olfactory attribute label, This represents a label-specific molecule. Representation of tag-specific molecules With the Each olfactory attribute label represents The inner product similarity score, A learnable scaling factor. For learnable olfactory attributes, Indicates the target molecule at the th Predicted scores on each olfactory attribute label.
[0022] Preferably, in S3, the olfactory attribute labels are obtained by constructing an olfactory attribute relationship graph and then using a label relationship graph attention encoding module based on the olfactory attribute relationship graph; wherein, the olfactory attribute relationship graph is constructed based on prior associations between olfactory attribute labels, and the prior associations include one or more of the following methods: Fixed relationships based on co-occurrence statistics of olfactory attribute labels; Introduce learnable residuals based on fixed relationships; Relationships are dynamically generated based on olfactory attribute tags to represent similarity.
[0023] This invention provides an olfactory perception prediction method based on the fusion of path enhancement and olfactory relationship. It can comprehensively utilize molecular structure information, path topological semantics, and olfactory attribute correlation to achieve effective prediction of molecular olfactory perception results, and has good practicality and promotion value.
[0024] Compared with existing technologies, the beneficial effects include at least the following: 1. This invention represents the structural information of the target molecule as a molecular graph and constructs atomic features, interatomic topological distances, molecular fragment relationships, and atom pair path structure features, so that the local chemical environment and overall topological information in the molecular structure can be modeled in a unified framework. The above-mentioned structural features contain important information related to olfactory perception, providing a sufficient structural expression basis for subsequent olfactory attribute prediction.
[0025] 2. This invention introduces distance bucket bias and molecular fragment bias in the self-attention modeling process of molecular graphs, so that while performing global structure modeling, the model can more reasonably focus on the atomic pair relationships within different topological distance ranges and enhance the characterization of the internal structural consistency of molecular fragments, which helps to improve the modeling effect of functional groups and local structural features.
[0026] 3. This invention introduces a path feature value enhancement mechanism to map the structural features of the shortest path between atomic pairs into path value vectors, and adaptively fuses them with the standard value vectors in the self-attention mechanism. This enables the model to effectively utilize the chemical semantic information contained in the path structure of molecules without significantly increasing computational complexity, thereby improving the ability to express long-range molecular structure dependencies.
[0027] 4. This invention constructs an olfactory attribute relationship graph and uses a label relationship graph attention encoding method to perform contextual modeling of olfactory attribute labels, so that the correlation between different olfactory attributes can be reflected in the prediction process, thereby improving the overall consistency and stability of prediction results in multi-label prediction scenarios.
[0028] 5. This invention adopts a label-conditional attention aggregation method, using olfactory attribute label representation as a condition to perform weighted aggregation of molecular node-level structural features, thereby obtaining label-specific molecular representations. This allows different olfactory attributes to focus on the structural regions in the molecule that are related to them, which helps to enhance the model's ability to distinguish between different olfactory attributes.
[0029] 6. By calculating the prediction score based on the label-specific molecular representation and the olfactory attribute label representation, a unified prediction output of the olfactory perception attributes of the target molecule is achieved. Attached Figure Description
[0030] 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.
[0031] Figure 1 This is a flowchart of the olfactory perception prediction method based on path enhancement and olfactory relationship fusion provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the olfactory perception prediction method provided in the embodiments of the present invention. Detailed Implementation
[0032] 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.
[0033] This invention discloses an olfactory perception prediction method based on the fusion of path enhancement and olfactory relationship, which is used to effectively characterize the correlation between molecular structure information, path topological semantics and olfactory attributes in multi-label olfactory prediction tasks, thereby achieving effective prediction of molecular olfactory perception results.
[0034] like Figure 1 As shown, the olfactory perception prediction method of this invention includes the following steps: S1. Determine the molecular graph input representation based on the molecular graph of the target molecule. The molecular graph input representation includes atomic features, interatomic distance information, molecular segment relationships, and atom pair path structure features. S2. Perform multi-layer self-attention modeling on the molecular graph input representation to obtain node-level structural features; the modeling introduces inter-atomic distance bucket bias, molecular fragment bias, and fused atom pair path structural features; S3. Using olfactory attribute labels as conditions, perform label-conditional attention aggregation on node-level structural features to obtain label-specific molecular representations; S4. Calculate the prediction score based on the label-specific molecule representation and the olfactory attribute label representation, and output the olfactory perception prediction result of the target molecule.
[0035] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0036] like Figure 2 As shown, the olfactory perception prediction model in this embodiment includes a molecular graph structure encoding branch and an olfactory attribute label encoding branch. The outputs of the two branches are fused in the label conditional attention aggregation module, and the prediction results of each olfactory attribute are output through the prediction score calculation module. Specifically, the molecular graph structure encoding branch takes atomic features as input and combines inter-atomic topological distance binning bias, molecular fragment bias, and atomic pair path structure features to obtain node-level structural feature representation through an N-layer fragment perception graph Transformer encoding layer. The olfactory attribute label encoding branch takes learnable label prototypes as initial representations and combines the label adjacency matrix, i.e., the matrix representation of the olfactory attribute relationship graph, in the label graph attention layer to perform contextual modeling and obtain a contextualized olfactory attribute label embedding representation. Subsequently, using the label embedding representation as a condition, attention aggregation is performed on the node-level structural features to obtain a label-specific molecular representation, and the prediction score is calculated based on the similarity between the label-specific molecular representation and the corresponding label representation to output the olfactory perception prediction result.
[0037] In one embodiment, S1: Determine the molecular graph input representation based on the molecular graph of the target molecule.
[0038] In this embodiment, the molecular structure information of the target molecule is first obtained, which can be represented as the string "SMILES". Then, the molecular structure information is parsed into a molecular graph structure, where atoms in the molecule are nodes and chemical bonds between atoms are edges.
[0039] Furthermore, the molecular graph input representation includes atomic features, interatomic distance information, molecular segment relationships, and atom pair path structure features.
[0040] In some implementations, for each atomic node, its corresponding atomic features are extracted. These atomic features include at least one or more of the following: atomic element type, atomic degree, number of hydrogen atoms, and valence state-related features. To facilitate subsequent batch processing, a maximum atomic number threshold can be set, and insufficient numbers can be padded with zeros to generate a corresponding node validity mask.
[0041] In some implementations, arbitrary atom pairs in the molecular diagram are calculated. The shortest path distance between atoms is used to obtain the inter-atom distance matrix, and the distance is then discretely bucketed to reduce the impact of the distance range on model stability. It should be noted that this embodiment uses existing bucketing methods; the core is not to improve the bucketing method itself, but to map the encoding result into a bias term and introduce it into self-attention modeling.
[0042] In some implementation schemes, the molecule is structurally segmented based on preset molecular fragmentation rules. For example, based on the BRICS fragmentation concept, chemical bonds in the molecule that meet preset breaking conditions are broken, dividing the molecule into several structural fragments; and each atom is labeled with the molecular fragment number to which it belongs. Further, a molecular fragment relationship mask is generated based on the molecular fragment number to which the atom belongs. When two atoms have the same molecular fragment number, their corresponding mask value is set to a first preset value; otherwise, it is set to a second preset value, used to indicate whether the atom pair belongs to the same molecular fragment.
[0043] At the same time, for each pair of atoms Atom-pair path structure features are constructed to describe the structural composition information of the shortest path between two atoms. These path structure features may include shortest path length encoding, statistical characteristics of different types of chemical bonds along the path, aromatic / conjugated / intracyclic bond ratio characteristics, aromaticity or conjugation indication characteristics, and donor-acceptor directionality characteristics. All of these path structure features can be automatically extracted based on molecular diagram structural information without the need for additional experimental measurements.
[0044] In one embodiment, S2: Perform multi-layer self-attention modeling on the molecular graph input representation to obtain node-level structural features; such as Figure 2 As shown, this process is implemented by an N-layer segment-aware graph Transformer encoding layer. The output node embedding representation of the encoding layer serves as the key vector and value vector input to the subsequent label conditional attention aggregation module. This embodiment includes a molecular graph self-attention encoding module for multi-layer self-attention modeling of the molecular graph input representation. During the modeling process, inter-atom distance bucketing bias, molecular segment bias, and fused atom pair path structure features are introduced.
[0045] In a preferred embodiment, for the first The layer self-attention module maps atomic features to node embeddings of a uniform dimension, and then maps them to query vectors. Key vector Sum value vector The calculation method is as follows:
[0046] In the formula, This represents the node representation output from the previous layer. 、 、 is a learnable parameter matrix.
[0047] In the attention calculation process, this application introduces interatomic distance binning bias and molecular fragment bias to enhance the modeling of structural correlations within molecular fragments. The attention weights can be expressed as:
[0048] In the formula, Indicates atomic pairs and Attention weights between and Indicates the atomic node index. Indicates the first The query vector of atomic feature mappings Indicates the first The bond vector of each atom feature map. Indicates transpose. This represents the feature dimensions of the query vector and the key vector. Indicates atomic pairs and The bias term of the inter-topological distance bucketing mapping is used to guide the self-attention mechanism to assign different attention weights to atomic pairs with different topological distances. Indicates atomic pairs and The bias term for the relationship between molecular segments is used to make the self-attention mechanism tend to form stronger structural associations within molecular segments.
[0049] Furthermore, a path feature value enhancement mechanism is introduced to fuse the path structure features of atomic pairs, thereby further integrating the path structure information of atomic pairs.
[0050] In this embodiment, the atom pair path structure features are first mapped to path value vectors. ;
[0051] in Indicates atomic pairs Path structure characteristics, It is a linear mapping function.
[0052] Subsequently, the standard attention value vector and the path value vector are fused using a gating function to obtain the enhancement vector; the formula is expressed as:
[0053]
[0054] In the formula, This represents the Sigmoid function. For learnable gating functions, Indicates that it consists of atomic nodes The standard attention value vector obtained by linear mapping of node-level features. Indicates a pair of atoms The gating weight coefficients generated from the path structure features are used to adjust the relative contributions of the standard attention value vector and the path value vector in the fusion process; Finally, node-level structural features are generated based on the enhancement vector according to the following formula:
[0055] In the formula, Indicates the first The atomic node at the th th Layer feature representation, This indicates the atom pairs after introducing interatomic distance binning bias and molecular fragment bias. and Attention weights between This represents the augmentation vector.
[0056] In one embodiment, S3: Using olfactory attribute labels as conditions, perform label-conditional attention aggregation on node-level structural features to obtain label-specific molecular representations; such as Figure 2 As shown, the olfactory attribute labels are first initialized by the label prototype, and then combined with the label adjacency matrix, i.e. the matrix representation of the olfactory attribute relationship graph, in the label graph attention layer to perform contextual modeling, and obtain the contextualized label embedding representation. The label embedding representation is used as a query vector input to the label conditional attention aggregation module.
[0057] In this embodiment, an olfactory attribute relationship graph is constructed based on the prior association between olfactory attribute tags, and an information interaction model is performed on the olfactory attribute tags through a tag relationship graph attention encoding module based on the olfactory attribute relationship graph to obtain an olfactory attribute tag representation that integrates tag association information.
[0058] Among them, the label relationship graph attention encoding module models the contextual interaction of olfactory attribute labels through a multi-head attention mechanism, and introduces the olfactory attribute relationship graph as a structural constraint during the attention calculation process.
[0059] Meanwhile, the prior association relationship includes one or more of the following methods: Fixed relationships constructed based on co-occurrence statistics of olfactory attribute tags; Introduce learnable residuals based on fixed relationships; Relationships are dynamically generated based on olfactory attribute tags to represent similarity.
[0060] For the The calculation method for the corresponding tag-specific molecule representation of each olfactory attribute tag is as follows:
[0061]
[0062] In the formula, For the first Individual atomic node-level structural features For the first An olfactory attribute label, Aggregate weights for attention. This represents a label-specific molecule.
[0063] In one embodiment, S4: Calculate a prediction score based on the tag-specific molecule representation and the olfactory attribute tag representation, and output the olfactory perception prediction result of the target molecule. The calculation process includes:
[0064]
[0065] In the formula, For the first An olfactory attribute label, This represents a label-specific molecule. Representation of tag-specific molecules With the Each olfactory attribute label represents The inner product similarity score, A learnable scaling factor. For learnable olfactory attributes, Indicates the target molecule at the th Predicted scores on each olfactory attribute label.
[0066] In one embodiment, the predicted scores corresponding to each olfactory attribute are obtained. Then, the predicted scores are mapped using a nonlinear mapping function to obtain the predicted probability of each olfactory attribute; preferably, the nonlinear mapping function is a sigmoid function. Further, the predicted probabilities are compared with a preset threshold. If the predicted probability of a certain olfactory attribute is greater than or equal to the preset threshold, the target molecule is determined to possess that olfactory attribute; otherwise, the target molecule is determined not to possess that olfactory attribute, thus obtaining the olfactory perception prediction result of the target molecule. Alternatively, the predicted probabilities of all olfactory attributes are sorted, and the top K olfactory attributes with the highest predicted probabilities are selected as the olfactory perception prediction results of the target molecule.
[0067] Based on the same inventive concept, this invention also provides an olfactory perception prediction device based on the fusion of path enhancement attention and olfactory relationship. Optionally, the device includes a molecular graph input representation determination module, a self-attention modeling module, an olfactory tag aggregation module, and an olfactory perception prediction module. The principle of the problem solved by the above modules is consistent with the aforementioned olfactory perception prediction method based on the fusion of path enhancement and olfactory relationship. Therefore, the implementation of this device can refer to the implementation of the aforementioned method, and the repeated parts will not be described again.
[0068] 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.
[0069] 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 prediction method based on path enhancement and olfactory relationship fusion, characterized in that, include: S1. Determine the molecular graph input representation based on the molecular graph of the target molecule. The molecular graph input representation includes atomic features, interatomic distance information, molecular segment relationships, and atom pair path structure features. S2. Perform multi-layer self-attention modeling on the molecular graph input representation to obtain node-level structural features; The modeling incorporates inter-atomic distance bucket bias, molecular fragment bias, and fused atom pair path structure features; S3. Using olfactory attribute labels as conditions, perform label-conditional attention aggregation on node-level structural features to obtain label-specific molecular representations; S4. Calculate the prediction score based on the label-specific molecular representation and the olfactory attribute label, and output the olfactory perception prediction result of the target molecule.
2. The olfactory perception prediction method according to claim 1, characterized in that, In S1, the atomic characteristics include at least one of the following: atomic element type, atomic degree, number of hydrogen atoms, and valence state related characteristics.
3. The olfactory perception prediction method according to claim 1, characterized in that, In S1, the inter-atomic distance information is obtained by calculating the topological shortest path distance between any pair of atoms in the molecular graph, and the distance is then discretely bucketed and encoded.
4. The olfactory perception prediction method according to claim 1, characterized in that, In S1, the molecular fragment relationship includes structural segmentation of molecules based on molecular fragment partitioning rules, generating a molecular fragment relationship mask to indicate whether atom pairs belong to the same molecular fragment.
5. The olfactory perception prediction method according to claim 1, characterized in that, In S1, the atom pair path structure feature includes at least one or more of the following features: Discrete coding features of shortest path length; Statistical characteristics of different types of chemical bonds along the shortest path; Characteristics of the proportion of aromatic bonds, conjugated bonds, or intracyclic bonds along the shortest path; Indicative features of aromatic or non-aromatic conjugated bonds on the shortest path; Path directionality characteristics based on the properties of atomic donors and acceptors.
6. The olfactory perception prediction method according to claim 1, characterized in that, In S2, the modeling incorporates interatomic distance binning bias and molecular fragment bias, including: In the formula, Indicates atomic pairs and Attention weights between them and Indicates the atomic node index. Indicates the first The query vector of atomic feature mappings Indicates the first The bond vector of each atom feature map. Indicates transpose. This represents the feature dimensions of the query vector and the key vector. Indicates atomic pairs and The bias term for the topological distance bucket mapping between nodes. Indicates atomic pairs and Bias term for the relationship between molecular fragments.
7. The olfactory perception prediction method according to claim 1 or 6, characterized in that, In S2, the fused atom pair path structure features include: The atomic pair path structure features are mapped to path value vectors; The enhancement vector is obtained by fusing the standard attention value vector and the path value vector through a gating function; Based on the enhancement vector, node-level structural features are generated according to the following formula: In the formula, Indicates the first The atomic node at the th th Layer feature representation, This indicates the atom pairs after introducing interatomic distance binning bias and molecular fragment bias. and Attention weights between them This represents the augmentation vector.
8. The olfactory perception prediction method according to claim 1, characterized in that, In S3, label-conditional attention aggregation includes: In the formula, For the first Individual atomic node-level structural features For the first An olfactory attribute label, Aggregate weights for attention. This represents a label-specific molecule.
9. The olfactory perception prediction method according to claim 1, characterized in that, In S4, the calculation process includes: In the formula, For the first An olfactory attribute label, This represents a label-specific molecule. Representation of tag-specific molecules With the Each olfactory attribute label represents The inner product similarity score, A learnable scaling factor. For learnable olfactory attributes, Indicates the target molecule at the th Predicted scores on each olfactory attribute label.
10. The olfactory perception prediction method according to claim 1, characterized in that, In S3, the olfactory attribute labels are obtained by constructing an olfactory attribute relationship graph and then using a label relationship graph attention encoding module based on the olfactory attribute relationship graph; wherein, the olfactory attribute relationship graph is constructed based on prior associations between olfactory attribute labels, and the prior associations include one or more of the following methods: Fixed relationships based on co-occurrence statistics of olfactory attribute labels; Introduce learnable residuals based on fixed relationships; Relationships are dynamically generated based on olfactory attribute tags to represent similarity.