Molecular property prediction method based on spectral position encoding and biomimetic lateral inhibition gating

Abstract

The application discloses a molecular property prediction method based on spectral position coding and bionic side inhibition gating, and relates to the field of olfactory perception. The method comprises the following steps: constructing a molecular structured input comprising atomic types, a Coulomb matrix and graph Laplacian eigenvector; fusing atomic chemical attributes and spectral domain topological positions through a feature extraction module to generate initial atomic features; using a multi-scale aggregation module, hierarchical neighborhood aggregation is carried out based on preset scale constraints to extract multi-level structure features covering chemical bonds, functional groups and molecular skeletons; with the help of a global interaction module of bionic side inhibition gating, local features are dynamically modulated by a global query signal to realize adaptive denoising and semantic sharpening; finally, through mask pooling and a decoupled multi-label prediction head, the scores of each odor attribute of the molecule are output. The application significantly improves the accuracy, structure perception ability and cross-task generalization performance of molecular property prediction.

Description

Technical Field

[0001] This invention relates to the field of olfactory perception technology, specifically to a molecular property prediction method based on spectral position encoding and biomimetic side-inhibition gating. Background Technology

[0002] Molecular property prediction is a fundamental problem in the fields of cheminformatics and computer olfaction, and has important applications in synthetic chemistry, the perfume industry, environmental monitoring, and artificial olfaction systems.

[0003] Traditional methods for predicting molecular properties mainly rely on artificially designed features and empirical models, such as quantitative structure-odor relationship prediction methods based on physicochemical properties or statistical learning models based on artificial feature extraction. While these methods perform reasonably well in specific tasks, their core drawback lies in their limited feature representation capabilities and inability to characterize the complex mechanisms of olfactory perception. Artificially designed features often fail to fully capture the nonlinear interaction patterns between molecular structure, electron distribution, and olfactory receptors. When faced with diverse odor molecules or unseen structural types, their predictive performance and generalization ability significantly decline, thus limiting their application value in real-world, complex olfactory perception scenarios.

[0004] In recent years, graph neural networks based on deep learning have demonstrated great potential, automatically learning relevant features from molecular structures using message passing mechanisms. However, existing graph methods lack precise quantification of the absolute positions of atoms in the overall molecular space, making it difficult to characterize the global symmetry and long-range structural relationships of molecules. Secondly, existing methods generally employ single-scale aggregation with limited receptive fields, making it difficult to model multi-level structural differences from "chemical bonds" to "molecular skeletons." In the information fusion stage, most methods simply perform summation or averaging, failing to highlight key features and leading to discrepancies between model predictions and human subjective olfactory perception.

[0005] Therefore, how to provide a method that can accurately predict molecular properties is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of the above problems, this invention is proposed to provide a molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating to overcome or at least partially solve the above problems. It can accurately quantify the global topological information of atoms in the spectral domain dimension, and use preset multi-scale constraints to collaboratively capture multi-level differential features from local chemical bonds to the global skeleton. Through biomimetic side-suppression logic, it achieves adaptive denoising and sharpening of key features, thereby accurately predicting molecular properties.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating, comprising the following steps: S1: Construct the structured input of the molecule to be predicted, which includes the atom type, the Coulomb matrix, and the Laplace eigenvectors characterizing the topological features of the molecule; S2: Input the atom type and Laplacian feature vector into the feature extraction module to obtain an initial atom feature matrix that integrates chemical properties and spectral domain topological positions; S3: Input the initial atomic feature matrix and the Coulomb matrix into the multi-scale aggregation module, and perform hierarchical neighborhood aggregation through preset scale constraint rules to generate an atomic representation containing multi-level structural information; S4: Input the atomic representation into the global interaction module of the bionic side suppression gate, extract the global query vector, and use the global query vector to perform cross-scale modulation on the local atomic features of the atomic representation to obtain denoised and semantically enhanced atomic deep semantic features. S5: Input the atomic deep semantic features into the multi-label prediction module, obtain the graph-level global representation through mask pooling, and output the prediction scores of each target odor attribute in parallel through multiple decoupled prediction heads.

[0009] Preferably, the feature extraction module in S2 performs the following steps: S21: Use the embedding layer to map the atom type to the initial embedding vector of the atom; S22: Use a multilayer perceptron to perform a nonlinear transformation on the Laplacian eigenvectors to generate learnable spectral position encoding vectors; S23: Concatenate the initial atomic embedding vector and the spectral position encoding vector along the feature dimension to form the initial atomic feature matrix.

[0010] Preferably, the multi-scale aggregation module in S3 performs the following steps: S31: Use Gaussian radial basis functions to perform high-dimensional nonlinear mapping on the interatomic distances in the Coulomb matrix to obtain the characteristics of interatomic interactions; S32: Based on the mapped inter-atomic interaction characteristics, dynamic connection weights between atoms are generated through a multilayer perceptron. S33: Based on the dynamic connection weights, under the constraints of a preset set of scale parameters, a sparse adjacency matrix of multiple scales is constructed for each atom. Different scales are used to explicitly capture chemical bonds, functional groups and global framework information, respectively. S34: Using the sparse adjacency matrices at each scale, the initial atomic feature matrices are aggregated to obtain the structural feature matrices corresponding to each scale; S35: Perform residual connection enhancement on the structural feature matrices at each scale in sequence, and output the atomic representation through a global path and local path information fusion mechanism.

[0011] Preferably, in S33, the preset multi-scale parameter K=[k1,k2,...,k N The sparse adjacency matrix at the k-th scale is constructed as follows:

[0012] in, This represents the dynamic connection weights between atoms. This represents the set of the top t largest weighted neighbor indices of atom i. Let represent the weight matrix at the k-th scale, where k∈[1,N].

[0013] Preferably, in S35, the global path and local path information fusion mechanism includes: First approach: Select a structural feature matrix with a single defined scale. As input, a multi-head self-attention mechanism is used to capture long-range dependencies across the entire graph, resulting in global interaction features. ; Second path: Structural feature matrices at all scales After concatenating the feature dimensions, a learnable linear projection layer is used for dimensionality reduction and fusion to obtain multi-scale fused features. ; Fusion and normalization steps: The global interaction features are then... Multi-scale fusion features Add them together and output the atomic representation. .

[0014] Preferably, the global interaction module for the bionic side suppression gating described in S4 performs the following steps: S41: Perform layer normalization on the input atomic representation; S42: Input the normalized atomic representation into the Transformer encoder to extract the global context, and generate a global query vector by performing masked average pooling and nonlinear transformation on the encoder output sequence. ; S43: Using the global query vector As a query, global guided features are obtained by performing cross-scale attention mapping with atomic representations as keys and values. S44: Based on the global guiding features and atomic representations, the suppression coefficient is dynamically calculated. The global guiding features and atomic representations are weighted and fused using the suppression coefficients through a gating network to output the atomic deep semantic features.

[0015] Preferably, the global interaction module for bionic side suppression gating is implemented through the following process:

[0016]

[0017]

[0018] in, Atom representation, Represents the global query vector. This indicates global guidance features. Indicates the inhibition coefficient. Represents the deep semantic features of atoms. and This represents two learnable weights. and This represents two bias coefficients. This represents the Sigmoid activation function. This indicates that corresponding elements are multiplied.

[0019] Preferably, the multi-label prediction module performs the following steps: Atomic deep semantic features utilize masked global average pooling to exclude padding atoms, resulting in a molecular-level global representation. ; The molecular-level global representation Input a shared multilayer perceptron for nonlinear feature transformation; The transformed features are input into multiple independent linear prediction layers, each corresponding to a specific odor descriptor, and their prediction scores are output in parallel.

[0020] Preferably, the multi-label prediction module is implemented through the following process:

[0021] in, Indicates molecular level representation, This represents the deep semantic feature of the i-th atom. This represents the atom mask value used when constructing the structured input, where valid atoms are 1 and fill atoms are 0. It represents a very small positive number.

[0022] Preferably, the method further includes a model training step, comprising: Collect and construct a training dataset containing structured inputs and their odor labels; Initialize the learnable parameters of the feature extraction module, multi-scale aggregation module, biomimetic side-inhibition gating global interaction module, and multi-label prediction module to form a complete model; Using the training data as input, the predicted score is calculated via forward propagation, and the prediction error is measured using a multi-label classification loss function. The learnable parameters of all modules are iteratively updated using the backpropagation algorithm until the model converges, resulting in a pre-trained model used to execute S1~S5.

[0023] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects: This invention utilizes spectral position encoding, covering 0-100 The Gaussian radial basis function with a mapping range quantifies the global topological information of atoms from the spectral domain dimension, overcoming the limitations of existing technologies that rely on three-dimensional absolute coordinates and are susceptible to interference from molecular rotation or geometric fine-tuning. Compared to traditional graph neural networks that rely on only a single scale, this method uses a multi-scale aggregation mechanism to pre-set constraints at different scales, explicitly capturing the features of atoms in local structure (chemical bonds), mid-range environment (functional groups), and global structure (molecular skeleton), thus improving the model's sensitivity to fine molecular structures.

[0024] This invention introduces a biomimetic lateral inhibition gating mechanism to achieve a shift from "blind feature stacking" to "adaptive information sharpening." By simulating the lateral inhibition mechanism in the animal olfactory system, it utilizes a globally semantically guided dynamic inhibition coefficient to weaken local redundant noise and sharpen key structural information. This alleviates the information overload problem in traditional models when processing high-noise molecular features, giving deep semantic features stronger discriminative power and making prediction results more consistent with physiological perception characteristics.

[0025] This invention addresses multi-label prediction tasks by employing a task-decoupled prediction architecture to achieve accurate prediction of molecular properties. Compared to existing dimensionality reduction prediction methods with uniform weights, the masking mechanism eliminates interference from invalid nodes, and the independent prediction head solves the negative transfer problem in multi-task learning, enhancing the model's prediction accuracy and generalization ability when facing molecules with complex structures and diverse labels. Attached Figure Description

[0026] 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.

[0027] Figure 1 A flowchart illustrating a molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating, provided as an embodiment of the present invention; Figure 2A schematic diagram of a molecular property prediction model based on spectral position encoding and biomimetic side-suppression gating is provided for an embodiment of the present invention. Figure 3 This is a schematic diagram of a molecular property prediction model based on spectral position encoding and biomimetic side-suppression gating, provided for an embodiment of the present invention. Detailed Implementation

[0028] 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.

[0029] This invention discloses a molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating, such as... Figure 1 As shown, it includes the following steps: S1: Construct the structured input of the molecule to be predicted. The structured input includes the atom type, the Coulomb matrix, and the Laplace eigenvectors that characterize the topological features of the molecule. S2: Input the atom type and Laplacian feature vector into the feature extraction module to obtain the initial atomic feature matrix that integrates chemical properties and spectral domain topological positions; S3: Input the initial atomic feature matrix and Coulomb matrix into the multi-scale aggregation module, and perform hierarchical neighborhood aggregation through preset scale constraint rules to generate atomic representations containing multi-level structural information; S4: Input the atomic representation into the global interaction module of the bionic side suppression gate, extract the global query vector, and use the global query vector to perform cross-scale modulation on the local atomic features of the atomic representation to obtain denoised and semantically enhanced atomic deep semantic features. S5: Input the atomic deep semantic features into the multi-label prediction module, obtain the graph-level global representation through mask pooling, and output the prediction scores of each target odor attribute in parallel through multiple decoupled prediction heads.

[0030] It should be noted that the Coulomb matrix and topological eigenvectors represent the global molecular structure from the perspectives of energy potential field distribution and spectral topology. Combined with atom type, these three elements must maintain strict chemical semantic mapping and atom index synchronization. This multidimensional data can be obtained through public datasets or relevant chemical computation software, ensuring that the heterogeneous features of the input model can characterize the same molecular object from three dimensions: energy field, connectivity, and chemical properties. This provides a foundation for subsequent coordinate-independent feature fusion and property prediction.

[0031] In molecular property prediction, different molecules have different numbers of atoms. To batch process a group of molecules simultaneously into a neural network, all molecules need to be padded to the same length (i.e., the maximum number of atoms) during the construction of the structured input. The padding positions are virtual atom positions introduced to align molecular maps of different sizes, and are also used to generate atom mask values ​​in the subsequent multi-label prediction module.

[0032] The following is combined with Figures 2-3 The execution method of each embodiment is explained in detail below: In one embodiment, the feature extraction module in S2 performs the following steps: S21: Use the embedding layer to map the atom type to the initial embedding vector of the atom; S22: Use a multilayer perceptron to perform a nonlinear transformation on the Laplacian eigenvectors to generate learnable spectral position encoding vectors; S23: Concatenate the initial atomic embedding vector and the spectral position encoding vector along the feature dimension to form the initial atomic feature matrix.

[0033] In the specific execution of this embodiment, the initial embedding features of atoms and learnable position-encoded features based on Laplacian feature vectors are extracted to capture the intrinsic topological connectivity of the molecular graph. This includes: obtaining the atomic numbers of atoms in the molecular graph and mapping them into high-dimensional dense vectors using an embedding layer to capture the basic chemical properties of atoms. To introduce geometric topological information independent of absolute spatial coordinates, the feature vectors obtained by eigenvalue decomposition of the graph Laplacian matrix are input into the MLP, and learnable position-encoded features are generated through linear projection and nonlinear activation. Finally, the features are concatenated to obtain the initial atomic feature representation, which serves as the input to the subsequent multi-scale aggregation module.

[0034] In one embodiment, S3, based on a multi-scale aggregation module, aggregates neighborhood information through preset scale constraints to calculate structural features covering different neighborhood ranges, from chemical bonds to the global framework. Specifically, the multi-scale aggregation module performs the following steps: S31: Use Gaussian radial basis functions to perform high-dimensional nonlinear mapping on the interatomic distances in the Coulomb matrix to obtain the characteristics of interatomic interactions; S32: Based on the mapped inter-atomic interaction characteristics, dynamic connection weights between atoms are generated through a multilayer perceptron. S33: Based on the dynamic connection weights, under the constraints of the preset scale parameter set, a sparse adjacency matrix of multiple scales is constructed for each atom. Different scales are used to explicitly capture chemical bonds, functional groups and global framework information, respectively. S34: Using the sparse adjacency matrices at each scale, the initial atomic feature matrices are aggregated to obtain the structural feature matrices corresponding to each scale; S35: Residual connections are sequentially performed on the structural feature matrices at each scale to enhance them, and atomic representations are output through a fusion mechanism of global and local path information.

[0035] In one embodiment, S31, the distance values ​​between atomic pairs in the Coulomb matrix are calculated using a range of 0-100. The Gaussian radial basis function of the mapping range is transformed to obtain nonlinear high-dimensional features characterizing interatomic interactions. This step is not a simple geometric distance encoding, but rather projects the distance scalar into a fixed-dimensional high-dimensional feature vector. (The vector is defined using numbers from 0 to 100.) Range mapping ensures that the model can capture not only short-range covalent bond information but also quantify the interrelationships between distant atoms. The formula is expressed as:

[0036] in, This represents the distance between atoms i and j in the Coulomb matrix. and Let these represent the center parameter and width parameter of the Gaussian radial basis function, respectively. This represents the nonlinear high-dimensional eigenvector obtained after the transformation.

[0037] In one embodiment, in S33, the preset multi-scale parameter K=[k1,k2,...,k N The sparse adjacency matrix at the k-th scale is constructed as follows:

[0038] in, This represents the dynamic connection weights between atoms. This represents the set of the top t largest weighted neighbor indices of atom i. Let represent the weight matrix at the k-th scale, where k∈[1,N].

[0039] In this embodiment, four scale parameters k=[2,5,10,all] are constructed, and a multi-scale sparse weight matrix is ​​built based on these four scale parameters. Different scales are used to explicitly capture chemical bonds, functional groups, and global skeletal information, respectively. For each scale k, a Top-t screening mechanism is used to force the weights of non-critical neighborhoods to be reset to zero, thereby constructing a sparse local sensing matrix. The selection of the scale parameter k strictly corresponds to the physical structure of the molecule: k=2 extracts the covalent bond features between atoms; k=5 and k=10 are used to obtain local functional groups with specific semantics (such as benzene rings, hydroxyl groups, etc.); k=all preserves global connectivity and captures molecular skeletal features.

[0040] In one embodiment, in S34, the multi-scale weight matrix aggregates the atomic features to calculate k structural features with different neighborhood ranges. Specifically, each scale weight matrix performs matrix multiplication with the atomic feature matrix, outputting k independent feature matrices with physical hierarchical meaning in parallel.

[0041] In one embodiment, S35, the global path and local path information fusion mechanism includes: First approach: Select a structural feature matrix with a single defined scale. As input, a multi-head self-attention mechanism is used to capture long-range dependencies across the entire graph, resulting in global interaction features. ; Second path: Structural feature matrices at all scales After concatenating the feature dimensions, a learnable linear projection layer is used for dimensionality reduction and fusion to obtain multi-scale fused features. ; Fusion and normalization steps: Incorporating global interaction features Multi-scale fusion features Add together, output atomic representation .

[0042] Understandably, this embodiment employs a dual-path fusion: one path utilizes multi-head self-attention to achieve global interaction; the other path integrates full-scale information through dimensionality reduction via splicing. The two paths are fused and smoothed to ensure that the final atomic representation possesses both a global perspective and multi-level structural details. The residual module enhances deep semantic expression and mitigates feature degradation and information loss during the transmission of multi-scale information. This dual-path fusion extracts long-range dependencies and global context of atoms across the entire graph through multi-head self-attention, and performs multi-scale integration to preserve multi-level details from microscopic bonding to the macroscopic skeleton. The formula is expressed as:

[0043]

[0044]

[0045] in, This represents the characteristic of k=2. This indicates the dimensionality reduction and fusion weights. This represents the global interaction features extracted by the multi-head self-attention mechanism. Indicates multi-scale fusion features, This represents the final atomic representation. This indicates a multi-head self-attention mechanism. This indicates a splicing operation. Presentation layer normalization operation.

[0046] In one embodiment, S4, based on a biomimetic side-inhibition gating global interaction module, achieves dynamic denoising, sharpening, and cross-scale fusion of local features by the globally enhanced query vector, specifically executing the following steps: S41: Perform layer normalization on the input atomic representation; S42: Input the normalized atomic representation into the Transformer encoder to extract the global context, and generate a high-purity global query vector by performing masked average pooling and nonlinear transformation on the encoder output sequence and query enhancement. ; S43: Using global query vectors As a query, global guided features are obtained by performing cross-scale attention mapping with atomic representations as keys and values. S44: Based on the global guiding features and atomic representation, the inhibition coefficient is dynamically calculated. A gating network is used to weight and fuse the global guiding features and atomic representation with the inhibition coefficient to output the atomic deep semantic features.

[0047] In this embodiment, the Transformer encoder comprises two core sub-layers: a multi-head self-attention mechanism and a feedforward network. Based on this, a query-enhanced transformation mechanism is introduced, utilizing masked average pooling and nonlinear activation to achieve semantic compression and denoising of global information, generating a high-purity query vector that represents the overall characteristics of the molecule. The formula is expressed as:

[0048]

[0049] in, This represents the global feature matrix output by the Transformer. Represents the atomic characteristic matrix, Represents the global query vector. and Indicates learnable weights, and Indicates bias. This indicates a mask average pooling operation. This represents the hyperbolic tangent activation function.

[0050] In this embodiment, the global interaction module of the biomimetic lateral inhibition gating utilizes a global query vector to achieve cross-scale attention mapping and combines it with a gating network to calculate dynamic inhibition coefficients. This weighted, complementary approach enhances global key information and suppresses local redundant noise, achieving adaptive denoising and sharpening of features and extracting atomic deep semantic features. Unlike traditional models that simply superimpose features, this step simulates the lateral inhibition mechanism ("enhancing strong signals and suppressing weak signals") in the animal olfactory system to achieve non-linear contrast enhancement, effectively improving the model's perceptual resolution and anti-interference capability. Specifically, this is achieved through the following process:

[0051]

[0052]

[0053] in, Atom representation, Represents the global query vector. This indicates global guidance features. Indicates the inhibition coefficient. Represents the deep semantic features of atoms. and This represents two learnable weights. and This represents two bias coefficients. This represents the Sigmoid activation function. This indicates that corresponding elements are multiplied.

[0054] In one embodiment, in S5, the multi-label prediction module calculates the molecular-level representation through masked global average pooling and utilizes multiple independent prediction heads to achieve supervised multi-label scoring. The specific execution logic includes: S51: Atomic deep semantic features are represented by masked global average pooling to eliminate padding atoms and obtain molecular-level global representations. ; S52: Representing the molecular level globally Input a shared multilayer perceptron for nonlinear feature transformation; S53: The transformed features are input into multiple independent linear prediction layers, each prediction layer corresponds to a specific odor descriptor, and their prediction scores are output in parallel.

[0055] In this embodiment, S51 uses a mask to eliminate interference from filling nodes, ensuring that molecular-level feature extraction focuses only on real atoms and eliminating numerical bias during heterogeneous molecule alignment. Unlike traditional non-masked ordinary pooling, a very small positive number is introduced here to avoid the denominator being zero, ensuring numerical stability during the pooling process. The generated molecular-level representation retains the contribution of the core atoms while possessing higher characterization purity. This is achieved through the following process:

[0056] in, Indicates molecular level representation, This represents the deep semantic feature of the i-th atom. This represents the atom mask value used when constructing the structured input, with 1 for valid atoms and 0 for padded atoms. It represents a very small positive number.

[0057] In this embodiment, in step S52, after the molecular-level representation is extracted using an MLP nonlinear feature extraction method, multiple independent linear layers are used to predict each odor descriptor in parallel. This decoupled prediction architecture eliminates negative transfer interference in multi-label learning, achieving supervised multi-label scoring. The MLP includes two fully connected layers and uses the GELU activation function to map the molecular representation to a lower-dimensional refined semantic space. Unlike the centralized prediction approach with unified weights used in traditional methods, this step, through structural decoupling, avoids negative interference between different odor labels, ensuring that each odor label can independently calculate the most accurate predicted score from shared semantic features.

[0058] In one embodiment, the method further includes a model training step, comprising: Collect and construct a training dataset containing structured inputs and their odor labels; The learnable parameters of the initialization feature extraction module, multi-scale aggregation module, biomimetic side-inhibition gating global interaction module, and multi-label prediction module are used to form a complete model. Using training data as input, the predicted score is calculated via forward propagation, and the prediction error is measured using a multi-label classification loss function. The learnable parameters of all modules are iteratively updated using the backpropagation algorithm until the model converges, resulting in a pre-trained model used to execute S1~S5.

[0059] The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating provided in this embodiment quantifies the geometric and topological information of molecules through learnable spectral encoding; constrains the aggregation of hierarchical structural features from microscopic chemical bonds and mid-range functional groups to the global skeleton by preset scale parameters; achieves adaptive denoising and sharpening of features through biomimetic side-suppression gating; and eliminates negative migration interference between multiple labels through a task decoupling prediction architecture, thereby achieving accurate molecular property prediction that conforms to the molecular structure features.

[0060] 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.

[0061] 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. A method for predicting molecular properties based on spectral position encoding and biomimetic side-suppression gating, characterized in that, Includes the following steps: S1: Construct the structured input of the molecule to be predicted, which includes the atom type, the Coulomb matrix, and the Laplace eigenvectors characterizing the topological features of the molecule; S2: Input the atom type and Laplacian feature vector into the feature extraction module to obtain an initial atom feature matrix that integrates chemical properties and spectral domain topological positions; S3: Input the initial atomic feature matrix and the Coulomb matrix into the multi-scale aggregation module, and perform hierarchical neighborhood aggregation through preset scale constraint rules to generate an atomic representation containing multi-level structural information; S4: Input the atomic representation into the global interaction module of the bionic side suppression gate, extract the global query vector, and use the global query vector to perform cross-scale modulation on the local atomic features of the atomic representation to obtain denoised and semantically enhanced atomic deep semantic features. S5: Input the atomic deep semantic features into the multi-label prediction module, obtain the graph-level global representation through mask pooling, and output the prediction scores of each target odor attribute in parallel through multiple decoupled prediction heads.

2. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 1, characterized in that, The feature extraction module described in S2 performs the following steps: S21: Use the embedding layer to map the atom type to the initial embedding vector of the atom; S22: Use a multilayer perceptron to perform a nonlinear transformation on the Laplacian eigenvectors to generate learnable spectral position encoding vectors; S23: Concatenate the initial atomic embedding vector and the spectral position encoding vector along the feature dimension to form the initial atomic feature matrix.

3. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 1, characterized in that, The multi-scale aggregation module described in S3 performs the following steps: S31: Use Gaussian radial basis functions to perform high-dimensional nonlinear mapping on the interatomic distances in the Coulomb matrix to obtain the characteristics of interatomic interactions; S32: Based on the mapped inter-atomic interaction characteristics, dynamic connection weights between atoms are generated through a multilayer perceptron. S33: Based on the dynamic connection weights, under the constraints of the preset scale parameter set, a sparse adjacency matrix of multiple scales is constructed for each atom. Different scales are used to explicitly capture chemical bonds, functional groups and global framework information, respectively. S34: Using the sparse adjacency matrices at each scale, the initial atomic feature matrices are aggregated to obtain the structural feature matrices corresponding to each scale; S35: Perform residual connection enhancement on the structural feature matrices at each scale in sequence, and output the atomic representation through a global path and local path information fusion mechanism.

4. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 3, characterized in that, In S33, the multi-scale parameter K=[k1,k2,...,k] is preset. N The sparse adjacency matrix at the k-th scale is constructed as follows: in, This represents the dynamic connection weights between atoms. This represents the set of the top t largest weighted neighbor indices of atom i. Let represent the weight matrix at the k-th scale, where k∈[1,N].

5. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 3, characterized in that, In S35, the global path and local path information fusion mechanism includes: First approach: Select a structural feature matrix with a single defined scale. As input, a multi-head self-attention mechanism is used to capture long-range dependencies across the entire graph, resulting in global interaction features. ; Second path: Structural feature matrices at all scales After concatenating the feature dimensions, a learnable linear projection layer is used for dimensionality reduction and fusion to obtain multi-scale fused features. ; Fusion and normalization steps: The global interaction features are then... Multi-scale fusion features Add them together and output the atomic representation. .

6. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 1, characterized in that, The global interaction module for the biomimetic side-suppression gating described in S4 performs the following steps: S41: Perform layer normalization on the input atomic representation; S42: Input the normalized atomic representation into the Transformer encoder to extract the global context, and generate a global query vector by performing masked average pooling and nonlinear transformation on the encoder output sequence. ; S43: Using the global query vector As a query, global guided features are obtained by performing cross-scale attention mapping with atomic representations as keys and values. S44: Based on the global guiding features and atomic representations, the suppression coefficient is dynamically calculated. The global guiding features and atomic representations are weighted and fused using the suppression coefficients through a gating network to output the atomic deep semantic features.

7. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 6, characterized in that, The global interaction module of the biomimetic side-inhibition gating is implemented through the following process: in, Atom representation, Represents the global query vector. Indicates global guidance features. Indicates the inhibition coefficient. Represents the deep semantic features of atoms. and This represents two learnable weights. and This represents two bias coefficients. This represents the Sigmoid activation function. This indicates that corresponding elements are multiplied.

8. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 1, characterized in that, The multi-label prediction module performs the following steps: Atomic deep semantic features utilize masked global average pooling to exclude padding atoms, resulting in a molecular-level global representation. ; The molecular-level global representation Input a shared multilayer perceptron for nonlinear feature transformation; The transformed features are input into multiple independent linear prediction layers, each corresponding to a specific odor descriptor, and their prediction scores are output in parallel.

9. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 8, characterized in that, The multi-label prediction module is implemented through the following process: in, Indicates molecular level representation, This represents the deep semantic feature of the i-th atom. This represents the atom mask value used when constructing the structured input, where valid atoms are 1 and fill atoms are 0. It represents a very small positive number.

10. The molecular property prediction method based on spectral position encoding and biomimetic side-suppression gating according to claim 1, characterized in that, The method also includes a model training step, comprising: Collect and construct a training dataset containing structured inputs and their odor labels; Initialize the learnable parameters of the feature extraction module, multi-scale aggregation module, biomimetic side-inhibition gating global interaction module, and multi-label prediction module to form a complete model; Using the training data as input, the predicted score is calculated via forward propagation, and the prediction error is measured using a multi-label classification loss function. The learnable parameters of all modules are iteratively updated using the backpropagation algorithm until the model converges, resulting in a pre-trained model used to execute S1~S5.

Images