Molecular chemistry and binding property prediction method based on Laplace-Beltrami multi-head attention spectrum analysis

Abstract

The invention discloses a molecular chemistry and binding property prediction method based on Laplace-Beltrami multi-head attention spectrum analysis. The method comprises the following steps: S1, performing triangulation on a molecular surface and removing an abnormal triangle; s2, a Laplace-Beltrami characteristic value is solved for the molecular surface after triangulation by using a finite element method, and a maximum characteristic value threshold parameter is set; s3, extracting geometric features and chemical features of the molecular surface and splicing the geometric features and the chemical features into a fusion feature matrix; and S4, completing propagation update in the spectral domain by using a conversion mechanism of the Laplace-Beltrami operator in the spatial domain and the spectral domain, and adjusting the maximum characteristic value threshold to optimize a prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a molecular chemistry and binding property prediction method based on Laplace-Beltrami multi-head attention spectrum analysis. BACKGROUND

[0002] Molecular property prediction, as an important means to solve new pharmaceuticals, drug efficacy prediction and environmental science problems, has attracted widespread attention from scholars at home and abroad. Research teams often use mathematical and computer field knowledge to model, focusing on the atomic species, relative position of the atoms and chemical properties of known molecules to complete the property prediction of unknown molecular structures.

[0003] Most of the current research on molecular property prediction uses E(3) variable graph neural network as the basic architecture, such as SEGNN and EGNN. This kind of method designs a network model that satisfies the rotation invariance, translation invariance and permutation invariance, and continuously updates the atomic vertex coordinates, edge weights and features in the spatial domain to complete a series of prediction tasks. However, in order to achieve rotation invariance, a large number of tensor multiplications and trigonometric function expansions are added to the network. The more layers, the higher the computational complexity of the algorithm used. At the same time, the transmission of neural information usually realizes long-distance communication by stacking deep layers or increasing the neighborhood radius of feature extraction, which makes it easy to lose local information when dealing with high molecular compounds due to excessive smoothing. Therefore, a processing method that takes into account both local information and overall information should be used to solve the problem.

[0004] Spectrum analysis is an important means to handle multi-resolution tasks. The Laplace-Beltrami operator can be used for molecular representation learning tasks due to its advantages in image processing. By using the operator, the molecular surface is regarded as an embedded manifold, and the scalar field is subjected to eigen-decomposition to obtain a series of eigenvalues and corresponding spectral bases, so as to map the feature information to the frequency domain and learn the molecular properties in the spectral domain. However, the geometric sensitivity of this method is still lacking, and the feature value relies too much on the extraction of high-curvature sites and concave pocket information. At the same time, the feature updating process in the spectral domain is not perfect enough, and the attention to local information is still lacking. There is still room for improvement in balancing the computational complexity and the proportion of molecular representation information, and it is difficult to balance the computational overhead and representation capacity.

[0005] In summary, the methods for molecular property prediction tasks can achieve good prediction of molecular properties to some extent, but there are still some deficiencies: on the one hand, most methods greatly increase the computational complexity by designing an invariant network due to excessive attention to rotational invariance, and do not effectively separate local information and global information; on the other hand, although the method using spectral analysis can alleviate this problem to some extent, there is still a lot of room for improvement in the selection of geometric features and the feature updating process in the spectral domain, and the difference between local information and global information should be further learned.

[0006] And the Chinese invention application with application number 202210944763.6 discloses a molecular property prediction method and system based on a graph neural network, the method comprising: obtaining a data file of a molecule to be predicted, and converting the data file into graph data, the graph data comprising a plurality of nodes and a plurality of edges, the nodes representing atoms constituting the molecule to be predicted, and the edges representing chemical bonds of the molecule to be predicted; inputting the graph data into a pre-trained molecular property prediction model to obtain the molecular characteristics of the molecule to be predicted; wherein the molecular property prediction model is trained according to a molecular graph sample, the molecular graph sample being an undirected graph converted from a data file sample, the nodes in the undirected graph representing atoms constituting the molecular sample, and the edges in the undirected graph representing chemical bonds of the molecular sample. SUMMARY

[0007] To solve the technical problems in the above background art, the present application provides a molecular chemical and binding property prediction method based on Laplace-Beltrami multi-head attention spectral analysis, and the technical solution adopted by the present application is: The first aspect of the present application provides a molecular chemical and binding property prediction method based on Laplace-Beltrami multi-head attention spectral analysis, the method comprising: S1: Triangulating the molecular surface and removing abnormal triangles; S2: Solving the Laplace-Beltrami eigenvalue of the triangulated molecular surface using the finite element method, and setting a maximum eigenvalue threshold parameter; S3: Extracting the geometric features and chemical features of the molecular surface and splicing them into a fusion feature matrix; S4: Using the conversion mechanism of the Laplace-Beltrami operator in the spatial domain and the spectral domain to complete the propagation update in the spectral domain, and adjusting the maximum eigenvalue threshold to optimize the prediction result.

[0008] As a preferred scheme, in step S1, the method of triangulating the molecular surface and removing abnormal triangles comprises: S11: Convert the coordinate position information of the preset molecular dataset into a format that can be triangulated by the MSMS program; S12: Triangulation of molecular surfaces using the MSMS program; S13: Remove abnormal triangles from the molecular surface after triangulation to complete the anomaly handling.

[0009] As a preferred embodiment, the abnormal triangle includes at least one of the following: A triangle whose acute angle does not meet the first preset condition; A triangle whose obtuse angle does not meet the second preset condition; Triangulation regions that cannot form triangles and whose number of connected vertices is less than a preset threshold; The first preset condition is: The second preset condition is .

[0010] As a preferred embodiment, in step S2, the method for solving the Laplace-Beltrami eigenvalues ​​of the triangulated molecular surface using the finite element method includes: set up Let be a function defined on the molecular surface, where Given the set of vertices of the triangulated mesh, solving the Laplace-Beltrami characteristic problem is equivalent to solving the equations. ,in and The expression is as follows:

[0011]

[0012] in, and They are respectively based on The areas of the two separated triangles Represents the vertex Solving for the set of adjacent vertices yields a non-negative eigenvalue matrix. and an eigenvector matrix ,and satisfy: and ,in It is an identity matrix.

[0013] As a preferred embodiment, in step S3, the method for extracting the geometric and chemical features of the molecular surface and splicing them into a fused feature matrix includes: S31: Calculate the shape index, Gaussian curvature, and mean curvature at each vertex, where the shape index... ; wherein are two principal curvatures, respectively, for reflecting local concave-convex near the vertex; S32: calculating the value of the heat kernel signature, according to the eigenvalues arranged in ascending order and the corresponding eigenvector matrix, selecting T time points in the time range in equal ratio, and using the following formula to calculate the value of the heat kernel signature:

[0014] wherein is the number of eigenvalues, is the time point; S33: concatenating the shape index, the curvature, the heat kernel signature and the Laplace-Beltrami eigenvalue to form a geometric property matrix ; S34: extracting the chemical properties of the molecular surface, wherein is the center node of interest, the k-nearest neighbor algorithm is used to find the nearest k nodes to the center node, and the reciprocal of the distance from the center node is taken as the weight, wherein represents the distance norm, and the chemical properties of the vertex are obtained by a multi-layer perception MLP weighted, and the chemical properties of all vertices are concatenated into a matrix ; S35: concatenating and to obtain a fusion feature matrix .

[0015] As a preferred solution, the step S3 further comprises: using a sparse self-attention module to enhance the extraction of local features of the fusion feature matrix, and the specific method comprises: first extracting initial features , and first forming an image form through a reshape operation, dividing the feature map into a plurality of local small blocks according to the number of atoms of the molecule, and performing 2D convolution enhancement, introducing a multi-head attention mechanism after layer normalization in each local block, and the formula is representing the query vector, the key vector and the value vector, respectively, d representing the dimension of each attention head, is a function of transferring to a probability distribution.

[0016] As a preferred solution, in step S4, the method for using the conversion mechanism of the Laplace-Beltrami operator in the spatial domain and the spectral domain to complete the propagation update in the spectral domain and adjusting the maximum eigenvalue threshold to optimize the prediction result comprises: ​S41: transform the fusion feature matrix to the spectral domain, and the specific formula is as follows:

[0017] wherein, is a Laplace-Beltrami feature vector matrix, is a quality matrix, is an input fusion feature matrix, transformed spectral domain feature; S42: adopt a multi-head attention propagation mechanism, and introduce a learnable propagation parameter ) calculate the propagation coefficient of each head and each dimension, and apply a GLU gating adjustment mechanism to complete feature updating; S43: perform inverse spectral transformation back to the spatial domain using the formula , and perform residual connection and pooling operation, wherein inverse transformed spatial domain feature; S44: complete the task according to different molecular prediction tasks by setting prediction indicators, and adjust the maximum eigenvalue threshold to optimize the prediction result.

[0018] As a preferred solution, in step S42, a multi-head attention propagation mechanism is adopted, and a learnable propagation parameter is introduced. ) The method for calculating the propagation coefficient of each head and each dimension and applying a GLU gating adjustment mechanism to complete feature updating includes: using a multi-head attention propagation mechanism in the spectral domain, using multiple propagation time heads to simulate diffusion, for each time head, introducing a learnable propagation parameter ) calculating the propagation coefficient of each head and each dimension, and using a GLU gating adjustment mechanism to first perform a first linear mapping on the input feature wherein, is a weight matrix, then dividing into two blocks in the last dimension then wherein, represents element-wise multiplication, performing batch normalization BatchNorm on each row to obtain , and finally using a SiLU function and performing a second linear mapping so that the output dimension remains unchanged; and using a Gaussian filter ​​​The processing of each feature value is to enhance the distinction of local features and overall features, wherein, a learnable center position parameter, a learnable bandwidth parameter, an input feature value; a corresponding loss function is set according to different molecular property prediction tasks to complete feature updating.

[0019] The second aspect of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the aforementioned molecular chemical and binding property prediction method based on Laplace-Beltrami multi-head attention spectral analysis.

[0020] The third aspect of the present application provides a computer device, which includes a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, and the computer program is executed by the processor to realize the steps of the aforementioned molecular chemical and binding property prediction method based on Laplace-Beltrami multi-head attention spectral analysis.

[0021] Compared with the prior art, the present application has the beneficial effects that: The present application introduces the Laplace-Beltrami operator, which has equivariance for rigid transformations in Euclidean space. This means that even if the molecular structure is slightly perturbed globally, it will not affect the results, thus naturally meeting the requirement of equivariance.

[0022] The present application increases the shape index at each vertex to reflect the concave and convex near the vertex, which helps to enhance the feature extraction effect.

[0023] The present application adopts a multi-head attention mechanism. Different time heads configure different parameters, so that the mechanism can take into account the information of the nearby position and the far position, and improve the efficiency and quality of information transmission.

[0024] The present application uses GLU gating mechanism and Gaussian filter to fully extract the molecular surface features. This helps to preserve local details before they are overwhelmed by global diffusion features, and distinguishes the information under different feature value corresponding resolution, thus effectively improving the prediction ability. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The present application provides a molecular chemical and binding property prediction method based on Laplace-Beltrami multi-head attention spectral analysis; Figure 2 The present application provides a data processing flowchart. DETAILED DESCRIPTION The accompanying drawings are for illustrative purposes only and should not be construed as limiting the invention. It should be understood that the described embodiments are merely some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of the embodiments of this application.

[0026] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the embodiments of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0027] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0028] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. The invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Example 1 Please refer to Figure 1 as well as Figure 2 This embodiment provides a method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy, the method comprising: S1: Triangulate the molecular surface and remove abnormal triangles; In one specific embodiment, the method of triangulating the molecular surface and removing abnormal triangles in step S1 comprises: S11: converting the coordinate position information of the preset molecular dataset into a format available for the MSMS program to perform triangulation; Specifically, the molecular dataset is a QM9 dataset or a protein molecular dataset. Specifically, the coordinate position information of the preset molecular dataset is converted into the.xyzrn format available for the MSMS program to perform triangulation. S12: triangulating the molecular surface using the MSMS program. S13: removing abnormal triangles from the triangulated molecular surface to complete the abnormal processing.

[0031] In one specific embodiment, the abnormal triangles include at least one of the following: triangles with an acute angle size not meeting a first preset condition; triangles with an obtuse angle size not meeting a second preset condition; edges that cannot form a triangle and triangulation regions with a number of connected vertices less than a preset threshold value; Specifically, the preset threshold value is determined according to the total number of atoms in the molecular dataset.

[0032] wherein the first preset condition is and the second preset condition is .

[0033] S2: solving the Laplace-Beltrami eigenvalue of the triangulated molecular surface using the finite element method and setting a maximum eigenvalue threshold parameter. In one specific embodiment, the method of solving the Laplace-Beltrami eigenvalue of the triangulated molecular surface using the finite element method in step S2 comprises: Let be a function defined on the molecular surface, where is a set of grid vertices after triangulation, and solving the Laplace-Beltrami eigenvalue problem is to solve the equation wherein and are expressed as follows:

[0034]

[0035] wherein and are the areas of two triangles separated by , respectively, denote a set of vertices adjacent to vertex , solve to obtain a non-negative eigenvalue matrix and an eigenvector matrix , and satisfy: and , where is an identity matrix.

[0036] S3: Extract the geometric and chemical features of the molecular surface and splice them into a fusion feature matrix; In a specific embodiment, in step S3, the method of extracting the geometric and chemical features of the molecular surface and splicing them into a fusion feature matrix includes: S31: Calculate the shape index, Gaussian curvature and mean curvature at each vertex, where the shape index ; wherein are the two principal curvatures, respectively, used to reflect the local concave-convex near the vertex; S32: Calculate the value of the heat kernel signature, according to the eigenvalues arranged in ascending order and the corresponding eigenvector matrix, select T time points in the time range in equal ratio, and use the following formula to calculate the value of the heat kernel signature:

[0037] wherein is the number of eigenvalues, is the time point; S33: Splice the shape index, curvature, heat kernel signature and Laplace-Beltrami eigenvalue to form a geometric property matrix ; S34: Extract the chemical properties of the molecular surface, let be the center node of interest, find the nearest k nodes to the center node by k-nearest neighbor algorithm, and take the reciprocal of the distance from the center node as the weight, wherein denotes the distance norm, obtain the chemical properties of the vertex by a multi-layer perceptron MLP weighting, and splice the chemical properties of all vertices into a matrix ; S35: Splice and by a multi-layer perceptron MLP to obtain a fusion feature matrix .

[0038] In an optional embodiment, the step S3 further includes: The sparse self-attention module is used to enhance the extraction of local features of the fusion feature matrix, and the specific method comprises: First, extract the initial features , First, reshape the image to form an image form, and divide the feature map into multiple local small blocks according to the number of molecules and atoms for 2D convolution enhancement, wherein if the total number of atoms is N , N , ; first perform layer normalization inside each local block and introduce a multi-head attention mechanism, and the formula is query vector, key vector and value vector, d , represents the dimension of each attention head, is a function of transferring to a probability distribution.

[0039] S4: Use the conversion mechanism of the Laplace-Beltrami operator in the spatial domain and the spectral domain to complete the propagation update in the spectral domain, and adjust the maximum eigenvalue threshold to optimize the prediction result; In one specific embodiment, in step S4, the method of using the conversion mechanism of the Laplace-Beltrami operator in the spatial domain and the spectral domain to complete the propagation update in the spectral domain, and adjusting the maximum eigenvalue threshold to optimize the prediction result comprises: S41: Transform the fusion feature matrix to the spectral domain, and the specific formula is as follows:

[0040] , wherein is the Laplace-Beltrami feature vector matrix, is the quality matrix, is the input fusion feature matrix, is the transformed spectral domain feature; S42: Adopt a multi-head attention propagation mechanism, and introduce a learnable propagation parameter , ) Calculate the propagation coefficient of each head and each dimension, and apply the GLU gating adjustment mechanism to complete the feature update; S43: Perform inverse spectral transformation back to the spatial domain using the formula , wherein is the inverse transformed spatial domain feature; S44: According to different molecular prediction tasks, set the prediction index to complete the task, and adjust the maximum eigenvalue threshold to optimize the prediction result.

[0041] In one specific embodiment, in step S42, a multi-head attention propagation mechanism is adopted, and a learnable propagation parameter , where ) The method for calculating the propagation coefficient of each head and each dimension and applying the GLU gating adjustment mechanism to complete the feature update comprises: Using a multi-head attention propagation mechanism in the spectral domain, using multiple propagation time heads Simulating diffusion, for each time head, introducing a learnable propagation parameter , where ) Calculate the propagation coefficient of each head and each dimension, and use the GLU gating adjustment mechanism to first linearly map the input feature , where is a weight matrix, , and then is divided into two blocks in the last dimension , then , where represents element-wise multiplication, and batch normalization is performed on each row to obtain , and finally a SiLU function is used , so that the output dimension is unchanged; a Gaussian filter is used to process each feature value to enhance the distinction between local features and overall features, where is a learnable center position parameter, is a learnable bandwidth parameter, is an input feature value; according to the different molecular property prediction tasks, a corresponding loss function is set to complete the feature update.

[0042] Specifically, the structure of the GLU (Gated Linear Unit) gating adjustment mechanism is Linear, GLU, BatchNorm, SiLU and Linear linked once.

[0043] Embodiment 2 The embodiment provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the molecular chemical and binding property prediction method based on Laplace-Beltrami multi-head attention spectral analysis in embodiment 1.

[0044] Embodiment 3 ​The embodiment provides a computer device, including a storage medium, a processor and a computer program stored in the storage medium and executable by the processor, when the computer program is executed by the processor, steps of a molecular chemistry and binding property prediction method based on Laplace-Beltrami multi-head attention spectrum analysis are implemented.

[0045] Embodiment 4.

[0046] The embodiment verifies and analyzes the method described in Embodiment 1, and more particularly: The embodiment uses three public data sets to show the implementation and implementation effect of the application. The three data sets are: TOX21 toxicity classification prediction data set, QM9 data set, and protein molecule data set.

[0047] The TOX21 data set is used to predict the toxicity (binary classification problem) on four receptors with fewer metal atoms (metal atoms have a greater impact on the geometric shape), and the result is presented by ROC-AUC value; the QM9 data set is used to predict various chemical properties, and the result is presented by mean absolute error; the protein molecule data set enables a sparse self-attention module and is used to predict which ligand (seven kinds in total) binds to the protein, that is, a seven-classification problem, and the result is presented by balanced accuracy (that is, the average of the prediction accuracy of the seven ligands), and experiments are performed on three types of chemical properties and geometric properties, only chemical properties and only geometric properties, and compared with the Masif-Ligand method, and the results are shown in Tables 1 to 3. Table 1. Toxicity prediction results of the application on the TOX21 data set for four receptors

[0048] Table 2. Comparison of average absolute errors of the application and other methods for predicting the QM9 data set

[0049] Table 3. Comparison of protein molecule ligand pocket classification prediction results of the application and the Masif-Ligand method

[0050] Obviously, the above embodiments of the application are only examples for clearly illustrating the application, and are not intended to limit the implementation modes of the application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the implementation modes are not required or can not be exhausted. Any modification, equivalent replacement and improvement made within the spirit and principle of the application should be included in the protection scope of the claims of the application.

Claims

1. A method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy, characterized in that, The method includes: S1: Triangulate the molecular surface and remove abnormal triangles; S2: Solve the Laplace-Beltrami eigenvalues ​​on the triangulated molecular surface using the finite element method, and set the maximum eigenvalue threshold parameter; S3: Extract the geometric and chemical features of the molecular surface and splice them into a fusion feature matrix; S4: The Laplace-Beltrami operator is used to perform propagation updates in the spectral domain by transforming between the spatial and spectral domains, and the maximum eigenvalue threshold is adjusted to optimize the prediction results.

2. The method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy according to claim 1, characterized in that, In step S1, the method for triangulating the molecular surface and removing anomalous triangles includes: S11: Convert the coordinate position information of the preset molecular dataset into a format that can be triangulated by the MSMS program; S12: Triangulation of molecular surfaces using the MSMS program; S13: Remove abnormal triangles from the molecular surface after triangulation to complete the anomaly handling.

3. The method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy according to claim 2, characterized in that, The abnormal triangle includes at least one of the following: A triangle whose acute angle does not meet the first preset condition; A triangle whose obtuse angle does not meet the second preset condition; Triangulation regions that cannot form triangles and whose number of connected vertices is less than a preset threshold; The first preset condition is: The second preset condition is .

4. The method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy according to claim 1, characterized in that, In step S2, the method for solving the Laplace-Beltrami eigenvalues ​​of the triangulated molecular surface using the finite element method includes: set up Let be a function defined on the molecular surface, where Given the set of vertices of the triangulated mesh, solving the Laplace-Beltrami characteristic problem is equivalent to solving the equations. ,in and The expression is as follows: in, and They are respectively based on The areas of the two separated triangles Represents the vertex Solving for the set of adjacent vertices yields a non-negative eigenvalue matrix. and an eigenvector matrix ,and satisfy: and ,in It is an identity matrix.

5. The method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy according to claim 1, characterized in that, In step S3, the method for extracting the geometric and chemical features of the molecular surface and splicing them into a fused feature matrix includes: S31: Calculate the shape index, Gaussian curvature, and mean curvature at each vertex, where the shape index... ;in These are two principal curvatures used to reflect the local concavity and convexity near the vertex; S32: Calculate the value of the hot core signature. Based on the ascendingly sorted feature values ​​and the corresponding feature vector matrix, select T time points in a proportional fashion within the time range, and use the following formula to calculate the value of the hot core signature: in It is the number of eigenvalues. It's a point in time; S33: Concatenate the shape index, curvature, heat core signature, and Laplace-Beltrami eigenvalues ​​to form a geometric property matrix. ; S34: Extract the chemical properties of the molecular surface, assuming... The central node of interest is used to find the distance using the k-nearest neighbor algorithm. Recent k There are 10 nodes, and they are ordered according to the reciprocal of their distance from the center node. As a weight, where The distance norm is used to obtain the chemical properties of a vertex through a weighted multilayer perceptron (MLP), and the chemical properties of all vertices are concatenated into a matrix. ; S35: Through a multilayer perceptron (MLP) and The fused feature matrix is ​​obtained by splicing. .

6. The method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy according to claim 1, characterized in that, Step S3 further includes: The extraction of local features from the fused feature matrix is ​​enhanced using a sparse self-attention module. Specific methods include: First extract initial features ,Will First, an image is formed through a reshape operation. The feature map is then divided into multiple local blocks according to the number of molecules and atoms, and enhanced by 2D convolution. Within each local block, layer normalization is performed before introducing a multi-head attention mechanism. The formula is as follows: These represent the query vector, key vector, and value vector, respectively. d This represents the dimension of each attention head. This is a function that transitions to the probability distribution.

7. The method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy according to claim 1, characterized in that, In step S4, the method of using the Laplace-Beltrami operator's transformation mechanism between the spatial and spectral domains to complete the propagation update in the spectral domain and adjusting the maximum eigenvalue threshold to optimize the prediction results includes: S41: Transform the fused feature matrix to the spectral domain, using the following formula: in, The Laplace-Beltrami eigenvector matrix, For the quality matrix, The input is the fusion feature matrix. Transformed spectral domain characteristics; S42: Employs a multi-head attention propagation mechanism and introduces a learnable propagation parameter. ,use ) The propagation coefficients for each head and each dimension are calculated, and a GLU gating adjustment mechanism is applied to complete the feature update. S43: Using the formula Perform inverse spectral transform back to the spatial domain, then perform residual connection and pooling operations, where... Spatial domain characteristics after inverse transformation; S44: Set prediction metrics according to different molecular prediction tasks to complete the task, and adjust the maximum eigenvalue threshold to optimize the prediction results.

8. The method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy according to claim 7, characterized in that, In step S42, a multi-head attention propagation mechanism is adopted, and a learnable propagation parameter is introduced. ,use ) Methods for calculating the propagation coefficients for each head and each dimension, and applying GLU gating mechanisms to complete feature updates include: In the spectral domain, a multi-head attention propagation mechanism is used, leveraging multiple propagation time heads. To simulate diffusion, a learnable propagation parameter is introduced for each time head. ,use ) The propagation coefficients for each head and each dimension are calculated, and a GLU gating mechanism is used to adjust the input features first. Perform the first linear mapping ,in This is the weight matrix. Then The last dimension is divided into two equal parts. ,Then ,in This represents element-wise multiplication, with the batch normalization operation BatchNorm applied to each row to obtain the result. Finally, the SiLU function is used. And perform a second linear mapping. To keep the output dimension unchanged; use a Gaussian filter. Each eigenvalue is processed to enhance the distinction between local and global features, whereby... For learnable center location parameters, For learnable bandwidth parameters, Input feature values; set corresponding loss functions according to different molecular property prediction tasks to complete feature updates.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy as described in any one of claims 1 to 8.

10. A computer device, characterized in that: The method includes a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of the method for predicting molecular chemistry and binding properties based on Laplace-Beltrami multi-head attention spectroscopy as described in any one of claims 1 to 8.

Images