Molecular property prediction method, system, device and equipment

Abstract

The invention relates to the field of artificial intelligence, in particular to a molecular property prediction method, system, device and equipment, and the method comprises the steps: firstly obtaining 3D molecular structure data containing atomic coordinate information; according to the set radius alpha and the filling value sigma, zero-dimensional, one-dimensional and two-dimensional simplex structures are extracted, and substructure position coding information is calculated; then, taking one-dimensional and two-dimensional simplex structures as atomic feature indexes, and fusing atomic topology and atomic features; on the basis of zero-dimensional simplex structure and substructure position coding, obtaining pairing position coding information of an atom-simplex pair; inputting the fused atomic features and pairing position coding information into a multi-head self-attention module to generate global molecular features; and finally, obtaining a molecular property prediction result according to the global molecular characteristics. According to the method, the defect of processing fine-grained substructure information in the prior art can be effectively solved, and the 3D molecular property prediction performance is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method, system, device and equipment for predicting molecular properties. Background Art

[0002] Molecular property prediction (MPP) is a key task in computational chemistry and molecular modeling. Existing molecular representation methods are mainly divided into three categories: one-dimensional, two-dimensional and three-dimensional (3D), each with its own advantages and disadvantages.

[0003] One-dimensional molecular representation methods use Simplified Molecular Input Linear Expressions (SMILES) to encode atoms as a sequence of symbols. While simple and efficient, this method ignores the spatial structure and local details of the molecule and only reflects the linear relationships between atoms. For large molecules, SMILES may require padding or truncation, resulting in data loss or increased noise, which can affect model performance.

[0004] Two-dimensional molecular representation methods use graph convolutional networks (GCNs) and image data to process the topological structure and spatial relationships of molecules. Although two-dimensional representations can better reflect molecular structure, existing methods often suffer from the problem of "over-smoothing", making it difficult to capture the detailed connections between different atoms in complex structures. In addition, when converting molecules into images for convolutional neural network (CNN) feature extraction, atomic characteristics and bond properties cannot be preserved, and blank areas in the image lack effective information, reducing prediction accuracy.

[0005] For three-dimensional molecular representation methods, 3D molecular data can truly reflect the spatial structure of molecules, including the precise 3D positions of atoms, and generally provides higher prediction accuracy than one-dimensional and two-dimensional data. However, existing 3D graph convolutional networks (3D-GCNs) and 3D grid methods have limitations. For example, they can only capture the pairing relationships between atoms or rely on predefined CNN convolution kernels, making it difficult to accurately model complex molecular structures. Summary of the Invention

[0006] To address the above technical problems, the present invention provides a molecular property prediction method, system, device, and apparatus. The method aims to enhance the expression of local substructure information by introducing topological data analysis (TDA) and a topological fusion network, and to combine global and local feature modeling to address the deficiencies of the existing technology and improve the accuracy and robustness of 3D molecular property prediction. The method comprises the following steps:

[0007] S1: Acquire 3D molecular structure data with atomic coordinate information;

[0008] S2: extracting the topological structure in the 3D molecular structure data according to the set radius α and filling value σ, including a zero-dimensional simplex structure, a one-dimensional simplex structure, and a two-dimensional simplex structure, and calculating the average value of the vertex spatial position of each extracted one-dimensional simplex structure and two-dimensional simplex structure to obtain substructure position encoding information;

[0009] Wherein, the zero-dimensional simplex structure represents a single atom; the one-dimensional simplex structure represents a covalent bond obtained within a specified radius; the two-dimensional simplex structure represents a functional group formed by connecting multiple atoms within a specified radius;

[0010] S3: For each atom, using the one-dimensional simplex structure and the two-dimensional simplex structure as indexes of the corresponding atomic features to obtain the atomic topological features, and fusing the atomic topological features with the corresponding atomic features to obtain the fused atomic features;

[0011] S4: obtaining pairing position encoding information of atom-simplex pairs based on the single atomic coordinate information in the zero-dimensional simplex structure and the substructure position encoding information;

[0012] S5: Inputting the fused atomic features and the paired position encoding information into a multi-head self-attention module, calculating the spatial relationship between the atomic features, and generating a global molecular feature;

[0013] S6: Based on the global molecular features, obtain molecular property prediction results.

[0014] In one embodiment of the present invention, in S2, the method for obtaining the substructure position encoding information is as follows:

[0015] By adjusting the value of the radius α, the one-dimensional simplex structure set and the two-dimensional simplex structure set of the 3D molecular structure data are obtained, respectively:

[0016]

[0017] Among them, S 1 represents a set of one-dimensional simplex structures, represents the j-th one-dimensional simplex structure, express The coordinate information of atom i in , j = 1, ..., T1, T1 represents the total number of one-dimensional simplex structures extracted; S 2 represents a set of two-dimensional simplex structures, represents the k-th two-dimensional simplex structure, express The coordinate information of atom i in , k = 1, ..., T2, T2 represents the total number of extracted two-dimensional simplex structures;

[0018] For each extracted one-dimensional simplex structure, calculate the average value of its vertex spatial position For each extracted two-dimensional simplex structure, calculate the average value of its vertex spatial positions

[0019] Summarize all and Substructure position encoding information

[0020] In one embodiment of the present invention, in S3, the method for obtaining the fused atomic features is as follows:

[0021] For each atom, a one-dimensional simplex structure S 1 and the two-dimensional simplex structure S 2 Aggregate the features of the one-dimensional simplex structure and the two-dimensional simplex structure as indices of the corresponding atomic features;

[0022] The two D-dimensional atomic features corresponding to the aggregated one-dimensional simplex structural features are connected and input into the multi-layer perceptron for processing to obtain the D-dimensional covalent bond substructure feature vector

[0023] When three D-dimensional atomic features correspond to a two-dimensional simplex structure feature, that is, a functional group feature, a D-dimensional functional group substructure feature vector is generated after multi-layer perceptron processing.

[0024] Will and Connect and get the atomic topological characteristics f S , the atomic topological features f S The corresponding atomic features are fused to obtain the fused atomic features, whose dimension is (N+T1+T2)×D.

[0025] In one embodiment of the present invention, in S4, the method for obtaining the pairing position encoding information of the atom-simplex pair is as follows:

[0026] Based on the coordinate information of a single atom in the zero-dimensional simplex structure and the substructure position encoding information, the Euclidean distance d between all atoms i and any substructure k is calculated. i,k ;

[0027] Based on all the calculated Euclidean distances d i,k , we get the Euclidean distance matrix with the dimension of (N+T1+T2)×(N+T1+T2), where N represents the number of atoms, T1 represents the number of one-dimensional simplex structures, and T2 represents the number of two-dimensional simplex structures;

[0028] For each element d i,k Perform an affine transformation and map it to the specified range to obtain the element d' after the affine transformation i,k ;

[0029] The element d' after the affine transformation i,k Input into Gaussian density function for preprocessing to obtain the pairing position encoding information f of atom-simplex pair i.k .

[0030] In one embodiment of the present invention, the pairing position encoding information f i.k :

[0031] f i.k ={G(d' i,k ,μ x ,σ x )|x∈[1,D]}

[0032] Among them, d' i,k =A(d i,k ,t i,k ; a, b), A(d i,k ,t i,k ; a, b) = a·d i,k +b represents the affine transformation function, a and b are both trainable parameters, t i,k represents the pairing type of atom i and substructure k, represents the Gaussian density function, μ x and σ x are the mean and standard deviation of the Gaussian density function, x represents the pairing index number consisting of atoms and substructures, and D represents the number of pairs.

[0033] In one embodiment of the present invention, the Euclidean distance d i,k The calculation method is as follows:

[0034]

[0035] Among them, (x i ,y i ,z i ) represents the atomic position coordinate information, (x k ,y k ,z k ) represents the position coordinate information of any substructure, including the position coordinate information of one-dimensional simplex structure and two-dimensional simplex structure.

[0036] In one embodiment of the present invention, the pairing position encoding information f i.k The update method is as follows:

[0037] The pairing position encoding information fi.k Through the projection matrix M∈R D×H Mapped into an encoding vector that meets the input of the multi-head self-attention module

[0038]

[0039] The encoding vector Updated as a bias term in the iterative calculation of the multi-head self-attention module:

[0040]

[0041] The multi-head self-attention module uses the following formula to update parameters:

[0042]

[0043] in, represents the pairing position encoding vector of atom i and substructure k in layer l, The pair position encoding vector of atom i and substructure k in the (l+1)th layer, Concat represents the connection operation, d represents the hidden representation dimension, represents the characteristics of atom i in layer l, represents the characteristics of substructure k in layer l; Q l,h , K l,h 、V l,h is the query, key, and value mapping function of the h-th attention head in the l-th layer, h∈[1,H], H represents the number of attention heads; softmax(·) represents the activation function.

[0044] Based on the same inventive concept, the present invention also provides a molecular property prediction system for implementing the steps of the molecular property prediction method. The molecular property prediction system includes the following modules:

[0045] Molecular data acquisition module, used to obtain 3D molecular structure data with atomic coordinate information;

[0046] A topological structure extraction and encoding module is used to extract the topological structure in the 3D molecular structure data according to the set radius α and filling value σ, including a zero-dimensional simplex structure, a one-dimensional simplex structure and a two-dimensional simplex structure, and for each extracted one-dimensional simplex structure and two-dimensional simplex structure, calculate the average value of the spatial position of its vertices to obtain substructure position encoding information;

[0047] Wherein, the zero-dimensional simplex structure represents a single atom; the one-dimensional simplex structure represents a covalent bond obtained within a specified radius; the two-dimensional simplex structure represents a functional group formed by connecting multiple atoms within a specified radius;

[0048] An atomic feature fusion module is used to obtain an atomic topological feature for each atom using a one-dimensional simplex structure and a two-dimensional simplex structure as an index of a corresponding atomic feature, and fuse the atomic topological feature with its corresponding atomic feature to obtain a fused atomic feature;

[0049] A position coding generation module, configured to obtain pairing position coding information of atom-simplex pairs based on the coordinate information of individual atoms in the zero-dimensional simplex structure and the substructure position coding information;

[0050] A global molecular feature generation module is used to input the fused atomic features and the pairing position encoding information into a multi-head self-attention module, calculate the spatial relationship between the atomic features, and generate a global molecular feature;

[0051] The molecular property prediction module is used to obtain molecular property prediction results based on the global molecular features.

[0052] The present invention also provides a molecular property prediction device, comprising the molecular property prediction system.

[0053] The present invention also provides an electronic device, which includes a processor, a memory and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the molecular property prediction method.

[0054] The above technical solution of the present invention has the following advantages over the prior art:

[0055] This paper proposes a topological fusion network that aims to integrate the advantages of topological data analysis (TDA) and the Transformer model to enhance the perception and extraction of local substructure information. By introducing one-dimensional and two-dimensional topological simplex features (corresponding to substructures such as covalent bonds and functional groups, respectively), it not only improves the accuracy of the representation of local molecular structural information, but also enables more precise processing of molecular microscopic details. This paper effectively overcomes the shortcomings of existing technologies in processing fine-grained substructure information and significantly improves the performance of 3D molecular property prediction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0057] Figure 1 1 is a flow chart of a molecular property prediction method provided in an embodiment of the present invention;

[0058] Figure 21 is a schematic diagram of a specific process of a molecular property prediction method provided in an embodiment of the present invention;

[0059] Figure 3 are simplex structures extracted in the embodiments of the present invention, wherein (a) represents a zero simplex structure, (b) represents a zero simplex structure, and (c) represents a zero simplex structure;

[0060] Figure 4 1 is a schematic diagram of the structure of a molecular property prediction system provided in an embodiment of the present invention;

[0061] Description of the accompanying drawings in the specification: 100, molecular data acquisition module; 200, topological structure extraction and encoding module; 300, atomic feature fusion module; 400, position code generation module; 500, global molecular feature generation module; 600, molecular property prediction module. DETAILED DESCRIPTION

[0062] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0063] Example 1

[0064] See also Figure 1 and Figure 2 As shown, the present invention provides a method for predicting molecular properties, which includes the following steps:

[0065] S1: Get atomic coordinate information 3D molecular structure data (i = 1, 2, ..., N, where N is the number of atoms), including atomic coordinates, chemical bond information, and molecular topology information;

[0066] S2: extracting the topological structure in the 3D molecular structure data according to the set radius α and filling value σ, including a zero-dimensional simplex structure, a one-dimensional simplex structure, and a two-dimensional simplex structure, and calculating the average value of the vertex spatial position of each extracted one-dimensional simplex structure and two-dimensional simplex structure to obtain substructure position encoding information;

[0067] Wherein, the zero-dimensional simplex structure represents a single atom; the one-dimensional simplex structure represents a covalent bond obtained within a specified radius; the two-dimensional simplex structure represents a functional group formed by connecting multiple atoms within a specified radius;

[0068] S3: For each atom, using the one-dimensional simplex structure and the two-dimensional simplex structure as indexes of the corresponding atomic features to obtain the atomic topological features, and fusing the atomic topological features with the corresponding atomic features to obtain the fused atomic features;

[0069] S4: obtaining pairing position encoding information of atom-simplex pairs based on the single atomic coordinate information in the zero-dimensional simplex structure and the substructure position encoding information;

[0070] S5: Inputting the fused atomic features and the paired position encoding information into a multi-head self-attention module, calculating the spatial relationship between the atomic features, and generating a global molecular feature;

[0071] S6: Based on the global molecular features, obtain molecular property prediction results.

[0072] It can be seen from the above technical solutions that this molecular property prediction method has significant advantages. First, in terms of data acquisition, it comprehensively covers atomic coordinates, chemical bonds and topological structure information, laying the foundation for accurate analysis; when extracting the topological structure, with the help of radius α and filling value σ, the system obtains zero-dimensional, one-dimensional and two-dimensional simplex structures, accurately grasping the molecular microscopic topology; in the atomic feature fusion link, simplex is used as index to achieve deep fusion of atomic topology and atomic features, enhancing feature expression; when generating pairing position coding information, the atomic and substructure position information is combined to improve the position information dimension; the multi-head self-attention module processes the fused features and coding information, effectively calculates the atomic spatial relationship, and generates global molecular features; finally, the molecular properties are predicted based on the global features. The whole process is closely linked, which not only enhances the capture of local substructure information, but also improves the accuracy of molecular property prediction, and can effectively solve the shortcomings of existing technologies in processing fine-grained substructure information.

[0073] Furthermore, if Figure 2 As shown, in S2, the 3D molecular structure data obtained by the simplex-based structure extraction module (SSEM) is processed, and for each extracted one-dimensional simplex structure and two-dimensional simplex structure, the average value of the vertex spatial position is calculated to obtain the substructure position encoding information as follows:

[0074] like Figure 3 As shown in , by adjusting the radius parameter value, the number of simplex structures can be precisely controlled. When the radius value approaches the minimum value, only the simplex structures such as Figure 3 The zero-dimensional simplex structure shown in (a) corresponds to a single atomic node in a molecule. As the radius value increases, the system begins to recognize and extract Figure 3 The one-dimensional simplex structure shown in (b) includes the covalent bonds connecting the atomic pairs. After further adjustment to the appropriate radius value, as shown in Figure 3 The two-dimensional simplex structure shown in (c) is revealed, which corresponds to the functional group structure in the molecule. In summary, the above method can obtain the zero-dimensional simplex structure set, the one-dimensional simplex structure set and the two-dimensional simplex structure set in the 3D molecular structure data respectively.

[0075] The one-dimensional simplex structure set and the two-dimensional simplex structure set of the 3D molecular structure data are respectively:

[0076]

[0077] Among them, S 1 represents a set of one-dimensional simplex structures, represents the j-th one-dimensional simplex structure, express The coordinate information of atom i in S, j = 1, ..., T1, T1 represents the total number of covalent bonds extracted; 2 represents a set of two-dimensional simplex structures, represents the k-th two-dimensional simplex structure, express The coordinate information of atom i in , k = 1, ..., T2, T2 represents the total number of functional groups extracted; for example, if α is set to 1.2 to control the radius of the sphere used for simplex construction, and δ is set to 18, the number of covalent bonds extracted is The number of functional groups

[0078] For each extracted one-dimensional simplex structure, calculate the average value of its vertex spatial position For each extracted two-dimensional simplex structure, calculate the average value of its vertex spatial positions

[0079] Summarize all and Substructure position encoding information

[0080] However, these existing methods often ignore the local substructure information in the molecule (such as key topological features such as covalent bonds and functional groups) during the processing process, which may lead to the model's inaccurate description of the local characteristics of the molecular structure. Therefore, this application proposes a topological fusion network (TFN) based on the Transformer architecture. The network enhances the characteristics of each atom through two dimensions: atomic-level fusion and position-level fusion, thereby organically integrating substructure information into atomic features to enhance the model's overall characterization of molecular structure and its ability to capture local details.

[0081] Given that the Transformer model itself does not have an inherent ability to perceive the order of the input sequence, this study extracts position encoding from 3D position information and embeds it into atomic features, thereby providing the model with explicit information about the spatial arrangement of atoms within the molecule. Subsequently, the Transformer model treats each atomic feature as an independent token and dynamically generates an attention score by calculating the relationship between the token and all other atomic tokens in the molecule. Finally, the atomic features are weighted and aggregated according to their corresponding attention weights and further nonlinearly transformed through a feedforward network (FFN) to generate the final output representation of the molecule.

[0082] Furthermore, in S3, for each atom, the one-dimensional simplex structure and the two-dimensional simplex structure are used as indexes of the corresponding atomic features to obtain the atomic topological features, and the atomic topological features are fused with the corresponding atomic features to obtain the fused atomic features, and the method is as follows:

[0083] For each atom, a one-dimensional simplex structure S 1 and the two-dimensional simplex structure S 2 As an index of the corresponding atomic features, the features of the one-dimensional simplex structure (covalent bonds) and the two-dimensional simplex structure (functional groups) are aggregated;

[0084] The two D-dimensional atomic features corresponding to the aggregated one-dimensional simplex structural features are connected and input into the multi-layer perceptron for processing to obtain the D-dimensional covalent bond substructure feature vector

[0085] When three D-dimensional atomic features correspond to a two-dimensional simplex structure feature, that is, a functional group feature, a D-dimensional functional group substructure feature vector is generated after multi-layer perceptron processing.

[0086] Will and Connect and get the atomic topological characteristics f S , the atomic topological features f S The corresponding atomic features are fused to obtain the fused atomic features, whose dimension is (N+T1+T2)×D.

[0087] Furthermore, in S4, based on the single atom coordinate information in the zero-dimensional simplex structure and the substructure position coding information, a method for obtaining the pairing position coding information of the atom-simplex pair is as follows:

[0088] S41: Based on the coordinate information of a single atom in the zero-dimensional simplex structure and the substructure position encoding information, calculate the Euclidean distance d between all atoms i and any substructure k i,k, which is calculated as follows:

[0089]

[0090] Among them, (x i ,y i ,z i ) represents the atomic position coordinate information, (x k ,y k ,z k ) represents the position coordinate information of any substructure, including the position coordinate information of a one-dimensional simplex structure and a two-dimensional simplex structure;

[0091] S42: Based on all calculated Euclidean distances d i,k , and obtain the Euclidean distance matrix with the dimension of (N+T1+T2)×(N+T1+T2), where N represents the number of atoms, T1 represents the number of one-dimensional simplex structures, and T2 represents the number of two-dimensional simplex structures; S43: for each element d i,k Perform an affine transformation and map it to the specified range to obtain the element d' after the affine transformation i,k ;

[0092] S44: transform the element d' after affine transformation i,k Input into Gaussian density function for preprocessing to obtain the pairing position encoding information f of atom-simplex pair i.k :

[0093] f i.k ={G(d' i,k ,μ x ,σ x )|x∈[1,D]}

[0094] Among them, d' i,k =A(d i,k ,t i,k ; a, b), A(d i,k ,t i,k ; a, b) = a·d i,k +b represents the affine transformation function, a and b are both trainable parameters, t i,k represents the pairing type of atom i and substructure k, represents the Gaussian density function, μ x and σ x are the mean and standard deviation of the Gaussian density function, x represents the pairing index number consisting of atoms and substructures, and D represents the number of pairs.

[0095] Specifically, the pairing position encoding information f i.k The method for updating in the iterative calculation of the multi-head self-attention module is as follows:

[0096] The pairing position encoding information f i.k Through the projection matrix M∈R D×H Mapped into an encoding vector that meets the input of the multi-head self-attention module

[0097]

[0098] The encoding vector Updated as a bias term in the iterative calculation of the multi-head self-attention module:

[0099]

[0100] The multi-head self-attention module uses the following formula to update parameters:

[0101]

[0102] in, represents the pairing position encoding vector of atom i and substructure k in layer l, The pair position encoding vector of atom i and substructure k in the (l+1)th layer, Concat represents the connection operation, d represents the hidden representation dimension, represents the characteristics of atom i in layer l, represents the characteristics of substructure k in layer l; Q l,h , K l,h 、V l,h is the query, key, and value mapping function of the h-th attention head in the l-th layer, h∈[1,H], H represents the number of attention heads; softmax(·) represents the activation function.

[0103] Furthermore, in step S5, the fused atomic features and the paired position encoding information are input into a 15-layer multi-head self-attention module (MHSA). This module introduces the [CLS] token and uses the mean coordinates of all atoms and their substructures as the spatial position information of the atoms to aggregate and generate global molecular features. In step S6, the linear head is used to fine-tune the model to perform downstream tasks such as classification or regression, thereby enabling the prediction of properties such as molecular solubility, toxicity, and activity.

[0104] Example 2:

[0105] Based on the same inventive concept as that of Example 1, the present invention also provides a molecular property prediction system for implementing the steps of the molecular property prediction method described in Example 1. Figure 4 As shown, the molecular property prediction system includes the following modules:

[0106] The molecular data acquisition module 100 is used to acquire 3D molecular structure data with atomic coordinate information;

[0107] A topological structure extraction and encoding module 200 is used to extract the topological structure in the 3D molecular structure data according to the set radius α and fill value σ, including a zero-dimensional simplex structure, a one-dimensional simplex structure, and a two-dimensional simplex structure, and for each extracted one-dimensional simplex structure and two-dimensional simplex structure, calculate the average value of the spatial position of its vertices to obtain substructure position encoding information;

[0108] Wherein, the zero-dimensional simplex structure represents a single atom; the one-dimensional simplex structure represents a covalent bond obtained within a specified radius; the two-dimensional simplex structure represents a functional group formed by connecting multiple atoms within a specified radius;

[0109] The atomic feature fusion module 300 is used to obtain an atomic topological feature for each atom using the one-dimensional simplex structure and the two-dimensional simplex structure as indexes of the corresponding atomic feature, and fuse the atomic topological feature with the corresponding atomic feature to obtain a fused atomic feature;

[0110] A position code generation module 400 is configured to obtain pairing position code information of atom-simplex pairs based on the coordinate information of individual atoms in the zero-dimensional simplex structure and the substructure position code information;

[0111] A global molecular feature generation module 500 is configured to input the fused atomic features and the pairing position encoding information into a multi-head self-attention module, calculate the spatial relationship between the atomic features, and generate a global molecular feature;

[0112] The molecular property prediction module 600 is used to obtain molecular property prediction results based on the global molecular features.

[0113] A molecular property prediction system proposed in this embodiment is used to implement the aforementioned molecular property prediction method. Therefore, the specific implementation methods of the molecular property prediction system can be seen in the embodiment part of the aforementioned molecular property prediction method. For example, the molecular data acquisition module 100, the topological structure extraction and encoding module 200, the atomic feature fusion module 300, the position encoding generation module 400, the global molecular feature generation module 500 and the molecular property prediction module 600 are respectively used to implement the corresponding steps S1, S2, S3, S4, S5 and S6 in the molecular property prediction method in Example 1. Therefore, its specific implementation methods can refer to the descriptions of the corresponding embodiments of each part. In order to avoid redundancy, they will not be repeated here.

[0114] Example 3:

[0115] The present invention also provides a molecular property prediction device, including the molecular property prediction system described in the second embodiment.

[0116] In terms of hardware architecture, the device can also be equipped with high-performance data processing units, such as multi-core processors and large-capacity memory, to ensure rapid processing of massive amounts of 3D molecular structure data. Furthermore, to store large amounts of raw data, intermediate results, and model parameters, the device is equipped with a large-capacity storage device. Furthermore, to facilitate data exchange with external devices, the device has a high-speed data interface, enabling rapid data transmission.

[0117] At the software level, the device's integrated molecular property prediction system employs a modular design. The system's molecular data acquisition module 100 automatically collects 3D molecular structure data with atomic coordinate information from a variety of data sources, including atomic coordinates, chemical bond information, and molecular topology information. It also includes data cleaning and preprocessing capabilities to ensure the quality of input data. The topology structure extraction and encoding module 200 accurately extracts zero-dimensional, one-dimensional, and two-dimensional simplex structures from 3D molecular structure data based on a set radius α and fill value σ, and calculates substructure position encoding information, providing a foundation for subsequent feature fusion.

[0118] The atomic feature fusion module 300 uses one-dimensional and two-dimensional simplex structures as indexes of atomic features, deeply fuses atomic topological features with atomic features, and generates more representative fused atomic features. The position coding generation module 400 generates pairing position coding information of atom-simplex pairs based on the single atomic coordinate information and substructure position coding information in the zero-dimensional simplex structure, and provides position information for the multi-head self-attention module. The global molecular feature generation module 500 receives the fused atomic features and pairing position coding information, and generates global molecular features by calculating the spatial relationship between atomic features. The final molecular property prediction module 600 uses advanced machine learning algorithms or deep learning models based on global molecular features to obtain accurate molecular property prediction results.

[0119] The device also features a user-friendly interface, allowing users to easily configure system parameters such as radius α and fill value σ, while also viewing data processing progress and prediction results in real time. Furthermore, the device supports multi-threaded processing and distributed computing, significantly improving prediction efficiency and making it suitable for processing large-scale molecular data. In practical applications, this molecular property prediction device can be widely used in fields such as drug development and materials science, providing researchers with powerful tools and accelerating the R&D process.

[0120] Example 4:

[0121] The present invention also provides an electronic device, which includes a processor, a memory and a bus system, wherein the processor and the memory are connected via the bus system, the memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the molecular property prediction method described in Example 1.

[0122] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0123] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0124] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0126] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for predicting molecular properties, characterized in that: The following steps are involved: S1: Acquire 3D molecular structure data with atomic coordinate information; S2: extracting the topological structure in the 3D molecular structure data according to the set radius α and filling value σ, including a zero-dimensional simplex structure, a one-dimensional simplex structure, and a two-dimensional simplex structure, and calculating the average value of the vertex spatial position of each extracted one-dimensional simplex structure and two-dimensional simplex structure to obtain substructure position encoding information; Wherein, the zero-dimensional simplex structure represents a single atom; the one-dimensional simplex structure represents a covalent bond obtained within a specified radius; the two-dimensional simplex structure represents a functional group formed by connecting multiple atoms within a specified radius; S3: For each atom, using the one-dimensional simplex structure and the two-dimensional simplex structure as indexes of the corresponding atomic features to obtain the atomic topological features, and fusing the atomic topological features with the corresponding atomic features to obtain the fused atomic features; S4: obtaining pairing position encoding information of atom-simplex pairs based on the single atomic coordinate information in the zero-dimensional simplex structure and the substructure position encoding information; S5: Inputting the fused atomic features and the paired position encoding information into a multi-head self-attention module, calculating the spatial relationship between the atomic features, and generating a global molecular feature; S6: Based on the global molecular features, obtain molecular property prediction results.

2. The molecular property prediction method according to claim 1, characterized in that In S2, the method for obtaining the substructure position encoding information is as follows: By adjusting the value of the radius α, the one-dimensional simplex structure set and the two-dimensional simplex structure set of the 3D molecular structure data are obtained, respectively: Among them, S 1 represents a set of one-dimensional simplex structures, represents the j-th one-dimensional simplex structure, express The coordinate information of atom i in , j = 1, ..., T1, T1 represents the total number of one-dimensional simplex structures extracted; S 2 represents a set of two-dimensional simplex structures, represents the k-th two-dimensional simplex structure, express The coordinate information of atom i in , k = 1, ..., T2, T2 represents the total number of extracted two-dimensional simplex structures; For each extracted one-dimensional simplex structure, calculate the average value of its vertex spatial position For each extracted two-dimensional simplex structure, calculate the average value of its vertex spatial positions Summarize all and Substructure position encoding information 3. The molecular property prediction method according to claim 1, characterized in that In S3, the method for obtaining the fused atomic features is as follows: For each atom, a one-dimensional simplex structure S 1 and the two-dimensional simplex structure S 2 Aggregate the features of the one-dimensional simplex structure and the two-dimensional simplex structure as indices of the corresponding atomic features; The two D-dimensional atomic features corresponding to the aggregated one-dimensional simplex structural features are connected and input into the multi-layer perceptron for processing to obtain the D-dimensional covalent bond substructure feature vector When three D-dimensional atomic features correspond to a two-dimensional simplex structure feature, that is, a functional group feature, a D-dimensional functional group substructure feature vector is generated after multi-layer perceptron processing. Will and Connect and get the atomic topological characteristics f S , the atomic topological features f S The corresponding atomic features are fused to obtain the fused atomic features, whose dimension is (N+T1+T2)×D.

4. The molecular property prediction method according to claim 1, characterized in that In S4, the method for obtaining the pairing position encoding information of the atom-simplex pair is as follows: Based on the coordinate information of a single atom in the zero-dimensional simplex structure and the substructure position encoding information, the Euclidean distance d between all atoms i and any substructure k is calculated. i,k ; Based on all the calculated Euclidean distances d i,k , we get the Euclidean distance matrix with the dimension of (N+T1+T2)×(N+T1+T2), where N represents the number of atoms, T1 represents the number of one-dimensional simplex structures, and T2 represents the number of two-dimensional simplex structures; For each element d i,k Perform an affine transformation and map it to the specified range to obtain the element d' after the affine transformation i,k ; The element d' after the affine transformation i,k Input into Gaussian density function for preprocessing to obtain the pairing position encoding information f of atom-simplex pair i.k .

5. The molecular property prediction method according to claim 4, characterized in that: The pairing position encoding information f i.k : f i.k ={G(d' i,k ,m x ,s x )|x∈[1,D]} Among them, d' i,k =A(d i,k ,t i,k ; a, b), A(d i,k ,t i,k ; a, b) = a·d i,k +b represents the affine transformation function, a and b are both trainable parameters, t i,k represents the pairing type of atom i and substructure k, represents the Gaussian density function, μ x and σ x are the mean and standard deviation of the Gaussian density function, x represents the pairing index number consisting of atoms and substructures, and D represents the number of pairs.

6. The molecular property prediction method according to claim 5, characterized in that: The Euclidean distance d i,k The calculation method is as follows: Among them, (x i ,y i ,z i ) represents the atomic position coordinate information, (x k ,y k ,z k ) represents the position coordinate information of any substructure, including the position coordinate information of one-dimensional simplex structure and two-dimensional simplex structure.

7. The molecular property prediction method according to claim 4, characterized in that: The pairing position encoding information f i.k The update method is as follows: The pairing position encoding information f i.k Through the projection matrix M∈R D×H Mapped into an encoding vector that meets the input of the multi-head self-attention module The encoding vector Updated as a bias term in the iterative calculation of the multi-head self-attention module: The multi-head self-attention module uses the following formula to update parameters: in, represents the pairing position encoding vector of atom i and substructure k in layer l, The pair position encoding vector of atom i and substructure k in the (l+1)th layer, Concat represents the connection operation, d represents the hidden representation dimension, represents the characteristics of atom i in layer l, represents the characteristics of substructure k in layer l; Q l,h , K l,h 、V l,h is the query, key, and value mapping function of the h-th attention head in the l-th layer, h∈[1,H], H represents the number of attention heads; softmax(·) represents the activation function.

8. A molecular property prediction system, characterized in that: For implementing the steps of the molecular property prediction method according to any one of claims 1 to 7, the molecular property prediction system comprises the following modules: Molecular data acquisition module, used to obtain 3D molecular structure data with atomic coordinate information; A topological structure extraction and encoding module is used to extract the topological structure in the 3D molecular structure data according to the set radius α and filling value σ, including a zero-dimensional simplex structure, a one-dimensional simplex structure and a two-dimensional simplex structure, and for each extracted one-dimensional simplex structure and two-dimensional simplex structure, calculate the average value of the spatial position of its vertices to obtain substructure position encoding information; Wherein, the zero-dimensional simplex structure represents a single atom; the one-dimensional simplex structure represents a covalent bond obtained within a specified radius; the two-dimensional simplex structure represents a functional group formed by connecting multiple atoms within a specified radius; An atomic feature fusion module is used to obtain an atomic topological feature for each atom using a one-dimensional simplex structure and a two-dimensional simplex structure as an index of a corresponding atomic feature, and fuse the atomic topological feature with its corresponding atomic feature to obtain a fused atomic feature; A position coding generation module, configured to obtain pairing position coding information of atom-simplex pairs based on the coordinate information of individual atoms in the zero-dimensional simplex structure and the substructure position coding information; A global molecular feature generation module is used to input the fused atomic features and the pairing position encoding information into a multi-head self-attention module, calculate the spatial relationship between the atomic features, and generate a global molecular feature; The molecular property prediction module is used to obtain molecular property prediction results based on the global molecular features.

9. A molecular property prediction device, characterized in that: Comprising the molecular property prediction system as claimed in claim 9.

10. An electronic device, characterized in that: The electronic device includes a processor, a memory and a bus system, the processor and the memory are connected through the bus system, the memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the molecular property prediction method described in any one of claims 1 to 7.

Images