Method for obtaining molecular representation data and molecular attribute acquisition method

By obtaining the electron orbital structure data of each atom of the molecule and generating a wave function atomic convolution kernel, the problems of insufficient accuracy of molecular representation and high consumption of computing resources in the prior art are solved, and an efficient understanding of amorphous small molecules and Attribute prediction has driven the discovery of new drugs and new materials.

CN114678075BActive Publication Date: 2024-11-22BEIJING INFORMATION SCI & TECH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210302098.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-24
Publication Date
2024-11-22
Estimated Expiration
2042-03-24

AI Technical Summary

Technical Problem

The existing molecular representation methods are mainly based on macro abstraction, and it is difficult to accurately describe the physical forms of molecules in the microscopic world. Especially the construction of three-dimensional models of amorphous small molecules has high computing resources and insufficient accuracy, which affects the generation of new drugs and the discovery of new materials. progress.

Method used

By obtaining the electron orbital structure data of each atom of the molecule, atomic convolution kernels are generated based on wave function, and convolution operations are performed to generate visual molecular representation data in voxel space, and the properties of molecules are predicted using pre-trained molecular attributes prediction networks.

Benefits of technology

It significantly improves the accuracy of molecular representation, reduces the consumption of computing resources, improves the understanding of amorphous small molecules, and promotes the discovery of new drugs and new materials.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114678075B_ABST
    Figure CN114678075B_ABST
Patent Text Reader

Abstract

This disclosure provides a method for obtaining molecular representation data based on molecular structure information, comprising: for each atom constituting the molecule, acquiring the electronic orbital structure data of the outer shell of a single atom, and generating electronic orbital structure data based on wavefunction representation, using the wavefunction-represented electronic orbital structure data as atomic structure data; combining the wavefunction-represented electronic orbital structure data in the voxel space of a single atom constituting the molecule to form a wavefunction-represented atomic convolution kernel; and performing a convolution operation on each atom constituting the molecule through the corresponding wavefunction-represented atomic convolution kernel in the entire voxel space of all atoms constituting the molecule, thereby generating voxel-based visualized molecular representation data. This disclosure also provides a method for obtaining molecular properties.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of quantum physics technology, and in particular to a method for obtaining molecular representation data and a method for obtaining molecular properties. Background Art

[0002] Visual media is a type of media that carries visual information. Traditional visual media include graphics, images, videos, etc. With the development of virtual reality technology, three-dimensional models have also become a common type of visual media. Regardless of the type of visual media, it greatly improves people's understanding of things by utilizing vision, the most dependent way of perception for humans, and has derived a variety of popular applications, such as autonomous driving and digital twins.

[0003] At present, common visual media mainly focus on macroscopic things, using visual perception information for understanding and analysis. As a unit that is ubiquitous around people, the visualization research on molecules is also becoming increasingly popular. Since molecules are members of the microscopic world, their volume is too small to be directly observed by the human eye. Therefore, how to turn microscopic things into visible things while retaining some characteristics of the microscopic world is the main problem that needs to be solved in order to represent microscopic molecules as visual models and understand and analyze them on this basis.

[0004] The current mainstream methods for representing molecules include string representation and graph representation, and the latter has achieved remarkable results with the development of graph neural networks. However, these representations are based on the macroscopic abstract representation of molecules, which is a simplification of the microscopic world from a macroscopic perspective, losing the characteristics of the microscopic world and making it difficult to accurately describe the physical form of molecules in the microscopic world.

[0005] There are some works trying to build three-dimensional models of molecules, such as RNA protein prediction or crystal molecule prediction. The former is usually a large molecule containing thousands of atoms. Since the number of atoms is large enough, it is sufficient to model the molecule at the atomic level without exploring the microscopic physical structure of a smaller scale such as a single atom. The properties of the latter depend on the regular grid structure of the crystal itself rather than the structure of a single molecule. The construction of three-dimensional models for non-crystalline small molecules often lacks an accurate representation of the microscopic world and is prone to the dilemma of relying on a large amount of computing resources during the training process.

[0006] If a microscopic visual model can be established to help understand and analyze the structural information of non-crystalline small molecules from a visual perspective, improve the understanding of non-crystalline small molecules, and avoid excessive computing resource overhead during training, it can effectively improve the progress of many related fields, such as new drug generation, new material discovery, and so on. These fields currently face a common bottleneck, which is the need for an existing molecular database for screening. It is a difficult task to construct a lead molecule for a specific target in the laboratory. By predicting the properties of alternative molecules in advance, the dependence on molecular databases can be reduced, thereby improving the efficiency of constructing molecules that meet the requirements. Summary of the invention

[0007] In order to solve at least one of the above technical problems, the present disclosure provides a method for acquiring molecular representation data based on molecular structure information, and a method for acquiring molecular properties.

[0008] According to one aspect of the present disclosure, a method for obtaining molecular representation data based on molecular structure information is provided, comprising:

[0009] For each atom constituting the molecule, each electron orbital structure data of the outer layer of a single atom is obtained, and each electron orbital structure data represented by a wave function is generated, and each electron orbital structure data represented by the wave function is used as the atomic structure data;

[0010] In the voxel space of a single atom constituting a molecule, each electronic orbital structure data represented based on the wave function is combined to form an atomic convolution kernel based on the wave function, wherein the atomic convolution kernel based on the wave function reflects the physical structure of the atom;

[0011] In the entire voxel space of all atoms constituting the molecule, a convolution operation is performed on each atom constituting the molecule through the atomic convolution kernel based on the wave function corresponding to each atom, and after each atom of the molecule is convoluted by the corresponding atomic convolution kernel based on the wave function, voxel-based visualized molecular representation data is generated;

[0012] Among them, the voxel space refers to establishing a coordinate system in three-dimensional space, taking the geometric center of the molecule as the origin, extracting the three axis directions of the coordinate system through PCA, rotating the molecule to each axis, and the position corresponding to the atom in the space is a unit. Each unit contains a vector, the vector length corresponds to the atomic type, and each vector represents the probability distribution of the electron cloud around the corresponding atomic type at the position.

[0013] According to the method for obtaining molecular representation data based on molecular structure information in at least one embodiment of the present disclosure, when the number of outer electron orbitals of the atom is two, the structure data of each electron orbital represented by the wave function includes:

[0014]

[0015]

[0016] The meanings of each expression are as follows:

[0017] r, represents the distance from the electron p in the i-th electron orbital to the nucleus p0, the value of r is ||p,p0||, and the value of i is 1 or 2;

[0018] Z, represents the nuclear charge number of the atom;

[0019] a0 is a constant.

[0020] According to the method for obtaining molecular representation data based on molecular structure information in at least one embodiment of the present disclosure, when the number of outer electron orbits of the atom is two, the atomic convolution kernel based on the wave function is expressed as:

[0021]

[0022] The meanings of each expression are as follows:

[0023] (x, y, z), represents the coordinates of each atom in voxel space;

[0024] Ψ i , each electron orbital represented by a wave function, i takes the value of 1 or 2;

[0025] l, represents the number of electron layers, l takes the value of 1 or 2;

[0026] r represents the distance from the electron p in the i-th electron orbit to the nucleus p0.

[0027] According to at least one embodiment of the method for obtaining molecular representation data based on molecular structure information of the present disclosure, the atomic convolution kernel based on the wave function of each atom performs a convolution operation, including:

[0028]

[0029] The meaning of each expression is as follows:

[0030] P, represents the atoms in the molecule;

[0031] δ x , δ y , δ z : The offset of the atom p in the three coordinate axes.

[0032] According to another aspect of the present disclosure, a method for obtaining molecular properties is provided, comprising:

[0033] Using the above method to generate voxel-based visualized molecular representation data for the molecules whose properties are to be predicted;

[0034] Inputting the voxel-based visualized molecular representation data into a molecular property prediction network to perform molecular property prediction, wherein the molecular property prediction network is obtained through pre-training;

[0035] The molecular property prediction network outputs the properties of the molecule whose properties are to be predicted.

[0036] According to the molecular property acquisition method of at least one embodiment of the present disclosure, the composition of the molecular property prediction network includes:

[0037] The molecular property prediction network is based on a neural network, including multiple convolution operation layers and multiple pooling operation layers;

[0038] The molecular property prediction network includes an average channel pooling layer, wherein the average channel pooling layer takes the average value of each voxel in all channels as the value of the voxel;

[0039] The channels reflect the distribution of atoms in space, and when the channels are different, they reflect the distribution of different atoms in space.

[0040] According to the molecular property acquisition method of at least one embodiment of the present disclosure, the training process of the molecular property prediction network includes:

[0041] Initialize the parameters of the attribute prediction network;

[0042] Divide the data set into training set, validation set and test set;

[0043] The voxel-based visualized molecular representation data and molecular properties in the training set are used as input to the molecular property prediction network for training;

[0044] Verify and test using the validation set and test set data respectively;

[0045] The training effect of the molecular property prediction network was evaluated by the mean absolute error, and training and verification were repeated until the expected training indicators were achieved.

[0046] According to the molecular property acquisition method of at least one embodiment of the present disclosure, the voxel-based visualized molecular representation data is normalized to ensure that each atom of the voxel-based visualized molecular representation data has a uniform size specification when input into the molecular property prediction network.

[0047] According to the molecular property acquisition method of at least one embodiment of the present disclosure, the molecular properties obtained by the molecular property prediction network include at least one property of rotation constant, dipole moment, and isotropic polarizability.

[0048] According to the molecular property acquisition method of at least one embodiment of the present disclosure, the voxel-based visualized molecule is processed by a diffusion strategy, in which each atom in the molecule is diffused by using the number zero or other positive integers as diffusion coefficients. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings illustrate exemplary embodiments of the present disclosure and together with the description serve to explain the principles of the present disclosure. These drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification.

[0050] Figure 1 It is a flowchart of a method for obtaining molecular representation data based on molecular structure information according to one embodiment of the present disclosure.

[0051] Figure 2 A schematic flow chart of a method for obtaining molecular properties according to one embodiment of the present disclosure.

[0052] Figure 3 It is a flowchart of a molecular property prediction network training method according to one embodiment of the present disclosure.

[0053] Figure 4 It is a schematic diagram of a molecular property prediction network structure according to an embodiment of the present disclosure.

[0054] Figure 5 It is a schematic diagram of a molecular property prediction network structure according to another embodiment of the present disclosure.

[0055] Figure 6 It is a schematic diagram of the structure of a device for acquiring molecular representation data based on molecular structure information according to one embodiment of the present disclosure.

[0056] Figure 7 It is a schematic diagram of the structure of a molecular property acquisition device according to an embodiment of the present disclosure.

[0057] Description of Reference Numerals

[0058] 1000 Device for obtaining molecular representation data based on molecular structure information

[0059] 2000 Molecular property acquisition device

[0060] 1002 Atomic structure data acquisition module

[0061] 1004 Convolution kernel generation module

[0062] 1006 Visual Molecular Representation Data Acquisition Module

[0063] 1008 Molecular property prediction module

[0064] 1100 Bus

[0065] 1200 processor

[0066] 1300 Memory

[0067] 1400 Other circuits. DETAILED DESCRIPTION

[0068] The present disclosure is further described in detail below in conjunction with the accompanying drawings and implementations. It is understood that the specific implementations described herein are only used to explain the relevant content, rather than to limit the present disclosure. It should also be noted that, for ease of description, only the parts related to the present disclosure are shown in the accompanying drawings.

[0069] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in the present disclosure can be combined with each other. The technical solution of the present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0070] Unless otherwise specified, the exemplary embodiments / embodiments shown will be understood as providing exemplary features of various details of some ways in which the technical concept of the present disclosure can be implemented in practice. Therefore, unless otherwise specified, the features of the various embodiments / embodiments can be combined, separated, interchanged and / or rearranged without departing from the technical concept of the present disclosure.

[0071] The use of cross-hatching and / or shading in the accompanying drawings is generally used to make the boundaries between adjacent components clear. As such, unless otherwise specified, the presence or absence of cross-hatching or shading does not convey or indicate any preference or requirement for the specific materials, material properties, dimensions, proportions, commonalities between the components shown, and / or any other characteristics, attributes, properties, etc. of the components. In addition, in the accompanying drawings, the sizes and relative sizes of the components may be exaggerated for clarity and / or descriptive purposes. When the exemplary embodiments can be implemented differently, the specific process sequence can be performed in a different order than described. For example, two successively described processes can be performed substantially simultaneously or in an order opposite to the described order. In addition, the same figure numbers represent the same components.

[0072] When a component is referred to as being "on," "over," "connected to," or "coupled to" another component, the component may be directly on, directly connected to, or directly coupled to the other component, or intervening components may be present. However, when a component is referred to as being "directly on," "directly connected to," or "directly coupled to" another component, there are no intervening components. For this purpose, the term "connected" may refer to a physical connection, an electrical connection, etc., with or without intervening components.

[0073] The terminology used in this embodiment is to describe the purpose of specific embodiments, and is not intended to be restrictive. As used herein, unless the context clearly indicates otherwise, the singular form "one (kind, person)" and "described (the)" are also intended to include plural forms. In addition, when the term "comprising" and / or "including" and their variations are used in this specification, it is explained that there are stated features, integral bodies, steps, operations, parts, assemblies and / or their groups, but one or more other features, integral bodies, steps, operations, parts, assemblies and / or their groups are not excluded from existence or addition. It should also be noted that, as used herein, the term "substantially", "approximately" and other similar terms are used as approximate terms and not as degree terms, so that they are used to explain the inherent deviation of the measured value, calculated value and / or the value provided that those of ordinary skill in the art will recognize.

[0074] Figure 1 It is a flowchart of a method for obtaining molecular representation data based on molecular structure information according to one embodiment of the present disclosure.

[0075] like Figure 1 As shown, the method S100 for acquiring molecular representation data based on molecular structure information includes the following steps.

[0076] In step S102, for each atom constituting the molecule, each electron orbital structure data of the outer layer of a single atom is obtained, and each electron orbital structure data represented by a wave function is generated, and each electron orbital structure data represented by a wave function is used as atomic structure data. When the number of outer electron orbits of an atom is two, each electron orbital structure data represented by a wave function includes:

[0077]

[0078]

[0079] Among them, the meaning of each expression is as follows: r represents the distance from the electron p of the i-th electron orbit to the nucleus p0, the value of r is ||p,p0||, and the value of i is 1 or 2; Z represents the nuclear charge number of the atom; a0 is a constant. This wave function reflects the properties of the electron orbit, the inner orbit is narrower, and the outer orbit is wider. In addition, the wave function also reflects the attenuation (shrinking) effect on the electron orbits of each layer as the nuclear charge number increases. In this embodiment, for the data of atoms within 2 layers, the QM9 data set that can be used only contains 5 types of atoms: H, C, N, O, and F, and the electron orbits of these atomic types do not exceed two layers.

[0080] In step S104, in the voxel space of a single atom constituting a molecule, the electronic orbital structure data represented by the wave function are combined to form an atomic convolution kernel based on the wave function, which reflects the physical structure of the atom. When the number of outer electron orbitals of an atom is two, the atomic convolution kernel based on the wave function is expressed as:

[0081]

[0082] The meanings of the expressions are as follows: (x, y, z) represents the coordinates of each atom in the voxel space; Ψ i , each electron orbital represented by a wave function, i is 1 or 2; l, represents the number of electron layers, l is 1 or 2; r, represents the distance from the electron p in the i-th electron orbital to the nucleus p0. The atomic convolution kernel based on the wave function reflects the physical structure of the atom by taking into account the influence of both the electron and nuclear charge numbers.

[0083] In step S106, in the entire voxel space of all atoms constituting the molecule, a convolution operation is performed on each atom constituting the molecule through the atomic convolution kernel based on the wave function corresponding to each atom. After the convolution operation of each atom of the molecule by the corresponding atomic convolution kernel based on the wave function, a voxel-based visualized molecular representation data is generated. Among them, the voxel space refers to the establishment of a coordinate system in a three-dimensional space, taking the geometric center of the molecule as the origin, extracting the three axis directions of the coordinate system through PCA, rotating the molecule to each axis, and the position corresponding to the atom in the space is a unit. Each unit contains a vector, the length of the vector corresponds to the atomic type, and each vector represents the probability distribution of the surrounding electron cloud of the corresponding atomic type at the position. In a specific implementation, the convolution operation based on the atomic convolution kernel of each atom includes:

[0084]

[0085] The meanings of the expressions are as follows: P represents the atom in the molecule; p(x, y, z) represents the position of atom p in the 3D space coordinate system; δ x , δ y , δ z : The offset of the atom p in the three coordinate axes.

[0086] According to quantum physics theory, atoms, as components of molecules, contain a nucleus P0 at the center and several electrons moving around the nucleus. Among them, the volume of the nucleus is very small and can be ignored compared to the atom. Electrons are distributed in different orbits outside the nucleus according to different energy levels, forming a spherical area around the nucleus, called an electron cloud. The density of the electron cloud indicates the probability of electrons appearing in different areas. When a point is used to describe an atom, it often refers to the position of the nucleus, or the center of the atom. In fact, for molecules, the motion state of electrons in different atoms has changed to form chemical bonds, thereby combining into a molecule. Therefore, by representing the distribution of electrons in atoms, the relationship between atoms in the molecule can be better reflected. The spherical area corresponding to the atom, that is, the diffusion range of the electron cloud, has a radius mainly affected by two factors: the number of electron layers and the number of nuclear charges. Generally speaking, the more electron layers there are, the smaller the nuclear charge number, the larger the atomic radius, that is, the larger the range of electron diffusion.

[0087] The molecular representation data obtained by the method for obtaining molecular representation data based on molecular structure information of this embodiment has the following technical advantages. The molecular representation data obtained by this embodiment significantly improves the effect of molecular representation and better simulates the atomic physics model. The atomic convolution kernel (wave transform kernel) of the wave function of this embodiment reflects the influence of the number of electron layers and nuclear charge numbers of different atoms on the atomic radius. The channel-by-channel convolution operation (convolution operation) effectively encodes the interference relationship between multiple atoms of the same type in the same voxel. This model not only expresses the difference between different atomic types by modeling the influence of nuclear charge number and electron layer number on atomic radius, but also reflects the relationship between atoms to a certain extent, so a more accurate representation of molecules can be obtained.

[0088] Figure 5Yes, the new model significantly improves the effect of molecular representation by better simulating the atomic physics model. Among them, the atomic convolution kernel of the wave function reflects the influence of the number of electron layers and nuclear charge numbers of different atoms on the atomic radius. The channel-by-channel convolution operation effectively encodes the interference relationship between multiple atoms of the same type in the same voxel. By modeling the influence of nuclear charge number and electron layer number on atomic radius, this model not only expresses the difference between different atomic types, but also reflects the relationship between atoms to a certain extent, so it can obtain a more accurate representation of molecules.

[0089] Figure 2 It is a schematic flow chart of a method for obtaining molecular properties according to at least one embodiment of the present disclosure.

[0090] like Figure 2 As shown, the molecular property acquisition method S200 includes the following steps.

[0091] In step S202, the molecules whose properties are to be predicted are subjected to the above method to generate voxel-based visualized molecular representation data.

[0092] In step S204, the voxel-based visualized molecular representation data is input into a molecular property prediction network to perform molecular property prediction, and the molecular property prediction network is obtained through pre-training.

[0093] In step S206, the molecular property prediction network outputs the property of the molecule to be predicted, and the molecular property includes multiple properties. In this embodiment, the molecular property obtained by the molecular property prediction network includes at least one property of rotation constant, dipole moment, and isotropic polarizability.

[0094] Figure 3 It is a flowchart of a molecular property prediction network training method according to one embodiment of the present disclosure.

[0095] like Figure 3 As shown, the training method S300 of the molecular property prediction network includes the following steps.

[0096] In step S302, the parameters of the attribute prediction network are initialized. In this embodiment, a batch training method is adopted, and each batch contains attribute data of 8 molecules. The learning rate is set to 1e-2, and the model is trained using the AdaDelta optimization algorithm, and batch normalization and dropout (dropout ratio = 0.5) are used to prevent overfitting.

[0097] In step S304, the data set is divided into a training set, a validation set and a test set. In this embodiment, the QM9 data set is used for training, and the data set contains the geometric information and physical properties of 133,885 molecules. In this embodiment, 15 properties of each molecule in the data set are predicted, and the properties include the rotation constant, dipole moment, isotropic polarizability, etc. of the molecule. The data set is divided into a training set, a validation set and a test set in a ratio of 8:1:1. Since the ranges between the attributes are different and the positive and negative are uncertain, the inverse tangent operation is used to normalize all attribute values.

[0098] In step S306, the voxel-based visualized molecular representation data and molecular properties in the training set are input as input to the molecular property prediction network for training.

[0099] In step S308, verification and testing are performed using the data of the verification set and the test set, respectively, wherein the verification set and the test set respectively include voxel-based visualized molecular representation data and molecular properties.

[0100] In step S310, the training effect of the molecular attribute prediction network is evaluated by the mean absolute error, and training, verification and testing are repeated until the expected training index is reached. In this embodiment, the molecular attribute prediction result is used as the evaluation target. On the basis of the 1024-dim vector output by the molecular attribute prediction network, a fully connected layer containing 15 nodes is used as the output layer, and each node outputs the regression result for an attribute. The mean absolute error (MAE) is used as the evaluation index of network performance.

[0101] In this implementation, the machine software and hardware system models for training are as follows: CentOS 7with CPU E5-2620v4@2.10GHz, 2NVIDIA GeForce GTX 1080Ti GPUs (22GBmemory in total).

[0102] In the above step S306, the voxel-based visualized molecular representation data can also be normalized to ensure that each atom of the voxel-based visualized molecular representation data has a uniform size specification when input into the molecular property prediction network. In this embodiment, the impact of using different norm strategies to process molecular voxels on the molecular property prediction task is compared. Since the size and shape of the molecules in the data set are different, the size of the voxel space occupied is also different, and it is necessary to perform a normalization process on the voxels of different molecules to ensure that each atom is unified into the same size before entering the network. The first strategy is based on the voxel perspective. For voxels that occupy less than half the standard size, the molecule is enlarged to twice the size, and then filled with 0. The second strategy directly fills the molecular voxels into the standard input size without scaling the molecules. The third strategy is based on the molecular perspective. The molecule is first scaled to the standard size, and then voxel division and 0 filling on the corresponding dimension are performed. The experimental results are shown in Table 2.

[0103] Table 2: Verification results of different normalization strategies

[0104] norm strategy train mae valid mae time(s) 1 0.1299652 0.1307257 10.979 2 0.1298442 0.1302136 9.047 3 0.1299810 0.1309107 10.977

[0105] From the experimental results in Table 2, we can see that scaling the molecule will lose some accuracy. This is because the size of the molecule and the distance between atoms have a certain influence on its properties, and scaling will weaken some of the information carried by the molecule size. At the same time, the norm strategy will also affect the effect of diffusion, causing some information to exceed the boundary and be lost.

[0106] In the above step S306, the voxel-based visualized molecules can also be processed by a diffusion strategy, in which each atom in the molecule is diffused by using the number zero or other positive integers as the diffusion coefficient. In this embodiment, the impact on the molecular property prediction task when different atomic diffusion strategies are used to represent the range in which the atoms diffuse is first compared. In this experiment, a total of five different atomic diffusion strategies were used. Among them, the first strategy does not diffuse any atoms, that is, each atom only occupies the voxel where its center is located. For the case of using Gaussian smooth diffusion, the diffusion coefficients are set to 1 and 4 to compare their different effects. In addition, another diffusion method is also compared. This method also modifies the Gaussian convolution kernel used for atomic diffusion, but the new kernel function used fails to effectively reflect the spatial structural characteristics of atoms. The experimental results are shown in Table 1.

[0107] Table 1: Verification results of different diffusion methods

[0108]

[0109]

[0110] In this embodiment, several molecules are randomly selected and different diffusion strategies are used to obtain the morphology of the molecules in three-dimensional space. From the experimental results, it can be found that the MAE of Gaussian smooth diffusion and ACS17 diffusion using different diffusion parameters are equivalent. That is, for molecular representation, atomic diffusion is continuous information. Without changing the spatial structure of atoms, simply adjusting the diffusion coefficient cannot improve the accuracy of various tasks.

[0111] Figure 5 It is a schematic diagram of a molecular property prediction network structure according to one embodiment of the present disclosure.

[0112] like Figure 5 As shown, the composition of the molecular property prediction network includes the following parts.

[0113] The molecular property prediction network is based on a neural network, including multiple convolution operation layers and multiple pooling operation layers. The average channel pooling layer uses the average value of each voxel in all channels as the value of the voxel. The channel reflects the distribution of atoms in space. When the channels are different, they reflect the distribution of different atoms in space.

[0114] Figure 4 and Figure 5 is a schematic diagram of a molecular property prediction network structure according to at least one embodiment of the present disclosure. Figure 4 As shown in Figure 5, the schematic diagram of the structure of network A. Network A and network B respectively include multiple convolutional layers and multiple pooling layers. In addition, network B also has an ACP layer (average channel pooling layer), which makes the average value of each voxel in all channels as the value of the voxel

[0115] In this embodiment, network A, network B, voxelnet and vismolA+B are selected as the molecular attribute prediction network structure, and the molecules under different networks are compared by normalizing and diffusing the molecules.

[0116] In this embodiment, SchNet and the method in acs17 are selected as comparative experiments. The former is a relatively mature graph structure network, which compares the effects of different representation methods on the prediction of molecular properties. The latter is a voxel-based three-dimensional convolutional network, which uses a modified Gaussian convolution kernel to diffuse molecules and extract features through continuous convolutional layers. In addition, this embodiment conducts a comparative experiment on the two proposed networks to compare the effects of spatial and electronic properties on property prediction. This embodiment also attempts to merge the two networks to jointly affect the molecular property prediction results from two aspects.

[0117] network train mae valid mae train time(s) voxelnet 0.1278553 0.1281986 20.634 vismol A 0.1298442 0.1302136 9.356 vismol B 0.1302792 0.1310272 8.084 vismol A+B 0.1322449 0.1324123 28.270

[0118] From the experimental results, it can be seen that Vismol reduces the number of parameters compared to VoxelNet through depthwise convolution, thereby shortening the training time without losing much accuracy. By comparing the two networks proposed in this embodiment, the effectiveness of processing molecular voxels from an electronic perspective is demonstrated.

[0119] Both Experiment 1 and Experiment 2 used the A network.

[0120] Experiment 1: Different diffusion (abbreviated as diff in other parts of this disclosure, with the same meaning) strategies: no diffusion at the center point, Gaussian 1, Gaussian 4, wave, and VisMol.

[0121] First group: Each atom occupies one voxel.

[0122] Second group: Gaussian 1 diffusion is performed on each atom to multiple voxels.

[0123] The third group: Gaussian 4 diffusion is performed on each atom to multiple voxels.

[0124] The fourth group: wave diffusion is performed on each atom to multiple voxels.

[0125] Group 5: Perform VisMol diffusion to multiple voxels for each atom.

[0126] From the above experiment 1, we can know that: first, different diffusion strategies have different prediction accuracy for molecular properties. Second, different diffusion strategies have different effects on the property prediction task.

[0127] Experiment 2: Different norm strategies: how to construct 32x32x32 input for molecules of different volumes.

[0128] The first group: further subdivide each voxel and select 32x32x32 as input.

[0129] Example 1: The molecule occupies the original space of 15x13x9. The original voxel size of 0.1A is further divided into 0.05A, so the space occupied by the molecule becomes 30x26x18, and then filled to 32x32x32.

[0130] Example 2: The molecule occupies the original space of 16x24x8, which is padded to 32x32x32.

[0131] The second group: directly fill to 32x32x32.

[0132] Example 1: The molecule occupies the original space of 15x13x9, so it is directly filled to 32x32x32.

[0133] Example 2: The molecule occupies the original space of 16x24x8, which is padded to 32x32x32.

[0134] The third group: calculates the new size according to the three dimensions and re-divides the voxels.

[0135] Example 1: The molecule occupies the original space of 15x13x9, and the X axis is the longest. The new size is calculated to be L=1A*15 / 32=0.47A. The voxels are divided according to 0.047A, and the space occupied by the molecule becomes 32x28x20, which is then filled to 32x32x32.

[0136] Example 2: The molecule occupies the original space of 16x24x8, and the Y axis is the longest. The new size is calculated to be L=1A*24 / 32=0.75A. The voxels are divided according to 0.075A, and the space occupied by the molecule becomes 22x32x11, which is then filled to 32x32x32.

[0137] From the above experiments, we can see that: first, the molecular property prediction accuracy of different normalization (norm) strategies is different. Second, the training speed of different norm strategies is different. Third, the number of zero padding used for filling is 3<1<2. Too much zero padding may affect the convergence speed and increase the number of training rounds, but guess 2 has the best accuracy because 2 does not change the scaling of the molecule.

[0138] Experiment 3, the attribute prediction results can be compared with classic ML, voxelnet, and 3d-LeNet networks with different parameters and strategies.

[0139] The first group: using the classic ML regression method.

[0140] The second group: uses VxoelNet (i.e. the existing technology ACS17 method) to extract features using continuous three-dimensional convolution.

[0141] The third group: the fifth group diff + the first group cellsize + the second group norm + A network.

[0142] Group 4: Group 3 method + B network.

[0143] Group 5: The third group method + A + B network.

[0144] From the above experiment 3, we can know: the property prediction accuracy of different methods is given. The training speed of different methods. Analysis: The accuracy of 3 is expected to be better than 4, which proves the effectiveness of learning the distribution of electrons in molecules; the effectiveness of the B network; the combination of A and B networks needs to be explored in the future.

[0145] Figure 6It is a schematic diagram of the structure of a device for acquiring molecular representation data based on molecular structure information according to one embodiment of the present disclosure.

[0146] like Figure 6 As shown, the apparatus 1000 for acquiring molecular representation data based on molecular structure information includes the following components.

[0147] The atomic structure data acquisition module 1002 acquires the electron orbital structure data of the outer layer of each atom constituting the molecule, generates the electron orbital structure data represented by the wave function, and uses the electron orbital structure data represented by the wave function as the atomic structure data.

[0148] The convolution kernel generation module 1004 combines the various electronic orbital structure data represented by the wave function in the voxel space of the single atom constituting the molecule to form an atomic convolution kernel based on the wave function, and the atomic convolution kernel based on the wave function reflects the physical structure of the atom.

[0149] The visualized molecular representation data acquisition module 1006 performs a convolution operation on each atom constituting the molecule through an atomic convolution kernel based on a wave function corresponding to each atom in the entire voxel space of all atoms constituting the molecule. After each atom of the molecule is convolved with the corresponding atomic convolution kernel based on a wave function, voxel-based visualized molecular representation data is generated.

[0150] It should be noted that, in the implementation of the present device, the specific implementation details of each module are consistent with the method for obtaining molecular representation data based on molecular structure information provided in the present disclosure, and will not be repeated here.

[0151] Figure 7 It is a schematic diagram of the structure of a molecular property acquisition device according to an embodiment of the present disclosure.

[0152] like Figure 7 As shown, the molecular property acquisition device 2000 includes the following components.

[0153] Atomic structure data acquisition module 1002, convolution kernel generation module 1004, visualized molecule representation data acquisition module 1006, and molecular property prediction module 1008. The molecular property prediction module 1008 receives the molecular representation data generated by the visualized molecule representation data acquisition module 1006 based on the molecular structure information and outputs the molecular properties. Atomic structure data acquisition module 1002, convolution kernel generation module 1004, visualized molecule representation data acquisition module 1006 Figure 6 In addition, the implementation details of each module of the device are consistent with the corresponding method and will not be repeated here.

[0154] Figure 6 to Figure 7An example diagram of an apparatus using a hardware implementation of a processing system is shown. The apparatus may include a corresponding module that performs each or several steps in the above flowchart. Therefore, each step or several steps in the above flowchart may be performed by a corresponding module, and the apparatus may include one or more of these modules. The module may be one or more hardware modules specifically configured to perform the corresponding steps, or implemented by a processor configured to perform the corresponding steps, or stored in a computer-readable medium for implementation by a processor, or implemented by some combination.

[0155] The hardware structure can be implemented using a bus architecture. The bus architecture can include any number of interconnecting buses and bridges, depending on the specific application and overall design constraints of the hardware. The bus 1100 connects various circuits including one or more processors 1200, memory 1300, and / or hardware modules together. The bus 1100 can also connect various other circuits 1400 such as peripherals, voltage regulators, power management circuits, external antennas, etc.

[0156] The bus 1100 may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the figure only uses one connecting line, but does not mean that there is only one bus or one type of bus.

[0157] Any process or method description in the flowchart or otherwise described herein can be understood as a module, fragment or portion of a code representing one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present disclosure includes other implementations, in which the functions may not be performed in the order shown or discussed, including performing the functions in a substantially simultaneous manner or in a reverse order according to the functions involved, which should be understood by a person skilled in the art of the present disclosure. The processor performs the various methods and processes described above. For example, the method implementation in the present disclosure can be implemented as a software program, which is tangibly contained in a machine-readable medium, such as a memory. In some embodiments, part or all of the software program can be loaded and / or installed via a memory and / or a communication interface. When the software program is loaded into the memory and executed by the processor, one or more steps in the method described above can be performed. Alternatively, in other embodiments, the processor can be configured to perform one of the above methods in any other appropriate manner (e.g., by means of firmware).

[0158] The logic and / or steps represented in the flowchart or otherwise described herein may be embodied in any readable storage medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other system that can fetch instructions from and execute instructions on an instruction execution system, apparatus, or device), or in combination with such instruction execution systems, apparatuses, or devices.

[0159] For the purposes of this specification, a "readable storage medium" may be any device that can contain, store, communicate, propagate or transmit a program for use with or in conjunction with an instruction execution system, device or apparatus. More specific examples of readable storage media (a non-exhaustive list) include the following: an electrical connection with one or more wirings (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and editable read-only memory (EPROM or flash memory), a fiber optic device, and a portable read-only memory (CDROM). In addition, the readable storage medium may even be paper or other suitable medium on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a memory.

[0160] It should be understood that the various parts of the present disclosure can be implemented in hardware, software, or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented in software stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0161] A person skilled in the art may understand that all or part of the steps of implementing the above-mentioned implementation method may be completed by instructing related hardware through a program, and the program may be stored in a readable storage medium, which, when executed, includes one or a combination of the steps of the implementation method.

[0162] In addition, each functional unit in each embodiment of the present disclosure may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a readable storage medium. The storage medium may be a read-only memory, a disk or an optical disk, etc.

[0163] In the description of this specification, the description with reference to the terms "one embodiment / method", "some embodiments / methods", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment / method or example are included in at least one embodiment / method or example of the present application. In this specification, the schematic representations of the above terms are not necessarily the same embodiment / method or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments / methods or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments / methods or examples described in this specification and the features of the different embodiments / methods or examples, unless they are contradictory.

[0164] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0165] Those skilled in the art should understand that the above embodiments are only for the purpose of clearly illustrating the present disclosure, and are not intended to limit the scope of the present disclosure. For those skilled in the art, other changes or modifications may be made based on the above disclosure, and these changes or modifications are still within the scope of the present disclosure.

Claims

1. A method for obtaining molecular representation data based on molecular structure information, characterized in that: include: For each atom constituting the molecule, each electron orbital structure data of the outer layer of a single atom is obtained, and each electron orbital structure data represented by a wave function is generated, and each electron orbital structure data represented by the wave function is used as the atomic structure data; In the voxel space of a single atom constituting a molecule, each electronic orbital structure data represented based on the wave function is combined to form an atomic convolution kernel based on the wave function, wherein the atomic convolution kernel based on the wave function reflects the physical structure of the atom; and In the entire voxel space of all atoms constituting the molecule, a convolution operation is performed on each atom constituting the molecule through the atomic convolution kernel based on the wave function corresponding to each atom, and after each atom of the molecule is convoluted by the corresponding atomic convolution kernel based on the wave function, voxel-based visualized molecular representation data is generated; Among them, the voxel space refers to establishing a coordinate system in three-dimensional space, taking the geometric center of the molecule as the origin, extracting the three axis directions of the coordinate system through PCA, rotating the molecule to each axis, and the position corresponding to the atom in the space is a unit. Each unit contains a vector, the vector length corresponds to the atomic type, and each vector represents the probability distribution of the electron cloud around the corresponding atomic type at the position.

2. The method for obtaining molecular representation data based on molecular structure information according to claim 1, characterized in that: When the number of outer electron orbitals of the atom is two, the structure data of each electron orbital represented by the wave function includes: The meanings of each expression are as follows: r, represents the distance from the electron p in the i-th electron orbital to the nucleus p0, the value of r is ||p,p0||, and the value of i is 1 or 2; Z, represents the nuclear charge number of the atom; a0 is a constant.

3. The method for obtaining molecular representation data based on molecular structure information according to claim 2, characterized in that: When the number of outer electron orbitals of the atom is two, the atomic convolution kernel based on the wave function is expressed as: The meanings of each expression are as follows: (x, y, z), represents the coordinates of each atom in voxel space; Ψ i , each electron orbital represented by a wave function, i takes the value of 1 or 2; l, represents the number of electron layers, l takes the value of 1 or 2; r represents the distance from the electron in the i-th electron orbit to the nucleus p0.

4. The method for obtaining molecular representation data based on molecular structure information according to claim 1, characterized in that: The atomic convolution kernel based on the wave function of each atom performs a convolution operation, including: The meaning of each expression is as follows: P, represents the atoms in the molecule; δ x , δ y , δ z : The offset of atom p in the three coordinate axes respectively.

5. A method for obtaining molecular properties, characterized in that: include: Generate voxel-based visualized molecular representation data for the molecule whose properties are to be predicted by the method described in any one of claims 1 to 4; Inputting the voxel-based visualized molecular representation data into a molecular property prediction network to perform molecular property prediction, wherein the molecular property prediction network is obtained through pre-training; and The molecular property prediction network outputs the properties of the molecule whose properties are to be predicted.

6. The method for obtaining molecular properties according to claim 5, characterized in that: The composition of the molecular property prediction network includes: The molecular property prediction network is based on a neural network, including multiple convolution operation layers and multiple pooling operation layers; and The molecular property prediction network includes an average channel pooling layer, wherein the average channel pooling layer takes the average value of each voxel in all channels as the value of the voxel; The channels reflect the distribution of atoms in space, and when the channels are different, they reflect the distribution of different atoms in space.

7. The method for obtaining molecular properties according to claim 5, characterized in that: The training process of the molecular property prediction network includes: Initialize the parameters of the attribute prediction network; Divide the data set into training set, validation set and test set; The voxel-based visualized molecular representation data and molecular properties in the training set are used as input to the molecular property prediction network for training; Verify and test using the validation set and test set data respectively; and The training effect of the molecular property prediction network was evaluated by the mean absolute error, and training and verification were repeated until the expected training indicators were achieved.

8. The method for obtaining molecular properties according to claim 7, characterized in that: The voxel-based visualized molecular representation data is normalized to ensure that each atom in the voxel-based visualized molecular representation data has a uniform size specification when input into the molecular property prediction network.

9. The method for obtaining molecular properties according to claim 7, characterized in that: The molecular properties obtained by the molecular property prediction network include at least one property of rotation constant, dipole moment, and isotropic polarizability.

10. The method for obtaining molecular properties according to claim 7, characterized in that: The voxel-based visualized molecules are processed by a diffusion strategy, in which each atom in the molecule is diffused by using the number zero or other positive integers as diffusion coefficients.