A molecular spectroscopy artificial intelligence prediction method, system, medium and device

Abstract

The present disclosure provides a molecular spectrum artificial intelligence prediction method, system, medium and device, which relates to the technical field of artificial intelligence prediction, extracts molecular structure data, optimizes the geometric structure of the molecule, establishes a molecular property and molecular spectrum database, obtains atomic coordinates and atomic types in the database, extracts various scalar, vector and tensor properties and spectral information of the molecule, constructs initial atomic features and atomic environment features, interacts the atomic environment features with the equivariant vector using a message passing function and an update function, uses an E(3) point group, guarantees the equivariance of the molecular properties through a tensor product and an irreducible representation, realizes the analysis of the vector and the tensor through gradient tracking, obtains the derivative information of the vector and the tensor, and realizes the prediction of various molecular spectra.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence prediction, in particular to a molecular spectrum artificial intelligence prediction method, system, medium and device. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute the prior art.

[0003] Molecular spectrum can provide "fingerprint information" of molecules or materials, realize trace measurement and configuration identification of molecules, and is widely used in physics, biology, medicine, environment, materials and chemical industry and other fields. There are two main ways to identify molecular structure or explore fine molecular configuration by using molecular spectrum: spectral database retrieval method and theoretical simulation trial and error method. The spectral database retrieval method determines the molecular structure corresponding to the target spectrum by calculating the similarity between the target spectrum and all spectra in the database. However, the spectral database retrieval method has the disadvantages of high construction cost and low coverage, which leads to the inability to identify molecules not included in the database. On the other hand, the trial and error method based on theoretical simulation can identify unknown molecular structure or explore fine molecular configuration through a self-consistent process of "building model → theoretical simulation → comparing spectrum → adjusting model → …". The trial and error method relies heavily on the experience and intuition of researchers, and involves expensive electronic structure calculation and complex spectrum simulation, resulting in low efficiency.

[0004] With the advent of the big data era, the significant improvement of computing power and the gradual maturity of algorithms make data-driven innovation a research method, which is the fourth research paradigm after experiment, theory and simulation. Since 2011, deep learning, a kind of algorithm based on data representation learning in machine learning, has made amazing progress in important information technology fields such as computer vision, speech recognition, natural language processing and bioinformatics. However, the inventors found that the existing artificial intelligence method still has the problems of high requirement for instruments and technical personnel, relatively long cycle and complicated theoretical simulation process. SUMMARY

[0005] In order to solve the above problems, the present disclosure provides a molecular spectrum artificial intelligence prediction method, system, medium and device, which uses E(3) equivariant message passing neural network model (DetaNet) to realize fast, accurate and complete molecular property prediction and molecular spectrum simulation.

[0006] According to some embodiments, the present disclosure adopts the following technical solutions:

[0007] A molecular spectrum artificial intelligence prediction method comprises:

[0008] Extract molecular structure data, optimize the geometric structure of the molecule, and establish a database of molecular properties and molecular spectra;

[0009] Obtain atomic coordinates and atomic types in the database, and extract scalar, vector, and tensor properties and spectral information of the molecule;

[0010] Construct initial atomic features and atomic environment features, interact atomic environment features with equivariant vectors using a message passing function and an update function, use an E(3) point group, ensure the equivariance of molecular properties through tensor product and irreducible representation, track gradients to realize the analysis of vectors and tensors, obtain derivative information of vectors and tensors, and realize the prediction of various molecular spectra.

[0011] According to some embodiments, the present disclosure adopts the technical solutions as follows:

[0012] A molecular spectrum artificial intelligence prediction system comprises:

[0013] A molecule extraction module extracts molecular structure data, optimizes the geometric structure of the molecule, and establishes a database of molecular properties and molecular spectra;

[0014] A feature extraction module obtains atomic coordinates and atomic types in the database, and extracts scalar, vector, and tensor properties and spectral information of the molecule;

[0015] A prediction module constructs initial atomic features and atomic environment features, interacts atomic environment features with equivariant vectors using a message passing function and an update function, uses an E(3) point group, ensures the equivariance of molecular properties through tensor product and irreducible representation, tracks gradients to realize the analysis of vectors and tensors,

[0016] obtains derivative information of vectors and tensors, and realizes the prediction of various molecular spectra.

[0017] According to some embodiments, the present disclosure adopts the technical solutions as follows:

[0018] A computer-readable storage medium stores a plurality of instructions, which are suitable for being loaded and executed by a processor of a terminal device to implement a molecular spectrum artificial intelligence prediction method.

[0019] According to some embodiments, the present disclosure adopts the technical solutions as follows:

[0020] A terminal device comprises a processor and a computer-readable storage medium, wherein the processor is configured to implement instructions, and the computer-readable storage medium is configured to store a plurality of instructions, which are suitable for being loaded and executed by the processor to implement a molecular spectrum artificial intelligence prediction method.

[0021] Compared with the prior art, the present disclosure has the beneficial effects that:

[0022] Compared with existing experimental measurement and theoretical calculation techniques, the developed DetaNet based on E(3) equivariant message passing neural network of the method of the present disclosure helps researchers to accurately and efficiently predict the infrared, Raman, ultraviolet-visible absorption, nuclear magnetic resonance spectrum of molecules, reduces the time cost, labor cost and test cost, and is expected to be applied in the fields of physics, chemistry, materials, medical treatment, health, biology, pharmaceuticals, aerospace, military industry and the like. BRIEF DESCRIPTION OF DRAWINGS

[0023] The accompanying drawings, which form a part of the present disclosure, are used to provide further understanding of the present disclosure, and the schematic embodiments of the present disclosure and the description thereof are used to explain the present disclosure, and do not constitute improper limitations on the present disclosure.

[0024] Figure 1 The system architecture diagram of the present disclosure.

[0025] Figure 2 The method flowchart and functional module diagram of the present disclosure. DETAILED DESCRIPTION

[0026] The present disclosure will be further described below in combination with the drawings and embodiments.

[0027] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs.

[0028] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form, and furthermore, it should be understood that when the terms "comprise" and / or "include" are used in the specification, they indicate the presence of a feature, step, operation, device, component and / or combination thereof.

[0029] Embodiment 1

[0030] In an embodiment of the present disclosure, a molecular spectrum artificial intelligence prediction method is provided, comprising:

[0031] Step 1: Extract the molecular structure data, optimize the geometric structure of the molecule, and establish the molecular property and molecular spectrum database;

[0032] Step 2: Obtain the atomic coordinates and atomic types in the database, and extract various scalar, vector and tensor properties and spectral information of the molecule;

[0033] Step three: constructing initial atomic features and atomic environment features, interacting atomic environment features with equivariant vectors using a message passing function and an update function, using E(3) point groups to ensure equivariance of molecular properties through tensor products, irreducible representations, gradient tracking to achieve analyticity of vectors and tensors, and obtaining derivative information of vectors and tensors to achieve prediction of various molecular spectra.

[0034] Building a database of molecular properties and molecular spectra. To address the lack of high-order tensor properties and molecular spectra in the current QM9 database, the present disclosure uses high-throughput density functional theory calculations to build a database containing molecular properties and molecular spectra. Specifically, it includes:

[0035] First, extract the structure files of 130,000 molecules in the QM9 database, use the Gaussian16 software package, and use the B3LYP / def-TZVP functional basis set to optimize the molecular geometry.

[0036] Second, obtain infrared spectra, Raman spectra, and UV-Vis spectra through vibration analysis and time-dependent perturbation calculations; calculate the nonlinear optical properties of molecules using the response field theory of the Dalton program package to obtain the multi-photon absorption spectrum of the molecule; calculate the ground state energy and complete core-hole energy of the molecule using the Demon-Stobe program to obtain the X-ray spectrum of the molecule.

[0037] Based on the dataset, extract atomic coordinates and atomic types as input for DetaNet, extract various scalar, vector, and tensor properties and spectral information of the molecule as target reference data for learning, and use 90% of the data as the training set, 5% as the test set, and 5% as the evaluation set. Use Adam and Amsgrad optimizers for training. The dataset is divided into 64 batches for learning, and the initial learning rate is set to 10-3. Test every 50 training, if the loss does not decrease after 10 consecutive tests, the learning rate will be reduced by 20%. When the loss is reduced to 10-5 or reaches 1000000 times, the training is completed. All training uses mean square error as the loss function for training.

[0038] As an embodiment, extracting various scalar, vector, and tensor properties of the molecule includes: molecular energy, charge, and orbital energy level, etc. scalar properties; dipole moment, dipole moment derivative, transition dipole moment, and oscillator strength, etc. vector properties; polarizability, polarizability derivative, multipole moment, and Hessian matrix, etc. tensor properties.

[0039] The initial atomic features are constructed by: compiling the atom type using Nucleus One-hot vectors, compiling the atom orbital charge using Boolean sequence vectors, passing through a learnable linear layer, and adding the Nucleus One-hot vectors and Boolean vectors to construct the atomic features.

[0040] Constructing atomic environment features involves: for each atom, fitting the features of neighboring atoms within the cutoff radius and the distance features to each atom using a linear layer and convolving them to form atomic environment features.

[0041] Specifically, for each atom, the features of neighboring atoms and distance features within a cutoff radius of 5.0A are fitted to the atom using a linear layer and convolved to form the atomic environment features.

[0042] Message passing process. The interaction between atomic environmental features and equivariant vectors is achieved using message passing and update functions. The equivariance of molecular properties is guaranteed using the E(3) point group through tensor product and irreducible representation.

[0043] Gradient tracing enables the analysis of vectors and tensors, allowing for the rapid acquisition of their derivative information.

[0044] As an example, ML algorithms for various molecular spectra are established. UV-Vis spectra are obtained by training the transition dipole moments; infrared spectra are obtained by training the derivatives of the dipole moments with respect to normal coordinates; Raman spectra are obtained by training the derivatives of polarizability with respect to normal coordinates; NMR spectra are obtained by training the chemical shift and shielding tensors; multiphoton absorption spectra are obtained by training the hyperpolarizability model; and X-ray spectra are obtained by training the hole energies of the molecular core states.

[0045] As one example, such as Figure 2 As shown, DetaNet uses atomic numbers. and Cartesian coordinates As input, the element type and position of the atom are represented. Similar to previous MPNNs, DetaNet uses interactive layers defined according to ResNet-style message and update functions to build structure-property relationships. Unlike most previous MPNNs, the atomic features of DetaNet's hidden state nodes are represented as irreducible representation (IRREps) tensors. Where the positive integers l∈0,1,2,3… and m∈[-l..l] represent the rotation degrees in the irreps and the ordinal numbers under each rotation degree. p∈[-1,1] represents odd-odd-even and even-odd-even. The special cases of l=0 and p=1 (scalars) are specially separated and represented by… Representation. Irreps features of the initial input. is set to 0, the scalar feature is calculated by the Embedding layer:

[0046]

[0047] where O(Z i ) is the common One-hot feature, representing the nuclear type. Q(Z i ) represents the intrinsic electronic structure of the isolated atom. The learnable linear layers L z and L Q map the nuclear and electronic features to the -dimensional atomic features, while L emb integrates the two parts together.

[0048] The message module generates atomic messages for atom i using the messages from the upper layer and the position vectors of all neighboring atoms within a cutoff radius . Specifically, the absolute distance r = ||r ij || is embedded as two different radial weights w k and w v for the next self-attention module. Then, the edge features e ij are constructed by the self-attention mechanism:

[0049]

[0050]

[0051] where M q , M k and M v represent the Query, Key and Value features of the self-attention mechanism, I Mq , L Mk , L Mv and L out are learnable linear layers, and F Mq is the dimension of Mq. The self-attention mechanism module enhances the features of the central atom and the radial interactions of the neighboring atoms in the form of a continuous filter. Subsequently, the output of the message layer is obtained by the invariant and equivariant convolutions:

[0052]

[0053] where and represent the residuals of the upper layer and the features, respectively, and L This is a spherical harmonic function that maps vectors to Irreps tensors and performs tensor products with edge features to produce Irreps features. After passing inter-atomic messages, an equivariant update module is used to describe the interactions between features with different degrees and parity. Similarly, an attention mechanism is used to obtain the residuals of equivariant and invariant features.

[0054]

[0055]

[0056] Among them U q U k and UV (T) yes Plus Generated through a linear layer. (UV) (S) Through Plus It is generated through a linear layer. A is the attention feature generated by the scalar obtained by passing the tensor product of the linear combination of two Irreps features through the softmax function. The residual result of the update process is obtained by multiplying the attention feature A with the linear combination of the Irreps feature and the scalar feature.

[0057] After reaching the maximum interaction layer N, the molecular properties are output using a multilayer perceptron (MLP):

[0058]

[0059]

[0060] Where φ is the output function that produces specific molecular properties.

[0061] Example 2

[0062] One embodiment of this disclosure provides a molecular spectroscopy artificial intelligence prediction system, comprising:

[0063] The molecular extraction module extracts molecular structure data, optimizes molecular geometry, and establishes a database of molecular properties and molecular spectra.

[0064] The feature extraction module obtains atomic coordinates and atom types from the database, and extracts various scalar, vector, tensor properties and spectral information of molecules;

[0065] The prediction module is configured to construct initial atomic features and atomic environment features, interact the atomic environment features with the equivariant vector by using a message passing function and an update function, use an E(3) point group to ensure the equivariance of the molecular properties by using a tensor product and an irreducible representation, track the gradient to realize the analysis of the vector and the tensor, obtain derivative information of the vector and the tensor, and realize the prediction of various molecular spectra.

[0066] Embodiment 3

[0067] In an embodiment of the present disclosure, a computer readable storage medium is provided, wherein a plurality of instructions are stored, the instructions being adapted to be loaded and executed by a processor of a terminal device to implement the artificial intelligence prediction method for molecular spectra.

[0068] Embodiment 4

[0069] In an embodiment of the present disclosure, a terminal device is provided, comprising a processor and a computer readable storage medium, the processor being configured to implement instructions, and the computer readable storage medium being configured to store a plurality of instructions, the instructions being adapted to be loaded and executed by the processor to implement the artificial intelligence prediction method for molecular spectra.

[0070] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0071] These computer program instructions can also be loaded onto a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable data processing device to produce a computer implemented process, so that the instructions executed by the computer or other programmable data processing device provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 a step that carries out the functions specified in one or more flows and / or blocks.

[0072] The specific embodiments of the present disclosure are described above with reference to the accompanying drawings, but are not intended to limit the protection scope of the present disclosure, and those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present disclosure without creative labor are still within the protection scope of the present disclosure.

Claims

1. A molecular spectroscopy artificial intelligence prediction method, characterized by, The method comprises the following steps: extracting molecular structure data, optimizing the geometric structure of the molecule, and establishing a database of molecular properties and molecular spectra; obtaining atomic coordinates and atomic types in the database, and extracting scalar, vector and tensor properties and spectral information of the molecule; extracting scalar, vector and tensor properties of the molecule, including scalar properties of the molecule's energy, charge and orbital energy level, vector properties of dipole moment, dipole moment derivative, transition dipole moment and oscillator strength, and tensor properties of polarizability, polarizability derivative, multipole moment and Hessian matrix; constructing initial atomic features and atomic environment features, using message passing functions and update functions to interact atomic environment features with equivariant vectors, using E(3) point groups, ensuring the equivariance of molecular properties through tensor product and irreducible representation, and realizing the analysis of vectors and tensors through gradient tracking to obtain derivative information of vectors and tensors and realize the prediction of various molecular spectra; realizing the prediction of various molecular spectra, including training transition dipole moment to obtain UV-Vis spectrum, training dipole moment derivative with respect to normal coordinates to obtain infrared spectrum, training polarizability derivative with respect to normal coordinates to obtain Raman spectrum, training chemical shift and shielding tensor to obtain NMR spectrum, training hyperpolarizability model to obtain multi-photon absorption spectrum, and training complete core-hole energy of the molecule to obtain X-ray spectrum of the molecule.

2. The method of claim 1, wherein the method is a molecular spectroscopy artificial intelligence prediction method. The method of constructing initial atomic features comprises: using Nucleus One-hot vector to compile atomic types, using Boolean sequence vector to compile atomic orbital charge, adding Nucleus One-hot vector and Boolean vector through a learnable linear layer to construct atomic features.

3. The method of claim 1, wherein the method is a molecular spectroscopy artificial intelligence prediction method. The method of constructing atomic environment features comprises: for each atom, fitting the near-neighbor atomic features and distance features within the cutoff radius to each atom through a linear layer to form atomic environment features.

4. The molecular spectroscopic artificial intelligence prediction method of claim 1, wherein, The method of extracting spectral information of the molecule comprises: obtaining infrared spectrum, Raman spectrum and UV-Vis spectrum through vibration analysis and time-dependent perturbation calculation; calculating nonlinear optical properties of the molecule through response field theory of Dalton to obtain multi-photon absorption spectrum of the molecule; and calculating ground state energy and complete core-hole energy of the molecule through Demon-Stobe to obtain X-ray spectrum of the molecule.

5. The method of claim 1, wherein the method is a molecular spectroscopy artificial intelligence prediction method. The method of optimizing the geometric structure of the molecule is to use B3LYP / def-TZVP functional basis set to optimize the geometric structure of the molecule.

6. A molecular spectroscopy artificial intelligence prediction system, specifically implementing a molecular spectroscopy artificial intelligence prediction method according to any one of claims 1-5, characterized in that, The method comprises the following steps: a molecule extraction module for extracting molecular structure data, optimizing the geometric structure of the molecule, and establishing a database of molecular properties and molecular spectra; a feature extraction module for obtaining atomic coordinates and atomic types in the database, and extracting scalar, vector and tensor properties and spectral information of the molecule; a prediction module for constructing initial atomic features and atomic environment features, using message passing functions and update functions to interact atomic environment features with equivariant vectors, using E(3) point groups, ensuring the equivariance of molecular properties through tensor product and irreducible representation, and realizing the analysis of vectors and tensors through gradient tracking to obtain derivative information of vectors and tensors and realize the prediction of various molecular spectra.

7. A computer readable storage medium characterized in that, A computer readable storage medium, in which a plurality of instructions are stored, the instructions being suitable for being loaded and executed by a processor of a terminal device to implement the method of any one of claims 1-5.

8. A terminal device, comprising: A computer readable storage medium, in which a plurality of instructions are stored, the instructions being suitable for being loaded and executed by a processor of a terminal device to implement the method of any one of claims 1-5.

Images