A small molecule structure prediction method based on isovariant flow matching modeling

Abstract

The present invention relates to the technical field of chemical informatics, and in particular to a method for predicting small molecule structures based on equivariant flow matching modeling. The present invention comprises obtaining a data set and real 3D molecular graph information in the data set; constructing a flow matching model, performing flow matching to obtain the flow matching time, the probability path of the real molecular coordinates, and the real vector field; constructing and training a Modified Equiformer model; randomly initializing initial molecular coordinates of the same size that satisfy a Gaussian distribution based on the number of atoms in the real 3D molecular graph information, inputting the initial molecular coordinates, the flow matching time, and the trained Modified Equiformer model into a black box ODE solver to generate predicted 3D molecular graph information, thereby obtaining a predicted molecular structure graph. The present invention improves the accuracy of small molecule structural information and significantly increases the speed and accuracy of structure prediction.

Description

Technical Field

[0001] The present invention relates to the fields of chemical informatics and computer application technology, and in particular to a small molecule structure prediction method based on isovariant flow matching modeling. Background Art

[0002] Molecular structure prediction is an important technology in the field of chemical informatics. It aims to use computational methods and algorithms to determine the 3D spatial arrangement of molecules. The 3D structure of a molecule determines how it interacts with other molecules or protein surfaces, thereby affecting its biological activity and function. This is widely used in the field of computational drug and material design.

[0003] Traditional molecular structure prediction methods, such as those based on classical force fields or quantum mechanics calculations, or those based on molecular dynamics or Markov chain Monte Carlo, have very high computational costs and may require a long time to run to obtain reliable results, and cannot guarantee the symmetry of molecules.

[0004] Flow matching is an emerging generative modeling technique with the advantages of simplicity and efficient training. This technique is often combined with generative models based on normalizing flows (NFs) to learn the process of transferring samples from a noise distribution to a data distribution. Continuous Normalizing Flows (CNFs) are an extension of NFs that use ordinary differential equations (ODEs) to create a continuous transfer process for modeling probability distributions. Flow matching is a method for training CNFs that trains the model by learning a vector field associated with the probability transfer path and uses an ODE solver for faster sampling. Therefore, the introduction of flow matching provides new ideas and methods for small molecule structure prediction.

[0005] Existing methods typically apply flow matching methods to the generation of small molecule structures, including the generation of atomic coordinates, atom types, edge types, and formal charges. For example, the EquiFM model, proposed in 2023 by Tsinghua University and Stanford University, applies flow matching methods to the generation of molecular atomic types and atomic coordinates; the MolFlow model, proposed in 2024 by Chalmers University of Technology, also applies flow matching methods to the generation of molecular atomic coordinates, atom types, edge types, and formal charges. However, neither of these methods pre-fixes the atomic types and cannot solve the problem of atomic coordinate prediction after the molecular atomic types are fixed. Summary of the Invention

[0006] The purpose of the present invention is to solve the problems of the prior art and to propose a small molecule structure prediction method based on isovariant flow matching modeling.

[0007] In order to achieve the above objectives, the technical solutions of this application are as follows:

[0008] A small molecule structure prediction method based on isovariant flow matching modeling, comprising:

[0009] Acquire and preprocess the data set to obtain the real 3D molecular graph information in the preprocessed data set;

[0010] Construct a flow matching model, set the standard deviation to 0-0.2, and perform flow matching to obtain the flow matching time, the probability path corresponding to the real molecular coordinates in the real 3D molecular graph information, and the real vector field;

[0011] Construct a Modified Equiformer model and train it using real 3D molecular graph information, flow matching time, probability paths of real molecular coordinates, and real vector fields;

[0012] Based on the number of atoms in the small molecule structure to be predicted, the initial molecular coordinates of the same size are randomly initialized to satisfy the Gaussian distribution. , the initial molecular coordinates , the flow matching time and the trained Modified Equiformer model are input into a black box ODE solver to generate predicted 3D molecular graph information, wherein the predicted 3D molecular graph information includes predicted molecular coordinates and predicted atomic coordinates, and finally a predicted molecular structure graph is obtained.

[0013] Furthermore, the preprocessing includes: removing information related to hydrogen atoms in each molecule in the data set, and aggregating different conformations of the same molecule together according to the smiles formula of each molecule.

[0014] Furthermore, the real 3D molecular graph information includes: real molecular coordinates, edge index, number of atoms, real atomic coordinates, atom type, edge type and molecular smiles type.

[0015] Furthermore, the flow matching is performed to obtain the flow matching time, the probability path corresponding to the real molecular coordinates in the real 3D molecular graph information, and the real vector field, including:

[0016] Map the atomic type features and edge type features of each molecule in the real 3D molecular graph information to the specified range;

[0017] The mapped real 3D molecular graph information is input into the flow matching model. Based on the total number of conformations of each molecule, a time feature that satisfies the Gaussian distribution and has the same length as the total number of conformations is randomly initialized as the flow matching time.

[0018] Based on the true molecular coordinates of different conformations of each molecule, the initial molecular coordinates of the same size and type that satisfy the Gaussian distribution are randomly initialized, and the true molecular coordinates and initial molecular coordinates of each conformation are subjected to zero center of mass operation respectively;

[0019] The optimal transfer operation is performed between the initial molecular coordinates x0 and the true molecular coordinates x1 after the zero center of mass operation of different conformations. The initial molecular coordinates x0 and the true molecular coordinates x1 are rotated and aligned using the Kabach algorithm. The transfer cost between the initial molecular coordinates x0 and the true molecular coordinates x1 of each conformation is calculated. The calculation formula is as follows:

[0020]

[0021] in, Represents the transmission cost between the initial molecular coordinates and the real molecular coordinates of the same conformation, N is the number of atoms in the current conformation, and They represent the initial atomic coordinates and true atomic coordinates of the i-th atom of the same conformation after the zero center of mass operation, respectively. The initial atomic coordinates are coordinates of the same size and type that satisfy the Gaussian distribution and are randomly initialized based on the true atomic coordinates;

[0022] The transmission cost between the initial molecular coordinates and the actual molecular coordinates of all conformations constitutes a transmission cost function matrix ,in, represents the initial molecular coordinates of all conformations, Represent the true molecular coordinates of all conformations, select the optimal transmission mapping with the lowest total transmission cost based on the transmission cost function matrix, and calculate the probability path of the initial molecular coordinates and true molecular coordinates corresponding to the optimal transmission mapping and the real vector field , the formula is as follows:

[0023]

[0024]

[0025] in, Indicates the time when the flow matches. represents the standard deviation, Represents a normal distribution.

[0026] Furthermore, the Modified Equiformer model includes an embedding layer, an equivariant graph attention module, a first feedforward module and a second feedforward module. The embedding layer includes an atomic type embedding block, a relative coordinate embedding block and a temporal feature embedding block. The equivariant graph attention module and the first feedforward module alternately form a 6-layer module.

[0027] Furthermore, the Modified Equiformer model is trained using real 3D molecular graph information, flow matching time, probability path of real atomic coordinates, and real vector field, including:

[0028] Calculate the molecular relative coordinate matrix based on the probabilistic path of the real molecular coordinates;

[0029] Input the atom type into the atom type embedding block, input the edge type and the molecular relative coordinate matrix into the relative coordinate embedding block, and input the flow matching time into the time feature embedding block to obtain the feature information x;

[0030] The feature information x is input into a 6-layer module consisting of an equivariant graph attention module and a first feed-forward module;

[0031] The obtained results are input into the second feedforward module to obtain the predicted vector field of the real molecular coordinates corresponding to the time of flow matching ;

[0032] Based on real vector fields and the prediction vector field The loss is calculated for training the Modified Equiformer model. The loss is calculated as follows:

[0033]

[0034] in, represents the loss function, It represents the expectation under the premise that T satisfies the standard normal distribution.

[0035] Furthermore, the calculation of the molecular relative coordinate matrix based on the probability path of the real molecular coordinates includes: subtracting the probability path of the real atomic coordinates based on the corresponding index of the edge to obtain the molecular relative coordinate matrix.

[0036] Furthermore, the atom type is input into the atom type embedding block, the edge type and the molecular relative coordinate matrix are input into the relative coordinate embedding block, and the flow matching time is input into the time feature embedding block to obtain the feature information x, including:

[0037] Input the atom type into the atom type embedding block for one-hot encoding to obtain the atom type encoding feature;

[0038] A constant 1 vector and the edge type are respectively encoded by the relative coordinate embedding block, the encoded edge type is tensor-producted with the molecule relative coordinate matrix, and encoded by the relative coordinate embedding block. The features obtained by the relative coordinate embedding block are encoded using the sum aggregation method to obtain aggregated features. The aggregated features are divided by the square root of the average number of edge indices in the training set to obtain relative coordinate encoded features.

[0039] The time of stream matching is input into the time feature embedding block to encode the time coding feature;

[0040] The feature information x is obtained by adding the atom type encoding feature, relative coordinate encoding feature and time encoding feature element by element.

[0041] Compared with the existing technology, the present invention has the following significant advantages: it utilizes the fast and stable advantages of the flow matching method and the powerful information extraction ability of the Modified Equiformer model to learn the vector field of molecular coordinates, thereby improving the accuracy of small molecule structural information; finally, it utilizes the efficient generation capability of the ODE solver to greatly improve the speed and accuracy of structure prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is an overall flow chart of a small molecule structure prediction method based on isovariant flow matching modeling of the present invention;

[0043] Figure 2 It is a specific flow chart of the Embedding Block of the Modified Equiformer model of the present invention;

[0044] Figure 3 This is an example diagram of the predicted structure of a molecule obtained by the present invention on the GEOM-QM9 data set;

[0045] Figure 4 This is a real structural example diagram of a molecule obtained by the present invention on the GEOM-QM9 data set. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] The present invention provides a small molecule structure prediction method based on isovariant flow matching modeling, comprising the following steps:

[0048] 1) Dataset Preprocessing: First, we downloaded the original GEOM-QM9 dataset from the official GEOM database. We then removed hydrogen atom information from each molecule and performed conformational aggregation. Different conformations of the same molecule were grouped together using the smiles formula (Simplified Molecular Input Line Entry System) for each molecule. Following the dataset splitting method proposed in the 2022 GeoDiff model, the training set consisted of 40,000 molecules, each with 5 conformations, for a total of 200,000 conformations. The validation set consisted of 5,000 molecules, each with 5 conformations, for a total of 25,000 conformations. The test set consisted of 200 molecules, for a total of 22,408 conformations.

[0049] 2) Obtain the real 3D molecular graph information in the dataset, including the relevant feature representations of different conformations of each molecule: real molecular coordinates, edge index (edge_index), number of atoms (num_nodes), real atomic coordinates (structure_positions), atom type (atom_type), edge type (bond_type) and molecular smiles (smiles).

[0050] 3) Map the atom type features and edge type features of each molecule from step 2) to a specified range to facilitate subsequent one-hot encoding. For the GEOM-QM9 dataset, after removing hydrogen atoms, there are four atom types and four edge types, all mapped to [0, 1, 2, 3].

[0051] 4) Set the parameters of the flow matching model: standard deviation = 0 ( is 0~0.2, with 0 being optimal).

[0052] 5) Perform flow matching, such as Figure 1 As shown, the process is as follows:

[0053] 5.1) Input the atom type features and edge type features processed in step 3) into the flow matching model. Based on the total number of conformations of each molecule, randomly initialize a time feature that satisfies a Gaussian distribution with the same length as the total number of conformations to ensure that atoms in a conformation correspond to the same time representation, which serves as the flow matching time T.

[0054] 5.2) Treat each molecule separately: Randomly initialize the initial molecular coordinates of the same size and type that satisfy the Gaussian distribution based on the true molecular coordinates of different conformations of the molecule.

[0055] 5.3) Perform a zero-center-of-mass operation on the initial molecular coordinates and true molecular coordinates of each conformation, bringing the coordinate mean to zero. This ensures translational invariance during the flow matching process. The zero-center-of-mass operation subtracts the mean from all feature values ​​to achieve a mean of 0. Hereinafter, the initial molecular coordinates and true molecular coordinates mentioned are those after the zero-center-of-mass operation.

[0056] 5.4) After the zero-center-of-mass operation, perform an optimal transport (OT) operation between the initial molecular coordinates x0 and the true molecular coordinates x1 of different conformations. Use the Kabach algorithm to rotate and align x0 and x1. Calculate the transport cost between the initial molecular coordinates x0 and the true molecular coordinates x1 for each conformation using the following formula:

[0057]

[0058] in, Represents the transmission cost between the initial molecular coordinates and the real molecular coordinates of the same conformation, N is the number of atoms in the current conformation, and They represent the initial atomic coordinates and true atomic coordinates of the i-th atom of the same conformation after the zero center of mass operation, respectively. The initial atomic coordinates are coordinates of the same size and type that satisfy the Gaussian distribution and are randomly initialized based on the true atomic coordinates.

[0059] The transmission cost between the initial molecular coordinates and the actual molecular coordinates of all conformations constitutes a transmission cost function matrix , represents the initial molecular coordinates of all conformations, Representing the true molecular coordinates of all conformations, based on Select the OT mapping that requires the lowest total transmission cost For example, if the molecule has 5 conformations, then the initial molecular coordinates and the true molecular coordinates are a total of 5 to 5 mappings. For a 5×5 matrix containing the transmission cost of each pair of mappings, select a set of mappings: such as left 1 to right 2, left 2 to right 3, left 3 to right 4, left 4 to right 5, left 5 to right 1 (a total of 5 factorial mappings), so that the sum of the costs of the five mappings is minimized.

[0060] 5.5) Based on the results obtained in step 5.4) The time T of the flow matching between the corresponding initial molecular coordinates and the true molecular coordinates and the corresponding conformation in step 5.1) is calculated. The probability path x corresponding to the initial molecular coordinates and the true molecular coordinates T and the real vector field u T , the formula is as follows:

[0061]

[0062]

[0063] in, Represents a normal distribution.

[0064] 6) Conduct training on the modified Equiformer model, such as Figure 1 As shown, the process is as follows:

[0065] 6.1) The probability path x of the true molecular coordinates obtained in step 5) is T Based on the corresponding subtraction of the edge index (edge_index), the molecular relative coordinate matrix rel_pos is obtained. The edge index (edge_index) represents the index of the existing edge, which is equivalent to a series of tuples, such as (1,2), (3,4), 1 and 2 represent the two nodes (atomic indices) connected by an edge, and (1,2) represents the edge connecting the two nodes. The subtraction is the x obtained in step 5) T , subtract according to the corresponding index of this tuple, such as x T1 -x T2 , representing the probabilistic path subtraction of the true molecular coordinates of the two nodes. This is then combined with the flow matching time T obtained in step 5) and the atom type feature atom_type and edge type feature bond_type from the mapped true 3D molecular graph obtained in step 3) through an embedding layer consisting of three embedding blocks to encode 480-dimensional feature information.

[0066] The three embedding blocks are: Atom type embedding block: directly passes the atom type through a linear layer to obtain the encoded features. Relative coordinate embedding block: After performing a tensor product operation (DTP) on the molecular relative coordinates and a constant 1 encoded through a linear layer, it passes through a linear layer and a scaling layer to obtain the features encoded by the molecular relative coordinates. After performing a tensor product operation on the molecular relative coordinates and the edge type feature encoded through a linear layer, it passes through a linear layer and a scaling layer to obtain a mixed feature of the edge type feature and the molecular relative coordinate encoding. Time feature embedding block: directly passes time through a linear layer to obtain the encoded features.

[0067] like Figure 2As shown, for the atomic type feature atom_type, the atomic type embedding block is used to convert the one-hot encoding of the atomic type to obtain the atomic type encoding feature; for the edge type feature bond_type and the molecular relative coordinate matrix rel_pos, first a constant 1 vector and the edge type feature bond_type are respectively encoded by the relative coordinate embedding block, and then a tensor product operation is performed on it and the molecular relative coordinate matrix rel_pos, and then the feature is encoded by the relative coordinate embedding block, and then the obtained feature is encoded using the sum aggregation method to obtain the aggregated feature, and finally the aggregated feature is scaled by dividing by the square root of the average number of edges in the training set (the number of edges is the number of edge indices) so that the standard deviation of the aggregated edge features is close to 1, and finally the relative coordinate encoding feature is obtained; for the time T of the flow matching, the time feature embedding block is used for encoding to obtain the time encoding feature; finally, the features obtained by each embedding block (atom type encoding feature, relative coordinate encoding feature and time encoding feature) are added element by element to obtain the 480-dimensional feature information x.

[0068] 6.2) The feature information x is input into the 6-layer module of the Modified Equiformer model, which is composed of an equivariant graph attention module and a first feedforward module alternating between them.

[0069] 6.3) Input the result obtained in step 6.2) into the second feedforward module to obtain the predicted vector field v of the true molecular coordinates corresponding to the final flow matching time T T .

[0070] 6.4) The vector field u of the real molecular coordinates obtained based on the flow matching model in step 5) T The predicted vector field v of the molecular coordinates predicted by the Modified Equiformer model in step 6.3 T Calculate the loss to train the ModifiedEquiformer model. The calculation formula is as follows:

[0071]

[0072] in, represents the loss function, It represents the expectation under the premise that T satisfies the standard normal distribution.

[0073] 7) Small molecule structure prediction: Randomly initialize the initial molecular coordinates of the same size that satisfy the Gaussian distribution based on the number of atoms in the small molecule structure to be predicted , together with the flow matching time T obtained in step 5) and the trained Modified Equiformer model in step 6), are input into a black box ODE solver to generate the predicted 3D molecular graph information. The predicted 3D molecular graph information includes the predicted molecular coordinates and predicted atomic coordinates. The hydrogen atoms are added back to meet the chemical legality, and the predicted molecular structure diagram can be obtained, as shown in Figure 3 、 Figure 4 Shown are examples of the predicted structure and the actual structure of a molecule, where green represents carbon atoms, blue represents nitrogen atoms, red represents oxygen atoms, and gray represents hydrogen atoms.

[0074] 8) Use the validation set and test set to calculate the evaluation index to evaluate the predictive ability of the trained Modified Equiformer model: the true coordinates x1 of the molecular structure input in step 2) are compared with the predicted coordinates of the molecular structure obtained in step 7) The rotation alignment was performed using the Kabsch algorithm, and then the evaluation indicators were calculated: Recall index of COV score (COV-R), Recall index of MAT score (MAT-R), Precision index of COV score (COV-P), Precision index of MAT score (MAT-P), where the COV score measures the percentage of one group of structures covered by another group, and coverage means that the root mean square deviation (RMSD) between the two conformations is within a certain threshold. The MAT score measures the average RMSD of the conformations in one set with their nearest neighbors in another set. The Precision metric reflects more on the accuracy, while the Recall metric focuses more on the diversity. Figure 3 As shown in the figure, in the structure prediction of the small molecule multi-conformation database GEOM-QM9, compared with the GeoDiff model in the current SOTA model, a small molecule structure prediction method based on isovariant flow matching modeling (Flow Matching for Molecular Conformation Prediction, MolFM) achieved better COV-P and MAT-P values ​​and higher prediction accuracy, but the COV-R and MAT-R values ​​were poor, and the diversity prediction still needs further improvement. The comparative data of MolFM and GeoDiff models are shown in Table 1:

[0075] Table 1 Comparison of evaluation indicators between MolFM and GeoDiff models

[0076]

[0077] Among them, GeoDiff-A and GeoDiff-C are two versions of the GeoDiff model, namely alignment and chain-rule.

[0078] The calculation formulas for the evaluation indicators COV-R, MAT-R, COV-P, and MAT-P are as follows:

[0079]

[0080]

[0081]

[0082]

[0083] in, represents the real molecular coordinates, represents the predicted molecular coordinates, RMSD for and Normalized Frobenius norm after alignment by Kabsch algorithm, represents the predicted molecular coordinates, represents the true molecular coordinates, RMSD for and Normalized Frobenius norm after alignment by Kabsch algorithm, and Represent the predicted molecular coordinate set and the real molecular coordinate set, respectively, and the threshold value is used on the small molecule multi-conformation dataset GEOM-QM9. (Unit: Angstrom One tenth of a nanometer, often used to describe the size of an atomic scale), each molecule Set to Twice the size.

[0084] The present invention calculates the loss on the validation set and compares it with the loss on the training set to prevent overfitting of the model training. By calculating the MAP-P value and finding the epoch with the smallest MAP-P value, the best trained model is obtained. Then, the evaluation index is calculated and the small molecule structure prediction is performed on the test set based on the trained model.

[0085] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A small molecule structure prediction method based on isotropic flow matching modeling, characterized in that: The small molecule structure prediction method based on isovariant flow matching modeling includes: Acquire and preprocess the data set to obtain the real 3D molecular graph information in the preprocessed data set; Construct a flow matching model, set the standard deviation to 0-0.2, and perform flow matching to obtain the flow matching time, the probability path corresponding to the real molecular coordinates in the real 3D molecular graph information, and the real vector field; Construct a Modified Equiformer model and train it using real 3D molecular graph information, flow matching time, probability paths of real molecular coordinates, and real vector fields; Based on the number of atoms in the small molecule structure to be predicted, the initial molecular coordinates of the same size are randomly initialized to satisfy the Gaussian distribution. , the initial molecular coordinates , the flow matching time and the trained Modified Equiformer model are input into a black box ODE solver to generate predicted 3D molecular graph information, wherein the predicted 3D molecular graph information includes predicted molecular coordinates and predicted atomic coordinates, and finally a predicted molecular structure graph is obtained; in, The flow matching is performed to obtain the flow matching time, the probability path corresponding to the real molecular coordinates in the real 3D molecular graph information, and the real vector field, including: Map the atomic type features and edge type features of each molecule in the real 3D molecular graph information to the specified range; The mapped real 3D molecular graph information is input into the flow matching model. Based on the total number of conformations of each molecule, a time feature that satisfies the Gaussian distribution and has the same length as the total number of conformations is randomly initialized as the flow matching time. Based on the true molecular coordinates of different conformations of each molecule, the initial molecular coordinates of the same size and type that satisfy the Gaussian distribution are randomly initialized, and the true molecular coordinates and initial molecular coordinates of each conformation are subjected to zero center of mass operation respectively; The optimal transfer operation is performed between the initial molecular coordinates x0 and the true molecular coordinates x1 after the zero center of mass operation of different conformations. The initial molecular coordinates x0 and the true molecular coordinates x1 are rotated and aligned using the Kabach algorithm. The transfer cost between the initial molecular coordinates x0 and the true molecular coordinates x1 of each conformation is calculated. The calculation formula is as follows: ； in, Represents the transmission cost between the initial molecular coordinates and the real molecular coordinates of the same conformation, N is the number of atoms in the current conformation, and They represent the initial atomic coordinates and true atomic coordinates of the i-th atom of the same conformation after the zero center of mass operation, respectively. The initial atomic coordinates are coordinates of the same size and type that satisfy the Gaussian distribution and are randomly initialized based on the true atomic coordinates; The transmission cost between the initial molecular coordinates and the actual molecular coordinates of all conformations constitutes a transmission cost function matrix ,in, represents the initial molecular coordinates of all conformations, Represent the true molecular coordinates of all conformations, select the optimal transmission mapping with the lowest total transmission cost based on the transmission cost function matrix, and calculate the probability path of the initial molecular coordinates and true molecular coordinates corresponding to the optimal transmission mapping and the real vector field , the formula is as follows: ； ； in, Indicates the time when the flow matches. represents the standard deviation, represents a normal distribution; The Modified Equiformer model includes an embedding layer, an equivariant graph attention module, a first feedforward module, and a second feedforward module. The embedding layer includes an atom type embedding block, a relative coordinate embedding block, and a temporal feature embedding block. The equivariant graph attention module and the first feedforward module alternately form a 6-layer module. The Modified Equiformer model is trained using real 3D molecular graph information, flow matching time, probability path of real molecular coordinates, and real vector field, including: Calculate the molecular relative coordinate matrix based on the probabilistic path of the real molecular coordinates; Input the atom type into the atom type embedding block, input the edge type and the molecular relative coordinate matrix into the relative coordinate embedding block, and input the flow matching time into the time feature embedding block to obtain the feature information x; The feature information x is input into a 6-layer module consisting of an equivariant graph attention module and a first feed-forward module; The obtained results are input into the second feedforward module to obtain the predicted vector field of the real molecular coordinates corresponding to the time of flow matching ; Based on real vector fields and the prediction vector field The loss is calculated for training the Modified Equiformer model. The loss is calculated as follows: ； in, represents the loss function, Indicates T Satisfies the expectation under the premise of standard normal distribution.

2. The small molecule structure prediction method based on isovariant flow matching modeling according to claim 1, characterized in that: The preprocessing includes: removing the information related to hydrogen atoms in each molecule in the data set, and aggregating different conformations of the same molecule according to the smiles formula of each molecule.

3. The small molecule structure prediction method based on isovariant flow matching modeling according to claim 1, characterized in that: The real 3D molecular graph information includes: real molecular coordinates, edge index, number of atoms, real atomic coordinates, atom type, edge type and molecular smiles type.

4. The small molecule structure prediction method based on isovariant flow matching modeling according to claim 1, characterized in that: The molecular relative coordinate matrix is ​​calculated based on the probability path of the real molecular coordinates, including: subtracting the probability path of the real atomic coordinates based on the corresponding index of the edge to obtain the molecular relative coordinate matrix.

5. The small molecule structure prediction method based on isovariant flow matching modeling according to claim 1, characterized in that: The atom type is input into the atom type embedding block, the edge type and the molecular relative coordinate matrix are input into the relative coordinate embedding block, and the flow matching time is input into the time feature embedding block to obtain the feature information x, including: Input the atom type into the atom type embedding block for one-hot encoding to obtain the atom type encoding feature; A constant 1 vector and the edge type are respectively encoded by the relative coordinate embedding block, the encoded edge type is tensor-producted with the molecule relative coordinate matrix, and encoded by the relative coordinate embedding block. The features obtained by the relative coordinate embedding block are encoded using the sum aggregation method to obtain aggregated features. The aggregated features are divided by the square root of the average number of edge indices in the training set to obtain relative coordinate encoded features. The time of stream matching is input into the time feature embedding block to encode the time coding feature; The feature information x is obtained by adding the atom type encoding feature, relative coordinate encoding feature and time encoding feature element by element.

Images