Molecular graph feature extraction method and device based on multi-scale information and medium

Abstract

The invention discloses a molecular map feature extraction method and device based on multi-scale information and a medium, and relates to the technical field of molecular activity prediction. The molecular map feature extraction method comprises the following steps: extracting atomic-scale features, bond-scale features and conformation features of target molecules; and fusing the atomic-level features, the bond-level features and the conformational features to obtain molecular map features of the target molecules. According to the method, the atomic-level features, the bond-level features and the conformation features of the molecules are extracted and fused to serve as input features of activity prediction, understanding of a model on the relationship between the atomic features, the bond features and the spatial features of the molecules and the activity is promoted, and therefore the molecular activity is predicted more comprehensively.

Description

Technical Field

[0001] The present invention relates to the technical field of molecular activity prediction, and in particular to a method, device and medium for extracting molecular graph features based on multi-scale information. Background Art

[0002] Molecular activity prediction is the use of computational methods to predict the biological activity of molecules interacting with biological targets. Molecular activity prediction allows for the rapid evaluation of the activity of a large number of compounds before actual synthesis and testing, identifying molecules with the desired biological activity while minimizing adverse side effects.

[0003] Quantitative Structure-Activity Relationship (QSAR) models are a common method for predicting molecular activity. Using mathematical and statistical methods, QSAR establishes a quantitative relationship between the structure of a chemical substance and its biological activity. By analyzing the chemical structure characteristics (such as molecular descriptors) of a molecule, QSAR predicts its activity against its biological target. This helps rapidly screen and optimize potential drug candidates, reduces experimental time and costs, and accelerates the discovery and design of new drugs.

[0004] Currently, quantitative structure-activity relationship (QSAR) methods are mainly divided into three categories: molecular sequence-based, molecular descriptor-based, and molecular structure-based QSAR methods. Molecular sequence-based QSAR methods typically use linear text representations of molecules (such as SMILES strings) to build models, and they learn features directly from the text representation of molecules. Molecular descriptor-based QSAR methods directly use pre-defined molecular physicochemical properties (i.e., molecular descriptors) to predict the biological activity or other chemical properties of molecules. Molecular graph structure-based QSAR methods view molecules as graphs, with atoms as nodes and chemical bonds as edges. Graph neural networks (GNNs) are typically used to process the molecular graph structure and extract features from it.

[0005] However, existing research on molecular activity prediction based on quantitative structure-activity relationship still has the following defects: (1) Molecular bonds have various properties, among which bond length and saturation affect the stability and activity of the molecule, and bond angle determines the spatial configuration of the molecule, affecting the binding of the molecule to the biological target. However, previous work has paid more attention to atomic information and has not paid more attention to the impact of bond characteristics on activity; (2) The biological activity of a molecule depends not only on its chemical composition and structure, but also on its spatial arrangement and shape. However, previous work on extracting spatial features was extremely sensitive to the rotation and translation of the molecule and its symmetric substructure, and could not provide useful molecular conformational features for activity prediction. (3) The reactivity and activity of a molecule are determined by the properties of its atoms, bonds, and stereo conformation. Previous work treated each feature as an independent individual. There was a lack of interaction between features at different scales, and features could not share their information. However, the interaction between features can greatly promote the model's understanding of molecular behavior, enabling the model to more comprehensively evaluate the biological activity of molecules.

[0006] In summary, the existing molecular activity prediction based on quantitative structure-activity relationship has the problem that the detection characteristics are not fully expressed, resulting in the lack of accuracy of the prediction results. Summary of the Invention The present invention provides a molecular graph feature extraction method, device and medium based on multi-scale information to solve the problem that the existing molecular activity prediction based on quantitative structure-activity relationship has insufficient expression of prediction features, which affects the accuracy of the prediction results.

[0007] The present invention is achieved through the following technical solutions: A first aspect of the present invention provides a molecular graph feature extraction method based on multi-scale information, comprising: Extract atomic-level features, bond-level features, and conformational features of target molecules; The atomic-level features, bond-level features and conformational features are integrated to obtain the molecular graph features of the target molecule.

[0008] The present invention integrates the atomic, bond, and conformational features of a molecule to create a more comprehensive fusion signature for molecular activity prediction. This facilitates the model's understanding of the relationship between the molecule itself and its activity, thereby improving prediction accuracy. Bond-level features allow the model to focus on the impact of bond characteristics on molecular stability and activity, while conformational features allow the model to focus on the impact of atomic position characteristics on activity, leading to a more comprehensive prediction of molecular activity.

[0009] In one embodiment, the method for extracting atomic-level features of a target molecule comprises: Convert the SMILES string of the target molecule into a two-dimensional molecular graph; Node features of the two-dimensional molecular graph are extracted through a message passing network. In each step of the message passing phase, each atom updates its own feature vector by aggregating messages from neighboring atoms. After the final update step, the atomic-level features of each atom in the two-dimensional molecular graph are obtained.

[0010] In one embodiment, the method for extracting bond-order features of a target molecule comprises: Reconstructing the two-dimensional molecular graph by using adjacent nodes of the two-dimensional molecular graph and edges therebetween as new nodes to obtain a bond graph; Reconstruct the bond graph by using adjacent nodes of the bond graph as new nodes to obtain a bond angle graph; The node feature matrices of the bond graph and the bond angle graph are subjected to feature extraction by a transformer model to obtain bond-level features.

[0011] In one embodiment, the feature extraction of the node feature matrix of the bond graph and the bond angle graph by using a transformer model includes: The shortest path information of each pair of nodes in the bond graph or bond angle graph is introduced into the multi-head attention layer to improve the attention matrix of the transformer model. The improved attention matrix is ​​expressed as:

[0012] in, represents the attention matrix, 、 represents the vector of calculated weights, represents the dimension of each self-attention head, is the shortest path encoding matrix of the bond graph or bond angle graph, Represents the softmax operation. In one embodiment, a generative self-supervised learning method is combined to extract the final key-level features, including: The original node graph of the two-dimensional molecular graph is used to obtain the molecular fingerprint using the ECFP method; Initialize the molecular fingerprint as a virtual node and interact with other nodes as a guidance node for mask prediction; Randomly select a portion of the nodes in the original node graph for masking, input the masked node feature matrix into the transformer model, and obtain the output features embedded with attention information; The masked nodes in the output features are predicted by a multi-layer perceptron to obtain the final key-level features. In one embodiment, the method for extracting the conformational features of the target molecule includes: fusing the atomic-level features with atomic coordinate information to obtain the conformational features.

[0013] In one embodiment, the method of fusing the atomic-level features, bond-level features, and conformational features to obtain the molecular graph features of the target molecule includes: Selectively transmitting information to interact the atomic-level features with the bond-level features to obtain first interaction information, and extracting features from the first interaction information through a graph message passing network to obtain first extracted features; Performing selective information transfer interaction between the first extracted feature and the atomic-level feature to obtain second interaction information, and performing feature extraction on the second interaction information through a graph message passing network to obtain a second extracted feature; Performing selective information transfer interaction between the first extracted feature and the conformation feature to obtain third interaction information, and performing feature extraction on the third interaction information through a graph message passing network to obtain a third extracted feature; Performing selective information transfer interaction between the third extracted feature and the first extracted feature to obtain fourth interaction information, and performing feature extraction on the fourth interaction information through a graph message passing network to obtain a fourth extracted feature; The second extracted feature, the third extracted feature and the fourth extracted feature are spliced ​​together to obtain the molecular graph feature of the target molecule.

[0014] The second aspect of the present invention provides a method for predicting molecular activity, comprising: extracting molecular graph features of the target molecule by a molecular graph feature extraction method based on multi-scale information as described in any one of the first aspects of the present invention, inputting the molecular graph features into a pre-trained multi-layer perceptron for prediction, and obtaining a molecular activity prediction result.

[0015] The third aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor, when executing the computer program, implements the molecular graph feature extraction method based on multi-scale information as described in any one of the first aspect of the present invention or the molecular activity prediction method as described in the second aspect.

[0016] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, it implements the molecular graph feature extraction method based on multi-scale information as described in any one of the first aspect of the present invention or the molecular activity prediction method as described in the second aspect.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects: by extracting and fusing the atomic-level features, bond-level features and conformational features of molecules, the model's understanding of the relationship between the molecule itself and activity is promoted, and the molecular activity can be evaluated more comprehensively; the atom-bond-atom of the two-dimensional molecular graph is used as a new node to construct a bond graph, and then the bond-bond of the bond graph is used as a new node to reconstruct the bond angle graph, and the bond-level features are extracted based on the bond graph and the bond angle graph, so that the model can focus on the various properties of the bond and the influence of the bond angle information on the activity; the atomic-level features of the molecule are fused with the atomic position features, and the topological structure and the spatial structure are uniformly encoded, so that the model can recognize the spatial structure of the molecule, thereby improving the molecular activity prediction ability; by interacting information between the atoms, bonds, and conformational information of the molecule, a fusion feature is obtained, which promotes the model's understanding of the relationship between the molecule itself and activity. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings: Figure 1 1 is a flow chart of a molecular graph feature extraction method according to an embodiment of the present invention; Figure 2 1 is a schematic diagram of a process for extracting atomic-level features according to an embodiment of the present invention; Figure 3 is a schematic diagram of conversion of a two-dimensional molecular graph according to an embodiment of the present invention; Figure 4 1 is a schematic diagram of a process for extracting key-level features according to an embodiment of the present invention; Figure 5 is a schematic diagram of reconstruction of a two-dimensional molecular graph according to an embodiment of the present invention; Figure 6 Schematic diagram of a key-level feature extraction method according to an embodiment of the present invention; Figure 7 is a schematic diagram of coding of a conformational feature of an embodiment of the present invention; Figure 8 This is a schematic diagram of feature fusion according to an embodiment of the present invention; Figure 9 It is a schematic diagram of selective transmission of a feature fusion process in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0020] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to other steps or units inherent in the device.

[0021] The terms used in various embodiments of the present invention are only used to describe the purpose of specific embodiments and are not intended to limit the various embodiments of the present invention. As used herein, the singular form is intended to also include the plural form, unless the context clearly indicates otherwise. Unless otherwise limited, all terms used here (including technical terms and scientific terms) have the same meaning as those of ordinary skill in the art generally understood by the various embodiments of the present invention. The terms (such as those defined in generally used dictionaries) will be interpreted as having the same meaning as the contextual meaning in the relevant technical field and will not be interpreted as having idealized meaning or too formal meaning, unless clearly defined in various embodiments of the present invention.

[0022] Embodiments of the present invention provide a molecular graph feature extraction method, a molecular activity prediction method, an apparatus, and a medium. By extracting multi-scale information including atomic-level features, bond-based reconstructed molecular graph features, and molecular 3D rotational-translational symmetry invariance features, the different scale features are fused to provide comprehensive information for molecular activity prediction.

[0023] See Figure 1 , Figure 1 The figure shows a flow chart of the molecular graph feature extraction method, which includes the following steps: S1. Multi-scale feature extraction: including extraction of atomic-level features, bond-level features, and conformational features of target molecules; S2. Fusion of the atomic-level features, bond-level features, and conformational features of the target molecule to obtain the molecular graph features of the target molecule.

[0024] In one embodiment, see Figure 2 The schematic diagram of atomic-level feature extraction shown in the figure shows that extracting the atomic-level features of the target molecule includes the following steps: S101, first converting the SMILES string representation of the target molecule into a two-dimensional molecular graph; S102. Using a message passing network to extract node features from a two-dimensional molecular graph to obtain atomic-level features.

[0025] SMILES string is a linear symbol used to input and represent molecular reactions and is encoded in ASCII, as shown in Table 1.

[0026] Table 1 Correspondence between SMILES strings and molecular names

[0027] A two-dimensional molecular graph describes the two-dimensional topological structure of a molecule, where each node (atom) is connected by edges (bonds). Figure 3The schematic diagram of two-dimensional molecular graph conversion is shown. The target molecule is 3-methyl-2-butene, and its SMILES string is CC=C(C)C. The RDKit tool is used to convert the SMILES string CC=C(C)C into a two-dimensional molecular graph representation.

[0028] In a preferred embodiment, the two-dimensional molecular graph is input into a module with three layers of information extraction at different scales to extract molecular features at the atomic level. v Assign an initial eigenvector The first layer of information extraction uses the Message Passing Neural Network (MPNN). In each step of the message passing phase, each atom v Aggregate information from its neighbors to update its own feature representation, which is expressed as follows:

[0029]

[0030]

[0031] in, is the message function, For the node update function, t is the time step of the run, It's an atom v Neighbor atoms w A collection of and is the feature transformation function, Represents logical operations, is the sigmoid function, is a nonlinear feature transfer function, The running time step is t Atom v The feature representation of The running time step is t Atom w The feature representation of It is composed of atoms v The generated transfer t+ 1 run time step message, It's an atom v exist t+ 1 running time step to update the feature representation, It's an atom v The final feature representation of .

[0032] In one embodiment, see Figure 4The process flow diagram of bond-level feature extraction shown in the figure includes the following steps: S111. Using atoms and bonds as nodes, reconstruct the two-dimensional molecular graph to obtain the bond graph; S112, using the adjacent nodes of the bond graph as new nodes, reconstructing the bond graph to obtain a bond angle graph; S113. Node features are extracted from the bond graph and the bond angle graph through a transformer network to obtain bond-level features.

[0033] In order to better obtain the characteristics of molecular bonds, the original molecular graph is reconstructed based on different bond information to obtain the molecular characteristics at the bond level. Figure 5 The diagram of the reconstruction of the two-dimensional molecular graph shown in the figure reconstructs the original two-dimensional molecular graph based on different bond information to obtain bond graphs and bond angle graphs, and finally obtains the molecular features at the bond level based on the Transformer model. 、 and the edge between them as a new node , for two new nodes, if they have common atoms, then a new edge is formed between them. Initial feature embedding of It is expressed as follows:

[0034] in, and represents the trainable projection matrix, represents the connection operator, 、 Respectively represent the formation of new nodes Two atoms of 、 The eigenvector of Indicates the formation of a new node Two atoms of 、 The edge between.

[0035] The bond angle graph is a molecular graph that is further reconstructed based on the bond graph, thus focusing on the information between bonds. The reconstruction method is similar to the bond graph. The bond angle graph regards two adjacent nodes of the bond graph as a new node. , and a new edge is formed between two new nodes with a common key graph node, and the new node Initial feature embedding of As shown below:

[0036] in, represents the trainable projection matrix, and Form new nodes respectively The eigenvectors of two nodes in the bond graph.

[0037] Transformer is a model based on the attention mechanism. In step S113, given the node feature matrix H of the key graph (H is the initial feature embedding vector of all nodes in the key graph The stacked feature matrix is ​​fed into a multi-head self-attention module for feature extraction. The multi-head self-attention module is expressed as follows:

[0038]

[0039]

[0040] in, Indicates the l -1 layer node feature matrix, , and Respectively represent l Layer k A trainable projection matrix for the head, 、 represents the vector of calculated weights, is the number of heads of multi-head attention, represents the dimension of each self-attention head, Represents the softmax operation.

[0041] Finally, the multi-head attention mechanism finally outputs Pass it to the feedforward neural network FFN to obtain the final feature representation , i.e., the bond-level molecular characteristics, are expressed as follows:

[0042]

[0043] in, Represents the LayerNorm operation, It is the new feature matrix obtained after the LayerNorm operation, which standardizes the statistical attributes while retaining most of the information of the original data. represents the GELU activation function, and Representative layer l The trainable projection matrix at .

[0044] The aforementioned bond-level feature extraction directly applies the classic Transformer architecture, which results in a significant loss of molecular structural information because it ignores the connectivity of the graph. To preserve the original connectivity information, this paper improves the classic Transformer architecture by using the shortest path as an additional embedding, thereby introducing structural information into the multi-head self-attention layer.

[0045] In one embodiment, an improved Transformer architecture is used to extract bond-level features. Specifically, for each pair of nodes in the molecular bond graph and , first derive the shortest path between them, and then encode the shortest path feature into the path attention matrix Attention scalar , as shown below:

[0046]

[0047] in, represents the shortest path function implemented by networkx, express and The shortest path between represents the length of the path, Indicates the shortest path n The characteristics of the nodes, Indicates the path n The trainable projection matrix of nodes, Represents a trainable projection matrix that projects the path embedding to the attention scalar.

[0048] The same operation is performed on the molecular bond angle graph to obtain its shortest path attention matrix.

[0049] Then, in order to introduce the encoded shortest path structure information into the model, the attention matrix The formula is rewritten as:

[0050] in, is the shortest path encoding matrix of the bond graph or bond angle graph. The node feature matrix H of the bond graph is extracted through the improved Transformer architecture.

[0051] Further, see Figure 6The diagram below shows a schematic diagram of the bond-level feature extraction method. This method uses a generative self-supervised learning scheme to obtain the final bond-level features of a molecule. First, the Extended-Connectivity Fingerprints (ECFP) method is used to obtain the molecular fingerprint of the original node graph (atomic node graph) of the two-dimensional molecular graph. The molecular fingerprint is initialized as a virtual node, which interacts with all other nodes and provides a guidance node for masked prediction, thereby better guiding the model to learn the features of the masked nodes. A portion of the nodes in the atomic node graph are then randomly selected for masking, effectively blocking the initial features of these nodes. The masked node feature matrix is ​​then input into a transformer model. Self-attention is used to calculate the association weights between unmasked and masked positions, resulting in output features that embed the attention information. Finally, the output features are passed through a multi-layer perceptron (MLP) to predict the masked nodes, resulting in the final bond-level features.

[0052] This method uses cross entropy loss to optimize the feature recovery of virtual nodes and masked nodes of molecular fingerprints, which is expressed as follows:

[0053] in, is the true label, which takes the value of 0 or 1. is the probability that the model predicts that the sample is class 1, is the natural logarithm.

[0054] In one embodiment, extracting the conformational features of the target molecule includes the following steps: combining the atomic-level features obtained in steps S101 to S102 with the coordinate information of the node. The atomic-level features are 2D features, and 3D features are obtained after combining with the coordinate information, and molecular features at the conformational level can be obtained.

[0055] G represents the 2D molecular graph, R represents the 3D conformation of the molecule, and the first i Rows represent nodes i To better combine the 2D and 3D information of molecules, the two are uniformly encoded and output as a unified representation. Figure 7 , Figure 7 The following is a schematic diagram of the encoding of conformational features. The conformational feature extraction model is represented as , This is achieved by stacking L identical blocks, whose encoding results are derived from the final block. In each block, the atomic representations (black nodes with numbers) are first updated using information from the relevant edge representations (arrows) and the global representation (black nodes without numbers). The bond representations are then updated using the relevant atomic representations and the global representation. Finally, the global representation is updated using the updated atomic and bond representations. Each block takes the output of the previous block as input, and the final block outputs the conformational features of the target molecule, expressed as:

[0056] in, Represents an atomic set Atoms in v The representation of , that is, the conformational characteristics, Represents the encoding model.

[0057] Model In the example, each block takes the output of the previous block as input. l- 1 block will output atoms Representation ,key Representation , predicted conformational features , and global representation of molecules , No. l The workflow of each block is as follows: (1) Merge l- 1 block outputs geometric information:

[0058]

[0059] in, represents a two-layer feedforward neural network with ReLU activation and batch normalization, are real coordinates or coordinates randomly sampled from [-1,1], these two cases appear randomly, It is l- 1 block i Atomic coordinates, ";" represents the cascade operation of the two vectors before and after, It is l- 1 block output atom The expression, Represents atoms combined with 3D information i The expression, It is l- 1 block output key The expression, Indicates the key after combining with 3D information Representation.

[0060] (2) Update atomic representation:

[0061]

[0062]

[0063] in, 、 and are the parameters to be learned, is the normalized exponential function, is a scaling factor, and MLP is a multi-layer perceptron.

[0064] (3) Update key representation:

[0065] in, The input is the concatenation of all input tensors, and the output is a n dimensional vector, where n is the dimension of the network's hidden state.

[0066] (4) Update the global representation:

[0067] in, Indicates the number of molecular nodes, Indicates the number of molecular edges.

[0068] (5) Predict new 3D conformational features: After obtaining the updated representation, Each block predicts a new 3D conformation, i.e., a unified representation of the conformation. , there are the following prediction methods:

[0069]

[0070] in, It is Atoms in new conformations predicted by each block The coordinate offset of It is The average coordinate shift of the molecular coordinates of the block, It's an atom Final predicted coordinates, final nodes Expressed as .

[0071] Furthermore, before encoding the coordinate information of the atomic-level features and the nodes, the molecular invariant representation is obtained by reconstructing the atoms and coordinates. This module has two different tasks: the first is the task of reconstructing the atoms of the 2D graph G , the second is the 3D conformation R reconstruction coordinate task . will output masked atomic features, The coordinates of the screened atoms will be output.

[0072] See also Figure 7 As shown, Figure 7 The atomic information of node 4 and the coordinates of node 5 are shielded. and Afterwards, they and Perform reconstruction.

[0073] Given a The node molecule G, each of which will have a probability is masked. In addition, each atom information and its coordinates are also masked independently. The set of indices of unmasked atoms and coordinates are represented as and The first task is to reconstruct the shielded atoms, and its loss function is:

[0074] in, i are all masked atom indices, Refers to the given unobscured atom and the unobstructed coordinates In the case of shadowed atoms, the prediction distribution.

[0075] Second Task Is to evaluate the real coordinates and the generated coordinates The difference between them, for 3D conformations, is that there may be symmetrical molecules in which the coordinates of the atoms can be exchanged and the conformation remains unchanged after the coordinate exchange, and even if some molecules are not symmetrical, they still have symmetrical substructures.

[0076] To solve this problem, the present invention uses substitution invariant loss, given a molecule G, let Indicates from First, a graph automorphism algorithm is used to identify symmetries in the molecular graph. This algorithm can find atoms in the molecule that can be interchanged without changing the overall structure of the molecule. Then, based on the identified symmetries, one or more symmetry mappings are generated. , each mapping is a permutation function that maps an atom index to itself or to another symmetric atom. Let represents the set of atomic mappings on the symmetric substructure, then the loss function of the symmetric substructure with permutation invariance is defined as follows:

[0077] In the loss function, for each predicted atomic coordinate , considering all possible symmetric mappings And apply them to evaluate the difference between the predicted coordinates and the actual coordinates after taking into account symmetry. At the same time, the conformation of the molecule should maintain rotational translation invariance. Rotational translation invariance means that if we rigidly rotate and translate the generated 3D conformation, the distance between the conformation and the ground truth conformation remains unchanged. The final loss function for the second task is defined as follows:

[0078] in, is a rotation and translation operation, which can be written as , is a 3×3 rotation matrix, is the translation vector.

[0079] The total loss of this module is:

[0080] in, The tasks are and The weight of the loss, .

[0081] In one embodiment, in order to fuse the features at the three different levels above and provide comprehensive molecular insights for activity prediction, the information at different scales is initially extracted through the above operations and then transferred downward and upward to fuse the information at the current scale. The fusion steps include the following: Selectively transferring information between the atomic-level features and the bond-level features to obtain first interactive information; Performing feature extraction on the first interaction information through a graph message passing network to obtain a first extracted feature; Selectively transferring information between the first extracted feature and the atomic-level feature to obtain second interactive information; Performing feature extraction on the second interaction information through a graph message passing network to obtain a second extracted feature; Selectively transferring information between the first extracted feature and the conformational feature to obtain third interactive information; Performing feature extraction on the third interaction information through a graph message passing network to obtain a third extracted feature; Selectively transferring information between the third extracted feature and the first extracted feature to obtain fourth interaction information; Performing feature extraction on the fourth interaction information through a graph message passing network to obtain a fourth extracted feature; The second extracted feature, the third extracted feature and the fourth extracted feature are spliced ​​together to obtain the molecular graph feature of the target molecule.

[0082] See also Figure 8 As shown in the feature fusion diagram, the atomic-level features, bond-level features, and conformational features are respectively passed through the selective transmission module and then input into the subsequent graph message passing network MPNN for processing. At the same time, the atomic-level molecular features not only need to be processed at the same level, but also need to be passed down to the selective transmission module 2 for the bond-level features. After the selective transmission module 2 interacts with the bond-level features, the first interactive information is obtained, thereby enhancing the feature representation at this scale. The first interaction is sent to the graph message passing network MPNN-2 for feature extraction to obtain the first extracted feature. The first extracted feature needs to continue to be operated by the selective transmission module 4 at the same level. At the same time, the first extracted feature needs to be passed up to the selective transmission module 1 to interact with the atomic-level features to obtain the second interactive information. The second interactive information is passed through the graph message passing network MPNN-1 at the same level for feature extraction to obtain the second extracted feature. The first extracted feature also needs to be passed down to the selective transmission module 3 to interact with the conformational features to obtain the third interactive information. The third interactive information is passed through the graph message passing network MPNN-3 at the same level for feature extraction to obtain the third extracted feature. The third extracted feature is passed upward to the selective transmission module 4, interacting with the first extracted feature to obtain the fourth interactive information. The fourth interactive information is subjected to feature extraction by the graph message passing network MPNN-4 at the same level to obtain the fourth extracted feature. Finally, the second extracted feature, the third extracted feature and the fourth extracted feature are spliced ​​together to obtain a multi-scale fusion feature. .

[0083] Furthermore, in order to avoid the influence of initial feature quality on fusion weights, a two-stage approach is used in the selective transfer stage. Figure 9 As shown in the schematic diagram of selective transfer, given two features of different scales X (current scale) and Y (other scales), the selected fusion feature Z is first obtained, and the formula is as follows:

[0084]

[0085] in, is the fusion weight, such as Figure 9As shown by the solid arrow pointing left, Indicated by the right-pointing dotted arrow, softmax is a normalized exponential function. is the scaling factor.

[0086] Then use the fusion feature Z as the input of another module of the same type to calculate the fusion weight And the final fusion feature, the formula is as follows:

[0087]

[0088] in, It is the fusion feature of features X and Y.

[0089] In summary, in the above-described embodiments of the present invention, the molecular structure is reconstructed, with atoms and bonds as nodes. The model then automatically learns bond features. Next, by updating the positional information between nodes, a 3D representation of the molecular graph is obtained, which is invariant to rotation and translation. Finally, a message-passing network is used to interact and fuse atomic, bond, and conformational features to obtain a holistic representation of the molecule, providing comprehensive information for activity prediction.

[0090] An embodiment of the present invention further provides a method for predicting molecular activity, comprising: S210, extracting atomic-level features, bond-level features, and conformational features of the target molecule; S220, fusing the atomic-level features, bond-level features, and conformational features in S210 to obtain molecular graph features of the target molecule; S310 , inputting the molecular graph features in S220 into a multi-layer perceptron established according to the prediction task, so as to predict the activity of a certain category of drug molecules.

[0091] The prediction process is expressed as follows:

[0092] in, is a multi-layer perceptron, is the molecular graph feature after multi-scale feature fusion, For the prediction results.

[0093] The loss function for the classification task uses the cross entropy loss function , the regression task uses the root mean square error (RMSE) loss function .

[0094]

[0095]

[0096] in, is the number of molecules in the current training batch, and represent the true and predicted values ​​of molecular properties, respectively.

[0097] An embodiment of the present invention further provides an electronic device comprising a processor and a memory, wherein the number of processors may be one or more. The memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor executes the software programs, instructions, and modules stored in the memory to perform various functional applications and data processing of the electronic device, thereby implementing the molecular graph feature extraction method or molecular activity prediction method based on multi-scale information according to any of the above embodiments of the present invention.

[0098] The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function; the data storage area may store data generated based on the use of the terminal, etc. Furthermore, the memory may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state memory device. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories may be connected to the electronic device via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0099] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the molecular graph feature extraction method or molecular activity prediction method based on multi-scale information according to any embodiment of the present invention.

[0100] The computer storage media of the embodiments of the present invention may employ any combination of one or more computer-readable media. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device, or component.

[0101] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0102] An embodiment of the present invention further provides a computer program product, which, when executed on a computer, enables the computer to execute the molecular graph feature extraction method or molecular activity prediction method based on multi-scale information according to any of the above embodiments of the present invention.

[0103] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A molecular graph feature extraction method based on multi-scale information, characterized in that: include: Extract atomic-level features, bond-level features, and conformational features of target molecules; The atomic-level features, bond-level features and conformational features are integrated to obtain the molecular graph features of the target molecule.

2. The molecular graph feature extraction method based on multi-scale information according to claim 1, characterized in that: The method for extracting atomic-level features of a target molecule comprises: Convert the SMILES string of the target molecule into a two-dimensional molecular graph; Node features of the two-dimensional molecular graph are extracted through a message passing network. In each step of the message passing phase, each atom updates its own feature vector by aggregating messages from neighboring atoms. After the final update step, the atomic-level features of each atom in the two-dimensional molecular graph are obtained.

3. The molecular graph feature extraction method based on multi-scale information according to claim 2, characterized in that: The method for extracting bond order features of a target molecule comprises: Reconstructing the two-dimensional molecular graph by using adjacent nodes of the two-dimensional molecular graph and edges therebetween as new nodes to obtain a bond graph; Reconstruct the bond graph by using adjacent nodes of the bond graph as new nodes to obtain a bond angle graph; The node feature matrices of the bond graph and the bond angle graph are subjected to feature extraction by a transformer model to obtain bond-level features.

4. The molecular graph feature extraction method based on multi-scale information according to claim 3, characterized in that: The step of extracting features from the node feature matrices of the bond graph and the bond angle graph using a transformer model includes: The shortest path information of each pair of nodes in the bond graph or bond angle graph is introduced into the multi-head attention layer to improve the attention matrix of the transformer model. The improved attention matrix is ​​expressed as: in, represents the attention matrix, 、 represents the vector of calculated weights, represents the dimension of each self-attention head, is the shortest path encoding matrix of the bond graph or bond angle graph, Represents the softmax operation.

5. The molecular graph feature extraction method based on multi-scale information according to claim 3 or 4, characterized in that: Combined with generative self-supervised learning methods, the final key-level features are extracted, including: The original node graph of the two-dimensional molecular graph is used to obtain the molecular fingerprint using the ECFP method; Initialize the molecular fingerprint as a virtual node and interact with other nodes as a guidance node for mask prediction; Randomly select a portion of the nodes in the original node graph for masking, input the masked node feature matrix into the transformer model, and obtain the output features embedded with attention information; The masked nodes in the output features are predicted by a multi-layer perceptron to obtain the final key-level features.

6. The molecular graph feature extraction method based on multi-scale information according to claim 1, characterized in that: The method for extracting the conformational features of the target molecule includes: fusing the atomic-level features with atomic coordinate information to obtain the conformational features.

7. The molecular graph feature extraction method based on multi-scale information according to claim 1, characterized in that: The method for fusing the atomic-level features, bond-level features, and conformational features to obtain the molecular graph features of the target molecule includes: Selectively transmitting information to interact the atomic-level features with the bond-level features to obtain first interaction information, and extracting features from the first interaction information through a graph message passing network to obtain first extracted features; Performing selective information transfer interaction between the first extracted feature and the atomic-level feature to obtain second interaction information, and performing feature extraction on the second interaction information through a graph message passing network to obtain a second extracted feature; Performing selective information transfer interaction between the first extracted feature and the conformation feature to obtain third interaction information, and performing feature extraction on the third interaction information through a graph message passing network to obtain a third extracted feature; Performing selective information transfer interaction between the third extracted feature and the first extracted feature to obtain fourth interaction information, and performing feature extraction on the fourth interaction information through a graph message passing network to obtain a fourth extracted feature; The second extracted feature, the third extracted feature and the fourth extracted feature are spliced ​​together to obtain the molecular graph feature of the target molecule.

8. A method for predicting molecular activity, characterized in that: include: The molecular graph features of the target molecule are extracted by the molecular graph feature extraction method based on multi-scale information as described in any one of claims 1 to 7, and the molecular graph features are input into a pre-trained multi-layer perceptron for prediction to obtain a molecular activity prediction result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for extracting molecular graph features based on multi-scale information according to any one of claims 1 to 7 or the method for predicting molecular activity according to claim 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for extracting molecular graph features based on multi-scale information according to any one of claims 1 to 7 or the method for predicting molecular activity according to claim 8 is implemented.

Images