Molecular property prediction method, device, medium and equipment

Abstract

The invention discloses a chemical property prediction method and device, a medium and equipment, and relates to the technical field of molecular property prediction. Constructing a graph neural network composed of a learner view branch and an anchoring view branch; taking atoms in molecules as nodes in graph data and chemical bonds between the atoms as edges in the graph data to obtain molecular graph data, inputting the molecular graph data into a graph neural network for self-supervised training, and obtaining a first molecular property label by a learner view branch; obtaining a second molecular property tag by anchoring the view branch; performing iterative updating on weight parameters in the graph neural network by taking a mean square error between the first molecular property tag and the second molecular property tag as a loss function; and performing molecular property prediction by using the trained graph neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of molecular property prediction, and particularly relates to a molecular property prediction method, device, medium and equipment. Background Art

[0002] Molecular representation is the core foundation for functional research in the field of chemistry and the innovative design of new compounds. In recent years, with the booming development of machine learning technology, this data-driven method automatically extracts and learns meaningful representations from a vast amount of chemical data through algorithms. Molecular representation is the core foundation for functional research in the field of chemistry and the innovative design of new compounds. However, given the astonishing diversity and extensiveness of potentially stable compounds, constructing a rich and generally applicable representation system that covers the entire chemical space is undoubtedly a daunting task. In past chemical research, traditional molecular representation methods, such as extended connectivity fingerprints, have played a crucial role in the development of computational chemistry by providing a standardized molecular feature encoding framework. These methods have laid a solid foundation for chemical research and promoted progress in multiple fields. In recent years, with the booming development of machine learning technology, the landscape of molecular representation learning has changed dramatically. This data-driven method automatically extracts and learns meaningful representations from a vast amount of chemical data through algorithms. These representations are not only optimized for specific tasks but also possess excellent generalization ability, thus attracting extensive interest and attention. They have shown great potential in fields such as chemical property prediction, complex chemical modeling, and innovative molecular design.

[0003] However, during the process of molecular representation learning, huge challenges still remain. On the one hand, the traditional string representations adopted by classical machine learning and some deep learning methods inherently cannot directly encode the key topological structure information inside molecules. Although graph models have certain advantages in dealing with molecular graphs, when facing large-scale and complex drug molecules, they still face problems of computational complexity and efficiency, and at the same time, the accuracy of feature description is also limited. Although methods such as ImageMol attempt to transform molecular structures into image representations and use computer vision and self-supervised learning techniques to extract semantic features, this method still ignores the inherent graph structure characteristics of molecules. On the other hand, the vastness of the chemical space has led to an astonishing number of potentially pharmacologically active molecules, which highlights the severe problem of data scarcity in molecular learning tasks. Since the experimental verification process of molecular properties is complex and time-consuming, the cost of obtaining labeled data is high. Therefore, machine learning models trained on limited data are prone to overfitting, thus weakening their generalization ability for new compound molecules.

[0004] To address these challenges, many studies in recent years have started using graph neural networks (GNNs) for molecular property prediction and achieved promising results. However, the current self-supervised pre-training frameworks for GNNs mainly focus on node-level or graph-level tasks, and these methods have insufficient generalization ability to capture molecular information in embedded graph data, resulting in the need to improve the accuracy of the predicted results of molecular properties. Summary of the Invention

[0005] Based on this, to solve the technical problems in the prior art, the present invention provides a method, device, medium, and equipment for molecular property prediction.

[0006] The present invention provides a method for molecular property prediction, including:

[0007] Construct a graph neural network composed of a learner view branch and an anchor view branch; wherein, the learner view branch includes a graph learner, an online encoder, and a first projection module connected in sequence; the anchor view branch includes a graph generator, a target encoder, and a second projection module connected in sequence;

[0008] Taking the atoms in the molecule as nodes in the graph data and the chemical bonds between atoms as edges in the graph data to obtain molecular graph data, and inputting the molecular graph data into the graph neural network for self-supervised training, including:

[0009] Enhance the molecular graph data through the graph learner to generate the first molecular initial encoding feature; encode the molecular initial encoding feature through the online encoder to obtain the online molecular encoding feature; project the online molecular encoding feature into the first molecular property label through the first projection module;

[0010] Guide the graph generator to generate or update the second molecular initial encoding feature through the first molecular initial encoding feature; encode the second molecular initial encoding feature through the target encoder to obtain the target molecular encoding feature; project the target molecular encoding feature into the second molecular property label through the second projection module;

[0011] Taking the mean square error between the first molecular property label and the second molecular property label as the loss function to iteratively update the weight parameters in the graph neural network, wherein the weight parameters of the online encoder are iteratively updated using the gradient descent algorithm based on the loss function; the weight parameters of the target encoder are indirectly iteratively updated by performing an exponential moving average on the weight parameters of the online encoder;

[0012] Use the trained graph neural network for molecular property prediction.

[0013] Further, the online encoder and the target encoder have the same structure, both including:

[0014] A local attention mechanism module for performing atomic-level local feature extraction on the features input to the online encoder or the target encoder to obtain initial atomic-level features;

[0015] A first graph pooling module for aggregating the initial atomic-level features to obtain enhanced atomic-level features;

[0016] A first global attention mechanism module for performing atomic-cluster-level global feature extraction on the enhanced atomic-level features to obtain initial atomic-cluster-level features;

[0017] A second graph pooling module for aggregating the initial atomic-cluster-level features to obtain enhanced atomic-cluster-level features;

[0018] A second global attention mechanism module for performing atomic-cluster-level feature extraction on the enhanced atomic-cluster-level features to obtain molecular-level features.

[0019] Further, the local attention mechanism module is formed by introducing a routing attention mechanism module into a weighted graph isomorphism network.

[0020] Further, the generation or update of the second molecular initial encoding feature by guiding the graph generator with the first molecular initial encoding feature is achieved through the following formula:

[0021] A a ← τA a +(1 - τ)S

[0022] where S is the first molecular initial encoding feature generated by the graph learner, τ is the decay rate, and A a is the second molecular initial encoding feature.

[0023] Further, the enhancement of the molecular graph data includes image rotation enhancement, image cropping enhancement, and random perturbation enhancement of the graph structure.

[0024] Further, the weight parameters in the second projection module are iteratively updated based on truncated gradients.

[0025] The present invention provides a molecular property prediction device, including:

[0026] A graph neural network construction module for constructing a graph neural network composed of a learner view branch and an anchor view branch; wherein, the learner view branch includes a graph learner, an online encoder, and a first projection module connected in sequence; the anchor view branch includes a graph generator, a target encoder, and a second projection module connected in sequence;

[0027] A graph neural network training module, which is used to obtain molecular graph data by taking atoms in a molecule as nodes in the graph data and chemical bonds between atoms as edges in the graph data, and input the molecular graph data into the graph neural network for self-supervised training, including:

[0028] Enhance the molecular graph data through a graph learner to generate the first initial molecular encoding features; encode the initial molecular encoding features through an online encoder to obtain online molecular encoding features; project the online molecular encoding features through a first projection module into the first molecular property labels;

[0029] Guide the graph generator to generate or update the second initial molecular encoding features through the first initial molecular encoding features; encode the second initial molecular encoding features through a target encoder to obtain target molecular encoding features; project the target molecular encoding features through a second projection module into the second molecular property labels;

[0030] Use the mean square error between the first molecular property labels and the second molecular property labels as the loss function to iteratively update the weight parameters in the graph neural network, where the weight parameters of the online encoder are iteratively updated using the gradient descent algorithm based on the loss function; the weight parameters of the target encoder are indirectly iteratively updated by performing an exponential moving average on the weight parameters of the online encoder;

[0031] A molecular property prediction module, which is used to predict molecular properties using the trained graph neural network.

[0032] The present invention provides a computer-readable storage medium, where the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned molecular property prediction method is implemented.

[0033] The present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the above-mentioned molecular property prediction method is implemented.

[0034] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0035] In the molecular property prediction method provided by the present invention, through the collaborative training mechanism of the dual-view architecture (learner view and anchor view), the adaptive feature enhancement and stable representation learning of molecular graph data are realized. The learner view can adaptively adjust the feature representation of the model for molecular graph data through dynamic gradient update, which enables the model to capture key information in the data more meticulously. The anchor view provides a smooth target signal through exponential moving average, and the two are aligned through mean square error constraint, ensuring that the model can learn consistent feature representations at different levels. The alignment process of this dual-view architecture enables the model to learn consistent property predictions in different feature spaces, which helps to improve the generalization ability of the model on unseen data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0037] Figure 1 Schematic diagram of the graph neural network structure provided by the present invention;

[0038] Figure 2 Schematic diagram of the data augmentation strategy and graph neural network training framework provided by the present invention;

[0039] Figure 3 Schematic diagram of the local routing attention mechanism structure provided by the present invention;

[0040] Figure 4 Schematic diagram of the global attention mechanism structure provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments and corresponding drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0042] Current self-supervised pre-training frameworks for GNNs mainly focus on node-level or graph-level tasks, and these methods still have deficiencies in capturing the rich information embedded in subgraphs or graph bases. For example, molecular functional groups usually carry indicative information about molecular properties, so effectively extracting functional group information becomes the key to evaluating model performance. Although significant progress has been made in molecular representation learning in recent years, the present invention still needs to continuously explore and innovate to overcome existing challenges and construct a more perfect, information-rich, and generalizable molecular representation system. The method closest to the present invention is the ImageMol method. The ImageMol method converts the molecular structure into an image representation and uses computer vision and self-supervised learning techniques to extract profound semantic features from a large number of unclassified molecular image datasets. Nevertheless, although the ImageMol method uses large-scale unlabeled molecular image data for unsupervised pre-training, reducing the dependence on data annotation, the quality and quantity of the data still have an important impact on the performance of the model. If there is noise or uneven data distribution in the dataset, it may affect the training effect and generalization ability of the model. And from the perspective of image texture features, the ImageMol method ignores the inherent graph structure properties of molecules.

[0043] The method of the present invention introduces a molecular image representation learning framework that innovatively decomposes molecular images into small blocks, conceptualizes them as graph nodes, and then conducts graph structure bootstrap learning. This composite neural network architecture not only proposes a bootstrap mode for the graph structure but also promotes hierarchical analysis of chemical information, from the atomic level, through chemical bonds, to the functional group level within the molecular structure. To train this model, the present invention designs a self-supervised framework that utilizes the powerful capabilities of convolutional neural networks to capture complex image features while delving into subtle chemical structure information across multiple molecular hierarchical levels. This overall approach enables the model of the present invention to generate more accurate and comprehensive molecular feature representations. Moreover, based on the new concept of constructing molecular images, the present invention introduces a groundbreaking molecular data representation method. This perspective provides a novel dimension for data processing. The present invention also develops tailored molecular data augmentation techniques, enabling bootstrap learning of molecular chemical structures and ultimately generating comprehensive and accurate molecular representations. In addition, the present invention proposes an innovative graph attention mechanism based on a weighted graph isomorphism network for processing graph structure data. It solves the problem that the ImageMol method ignores the inherent graph structure properties of molecules. This method reflects the intuitive progress of image understanding, transitioning from fine-grained analysis to holistic understanding. Importantly, it incorporates inductive bias into a multi-level self-attention framework, effectively reducing complexity, alleviating training challenges, thereby enhancing the overall model performance and interpretability, and solving the drawbacks of existing methods.

[0044] To make the above objects, features, and advantages of the present application more apparent and understandable, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] In an embodiment of the present application, the molecular property prediction method includes the following steps:

[0046] S1: Construct a graph neural network composed of a learner view branch and an anchor view branch; wherein, the learner view branch includes a graph learner, an online encoder, and a first projection module connected in sequence; the anchor view branch includes a graph generator, a target encoder, and a second projection module connected in sequence.

[0047] S2: Obtain molecular graph data with atoms in the molecule as nodes in the graph data and chemical bonds between atoms as edges in the graph data, and input the molecular graph data into the graph neural network for self-supervised training, including:

[0048] Enhance the molecular graph data through the graph learner to generate a first molecular initial encoding feature; encode the first molecular initial encoding feature through the online encoder to obtain an online molecular encoding feature; project the online molecular encoding feature into a first molecular property label through the first projection module; generate or update a second molecular initial encoding feature through the first molecular initial encoding feature to guide the graph generator; encode the second molecular initial encoding feature through the target encoder to obtain a target molecular encoding feature; project the target molecular encoding feature into a second molecular property label through the second projection module; use the mean square error between the first molecular property label and the second molecular property label as a loss function to iteratively update the weight parameters in the graph neural network, wherein the weight parameters of the online encoder are iteratively updated using a gradient descent algorithm based on the loss function; the weight parameters of the target encoder are indirectly iteratively updated by performing an exponential moving average on the weight parameters of the online encoder;

[0049] S3: Use the trained graph neural network for molecular property prediction.

[0050] In a specific embodiment, the graph neural network constructed in step S1 is as Figure 1 shown.

[0051] The method proposed by the present invention is different from previous attempts and is a graph-structured self-bootstrapping learning method. It seamlessly combines the powerful function of a convolutional neural network (CNN) for extracting complex local visual features, providing a unique and enhanced perspective for molecular representation and analysis. The constructed neural network model is as Figure 1 shown.

[0052] S101. Structure Bootstrapping Self-supervised Learning: A self-supervised learning framework is proposed, which includes three steps: dual-view construction, neural network construction, and loss function definition and optimization.

[0053] To address the problem that maximizing the consistency between two views often leads to the overfitting of the learned graph structure to the fixed anchor view structure, thus hindering the acquisition of accurate graph representations, the present invention introduces a structure guiding mechanism. The core of this mechanism lies in the gradual update of the anchor view structure A a , guided by the insight of dynamically learning the graph structure s. Given the decay rate τ, the present invention has: A a ← τA a + (1 - τ)S. Regarding data augmentation, the present invention adopts three augmentation strategy techniques on the input molecular images, such as Figure 2 shown, including:

[0054] 1) Image Rotation: Rotate the molecular image at different angles, then divide the image into completely different nodes through a grid, and obtain the node feature representation through convolution. At the same time, the neural network integrates and extracts a coherent feature representation from two angles, each angle containing a highly diverse set of graph nodes, and finally trains the model to recognize and understand the chemical structure of the molecule.

[0055] 2) Image Cropping: The present invention applies the cropping technique to the image and trains the neural network to accurately predict and reconstruct the missing part of the cropped molecular image. This method enhances the model's ability to understand and interpret complex molecular structures.

[0056] 3) Random Perturbation of Graph Structure: The present invention incorporates random perturbation into the input of the initial spatial adjacency matrix of the graph learner. The matrix is modified by changing non-zero elements to zero or vice versa, changing zero elements to non-zero elements, and training the graph learner to learn the ability to eliminate incorrect spatial edge connections and establish meaningful connections within an unsupervised framework. Thereby, the robustness of the model when processing the initial spatial adjacency matrix as input is enhanced, and the ability of graph structure reasoning is also improved.

[0057] S102: Hierarchical Attention Mechanism Based on Graph Isomorphism Network.

[0058] 1) Weighted Graph Isomorphism Networks: In standard Graph Isomorphism Networks (GINs), node features are aggregated from their neighbors using a predefined aggregation function. The model of the present invention introduces a weighted node aggregation method that is suitable for the continuous values of the learned graph structure matrix s. These values represent the distance and strength of the connections between nodes. However, matrix S is dense, representing a fully connected graph, and often requires a large computational cost. Therefore, the preprocessing matrix must be sparsified before the message passing phase in the graph isomorphism network. The matrix is refined by retaining only the λ strongest adjacent edges, thereby improving computational efficiency and focusing on the most important connections within the graph structure.

[0059] 2) Routing attention: The purpose of routing attention is to limit the attention scope to the graph nodes with a higher connectivity score defined by the graph structure matrix in the given graph data. As Figure 1 shown, Figure 1 it has two view branches. In the anchor view, since the graph structure matrix is initialized as the spatial adjacency matrix, it is approximately equal to the initial value Ai in the initial stage of training. By selecting the k nearest neighbors (k-NN) and ensuring k≥8, it is ensured that each node captures its spatial adjacent modules in 8 directions, effectively using the "gather" operation to gather useful nodes under the attention mechanism. This constitutes a spatial local attention mechanism, which promotes the spatial clustering of fragmented molecular image segments and basically solves the problem of ignoring local key information with global attention.

[0060] The specific process is as Figure 3 shown, and it involves using the Graph Isomorphism Network (GIN) to calculate the key, query, and value (KQV) characterizations. Subsequently, node collection is performed based on the graph structure matrix, and the final graph node feature representation is derived through the matrix inner product. To alleviate the vanishing gradient problem in the deep network, the attention layer incorporates a residual connection structure, and its input is added to the output to ensure the effective propagation of valuable gradients throughout the network. This method harmoniously integrates the graph structure information with the attention mechanism, enhancing the model's ability to capture local patterns while maintaining the integrity of the context.

[0061] Graph attention: A global attention mechanism that first uses the molecular image segmentation graph and the graph structure matrix, and uses the Graph Isomorphism Network (GIN) to calculate the key and query, thereby generating an attention matrix. Subsequently, again using GIN, based on the generated matrix and the original molecular image segmentation graph, the value is calculated using the residual connection, as Figure 4 shown.

[0062] To address the low-rank bottleneck problem, the present invention adopts the "talking-heads" method proposed by Shazeer et al., which combines linear projections across the entire attention head dimension. This method can alleviate the low-rank bottleneck problem and promote rich interactions between attention heads, enhancing their collective ability to capture complex dependencies and nuances in molecular graph data.

[0063] 4) Graph-pooling: As Figure 1 shown, two graph pooling layers are used between the routing attention layer and the graph attention layer (LocalRouting Attention). The role of these pooling layers is to aggregate molecular image fragments into atomic and chemical bond information, and further aggregate it into functional group information. This hierarchical aggregation of semantic information enables hierarchical analysis of the molecular chemical structure, thus allowing for a more concise and hierarchical representation of the graph's structural properties.

[0064] In a specific embodiment, the process of training the graph neural network in step S2 is as Figure 2 shown.

[0065] The neural network architecture includes three different neural networks: a Graph Learner, an Online Network, and a Target Network. The Graph Learner adopts the same neural architecture as the global graph attention component within the encoder network. The Online Network includes three stages: an encoder g θ , a projector P θ , and a predictor q θ . The Target Network consists of an encoder g ξ and a projector p ξ , providing the regression target for training the Online Network, and its parameters are the exponential moving average of the Online Network parameters. This asymmetric structural design can prevent the model from collapsing during training.

[0066] Finally, the loss function is defined as the mean squared error between the representations generated by the two branches. The Online Network is updated by minimizing the loss function through gradient descent, while at the backend of the Target Network, the gradient calculation is truncated. This setting enables the framework to learn a robust representation of molecular images in an unsupervised manner, leveraging the inherent structure and properties of the data through different views generated by the Graph Learner and augmentation.

[0067] In addition, to prove the effectiveness of the method of the present invention, the present invention benchmarks the performance of multiple challenging classification and regression tasks in MoleculeNet. For all datasets except QM9, the present invention adheres to the scaffold splitting strategy advocated in the strategy of graph model and graph neural network pre-training, and designs 80 / 10 / 10 splits for training, validation, and testing respectively. The main results observed by the present invention from these results are as follows: By combining self-supervised pre-training, the method of the present invention shows significant robustness and outperforms all supervised methods. Even on the remaining dataset ClinTox, the method of the present invention also maintains a competitive advantage.

[0068] To address the problems of existing methods in processing molecular data, such as traditional string representations being unable to encode molecular topologies, graph models having complex calculations and limited feature descriptions, methods like ImageMol ignoring graph structures, insufficient information capture in the GNN self-supervised pre-training framework, and overfitting of machine learning models on limited data, the present invention has studied a new method. The present invention introduces a molecular image representation learning framework that innovatively decomposes molecular images into small pieces, conceptualizes them as graph nodes, and then performs graph structure bootstrap learning. This method, based on the principles of weighted graph isomorphism networks and graph pooling functions, innovatively combines graph learners with a hybrid mechanism composed of local routing and global graph attention. This composite neural network architecture not only proposes a bootstrap mode for graph structures but also promotes hierarchical analysis of chemical information, from the atomic level, through chemical bonds, to the functional group level within molecular structures.

[0069] To train the model, the present invention designs a self-supervised framework that uses a contrastive learning strategy to co-train the anchor view and the learner view that enhance molecular representations. This framework utilizes the powerful capabilities of convolutional neural networks (CNNs) to capture complex image features while delving into subtle chemical structure information across multiple molecular hierarchical levels. This overall method enables the model of the present invention to generate more accurate and comprehensive molecular feature representations.

[0070] Furthermore, based on a new concept for constructing molecular images, the present invention introduces a groundbreaking method for representing molecular data. This perspective provides a novel dimension for data processing. Secondly, the present invention formulates a novel fusion training paradigm that integrates a graph learner with branches of anchor views and learner views, all within a robust BYOL self-supervised learning framework. The present invention also develops tailored molecular data augmentation techniques that enable bootstrap learning of molecular chemical structures and ultimately generate comprehensive and accurate molecular representations. In addition, the present invention proposes an innovative graph attention mechanism based on a weighted graph isomorphism network for processing graph-structured data. This mechanism can precisely adjust its focus at different information levels, from superpixel segments to individual atoms, chemical bonds, and ultimately functional groups. This approach reflects the intuitive progress of image understanding, transitioning from fine-grained analysis to holistic understanding. Importantly, it incorporates inductive biases into a multi-level self-attention framework, effectively reducing complexity and alleviating training challenges, thereby enhancing overall model performance and interpretability and addressing the drawbacks of many existing methods.

[0071] The main advantage of the model of the present invention lies in its integration within a self-supervised framework that leverages the powerful capabilities of convolutional neural networks to capture complex image features while delving into subtle chemical structure information across multiple molecular hierarchical levels. This holistic approach enables the model of the present invention to generate more precise and comprehensive molecular feature representations. Furthermore, based on a new concept for constructing molecular images, the present invention introduces a groundbreaking method for representing molecular data. This perspective provides a novel dimension for data processing. Secondly, the present invention formulates a novel fusion training paradigm, develops tailored molecular data augmentation techniques that enable bootstrap learning of molecular chemical structures and ultimately generate comprehensive and accurate molecular representations. In addition, the present invention proposes an innovative graph attention mechanism based on a weighted graph isomorphism network for processing graph-structured data, which reflects the intuitive progress of image understanding, transitioning from fine-grained analysis to holistic understanding. It effectively reduces complexity and alleviates training challenges, thereby enhancing overall model performance and interpretability. The method of the present invention exhibits remarkable robustness, outperforming all supervised methods. And tests have shown that the GIN of the method of the present invention has demonstrated excellent performance on datasets such as BACE and MUV. Even on the remaining dataset ClinTox, the method of the present invention maintains a competitive edge.

[0072] The above is the method for predicting molecular properties provided by one or more embodiments of the present invention. Based on the same concept, the present invention also provides a corresponding device for predicting molecular properties, including:

[0073] A graph neural network construction module for constructing a graph neural network composed of a learner view branch and an anchor view branch; wherein, the learner view branch includes a graph learner, an online encoder, and a first projection module connected in sequence; the anchor view branch includes a graph generator, a target encoder, and a second projection module connected in sequence.

[0074] A graph neural network training module for obtaining molecular graph data with atoms in a molecule as nodes in the graph data and chemical bonds between atoms as edges in the graph data, and inputting the molecular graph data into the graph neural network for self-supervised training, including:

[0075] Enhancing the molecular graph data through the graph learner to generate a first molecular initial encoding feature; encoding the first molecular initial encoding feature through the online encoder to obtain an online molecular encoding feature; projecting the online molecular encoding feature into a first molecular property label through the first projection module; guiding the graph generator to generate or update a second molecular initial encoding feature through the first molecular initial encoding feature; encoding the second molecular initial encoding feature through the target encoder to obtain a target molecular encoding feature; projecting the target molecular encoding feature into a second molecular property label through the second projection module; using the mean square error between the first molecular property label and the second molecular property label as a loss function to iteratively update the weight parameters in the graph neural network, wherein the weight parameters of the online encoder are iteratively updated using a gradient descent algorithm based on the loss function; the weight parameters of the target encoder are indirectly iteratively updated by performing an exponential moving average on the weight parameters of the online encoder.

[0076] A molecular property prediction module for predicting molecular properties using the trained graph neural network.

[0077] For the specific limitations of the molecular property prediction device, reference can be made to the limitations on the molecular property prediction method in the above text, which will not be elaborated here. Each module in the above molecular property prediction device can be implemented in whole or in part through software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0078] The present invention also provides a computer-readable storage medium storing a computer program, which can be used to execute the above Figure 1 provided molecular property prediction method.

[0079] The present invention also provides the structure of a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above Figure 1 provided molecular property prediction method.

[0080] Those of ordinary skill in the art can understand that all or part of the processes in the above-described embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-described method embodiments. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided by the present invention can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0081] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded by the present invention.

Claims

1. A method for predicting molecular properties, characterized in that, Including: Construct a graph neural network composed of a learner view branch and an anchor view branch; wherein, the learner view branch includes a graph learner, an online encoder, and a first projection module connected in sequence; the anchor view branch includes a graph generator, a target encoder, and a second projection module connected in sequence; Taking the atoms in the molecule as the nodes in the graph data and the chemical bonds between the atoms as the edges in the graph data to obtain molecular graph data, and inputting the molecular graph data into the graph neural network for self-supervised training, including: Enhancing the molecular graph data through the graph learner to generate a first molecular initial encoding feature; encoding the molecular initial encoding feature through the online encoder to obtain an online molecular encoding feature; projecting the online molecular encoding feature into a first molecular property label through the first projection module; Guiding the graph generator to generate or update a second molecular initial encoding feature through the first molecular initial encoding feature; encoding the second molecular initial encoding feature through the target encoder to obtain a target molecular encoding feature; projecting the target molecular encoding feature into a second molecular property label through the second projection module; Taking the mean square error between the first molecular property label and the second molecular property label as the loss function to iteratively update the weight parameters in the graph neural network, wherein, the weight parameters of the online encoder are iteratively updated using the gradient descent algorithm based on the loss function; the weight parameters of the target encoder are indirectly iteratively updated by performing an exponential moving average on the weight parameters of the online encoder; Using the trained graph neural network for molecular property prediction.

2. The molecular property prediction method according to claim 1, wherein The online encoder and the target encoder have the same structure, and both include: A local attention mechanism module for performing atomic-level local feature extraction on the features input to the online encoder or the target encoder to obtain an initial atomic-level feature; A first graph pooling module for aggregating the initial atomic-level features to obtain enhanced atomic-level features; A first global attention mechanism module for performing atomic group-level global feature extraction on the enhanced atomic-level features to obtain an initial atomic group-level feature; A second graph pooling module for aggregating the initial atomic group-level features to obtain enhanced atomic group-level features; A second global attention mechanism module for performing atomic group-level feature extraction on the enhanced atomic group-level features to obtain a molecular-level feature.

3. The molecular property prediction method according to claim 2, wherein The local attention mechanism module is formed by introducing a routing attention mechanism module into a weighted graph isomorphism network.

4. The molecular property prediction method according to claim 1, wherein The guiding the graph generator to generate or update a second molecular initial encoding feature through the first molecular initial encoding feature is achieved by the following formula: A a ← τA a +(1 - τ)S Among them, S is the first molecular initial encoding feature generated by the graph learner, τ is the decay rate, and A a is the second molecular initial encoding feature.

5. The molecular property prediction method according to claim 1, wherein The enhancing the molecular graph data includes image rotation enhancement, image cropping enhancement, and random perturbation enhancement of the graph structure.

6. The molecular property prediction method according to claim 1, wherein The weight parameters in the second projection module are iteratively updated based on truncated gradients.

7. A molecular property prediction device, characterized in that Including: A graph neural network construction module for constructing a graph neural network composed of a learner view branch and an anchor view branch; wherein, the learner view branch includes a graph learner, an online encoder, and a first projection module connected in sequence; the anchor view branch includes a graph generator, a target encoder, and a second projection module connected in sequence; A graph neural network training module, which is used to obtain molecular graph data by taking atoms in a molecule as nodes in graph data and chemical bonds between atoms as edges in graph data, and input the molecular graph data into a graph neural network for self-supervised training, including: Enhancing the molecular graph data through a graph learner to generate a first initial molecular encoding feature; encoding the first initial molecular encoding feature through an online encoder to obtain an online molecular encoding feature; projecting the online molecular encoding feature into a first molecular property label through a first projection module; Guiding a graph generator to generate or update a second initial molecular encoding feature through the first initial molecular encoding feature; encoding the second initial molecular encoding feature through a target encoder to obtain a target molecular encoding feature; projecting the target molecular encoding feature into a second molecular property label through a second projection module; Taking the mean square error between the first molecular property label and the second molecular property label as a loss function to iteratively update the weight parameters in the graph neural network, wherein the weight parameters of the online encoder are iteratively updated by using a gradient descent algorithm based on the loss function; the weight parameters of the target encoder are indirectly iteratively updated by performing an exponential moving average on the weight parameters of the online encoder; A molecular property prediction module, which is used to predict molecular properties by using the trained graph neural network.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and when the computer program is executed by a processor, the method described in any one of claims 1 to 6 above is implemented.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the method described in any one of claims 1 to 6 above is implemented.

Images