Molecular property prediction method, system, apparatus, and storage medium

Abstract

The application discloses a molecular property prediction method, system, device and storage medium. The method comprises the following steps: obtaining a first molecule; extracting a first feature of the first molecule through a dual-graph network model; in a pre-training process, the dual-graph network model aggregates the feature representation of one molecule through multiple neural networks based on dual-graph neural network contrast learning, and performs supplementary fusion processing on the molecular information obtained by different neural networks; and predicting the molecular property through the first feature. The embodiment of the application performs molecular feature representation based on the dual-graph network model, makes the molecular representations obtained by different networks as similar as possible, and tries to achieve the goal of making the molecular information extracted by different networks complementary to each other. The method is favorable for avoiding the process of obtaining molecular features through an enhanced graph, favorable for improving the accuracy of extracting molecular features, and further improving the accuracy of molecular property prediction, and can be applied to the field of artificial intelligence technology.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, system, device and storage medium for predicting molecular properties. Background Technology

[0002] Molecular property prediction is one application area of ​​molecular representation learning. Researchers use features extracted from molecular representation learning to predict various chemical and biological properties of molecules, such as solubility, toxicity, and protein target affinity. These molecular property predictions can provide crucial information for laboratory research, guiding experimental design, reducing research costs, and accelerating the research process. Among related technologies, graph neural networks are used for molecular property prediction. Specifically, contrastive learning methods in the field of molecular property prediction focus on how to generate effective augmented molecular graphs.

[0003] Contrastive learning methods, such as GraphCL and MolCLR, which focus on generating enhanced graphs, lack an effective method for generating enhanced graphs based on common prior knowledge. Because small molecules have a relatively small number of atoms and bonds, the presence of each atom and bond can significantly affect the properties of the molecule. This makes it difficult to directly apply traditional graph enhancement techniques (such as node masking and edge deletion) to molecular graphs, thus affecting the accuracy of molecular property predictions. Summary of the Invention

[0004] The main objective of this application is to provide a highly accurate method, system, device, and storage medium for predicting molecular properties.

[0005] To achieve the above objectives, one aspect of this application proposes a molecular property prediction method. The method includes: acquiring a first molecule; inputting the first molecule into a dual-graph network model to obtain a first feature of the first molecule; the dual-graph network model being used during pre-training to aggregate the feature representation of a molecule through multiple neural networks based on dual-graph neural network contrastive learning, and to supplement and fuse the molecular information obtained from different neural networks; acquiring several molecular fingerprints and fusing the several molecular fingerprints to obtain a second feature; combining the first feature with the second feature to obtain a third feature; and inputting the third feature into a prediction model to obtain a target result predicting the molecular properties of the first molecule. This application embodiment uses a dual-graph network model for molecular feature representation. During pre-training, the dual-graph neural network model aggregates the feature representation of a molecule through multiple neural networks based on dual-graph neural network contrastive learning, making the molecular representations obtained from different networks as similar as possible. The goal is to make the molecular information extracted by different networks complementary to each other; this helps to alleviate the process of obtaining molecular features through enhanced graphs, improves the accuracy of molecular feature extraction, and thus improves the accuracy of molecular property prediction.

[0006] In some embodiments, the method provided in this application includes a dual-graph network model comprising a graph isomorphic network and a graph attention network, and the dual-graph network model is trained through the following steps:

[0007] The first molecular sample is processed through the graph isomorphic network to determine the first latent vector sample;

[0008] The first molecular sample is processed through a graph attention network to determine the second latent vector sample;

[0009] The similarity between the first latent vector sample and the second latent vector sample is determined, and a contrastive loss function is constructed based on the similarity. The parameters of the dual-graph network model are updated based on the contrastive loss function to obtain the trained dual-graph network model.

[0010] In some embodiments, the method provided in this application further includes:

[0011] The first latent vector sample and the second latent vector sample are identified as positive sample pairs;

[0012] The second molecular sample is processed through the graph isomorphic network to determine the third latent vector sample;

[0013] The second molecular sample is processed through a graph attention network to determine the fourth latent vector sample;

[0014] The first latent vector sample and the third latent vector sample, and the first latent vector sample and the fourth latent vector sample are identified as negative sample pairs;

[0015] The dual-graph network model is updated based on the positive and negative sample pairs.

[0016] In some embodiments, the prediction model provided in this application is trained through the following steps:

[0017] Based on the classification of valid samples in the training set, a first prediction weight and a second prediction weight are determined; the first prediction weight is the weight of the positive class, and the second prediction weight is the weight of the negative class.

[0018] The feature samples in the training set are processed by the prediction model to determine the prediction set;

[0019] The prediction loss function is determined based on the first prediction weight, the second prediction weight, the prediction set, and the labels of the feature samples.

[0020] The parameters of the prediction model are updated according to the prediction loss function to obtain the trained prediction model.

[0021] In some embodiments, the method provided in this application provides a dual-graph network model including a first graph neural network and a second graph neural network. The prediction result of the first graph neural network on the first molecule is first molecule information, and the prediction result of the second graph neural network on the first molecule is second molecule information. The step of combining the first feature and the second feature to obtain a third feature includes:

[0022] Obtain extended connectivity fingerprint;

[0023] The extended connectivity fingerprint is processed by a first multilayer sensor to determine a first molecule weight; the first molecule weight is the weight of the first molecule information during the combination process.

[0024] The extended connectivity fingerprint is processed by a second multilayer perceptron to determine a second molecular weight; the second molecular weight is the weight of the second molecular information during the binding process.

[0025] The extended connectivity fingerprint is processed by a third multilayer perceptron to determine a third molecular weight; the third molecular weight is the weight of the second feature during the combination process; wherein the sum of the first molecular weight, the second molecular weight and the third molecular weight is 1;

[0026] The third feature is determined based on the first molecule information, the second molecule information, the second feature, the first molecule weight, the second molecule weight, and the third molecule weight.

[0027] In some embodiments of the present application, before inputting the first molecule into a dual-graph network model to obtain the first feature of the first molecule, the method further includes:

[0028] The node features and edge features of the first molecule are initialized based on molecular properties; the node features are related to the atomic properties of the first molecule, and the edge features are related to the interatomic bond properties of the first molecule.

[0029] In some embodiments, the method provided in this application, wherein obtaining a plurality of molecular fingerprints and fusing the plurality of molecular fingerprints to obtain a second feature, includes:

[0030] The first fingerprint was identified as the MACCS fingerprint, the second fingerprint as the PubChem fingerprint, and the third fingerprint as the Pharmacophore ErG fingerprint.

[0031] The MACCS fingerprint, the PubChem fingerprint, and the Pharmacophore ErG fingerprint are spliced ​​together to determine the mixed fingerprint;

[0032] The hybrid fingerprint is processed through a fully connected network to obtain a second feature.

[0033] To achieve the above objectives, another aspect of this application proposes a molecular property prediction system, the system comprising:

[0034] The first module is used to obtain the first molecule;

[0035] The second module is used to input the first molecule into the dual-graph network model to obtain the first feature of the first molecule; the dual-graph network model is used to aggregate the feature representation of a molecule through multiple neural networks based on the contrastive learning of the dual-graph neural network during the pre-training process, and to supplement and fuse the molecular information obtained from different neural networks.

[0036] The third module is used to acquire several molecular fingerprints and fuse the several molecular fingerprints to obtain a second feature;

[0037] The fourth module is used to combine the first feature and the second feature to obtain the third feature;

[0038] The fifth module is used to input the third feature into the prediction model to obtain the target result of predicting the molecular properties of the first molecule.

[0039] To achieve the above objectives, another aspect of this application provides a molecular property prediction device, the device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method.

[0040] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0041] The embodiments of this application include at least the following beneficial effects: The method provided in this application includes: obtaining a first molecule; inputting the first molecule into a dual-graph network model to obtain a first feature of the first molecule; the dual-graph network model is used to aggregate the feature representation of a molecule through multiple neural networks based on dual-graph neural network contrastive learning during pre-training, and to supplement and fuse the molecular information obtained from different neural networks; obtaining several molecular fingerprints, and fusing the several molecular fingerprints to obtain a second feature; combining the first feature with the second feature to obtain a third feature; inputting the third feature into a prediction model to obtain a target result for predicting the molecular properties of the first molecule. The embodiments of this application are based on a dual-graph network model for molecular feature representation. During pre-training, the dual-graph neural network model aggregates the feature representation of a molecule through multiple neural networks based on dual-graph neural network contrastive learning, making the molecular representations obtained by different networks as similar as possible. The goal is to make the molecular information extracted by different networks complement each other; this helps to alleviate the process of obtaining molecular features through augmented graphs, and helps to improve the accuracy of molecular feature extraction, thereby improving the accuracy of molecular property prediction. Attached Figure Description

[0042] Figure 1 This is a flowchart of an embodiment of the molecular property prediction method provided in this application;

[0043] Figure 2 This is a schematic diagram of the structure of one embodiment of the dual-graph network model provided in this application;

[0044] Figure 3 This is a schematic diagram of the structure of one embodiment of the prediction process provided in this application;

[0045] Figure 4 This is a schematic diagram of another embodiment of the prediction process provided in this application;

[0046] Figure 5 This is a schematic diagram of the molecular property prediction system provided in the embodiments of this application;

[0047] Figure 6This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0049] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0050] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0052] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0053] Molecular property prediction is one of the applications of molecular representation learning. Researchers use features extracted from molecular representation learning to predict various chemical and biological properties of molecules, such as solubility, toxicity, and protein target affinity. The results of these molecular property predictions can provide crucial information for laboratory research, thereby guiding experimental design, reducing research costs, and accelerating the research process. Molecular property prediction is a key research direction in cheminformatics, with researchers from different fields employing various techniques and methods to predict the physicochemical properties of molecules. Currently, methods for molecular property prediction can be mainly classified into the following categories:

[0054] Wet-lab-based methods: Traditionally, the prediction of molecular properties relies primarily on wet-lab experiments. These methods directly measure the physicochemical properties of molecules under specific conditions. The advantages of this approach are accurate and reliable results, as it is applicable to most molecules regardless of size and complexity, and it provides experimentally validated data, serving as a benchmark for other prediction methods. However, these methods are costly and time-consuming, and may have limitations in predicting the properties of complex molecules, making them unsuitable for predicting properties from large-scale molecular data and exhibiting poor scalability.

[0055] Density Functional Theory (DFT) is a core tool in computational chemistry, aiming to predict molecular properties by utilizing electron density rather than wavefunctions. DFT is built on a solid theoretical foundation and is typically implemented using the Kohn-Sham equations, which simplify complex many-body problems into the motion of non-interacting electrons in an effective potential field composed of an external potential and electron-electron interactions. Handling electron-electron interactions is the main challenge of Kohn-Sham DFT, and an ideal exact solution method is still lacking; the Local Density Approximation (LDA) is commonly used for approximate solutions. Previously, DFT was widely used in calculating molecular physical properties due to its significant cost advantage, but its early accuracy was limited. Since the 1990s, with improvements in approximation methods, DFT has become the preferred method for calculating electronic structure, capable of simulating the electronic structure of various molecules and providing relatively accurate predictions. Theoretically, DFT can simulate the electronic structure of all molecules, thus providing relatively accurate predictions, and can also predict the properties of unknown molecules without experimental data. However, DFT calculations are expensive, especially when dealing with large molecular systems, where high-precision calculations may require a long computation time.

[0056] String-based methods: After converting molecules into strings, they can be modeled using natural language processing, predicting their biochemical properties based on the cheminformatics principle that similar molecules have similar characteristics. Commonly used string representations include: 1) Registry systems: Various registry systems exist, sharing an arbitrary but unique number assigned to new molecules not found in the database, such as CAS Registry Number (RN), PubChem CID, ChemSpider, and ChemBL. This uniqueness facilitates communication, but decoding these numbers requires referencing relevant databases. 2) International Union of Pure and Applied Chemistry (IUPAC) nomenclature: This uses words to represent functional groups, differing significantly from other letter- and number-based representations. Although IUPAC nomenclature may be considered standardized, it is not considered a truly standardized representation due to the frequent use of alternative names, especially reserved names. Furthermore, another problem with IUPAC nomenclature is that, despite the existence of interpreters, computers still face difficulties in parsing it. 3) Simplified Molecular Input Line Entry Specification (SMILES): This technique uses a mapping algorithm from molecular diagrams to text, simplifying chemical structures with a strict syntax, enabling efficient input and storage of molecular information. The conversion from molecular structure to text makes SMILES easy for computers to process for training machine learning models. However, on the one hand, SMILES fails to effectively capture the similarities between molecular structures, leading to significant differences in SMILES strings even with minor structural variations; on the other hand, SMILES strings are not unique, meaning the same molecule may correspond to multiple different SMILES representations. 4) Self-referencing Embedded Strings (SELFIES): This aims to overcome the limitations of traditional SMILES representations. SELFIES is based on a novel encoding rule, possessing 100% robustness and automatically ensuring the chemical rationality of the generated molecular structures. Furthermore, each molecule can be represented in the form of SELFIES. The design of SELFIES not only improves the accuracy and flexibility of molecular structure representation but also significantly enhances the processing power of chemical data and the reliability of molecular generation models.

[0057] Graph Neural Network (GNN) based methods: This innovative technique in deep learning research is specifically designed for processing graph-structured data. GNNs operate directly on the graph structure, including nodes, edges, and their attributes. By aggregating information from neighboring nodes to update the state of each node, they effectively capture complex graph topological relationships and interactions between nodes, thereby learning representations of nodes and the entire graph. With further research, various GNN variants have been developed, such as Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs), which improve the aggregation and propagation of information through different mechanisms. In molecular data analysis, GNN-based methods are particularly suitable for mining complex molecular structural information. Current research innovates GNN models through two main approaches: first, by introducing richer initial features to expand the representation of molecular information; and second, by optimizing the graph neural network architecture to more effectively utilize existing molecular information. These advances provide powerful new tools for accurately simulating molecular properties and predicting unknown molecular properties, demonstrating the broad potential of GNN-based methods in bioinformatics applications.

[0058] Understandably, the rapid growth and complexity of big data in the biosciences present both new challenges and opportunities for research. Particularly in molecular biology and systems biology, labeled data constitutes only a small percentage, while a vast amount of unlabeled biological data contains undiscovered biological knowledge and patterns. Contrastive Learning (CL), as a powerful self-supervised learning framework, enhances information extraction and processing capabilities by guiding models to learn the inherent similarities and differences in data, and has already demonstrated outstanding performance across various fields. For example, in computer vision, CL improves the accuracy of image classification and object recognition by optimizing feature extraction; in natural language processing, CL effectively improves the performance of machine translation and text translation by comparing and analyzing text fragments. Simultaneously, contrastive learning has also achieved breakthroughs in several biological challenges, including protein structure prediction, gene expression analysis, and molecular property prediction. Contrastive learning methods not only improve data utilization but also uncover new biological insights, bringing new innovative vitality to fields such as disease diagnosis, drug discovery, and biomarker identification.

[0059] Current contrastive learning methods in the field of molecular property prediction focus primarily on generating effective augmented molecular graphs. Generally, molecules have relatively few nodes and edges, and contrastive learning (CL) lacks an effective method that preserves chemical properties when generating augmented graphs. Furthermore, CL's method of bringing similar representations closer together is not always suitable for molecular learning; for example, the molecular activity cliff phenomenon occurs, where similar molecules possess completely different properties.

[0060] One type of contrastive learning is widely used for two-dimensional molecular graphs. A key research focus is on generating appropriate augmented molecular graphs to enhance the model's understanding of molecular structure and properties. Against this backdrop, researchers have made various attempts to optimize this learning process. MM-Deacon uses two separate Transformers to encode the SMILES and IUPAC of molecules, thereby capturing information from different aspects of the molecule; GraphCL, based on 2D graphs, uses node dropping, edge perturbation, attribute masking, and subgraph extraction methods to generate enhanced graphs for contrastive learning, attempting to strengthen the model's ability to capture key information in molecular graphs; MolCLR proposes three molecular graph enhancement strategies—atomic masking, bond deletion, and subgraph removal—to improve model learning performance; MoCL incorporates biological knowledge, proposing methods such as isomorphic substitution and adding / removing trivial carbon atoms to generate enhanced graphs of 2D molecular graphs, attempting to preserve the property similarity between the original graph and the enhanced graph as much as possible; ATMOL uses GAT encoding to mask nodes or edges with high attention scores in molecular graphs, aiming to more effectively capture key molecular structural and semantic information; 3DGCL uses RDKit to generate multiple 3D conformations from the same SMILES representation, realizing contrastive learning in 3D space.

[0061] Another type of contrastive learning aims to complement information from different dimensions, enhancing the model's information acquisition capabilities. This approach is based on the premise that molecular representations of different dimensions can provide diverse information to the model, thereby promoting more accurate predictions of molecular properties. DMP compares SMILES with molecular graphs as one of its tasks, finding that the Transformer model performs poorly in distinguishing molecules with more than three connected loops, while GNNs have limitations in distinguishing molecules with long chains. GeomGCL and GraphMVP both treat two-dimensional and three-dimensional molecular graphs as two enhanced views of the molecule. GeomGCL designed a dual-view geometric information transfer network, while GraphMVP utilizes existing, highly expressive graph networks to represent and fuse molecular graphs for generation tasks. CLOOME proposes comparing the fingerprint and image of the same molecule, achieving information comparison by comparing the molecular fingerprint of a molecule with its image views from different angles.

[0062] It is evident that in the specific application scenario of molecular property prediction, contrastive learning in related technologies focuses on data augmentation, which faces certain challenges:

[0063] First, contrastive learning methods, such as GraphCL and MolCLR, which focus on generating enhanced graphs, lack an effective method for generating enhanced graphs based on common prior knowledge. Since small molecules have a relatively small number of atoms and bonds, the presence of each atom and bond can significantly affect the properties of the molecule. This makes it difficult to directly apply traditional graph enhancement techniques (such as node masking and edge deletion) to molecular graphs. Although the strategy of replacing functional groups alleviates this problem to some extent, given that the properties of compounds are usually determined by the interactions of multiple functional groups, and that there may be subtle differences in properties between different functional groups, this method cannot guarantee the absolute invariance of the properties of small molecules.

[0064] Second, while contrastive learning methods, such as GeomGCL and GraphMVP, which utilize different views, can indeed avoid generating augmented graphs, three-dimensional structures usually contain information about two-dimensional structures. In the process of trying to make the molecular representations generated by two-dimensional and three-dimensional graph networks as similar as possible, the two-dimensional structure is insufficient to provide more information about the three-dimensional structure, which may reduce the performance of the three-dimensional graph network in representation learning. This limits the possibility of the model improving its information capture ability through contrastive learning.

[0065] The molecular property prediction method provided in this application relates to the field of artificial intelligence technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the molecular property prediction method, but is not limited to the above forms.

[0066] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0067] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0068] Figure 1 This is an optional flowchart of the molecular property prediction method provided in the embodiments of this application; Figure 1 The method may include, but is not limited to, steps S100 to S400.

[0069] Step S100: Obtain the first molecule;

[0070] Step S200: Input the first molecule into the dual-graph network model to obtain the first feature of the first molecule; the dual-graph network model is used to aggregate the feature representation of a molecule through multiple neural networks based on the contrastive learning of dual-graph neural networks during the pre-training process, and to supplement and fuse the molecular information obtained from different neural networks.

[0071] Step S300: Obtain several molecular fingerprints and fuse them to obtain a second feature;

[0072] Step S400: Combine the first feature with the second feature to obtain the third feature;

[0073] Step S500: Input the third feature into the prediction model to obtain the target result of predicting the molecular properties of the first molecule.

[0074] This application proposes a Dual-Graph Neural Networks Contrastive Learning (DGCL) method, which focuses on changing the way molecular features are extracted, avoiding the step of generating enhanced graphs. During the pre-training phase, to preserve the properties of the molecules themselves, the method of generating enhanced graphs is not chosen; instead, the method of extracting molecular features is considered. The basic idea is to aggregate the feature representations of the same molecule using different networks (i.e., the first and second graph neural networks of the dual-graph network model in this application), making the molecular representations obtained by different networks as similar as possible, with the goal of making the molecular information extracted by different networks complementary. The key to this method is finding two networks with different focuses but relatively strong representational capabilities. In the downstream task stage, since graph networks can mostly only extract the topological structure of molecules, carefully selected hybrid molecular fingerprints are used for supplementation. Molecular fingerprints can represent the functional group properties within a few steps to a certain extent. Selecting molecular fingerprints with more comprehensive information and combining them with molecular embeddings generated by the trained encoder further enhances the ability of molecular representation. Meanwhile, by using extended connectivity fingerprints to weight different molecular representations instead of simply splicing them together, the model's performance on molecular property prediction tasks is effectively improved.

[0075] Therefore, this embodiment avoids the augmented graph generation step relied upon by traditional contrastive learning. Instead, it directly optimizes molecular feature extraction and utilizes two high-performance graph neural networks with different focuses for contrastive learning. The aim is to make the representations of the same molecule as similar as possible under different graph encoders, thereby enabling the utilization of large-scale unlabeled data, improving pre-training efficiency, and enhancing the model's generalization ability. This embodiment further enhances the model's predictive performance by adding carefully selected hybrid molecular fingerprints to the model.

[0076] The method provided in this application includes: obtaining a first molecule; inputting the first molecule into a dual-graph network model to obtain a first feature of the first molecule; the dual-graph network model is used to aggregate the feature representation of a molecule through multiple neural networks based on dual-graph neural network contrastive learning during pre-training, and to supplement and fuse the molecular information obtained from different neural networks; obtaining several molecular fingerprints and fusing the several molecular fingerprints to obtain a second feature; combining the first feature with the second feature to obtain a third feature; and inputting the third feature into a prediction model to obtain a target result for predicting the molecular properties of the first molecule. This application embodiment uses a dual-graph network model for molecular feature representation. During pre-training, the dual-graph neural network model aggregates the feature representation of a molecule through multiple neural networks based on dual-graph neural network contrastive learning, making the molecular representations obtained from different networks as similar as possible. The goal is to make the molecular information extracted by different networks complementary to each other; this helps to alleviate the process of obtaining molecular features through augmented graphs, improves the accuracy of molecular feature extraction, and thus improves the accuracy of molecular property prediction.

[0077] In some embodiments, the method provided in this application provides a dual-graph network model that includes a graph isomorphic network and a graph attention network. The dual-graph network model is trained through the following steps:

[0078] The first molecular sample is processed through a graph isomorphic network to determine the first latent vector sample;

[0079] The first molecular sample is processed through a graph attention network to determine the second latent vector sample;

[0080] The similarity between the first latent vector sample and the second latent vector sample is determined, and a contrastive loss function is constructed based on the similarity. The parameters of the dual-graph network model are then updated based on the contrastive loss function to obtain the trained dual-graph network model.

[0081] In some possible implementations, this application embodiment selects a Graph Isomorphism Network (GIN) and a Graph Attention Network (GAT) that simultaneously integrate node and edge features for comparative learning during the pre-training phase. The GIN can distinguish different graph isomorphism classes and is the strongest isomorphic network among graph networks, with performance comparable to the WL graph isomorphism test. The GAT introduces an attention mechanism to capture local dependencies in the graph structure, focusing on neighboring nodes with closer connections to the target node when aggregating node representations. The goal of the pre-training phase is to bring the molecular representation vectors output by the GIN and GAT as close as possible, allowing the GIN to learn attention-related information from the GAT, while improving the GAT's performance in graph isomorphism. It is understood that this application embodiment does not limit the specific type of graph neural network in the dual-graph network model.

[0082] Regarding the selection of dataset and evaluation metrics in this embodiment: The dataset used for pre-training was downloaded from ZINC and includes SMILES descriptors for 306,347 biologically active small molecule compounds. The open-source tool RDKit was used to convert each SMILES descriptor into a molecular graph, where each node represents an atom and each edge represents a chemical bond. Node and edge features were extracted using molecular properties.

[0083] Nine downstream tasks were selected from the MoleculeNet dataset, including six classification tasks and three regression tasks. The classification tasks are BBBP, SIDER, ClinTox, Tox21, BACE, and HIV, with ROC-AUC as the evaluation metric. The regression tasks are ESOL, FreeSolv, and Lipo, with RMSE as the evaluation metric. The selected tasks cover various molecular properties, including membrane permeability, toxicity, bioactivity, and solubility, making it the most commonly used and widely accepted dataset in the field of molecular property prediction.

[0084] Regarding the DGCL method framework in the embodiments of this application:

[0085] In the pre-training phase, such as Figure 2 As shown, this model transforms the same molecular graph generated by SMILES into latent vectors through a graph isomorphism network and a graph attention network, respectively. and And it is mapped to using a nonlinear projection head. and Then, the similarity between the two projections is calculated. Following the SimCLR design for the loss function, NT-Xent Loss is used as the contrastive loss function. In this model, cosine distance is used to evaluate the similarity between two representations of the same molecule, and the contrastive loss function is shown below:

[0086]

[0087] Among them 1 [k≠i] ∈{0,1} is an indicator function, representing 1 when k≠i and 0 when k=i; τ represents the temperature parameter; N represents the size of a batch. In this model, cosine distance is used to evaluate the similarity between two representations of the same molecule, which can be expressed as:

[0088] In some embodiments, the method provided in this application further includes:

[0089] The first latent vector sample and the second latent vector sample are identified as positive sample pairs.

[0090] The second molecular sample is processed through a graph isomorphic network to determine the third latent vector sample;

[0091] The second molecular sample is processed through a graph attention network to determine the fourth latent vector sample;

[0092] The first latent vector sample and the third latent vector sample, and the first latent vector sample and the fourth latent vector sample are identified as negative sample pairs;

[0093] Update the dual-graph network model based on positive and negative sample pairs.

[0094] Among the possible implementations, DGCL innovates the strategy for constructing positive and negative sample pairs. First, it considers molecular representations obtained from the same molecular graph through different graph networks (i.e., the first latent vector sample and the second latent vector sample) as positive sample pairs. This method effectively utilizes intrinsic molecular structural information, capturing multi-dimensional features of molecules through different graph networks to enhance the model's representational ability. Second, during training, this framework treats the remaining samples in a mini-batch as negative samples through both representations by the encoder, significantly expanding the number and diversity of negative samples. This helps strengthen the model's performance in distinguishing between positive and negative samples, prompting the model to focus more on capturing key information from molecular representations during learning, thus avoiding the risk of overfitting due to a small number of negative samples.

[0095] In some embodiments, the prediction model provided in this application is trained through the following steps:

[0096] Based on the classification of valid samples in the training set, determine the first prediction weight and the second prediction weight; the first prediction weight is the weight of the positive class, and the second prediction weight is the weight of the negative class.

[0097] The feature samples in the training set are processed by the prediction model to determine the prediction set;

[0098] The prediction loss function is determined based on the first prediction weight, the second prediction weight, the prediction set, and the labels of the feature samples.

[0099] The parameters of the prediction model are updated based on the prediction loss function to obtain the trained prediction model.

[0100] In some possible implementations, in the downstream task stage of the DGCL model, this embodiment chooses to freeze the model weights obtained during the pre-training stage. While this saves time and space costs, it may limit the model's ability to further learn molecular representations. To compensate for the performance loss caused by weight freezing, this embodiment carefully designs the classification and regression heads of the MGN-CL model, including batch normalization and activation functions. This approach aims to enhance the model's performance in downstream tasks. Since most classification datasets face the challenge of label imbalance—specifically, the number of samples labeled "0" is far greater than that labeled "1"—this imbalance may cause the model to favor the majority class, thus affecting the prediction accuracy for the minority class. To alleviate this problem, this embodiment employs a weighted loss function, adjusting the loss function by assigning different weights to samples of different classes to compensate for the impact of label imbalance. The formula is as follows:

[0101]

[0102] in, It is the loss function, where N is the total number of samples, and y is the loss function. i It is the actual label of the i-th sample. is the predicted probability of the i-th sample (i.e., the probability that the model predicts it as label 1), w1 is the weight of the positive class, and w0 is the weight of the negative class.

[0103] The calculation of these weights relies on the "Balancing Transformer" method from the DeepChem library. This is a transformer specifically designed to handle class imbalance in classification tasks. It automatically adjusts the sample weights in the dataset, typically giving higher weights to the minority class, allowing the model to focus more on the smaller class during training. This method adjusts the weights primarily based on the inverse of the number of samples in each class, using the following formula:

[0104]

[0105] Among them, w c N is the weight factor for category c. total N is the total number of all valid samples in the dataset. c It is the number of valid samples in category c, and C is the number of all categories. For a binary classification problem, C = 2.

[0106] In some embodiments, the method provided in this application includes a dual-graph network model comprising a first graph neural network and a second graph neural network, wherein the prediction result of the first graph neural network on the first molecule is first molecule information, and the prediction result of the second graph neural network on the first molecule is second molecule information; combining the first feature and the second feature to obtain a third feature; including:

[0107] Obtain extended connectivity fingerprint;

[0108] The extended connectivity fingerprint is processed by a first multilayer sensor to determine the first molecule weight; the first molecule weight is the weight of the first molecule information during the combination process.

[0109] The extended connectivity fingerprint is processed by a second multilayer perceptron to determine the second molecule weight; the second molecule weight is the weight of the second molecule information during the combination process.

[0110] The extended connectivity fingerprint is processed by a third multilayer perceptron to determine the third molecular weight; the third molecular weight is the weight of the second feature during the combination process; wherein, the sum of the first molecular weight, the second molecular weight and the third molecular weight is 1;

[0111] The third feature is determined based on the first molecule information, the second molecule information, the second feature, the first molecule weight, the second molecule weight, and the third molecule weight.

[0112] In some possible implementations, in the improved parts, such as Figure 3 As shown, this embodiment of the application uses ECFP as a basis, and performs weighted processing on the hybrid molecular fingerprint representation processed by the graph encoder and the fully connected layer to replace direct splicing. The final spliced ​​result can be expressed as follows:

[0113] α=MLP1(ECFP), β=MLP2(ECFP), γ=MLP3(ECFP),

[0114] X = CONCAT(αX) GAT ,βX GIN γX MFP ),

[0115] Where α+β+γ=1.

[0116] In some embodiments, before inputting the first molecule into a dual-graph network model to obtain the first feature of the first molecule, the method further includes:

[0117] Initialize the node features and edge features of the first molecule based on molecular properties; the node features are related to the atomic properties of the first molecule, and the edge features are related to the interatomic bond properties of the first molecule.

[0118] In some possible implementations, embodiments of this application propose a molecular characterization step. Before inputting the compound into the graph network model, embodiments of this application initialize the node features and edge features in the molecular graph using molecular attributes. Node features are composed of atomic attributes within the molecule, while edge features are determined by the bond attributes between atoms. In this method, embodiments of this application employ a comprehensive set of features to initialize the molecular graph. These features not only cover the basic chemical properties of atoms and bonds but also include more complex structural information, such as atomic degrees, implicit valences, and bond stereochemistry. Detailed descriptions of these features are shown in Table 1. By setting the node and edge features, this approach aims to enrich the information obtained by the graph network and enhance the model's ability to perceive the details of the compound's molecular structure.

[0119]

[0120] Table 1 Molecular Characterization

[0121] In some embodiments, the method provided in this application, which acquires several molecular fingerprints and fuses them to obtain a second feature, includes:

[0122] The first fingerprint was identified as the MACCS fingerprint, the second fingerprint as the PubChem fingerprint, and the third fingerprint as the Pharmacophore ErG fingerprint.

[0123] The MACCS fingerprint, PubChem fingerprint, and Pharmacophore ErG fingerprint were spliced ​​together to determine the mixed fingerprint.

[0124] The hybrid fingerprint is processed through a fully connected network to obtain a second feature.

[0125] The molecular fingerprint selection process provided in this application embodiment is as follows:

[0126] Molecular fingerprints can be broadly categorized into three types: fingerprints based on substructure keys, fingerprints based on topology or path, and circular fingerprints. This model, referencing this classification, selects three complementary fingerprints and combines them: MACCS (Molecular Access System) fingerprints, PubChem fingerprints, and Pharmacophore ErG (Pharmacophore Extended Reduced Graph) fingerprints. To demonstrate the effectiveness of these three fingerprints, they are compared with the most basic Extended Connectivity Fingerprint (ECFP).

[0127] ECFP: This fingerprint considers the topological environment of atoms in a molecule, encoding this information as a binary vector, typically 1024-dimensional or 2048-dimensional, with different radii specified. It focuses on the local chemical environment and topological information of the molecule.

[0128] MACCS fingerprint: Based on 166 predefined chemical substructures (following SMARTS rules), a binary vector is generated by matching the frequency of occurrence of these substructures. This fingerprint facilitates rapid screening of chemical data and comparison of intermolecular similarity.

[0129] PubChem fingerprints are generated based on the substructures and chemical properties of known compounds in the PubChem database. Similar to MACCS, the method matches the molecular structure with predefined substructure patterns to obtain an 881-bit binary vector. This fingerprint is suitable for structure searches and chemical space exploration based on large databases, reflecting the comprehensive chemical characteristics of compounds.

[0130] Pharmacophore ErG fingerprints are generated based on pharmacophores and their topological relationships within a molecule. By identifying the pharmacophore elements of a molecule and combining them according to their relative positions to form a topological pharmacophore map, the final result is encoded as a 441-dimensional binary vector. This fingerprint focuses on characterizing the key interactions between drug molecules and their biological targets, and is of significant value for drug design and mechanism-of-action research.

[0131] It should be noted that the selection of molecular fingerprints described above is merely an example. Those skilled in the art can set specific selections for mixed molecular fingerprints according to actual needs, and this application does not impose any specific limitations.

[0132] Understandably, since graph networks can mostly only extract the topological structure of molecules, in downstream tasks, this embodiment uses molecular fingerprints to supplement the molecular representation output by the graph network. Molecular fingerprints can represent the functional group properties within a few steps to a certain extent. Selecting a molecular fingerprint with more comprehensive information and combining it with the molecular embedding generated by the trained encoder further enhances the ability of molecular characterization. In this model, three complementary fingerprints (MACCS fingerprint, PubChem fingerprint, and Pharmacophore ErG fingerprint) are concatenated according to the following formula:

[0133] MFP = CONCAT(FP) MACCS FP PubChem FP Pharmacophore ErG ),

[0134] This is called a mixed molecular fingerprint (MFP), and the resulting mixed fingerprint has 1489 dimensions. In downstream tasks, such as... Figure 4 As shown, the hybrid fingerprint is transformed into the same dimension as the graph network output through a fully connected network, and then combined with the molecular representation output by the graph network to jointly predict downstream tasks.

[0135] The following is a detailed description and explanation of the solutions in the embodiments of the present invention, using specific application examples:

[0136] In the pre-training phase, contrastive learning is performed on large-scale unlabeled data to obtain the molecular representation of each molecule and a trained encoder suitable for downstream tasks. The GIN encoder has three layers, each updating nodes by combining node and edge features, and the last layer uses mean pooling for aggregation. The GAT encoder has two layers; the first layer uses ten attention heads, and the second layer uses mean pooling for aggregation. The loss function is optimized using the batch gradient descent algorithm of the Adam optimizer, with a learning rate of 1e-4, a batch size of 128, and 200 pre-training epochs.

[0137] In the transfer learning process for molecular property prediction, the weights of the pre-trained encoder are frozen and treated as a feature extractor. A prediction head is then added, and only the prediction head is optimized during the fine-tuning phase. For classification tasks, weighted binary cross-entropy (BCE) loss is used, while for regression tasks, mean squared error (MSE) loss is used. The Adam optimizer is used in the training phase for all downstream tasks, with a batch size of 256. The evaluation metric for classification tasks is ROC-AUC, and for regression tasks, it is RMSE.

[0138] Meanwhile, in the downstream task phase, an early stopping strategy was further employed to optimize the model training process, aiming to prevent overfitting and improve training efficiency. Specifically, using the performance metrics on the validation set as a benchmark, if the performance on the validation set does not improve within 100 consecutive training epochs, the training process is terminated early. This strategy can, to some extent, ensure that the model stops training when it reaches optimal performance, and then use the model parameters in this state to evaluate the test set.

[0139] Each dataset used for molecular property prediction is split into training, validation, and test datasets in an 8:1:1 ratio according to the molecular scaffold.

[0140] In the field of drug discovery, high-quality labeled data is scarce. Although the development of contrastive learning has facilitated the utilization of large-scale unlabeled data, the application of these methods to molecular data remains limited. To address this issue, a novel molecular representation learning model is proposed, aiming to overcome the limitations of traditional contrastive learning methods in generating molecular enhancement maps. DGCL avoids the step of generating enhancement maps by changing the molecular feature extraction method. During the pre-training stage, it aggregates feature representations of the same molecule using different networks, ensuring that these representations are as similar as possible. Furthermore, in the downstream task stage, DGCL introduces carefully selected hybrid molecular fingerprints to further improve the model's performance. Compared with the eight most competitive self-supervised learning methods, DGCL achieves the best average performance on the classification task, improving by 3.73% compared to the second-best model; and outperforms the second-best method by 0.126 on the regression task Lipo. In addition, we explore the impact of network fusion, molecular fingerprints, feature dimensions, and pre-training on the performance of the DGCL method. Visual analysis shows that the encoder can learn molecular representations beneficial to downstream tasks after pre-training. In the improvement section, we weighted the three feature representations based on ECFP, which further enhances the performance of the DGCL method. Compared with other self-supervised learning models, DGCL achieves the same or better performance with a relatively smaller pre-training dataset and requires less time per epoch, suggesting that DGCL has the potential to be extended to models with stronger representational capabilities in the future.

[0141] This application overcomes the limitations of traditional contrastive learning methods in enhancing graph generation by changing the way molecular features are extracted, thus avoiding the step of generating enhanced graphs. In the downstream task stage, the model's ability to predict molecular properties is further enhanced by fusing hybrid molecular fingerprints. Compared with the eight most competitive self-supervised learning methods currently available, MGN-CL achieves the best average performance on classification tasks, improving upon the suboptimal model by 3.73%; and outperforms the suboptimal method by 0.126 on the regression task Lipo.

[0142] Compared with other self-supervised learning models, MGN-CL achieves the same or better performance as other models with a relatively smaller pre-training dataset and requires relatively less time per training epoch. This suggests that MGN-CL has the potential to be extended to models with stronger representation capabilities in the future.

[0143] Please see Figure 5 This application also provides a molecular property prediction system that can implement the above-described molecular property prediction method. The system includes:

[0144] The first module 510 is used to obtain the first molecule;

[0145] The second module 520 inputs the first molecule into the dual-graph network model to obtain the first feature of the first molecule; the dual-graph network model is used to aggregate the feature representation of a molecule through multiple neural networks based on the contrastive learning of the dual-graph neural network during the pre-training process, and to supplement and fuse the molecular information obtained from different neural networks.

[0146] The third module 530 is used to acquire several molecular fingerprints and fuse them to obtain a second feature.

[0147] The fourth module 540 is used to combine the first feature and the second feature to obtain the third feature;

[0148] The fifth module 550 is used to input the third feature into the prediction model to obtain the target result of predicting the molecular properties of the first molecule.

[0149] In some embodiments, the system provided in this application includes a first graph neural network as a graph isomorphic network and a second graph neural network as a graph attention network. The system also includes a sixth module for training a dual-graph network model, processing the first molecular sample through the graph isomorphic network, and determining the first latent vector sample.

[0150] The first molecular sample is processed through a graph attention network to determine the second latent vector sample;

[0151] The similarity between the first latent vector sample and the second latent vector sample is determined, and a contrastive loss function is constructed based on the similarity. The parameters of the dual-graph network model are then updated based on the contrastive loss function to obtain the trained dual-graph network model.

[0152] In some embodiments, the system provided in this application further includes a seventh module, used to determine that the first latent vector sample and the second latent vector sample are positive sample pairs;

[0153] The second molecular sample is processed through a graph isomorphic network to determine the third latent vector sample;

[0154] The second molecular sample is processed through a graph attention network to determine the fourth latent vector sample;

[0155] The first latent vector sample and the third latent vector sample, and the first latent vector sample and the fourth latent vector sample are identified as negative sample pairs;

[0156] Update the dual-graph network model based on positive and negative sample pairs.

[0157] In some embodiments, the system provided in this application further includes an eighth module for training a prediction model:

[0158] Based on the classification of valid samples in the training set, determine the first prediction weight and the second prediction weight; the first prediction weight is the weight of the positive class, and the second prediction weight is the weight of the negative class.

[0159] The feature samples in the training set are processed by the prediction model to determine the prediction set;

[0160] The prediction loss function is determined based on the first prediction weight, the second prediction weight, the prediction set, and the labels of the feature samples.

[0161] The parameters of the prediction model are updated based on the prediction loss function to obtain the trained prediction model.

[0162] In some embodiments, the system provided in this application further includes a ninth module, used to initialize the node features and edge features of the first molecule through molecular properties; the node features are related to the atomic properties of the first molecule, and the edge features are related to the interatomic bond properties of the first molecule.

[0163] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0164] This application also provides a molecular property prediction device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described molecular property prediction method. This molecular property prediction device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0165] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0166] Please see Figure 6 , Figure 6 The hardware structure of an electronic device / apparatus according to another embodiment is illustrated. The electronic device includes:

[0167] The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0168] The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and called and executed by the processor 901 using the molecular property prediction method of the embodiments of this application.

[0169] The input / output interface 903 is used to implement information input and output;

[0170] The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0171] Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904);

[0172] The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0173] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described molecular property prediction method.

[0174] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0175] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0176] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0177] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0178] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0179] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0180] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0181] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0182] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0183] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0184] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0185] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0186] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for predicting molecular properties, characterized in that, The method includes: Obtain the first molecule; The first molecule is input into the dual-graph network model to obtain the first feature of the first molecule; the dual-graph network model is used to aggregate the feature representation of a molecule through multiple neural networks based on the contrastive learning of dual-graph neural networks during the pre-training process, and to supplement and fuse the molecular information obtained from different neural networks. Several molecular fingerprints are obtained, and the molecular fingerprints are fused to obtain a second feature; The first feature and the second feature are combined to obtain the third feature; The third feature is input into the prediction model to obtain the target result for predicting the molecular properties of the first molecule; The dual-graph network model includes a first graph neural network and a second graph neural network. The prediction result of the first graph neural network for the first molecule is the first molecule information, and the prediction result of the second graph neural network for the first molecule is the second molecule information. The step of combining the first feature and the second feature to obtain a third feature includes: Obtain extended connectivity fingerprint; The extended connectivity fingerprint is processed by a first multilayer sensor to determine a first molecule weight; the first molecule weight is the weight of the first molecule information during the combination process. The extended connectivity fingerprint is processed by a second multilayer perceptron to determine a second molecular weight; the second molecular weight is the weight of the second molecular information during the binding process. The extended connectivity fingerprint is processed by a third multilayer perceptron to determine a third molecular weight; the third molecular weight is the weight of the second feature during the combination process; wherein the sum of the first molecular weight, the second molecular weight and the third molecular weight is 1; The third feature is determined based on the first molecule information, the second molecule information, the second feature, the first molecule weight, the second molecule weight, and the third molecule weight.

2. The method according to claim 1, characterized in that, The dual-graph network model includes a graph isomorphic network and a graph attention network, and is trained through the following steps: The first molecular sample is processed through the graph isomorphic network to determine the first latent vector sample; The first molecular sample is processed through a graph attention network to determine the second latent vector sample; The similarity between the first latent vector sample and the second latent vector sample is determined, and a contrastive loss function is constructed based on the similarity. The parameters of the dual-graph network model are updated based on the contrastive loss function to obtain the trained dual-graph network model.

3. The method according to claim 2, characterized in that, The method further includes: The first latent vector sample and the second latent vector sample are identified as positive sample pairs; The second molecular sample is processed through the graph isomorphic network to determine the third latent vector sample; The second molecular sample is processed through the graph attention network to determine the fourth latent vector sample; The first latent vector sample and the third latent vector sample, and the first latent vector sample and the fourth latent vector sample are identified as negative sample pairs; The dual-graph network model is updated based on the positive and negative sample pairs.

4. The method according to claim 1, characterized in that, The prediction model is trained through the following steps: Based on the classification of valid samples in the training set, a first prediction weight and a second prediction weight are determined; the first prediction weight is the weight of the positive class, and the second prediction weight is the weight of the negative class. The feature samples in the training set are processed by the prediction model to determine the prediction set; The prediction loss function is determined based on the first prediction weight, the second prediction weight, the prediction set, and the labels of the feature samples. The parameters of the prediction model are updated according to the prediction loss function to obtain the trained prediction model.

5. The method according to claim 1, characterized in that, Before inputting the first molecule into the dual-graph network model to obtain the first feature of the first molecule, the method further includes: The node features and edge features of the first molecule are initialized based on molecular properties; the node features are related to the atomic properties of the first molecule, and the edge features are related to the interatomic bond properties of the first molecule.

6. The method according to claim 1, characterized in that, The process of acquiring several molecular fingerprints and fusing them to obtain a second feature includes: The first fingerprint was identified as the MACCS fingerprint, the second fingerprint as the PubChem fingerprint, and the third fingerprint as the Pharmacophore ErG fingerprint. The MACCS fingerprint, the PubChem fingerprint, and the Pharmacophore ErG fingerprint are spliced ​​together to determine the mixed fingerprint; The hybrid fingerprint is processed through a fully connected network to obtain a second feature.

7. A molecular property prediction system, characterized in that, The system includes: The first module is used to obtain the first molecule; The second module is used to input the first molecule into the dual-graph network model to obtain the first feature of the first molecule; the dual-graph network model is used to aggregate the feature representation of a molecule through multiple neural networks based on the contrastive learning of the dual-graph neural network during the pre-training process, and to supplement and fuse the molecular information obtained from different neural networks. The third module is used to acquire several molecular fingerprints and perform fusion processing on the molecular fingerprints to obtain a second feature; The fourth module is used to combine the first feature and the second feature to obtain the third feature; The fifth module is used to input the third feature into the prediction model to obtain the target result of predicting the molecular properties of the first molecule; The dual-graph network model includes a first graph neural network and a second graph neural network. The prediction result of the first graph neural network for the first molecule is the first molecule information, and the prediction result of the second graph neural network for the first molecule is the second molecule information. The fourth module is specifically used for: Obtain extended connectivity fingerprint; The extended connectivity fingerprint is processed by a first multilayer sensor to determine a first molecule weight; the first molecule weight is the weight of the first molecule information during the combination process. The extended connectivity fingerprint is processed by a second multilayer perceptron to determine a second molecular weight; the second molecular weight is the weight of the second molecular information during the binding process. The extended connectivity fingerprint is processed by a third multilayer perceptron to determine a third molecular weight; the third molecular weight is the weight of the second feature during the combination process; wherein the sum of the first molecular weight, the second molecular weight and the third molecular weight is 1; The third feature is determined based on the first molecule information, the second molecule information, the second feature, the first molecule weight, the second molecule weight, and the third molecule weight.

8. A molecular property prediction device, characterized in that, The device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images