Zero model-based molecular network generation method and system
The method of constructing molecular networks using a zero model solves the problems of overfitting and underfitting in molecular network generation. The generated samples conform to chemical rules and topological properties, improving the accuracy of molecular property prediction and can be applied to new compound discovery and new drug generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2022-09-06
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies suffer from overfitting and underfitting problems in molecular network generation, and the generated samples lack interpretability and fail to meet the requirements of chemical rules and topological properties.
A zero-model molecular network generation method is adopted to generate new molecular networks by breaking and reconnecting edges. Combined with the rationality judgment of molecular structure, isomers with the same number of nodes and elements as the original molecular network are generated, and important topological properties such as average degree are preserved.
The generated samples conform to chemical rules, preserve the topological properties of the network, improve the accuracy of molecular property prediction, and can be used for new compound discovery and new drug generation.
Smart Images
Figure CN115424678B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of chemistry and machine learning, and mainly to a method and system for generating molecular networks based on a zero model. Background Technology
[0002] With the rapid development of artificial intelligence (AI), more and more fields are combining AI with their applications, such as chemical molecule property prediction and new compound generation. Chemical molecule property prediction and new compound generation are key issues in chemistry, biology, pharmacy, and other fields. In the field of AI, chemical molecule property prediction refers to constructing molecular networks from given molecules, analyzing their topology and characteristics, and predicting their properties, such as toxicity, activity, and protein affinity. This can significantly reduce the investment in property prediction. New compound generation generally refers to establishing conditional generation models to generate molecules that may possess specific properties, helping to advance new drug development, etc.
[0003] A network is a mathematical structure used to study pairwise relationships between objects and entities. It possesses powerful representational capabilities, capable of representing the topological information of molecules in Euclidean space, where nodes represent atoms and edges represent chemical bonds connecting different elements within the molecule. Researchers have shown great interest in tasks such as molecular property prediction and new drug development using molecular networks.
[0004] The application of using networks to represent chemical molecules for molecular property prediction is increasingly prevalent. However, limitations such as high experimental costs and the difficulty in obtaining high-purity molecules hinder the implementation of molecular attribute label matching, thus restricting the amount of samples available for training attribute prediction models. In small-sample training tasks, models trained with a limited number of samples are prone to overfitting to the small sample size and underfitting to the target task. Data augmentation can help models obtain more data samples to address the overfitting problem. Currently, network-based data augmentation methods mainly include feature representation-level augmentation and data augmentation-level augmentation. Feature representation-level data augmentation involves building graph representation learning models to extract optimal features or higher-order semantic information for downstream tasks. Data augmentation-level data augmentation, on the other hand, generates more reliable training samples through appropriate models to help downstream tasks train better models. Current data augmentation methods mostly involve simply changing the network topology to generate samples using neural network models. These methods tend to focus on the model preferences of downstream tasks and lack interpretability. For example, in molecular property prediction tasks, unrestricted data augmentation may generate many samples that do not conform to chemical rules or are impossible to exist. The network null model is a network generative model that gradually approximates the original network from coarse to precise based on its basic characteristics. It can generate new networks while preserving the network features and is interpretable. Summary of the Invention
[0005] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provide a molecular network generation method and system based on a zero model.
[0006] This invention combines the concept of the null model with a method for reconnecting broken edges in a new network, and adds a molecular structure rationality judgment to generate a new network with the same number and elements as the original molecular network, i.e., isomers of the original molecule. Samples generated in this way not only conform to chemical rules but also retain important topological properties in network science, such as average degree. To achieve the above objectives, the technical solution of the molecular network generation method proposed in this invention is as follows:
[0007] The molecular network generation method based on the zero model includes the following steps:
[0008] S1. Read the dataset and obtain molecular network features;
[0009] S2. Construct zero models of molecular networks of order 0, order 1, and order 2 to generate new molecular networks, and calculate the parameters of the new molecular networks.
[0010] S3. Screen new molecular networks and output them.
[0011] Preferably, in step S1: the dataset is read, which contains multiple molecules and can construct multiple molecular networks. Each molecular network can be represented as G = (V, E, C, H, ε), where V represents the set of nodes in the network, E represents the set of edges in the network, C represents the set of formal charges carried by each node in the network, H represents the set of hydrogen atoms carried by each node in the network, and ε represents the set of chemical bonds of the edges in the network.
[0012] Preferably, the specific process for generating the zero-order null model of the original molecular network in step S2 is as follows: The network for generating the zero-order null model includes input parameters such as the original graph edge change ratio α; firstly, a connection e is randomly disconnected from the original network. del ={(v i v j )|ε del}, where ε del For the chemical bonds whose edges have been deleted, randomly select two unconnected nodes v from the original network. k v l Reconnect and set ε del The chemical bonds that act as reconnecting links are assigned to the new molecular network, i.e., e add ={(v k v l )|ε del}, and at the same time modify the characteristics of the node, node v iThe number of hydrogen atoms h i The update formula for ′ is as follows:
[0013] h′ i =h i +ε del (1)
[0014] node v j The update of the number of hydrogen atoms is the same as that of node v. i Node v k The number of hydrogen atoms h k The modified formula for ′ is as follows:
[0015] h′ k =h k -ε del (2)
[0016] If the number of hydrogen atoms is less than 0, then set the number of hydrogen atoms to 0 and update node v. k The formula for updating the formal charge number is:
[0017] c′ k =c k -ε del +h k (3)
[0018] node v l The formal charge number and hydrogen atom number are updated at the same node v. k The above only refers to the modification of one connection in the network graph. This operation process will occur several times during the generation of a new molecular network until the set ratio of the connection α is changed.
[0019] Preferably, the specific process for generating the first-order null model of the original molecular network in step S2 is as follows: The network for generating the first-order null model includes input parameters such as the original graph edge change ratio α; firstly, two edges e1 = {(v i v j )|ε1},e2={(v k v l If v | ε2}, i With v l Not connected, v k With v j If not connected, disconnect e1 and e2, and reconnect e1′={(v i v l )|ε1},e2′={(v k v j If the node v is not selected, the node's characteristics are updated; otherwise, the edges are reselected and the above process is repeated. l The formula for updating the number of hydrogen atoms is as follows:
[0020] h′ l =h l +ε1-ε2 (4)
[0021] If the number of hydrogen atoms is h′ l If the value is less than 0, then set the number of hydrogen atoms h′. l Set the value to 0 and update node v. l The formula for updating the formal charge number is:
[0022] c′ l =c l +h l +ε1-ε2 (5)
[0023] node v j The formula for updating the number of hydrogen atoms is as follows:
[0024] h′ j =h j +ε2-ε1 (6)
[0025] If the number of hydrogen atoms is h′ j If the value is less than 0, then set the number of hydrogen atoms h′. j Set the value to 0 and update node v. j The formula for updating the formal charge number is:
[0026] c j ′=c j +h j +ε2-ε1 (7)
[0027] The above only applies to the modification of one edge in the network graph. Such operations will occur several times during the data generation process until the set edge ratio α is reached or the number of failures exceeds the number of edges in the network graph, at which point the network generation ends.
[0028] Preferably, the specific process for generating the second-order null model of the original molecular network in step S2 is as follows: The network for generating the second-order null model includes input parameters such as the original graph and the edge change ratio α; First, two nodes v with the same degree value are randomly selected from the original network. i v j Then from v i v j Select two nodes v from the neighboring nodes. k v l Therefore, there exists an edge e1 = {(v i v k )|ε1},e2={(v j v l If node v i With v l Unconnected and vk With v j If they are not connected, then the edges are swapped to form e′1={(v j v k )|ε1},e′2={(v i v l If the node v is selected, its characteristics are updated; otherwise, the edges are reselected and the above process is repeated. i The formula for updating the number of hydrogen atoms is:
[0029] h′ i =h i +ε1+ε2 (8)
[0030] If the number of hydrogen atoms is h′ i If the value is less than 0, then set the number of hydrogen atoms h′. i Set the value to 0 and update node v. i The formula for updating the formal charge number is:
[0031] c′ i =c i +h i -ε1+ε2 (9)
[0032] node v j The formula for updating the number of hydrogen atoms is as follows:
[0033] h′ j =h j -ε2+ε1 (10)
[0034] If the number of hydrogen atoms is h′ j If the value is less than 0, then set the number of hydrogen atoms h′. j Set the value to 0 and update node v. j The formula for updating the formal charge number is:
[0035] c′ j =c j +h j -ε2+ε1 (11)
[0036] The above only applies to the modification of one edge in the network graph. Such operations will occur several times during the data generation process until the set edge ratio α is reached and the current network generation ends.
[0037] Preferably, in step S3: after performing network generation operations on all original networks in step S2, the newly generated molecular networks are screened for chemical bonds and atomic valence state rules to improve the probability and rationality of the existence of the final output molecules, so that the generated network carries the corresponding original network's label and outputs the final generated network.
[0038] A system for implementing the zero-model-based molecular network generation method of the present invention includes: a network reading module, a network generation module, a network filtering module, and a network output module;
[0039] The network reading module reads the dataset and constructs a molecular network, and extracts features such as the formal charge and chemical bonds of the atoms in the molecular network for easy use in the network generation module.
[0040] The network generation module generates a new molecular network from the molecular network obtained by the network reading module, and generates a new molecular network with the same average degree and other network characteristics as the original network.
[0041] The network screening module uses basic rules such as chemical bonds and atomic valence states to screen new molecular networks;
[0042] The network output module enables the generated network to carry the corresponding label of the original network and output it.
[0043] The network reading module, network generation module, network filtering module, and network output module are connected in sequence.
[0044] The beneficial effects of this invention are as follows:
[0045] (1) Compared with other network generation models, the molecular network generation method and system based on the zero model of the present invention is more targeted to molecular networks and has better interpretability.
[0046] (2). The molecular network generation method and system based on the zero model of the present invention generates more molecular network samples based on the principle of isomers, and maintains consistency with the original molecular network in important network features such as average degree and degree distribution. The generated network not only has the chemical properties of this type of compound, but also has the topological features of the original network.
[0047] (3). The molecular network generation method and system based on the zero model of the present invention can also be applied to data augmentation and molecular property prediction, which can improve the accuracy of molecular property prediction. It can also be used for new compound discovery and new drug generation. Attached Figure Description
[0048] Figure 1 This is a flowchart of the molecular network generation process based on the zero-order zero model of the present invention;
[0049] Figure 2 This is a flowchart of the molecular network generation process based on the first-order zero model of the present invention;
[0050] Figure 3 This is a flowchart of the molecular network generation process based on the second-order zero model of the present invention;
[0051] Figure 4 This is a system structure diagram of the present invention;
[0052] Figure 5 This is a diagram showing the application effect of the property prediction of the molecular network generation method based on the zero model of the present invention. Detailed Implementation
[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0055] Reference Figure 1 2, 3, 4. This invention provides a molecular attribute prediction method using a zero-model molecular network. It primarily combines the proposed method with data augmentation for molecular attribute prediction, employing the MUTAG dataset. MUTAG is a dataset of chemical molecules and compounds, where atoms act as nodes and chemical bonds as edges to form a network graph. The MUTAG dataset contains 188 nitro compounds, and its labels can determine whether a compound is aromatic or heteroaromatic. In the example training, the dataset is divided into training and testing data. The training data consists of data with known labels, and a small portion of the training dataset can be selected to validate classification accuracy. This example uses a GCN network feature extractor and a softmax classifier, and the existing labeled dataset is divided into an 8:2 ratio for training and testing the model.
[0056] This example basically consists of the following steps:
[0057] Step 1: Import the dataset, read the molecular network, and divide the dataset;
[0058] Step 2: Train an initial classifier using the existing labeled dataset;
[0059] Step 3: Expand the existing labeled dataset using a molecular network generation method and system based on a zero model, and retrain the molecular attribute prediction classifier to obtain an enhanced classifier for classifying the dataset to be tested.
[0060] Step 4: Input the target molecule features into the enhancement model obtained in Step 3 to classify the target molecules and achieve accurate molecular property prediction.
[0061] The specific steps of using the zero-model-based molecular network generation method and system are as follows:
[0062] S1: Import the dataset and construct a molecular network. Each molecular network can be represented as G = (V, E, C, H, ε), where V represents the set of nodes in the network, E represents the set of edges in the network, C represents the set of formal charges carried by each node in the network, H represents the set of hydrogen atoms carried by each node in the network, and ε represents the set of chemical bonds in the edges of the network.
[0063] S2: Perform network generation operation on the molecular network obtained in step S1 using a zero-model-based molecular network generation method;
[0064] S2.1: First, generate the molecular network of the 0th-order null model. The network of the 0th-order null model includes input parameters such as the original graph's edge modification ratio α. First, randomly disconnect an edge e from the original network. del ={(v i v j )|ε del Then, the network connectivity is checked. If the network is not connected, the operation is canceled and a new selection is made; if the network is connected, the process continues, where ε... del For the chemical bonds whose edges have been deleted, randomly select two unconnected nodes v from the original network. k v l Reconnect and set ε del The chemical bonds that act as reconnecting links are assigned to the new molecular network, i.e., e add ={(v k v l )|ε del}, and at the same time modify the characteristics of the node, node v i The number of hydrogen atoms h i The update formula for ′ is as follows:
[0065] h′ i =h i +ε del (12)
[0066] node v j The update of the number of hydrogen atoms is the same as that of node v. i Node v k The number of hydrogen atoms h k The modified formula for ′ is as follows:
[0067] h′ k =h k -ε del (13)
[0068] If the number of hydrogen atoms is less than 0, then set the number of hydrogen atoms to 0 and update node v.k The formula for updating the formal charge number is:
[0069] c′ k =c k -ε del +h k (14)
[0070] node v l The formal charge number and hydrogen atom number are updated at the same node v. k The above only refers to the modification of one connection in the network graph. This operation process will occur several times during the generation of a new molecular network until the set ratio of the connection α is reached.
[0071] S2.2: Next is the generation of the molecular network for the first-order null model. The network for generating the first-order null model includes input parameters such as the original graph's edge changes and the scaling factor α. First, two edges are randomly selected: e1 = {(v i v j )|ε1},e2={(v k v l If v | ε2}, i With v l Not connected, v k With v j If not connected, disconnect e1 and e2, and reconnect e1′={(v i v l )|ε1},e2′={(v k v j Then, the connectivity of the network is checked. If the network is not connected, the operation is canceled and a new selection is made. If the network is connected, the node characteristics are updated. Otherwise, a new selection is made. l The formula for updating the number of hydrogen atoms is as follows:
[0072] h′ l =h l +ε1-ε2 (15)
[0073] If the number of hydrogen atoms is h′ l If the value is less than 0, then set the number of hydrogen atoms h′. l Set the value to 0 and update node v. l The formula for updating the formal charge number is:
[0074] c′ l =c l +h l +ε1-ε2 (16)
[0075] node v j The formula for updating the number of hydrogen atoms is as follows:
[0076] h′j =h j +ε2-ε1 (17)
[0077] If the number of hydrogen atoms is h′ j If the value is less than 0, then set the number of hydrogen atoms h′. j Set the value to 0 and update node v. j The formula for updating the formal charge number is:
[0078] c j ′=c j +h j +ε2-ε1 (18)
[0079] The above only applies to the modification of a single edge in the network graph. This process will occur several times during the data generation process until the set edge ratio α is reached or the number of failures exceeds the number of edges in the network graph, at which point the network generation operation ends.
[0080] S2.3: The network for generating the second-order null model includes input parameters, the original graph, and the edge change ratio α. First, two nodes v with the same degree value are randomly selected from the original network. i v j Then from v i v j Select two nodes v from the neighboring nodes. k v l Therefore, there exists an edge e1 = {(v i v k )|ε1},e2={(v j v l If node v i With v l Unconnected and v k With v j If they are not connected, then the edges are swapped to form e′1={(v j v k )|ε1},e′2={(v i v l Then, the connectivity of the network is checked. If the network is not connected, the operation is canceled and a new selection is made. If the network is connected, the node features are updated. Otherwise, a new selection is made. i The formula for updating the number of hydrogen atoms is:
[0081] h′ i =h i +ε1+ε2 (19) If the number of hydrogen atoms h′ i If the value is less than 0, then set the number of hydrogen atoms h′. i Set the value to 0 and update node v. i The formula for updating the formal charge number is:
[0082] c′ i =c i +h i -ε1+ε2 (20)
[0083] node v j The formula for updating the number of hydrogen atoms is as follows:
[0084] h′ j =h j -ε2+ε1 (21)
[0085] If the number of hydrogen atoms is h′ j If the value is less than 0, then set the number of hydrogen atoms h′. j Set the value to 0 and update node v. j The formula for updating the formal charge number is:
[0086] c′ j =c j +h j -ε2+ε1 (22)
[0087] The above only applies to the modification of one edge in the network graph. This operation process will occur several times during the data generation process until the set edge ratio α is reached and the network generation operation ends.
[0088] S3: After completing the network generation operation for all original networks in step S2, the newly generated molecular networks are screened for chemical bonds and atomic valence state rules to improve the probability and rationality of the existence of molecules in the final output. The generated network carries the label of the corresponding original network and outputs the final generated network. The screened molecular networks are expanded to the original dataset. The expanded dataset is used to retrain the molecular property prediction model to obtain an enhanced prediction model.
[0089] In this application example, the edge modification ratio is 0.2. Figure 5 This is a bar chart comparing the classification accuracy of the dataset after five divisions using the above method. The horizontal axis represents the groups into which the dataset was divided. As the chart shows, compared to the original classifier's classification accuracy, the data generation method presented in this paper effectively improves the attribute prediction accuracy.
[0090] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A molecular network generation method based on a zero model, characterized in that: Includes the following steps: S1. Read the dataset and obtain molecular network features; S2. Construct zero models of molecular networks of order 0, order 1, and order 2 to generate new molecular networks, and calculate the parameters of the new molecular networks. The specific process for generating the zero-order null model of the original molecular network is as follows: The network for generating the zero-order null model includes input parameters, the original graph, edge connections, and scaling ratios. First, randomly disconnect one edge from the original network. ,in For the chemical bonds whose edges have been deleted, randomly select two unconnected nodes from the original network. Reconnect and The chemical bonds that act as reconnecting links are assigned to the new molecular network, i.e. Simultaneously update the node's characteristics, node Number of hydrogen atoms The update formula is as follows: （1） node The update of the number of hydrogen atoms is the same as that of the node. ,node Number of hydrogen atoms The modified formula is as follows: （2） If the number of hydrogen atoms is less than 0, set the number of hydrogen atoms to 0 and update the node. The formula for updating the formal charge number is: （3） node The formal charge number and hydrogen atom number are updated at the same node. The above only describes the modification of a single connection in the network graph. This process will occur several times during the generation of a new molecular network until the desired connection ratio is achieved. ; The specific process for generating the first-order null model of the original molecular network is as follows: The network for generating the first-order null model includes input parameters, the original graph, edge connections, and scaling ratios. First, randomly select two edges to connect. ,like and Not connected and If not connected, then disconnect. and reconnect Simultaneously update the node's characteristics; otherwise, reselect the edge and repeat the above process. The node... The formula for updating the number of hydrogen atoms is as follows: （4） If the number of hydrogen atoms If it is less than 0, then set its number of hydrogen atoms. Set to 0 and update the node. The formula for updating the formal charge number is: （5） node The formula for updating the number of hydrogen atoms is as follows: （6） If the number of hydrogen atoms If it is less than 0, then set its number of hydrogen atoms. Set to 0 and update the node. The formula for updating the formal charge number is: （7） The above only describes the modification of a single edge in the network graph. This process will occur several times during data generation until the desired edge ratio is achieved. The network generation operation will terminate if the number of failures exceeds the number of edges in the network graph. The specific process for generating the second-order null model of the original molecular network is as follows: The network for generating the second-order null model includes input parameters, the original graph, edge connections, and scaling ratios. First, randomly select two nodes with the same degree value from the original network. , and then from Select two nodes from the neighboring nodes. Therefore, there are connected edges. If node and Not connected and and If they are not connected, then swap the edges. If the node is not selected, its characteristics are updated; otherwise, the connection is reselected, and the above process is repeated. The formula for updating the number of hydrogen atoms is: （8） If the number of hydrogen atoms If it is less than 0, then set its number of hydrogen atoms. Set to 0 and update the node. The formula for updating the formal charge number is: （9） node The formula for updating the number of hydrogen atoms is as follows: （10） If the number of hydrogen atoms If it is less than 0, then set its number of hydrogen atoms. Set to 0 and update the node. The formula for updating the formal charge number is: （11） The above only describes the modification of a single edge in the network graph. This process will occur several times during data generation until the desired edge ratio is achieved. The current network generation operation ends here. S3. Screen new molecular networks and output them.
2. The molecular network generation method based on a zero model as described in claim 1, characterized in that: In step S1: The dataset is read. The dataset contains multiple molecules, and multiple molecular networks can be constructed. Each molecular network can be represented as... ,in Represents the set of nodes in a network. Represents the set of edges in the network. This represents the set of formal charges carried by each node in the network. This represents the set of the number of hydrogen atoms carried by each node in the network. This represents the set of chemical bonds connecting the edges in the network.
3. The molecular network generation method based on a zero model as described in claim 1, characterized in that: In step S3: After step S2, when all the original networks have been generated, the newly generated molecular networks are screened for chemical bonds and atomic valence rules to improve the probability and rationality of the existence of the final output molecules, so that the generated network carries the corresponding original network label and outputs the final generated network.
4. A molecular network generation system based on a null model, characterized in that, include: Network reading module, network generation module, network filtering module, network output module; Used to implement the molecular network generation method based on a zero model as described in claim 1; The network reading module reads the dataset and constructs a molecular network, and extracts the formal charge and chemical bond features of the atoms in the molecular network for easy use in the network generation module. The network generation module generates a new molecular network from the molecular network obtained by the network reading module, and generates a new molecular network with the same average degree network characteristics as the original network. The network screening module uses the basic rules of chemical bonds and atomic valence states to screen new molecular networks; The network output module enables the generated network to carry the corresponding label of the original network and output it. The network reading module, network generation module, network filtering module, and network output module are connected in sequence.