A self-supervised molecular classification method based on a deep learning model

Abstract

The application is suitable for the technical field of molecular classification, and provides a self-supervised molecular classification method based on a deep learning model, which comprises the following steps: step 1, an original molecule is processed into a form represented by an adjacency matrix and a feature matrix, and then represented by a graph; step 2, the graph is taken as an input of a graph neural network module, local features of the molecule are obtained, and the properties of the molecule are predicted; and step 3, a standard binary classification cross-entropy loss function is used to adjust a self-supervised learning task between a positive sample pair and a negative sample pair. The method has important significance for screening candidate drugs for specific diseases, especially for a new molecular data set without labels, and can achieve the purpose of rapid classification. Not only can the method replace time-consuming manual labeling of researchers and shorten the time to a short range, but also can control errors to a certain extent, so that real-time performance and accuracy are ensured.

Description

Technical Field

[0001] This invention belongs to the field of molecular classification technology, and in particular relates to a self-supervised molecular classification method based on a deep learning model. Background Technology

[0002] Throughout human development, disease has always been a constant companion, threatening human health. Disease not only affects human well-being but can also impact social stability. Therefore, developing drugs to target specific diseases and mitigate their adverse effects is of paramount importance. The properties of molecules can help us develop drugs for specific diseases and design molecules with corresponding functions. Determining the various properties of molecules is a crucial step in drug discovery. Computer-aided drug design technology has been incorporated into the computational prediction of molecular properties and has become one of the main research directions in bioinformatics.

[0003] Molecular classification is crucial for screening candidate drugs for specific diseases. Traditional machine learning algorithms can classify molecules, but molecules cannot be directly used as input to machine learning models; extensive experiments are required to obtain a series of molecular properties. These hand-generated features rely to some extent on the experience of the experimenters. Traditional feature extraction strategies typically calculate based on the three-dimensional structure of molecules, predicting molecular properties through quantitative structure-activity relationship models—a very time-consuming process. Later, with the rise of deep learning, researchers, inspired by convolutional neural networks (CNNs), proposed graph convolutional neural networks (GCNNs). These networks use chemical bonds as edges and atoms as nodes, naturally treating molecules as a graph, which can then be directly used as input to the GCNN model. GCNNs learn representations of molecular structures directly from molecular datasets, thereby predicting molecular properties. Compared to traditional machine learning algorithms, prediction performance has been significantly improved.

[0004] Most current molecular classification methods based on graph neural networks are supervised, meaning that the input to the graph neural network requires not only the adjacency matrix and features of the molecules, but also precise labels corresponding to the molecules. The graph neural network uses the adjacency matrix and features of the molecules to make reasonable predictions, then compares the prediction results with the labels using a loss function to identify the error in the network's prediction. This error is then backpropagated to the parameters of each layer of the neural network through methods such as gradient descent, readjusting the network's predictions to make them increasingly similar to the true labels. Once the error of the loss function converges to an acceptable range, it can be used to classify new molecular data of the same type.

[0005] Compared to manual annotation by researchers, supervised graph neural network models can classify new molecular data in a shorter time. However, their training still requires a large amount of labeled data from researchers as learning experience to continuously optimize themselves, and acquiring this labeled data is very time-consuming. In existing molecular datasets, the amount of unlabeled data far exceeds the amount of labeled data. Therefore, it is very meaningful to study how to make full use of unlabeled data for training self-supervised deep learning models. Summary of the Invention

[0006] The purpose of this invention is to provide a self-supervised molecular classification method based on a deep learning model, which aims to solve the problems mentioned in the background art.

[0007] This invention is implemented as follows: a self-supervised molecular classification method based on a deep learning model, wherein the deep learning model includes a graph neural network module and a graph contrastive learning module. For unlabeled molecular datasets, the method utilizes the rich information contained in the molecular structure itself to learn the high-dimensional features of the molecules for classification. The molecular classification method includes the following steps:

[0008] Step 1: Process the original molecules into a form represented by adjacency matrices and characteristic matrices, and then represent them using a graph;

[0009] Step 2: Use the graph as input to the graph neural network module to obtain the local features of the molecule, thereby predicting the properties of the molecule;

[0010] For unlabeled molecular data, the feature matrix is ​​augmented by row transformation to obtain a replacement map. Then, the local features of the original map and the replacement map are transformed into global features. The global features and the local features of the original network are called positive sample pairs, and the global features and the local features of the replacement network are called negative sample pairs.

[0011] Step 3: The standard binary cross-entropy loss function is used to adjust the self-supervised learning task between positive and negative sample pairs. By maximizing the mutual information between local and global features, the model can better learn features that are more similar to the original image, thereby completing the molecular classification task.

[0012] A further technical solution is that the prediction process of the molecular classification method includes a training phase and an application phase.

[0013] During the training phase, data augmentation is performed by row operations on the feature matrix of the molecules. The code for this method is based on the PyTorch framework, and the entire model runs on a Windows 10 operating system equipped with an RTX 3080 graphics card.

[0014] In the application phase, directly using a trained, predictive model to predict new molecular data outside the training set can be completed in a relatively short time, which is of great help in screening candidate drugs for specific diseases.

[0015] A further technical solution involves constructing the network architecture of the deep learning model, which includes the following steps:

[0016] Step 1: (X, A) is the graph obtained after processing the original molecule, where X is the feature matrix of the molecule and A is the adjacency matrix of the molecule; the feature matrix X is generated by row operations. Negative samples in contrastive learning are obtained by keeping the adjacency matrix unchanged.

[0017] Step 2: Combine (X,A) and As input to the model, a multi-layer graph neural network is used to obtain high-dimensional features of positive and negative samples. Then, graph pooling is used to transform the high-dimensional features into graph-level features. The properties of molecules are obtained through graph-level features, thereby classifying molecules.

[0018] Step 3: The global features and the local features of the original network are called positive sample pairs, and the global features and the local features of the replacement network are called negative sample pairs. Then, the standard binary cross-entropy loss function is used to adjust the self-supervised learning task between the positive and negative sample pairs. By maximizing the mutual information between local and global features, the model can learn features that are more similar to the original graph.

[0019] Further technical solutions suggest that the generation effect of the network model is mainly constrained by three aspects of the loss function:

[0020]

[0021] label i For the actual properties of the i-th molecule, pred i Let represent the predicted properties of the i-th molecule, and n be the total number of molecules used for verification.

[0022]

[0023] Where V and V′ are the molecular sets of positive and negative samples, respectively, and h u and ...

[0024]

[0025] Here This represents the graph-level features of the surrogate network. In other words, this loss function maximizes the mutual information between the local and global features of the surrogate graph. On one hand, it explores more information from the data itself, enhancing the model's generalization ability. On the other hand, it considers both the original graph network and the data-augmented surrogate graph network, making the entire model more robust. The final total loss function is shown below:

[0026] loss=αloss1+βloss2+γloss3

[0027] Where α, β, and γ are hyperparameters, representing the weights assigned to different tasks.

[0028] This invention provides a self-supervised molecular classification method based on a deep learning model. This method is of great significance for screening candidate drugs for specific diseases, especially for new molecular datasets without labels, enabling rapid classification. It not only replaces time-consuming manual labeling by researchers, reducing the time required, but also controls errors to a certain extent, ensuring both real-time performance and accuracy. Furthermore, in the absence of segmentation labels, this invention outperforms common supervised segmentation models. Attached Figure Description

[0029] Figure 1 A prediction flowchart for a self-supervised molecular classification method based on a deep learning model provided in an embodiment of the present invention;

[0030] Figure 2 This is a network architecture diagram of a self-supervised molecular classification method based on a deep learning model, provided for an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0033] like Figure 1 and 2 As shown, this invention provides a self-supervised molecular classification method based on a deep learning model. The deep learning model includes a graph neural network module and a graph contrastive learning module. For unlabeled molecular datasets, it utilizes the rich information contained within the molecular structure itself to learn the high-dimensional features of the molecules for classification. The molecular classification method includes the following steps:

[0034] Step 1: Process the original molecules into a form represented by adjacency matrices and characteristic matrices, and then represent them using a graph;

[0035] Step 2: Use the graph as input to the graph neural network module to obtain the local features of the molecule, thereby predicting the properties of the molecule;

[0036] For unlabeled molecular data, the feature matrix is ​​augmented by row transformation to obtain a replacement map. Then, the local features of the original map and the replacement map are transformed into global features. The global features and the local features of the original network are called positive sample pairs, and the global features and the local features of the replacement network are called negative sample pairs.

[0037] Step 3: The standard binary cross-entropy loss function is used to adjust the self-supervised learning task between positive and negative sample pairs. By maximizing the mutual information between local and global features, the model can better learn features that are more similar to the original image, thereby completing the molecular classification task.

[0038] In this embodiment of the invention, the prediction process of the molecular classification method includes a training phase and an application phase.

[0039] During the training phase, data augmentation is performed by row operations on the feature matrix of the molecules. The code for this method is based on the PyTorch framework, and the entire model runs on a Windows 10 operating system equipped with an RTX 3080 graphics card.

[0040] In the application phase, directly using a trained, predictive model to predict new molecular data outside the training set can be completed in a relatively short time, which is of great help in screening candidate drugs for specific diseases.

[0041] like Figure 2 As shown, in a preferred embodiment of the present invention, in the network architecture of the deep learning model, (X, A) is the graph obtained after processing the original molecule, where X is the feature matrix of the molecule and A is the adjacency matrix of the molecule. The feature matrix X is generated by row operations. Negative samples in contrastive learning are obtained by keeping the adjacency matrix unchanged. Then (X, A) and As input to the model, a multi-layered graph neural network is used to obtain high-dimensional features for positive and negative samples. Then, graph pooling is used to transform these high-dimensional features into graph-level features. The properties of molecules are then obtained through these graph-level features, enabling molecule classification. Global features and local features of the original network are termed positive sample pairs, while global features and local features of the replacement network are termed negative sample pairs. A standard binary cross-entropy loss function is then applied between positive and negative sample pairs to adjust the self-supervised learning task. By maximizing the mutual information between local and global features, the model can learn features more similar to the original graph. Thus, the graph neural network model incorporating contrastive learning is complete.

[0042] In a preferred embodiment of the present invention, the generation effect of the network model is mainly constrained by three aspects of the loss function:

[0043]

[0044] label i For the actual properties of the i-th molecule, pred i Let represent the predicted properties of the i-th molecule, and n be the total number of molecules used for verification.

[0045]

[0046] Where V and V′ are the molecular sets of positive and negative samples, respectively, and h v and ...

[0047]

[0048] Here This represents the graph-level features of the surrogate network. In other words, this loss function maximizes the mutual information between the local and global features of the surrogate graph. On one hand, it explores more information from the data itself, enhancing the model's generalization ability. On the other hand, it considers both the original graph network and the data-augmented surrogate graph network, making the entire model more robust. The final total loss function is shown below:

[0049] loss=αloss1+βloss2+γloss3

[0050] Where α, β, and γ are hyperparameters, representing the weights assigned to different tasks.

[0051] Example: BBBP molecular dataset prediction

[0052] The BBBP dataset comes from a recent study on modeling and predicting the blood-brain barrier permeability. This dataset includes binary labels for the permeability characteristics of over 2000 compounds. The dataset was partitioned using the Scaffold method, dividing it into training, validation, and test sets in an 8:1:1 ratio. Accuracy, precision, recall, F1-score, and AUC were used as evaluation metrics. The model was trained using the above method, stopping training at the 200th epoch. The prediction performance of the final model is compared with that of common supervised graph neural network models, as shown in Table 1.

[0053] Table 1: Prediction results of the BBBP dataset under various models

[0054] Model accuracy Accuracy Recall rate F1 score AUC GCN 0.826 0.911 0.860 0.885 0.865 GAT 0.812 0.928 0.824 0.873 0.873 SAGEConv 0.823 0.893 0.804 0.846 0.846 RGGC 0.836 0.902 0.886 0.894 0.870 TrimNet 0.814 0.917 0.837 0.875 0.850 Ours 0.899 0.937 0.931 0.934 0.916

[0055] Example: SIDER molecular dataset prediction

[0056] The SIDER dataset is a database of marketed drugs and adverse drug reactions, containing 1427 drugs divided into two categories. The dataset was randomly partitioned into training, validation, and test sets in an 8:1:1 ratio. The model described above was trained, stopping at the 200th epoch. The prediction performance of the final model is compared with that of common supervised graph neural network models, as shown in Table 2.

[0057] Table 2: Prediction results of the SIDER dataset under various models

[0058] Model accuracy Accuracy Recall rate F1 score AUC GCN 0.637 0.622 0.653 0.637 0.633 GAT 0.664 0.639 0.659 0.649 0.670 SAGEConv 0.639 0.625 0.593 0.609 0.642 RGGC 0.655 0.635 0.639 0.637 0.662 TrimNet 0.637 0.627 0.696 0.660 0.657 Ours 0.682 0.641 0.703 0.671 0.688

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-supervised molecular classification method based on a deep learning model, characterized in that, The deep learning model includes a graph neural network module and a graph contrastive learning module. For unlabeled molecular datasets, it uses the rich information contained in the molecular structure itself to learn the high-dimensional features of molecules for classification. The molecular classification method includes the following steps: Step 1: Process the original molecules into a form represented by adjacency matrices and characteristic matrices, and then represent them using a graph; Step 2: Use the graph as input to the graph neural network module to obtain the local features of the molecule, thereby predicting the properties of the molecule; For unlabeled molecular data, the feature matrix is ​​augmented by row transformation to obtain a replacement map. Then, the local features of the original map and the replacement map are transformed into global features. The global features and the local features of the original network are called positive sample pairs, and the global features and the local features of the replacement network are called negative sample pairs. Step 3: The standard binary cross-entropy loss function is used to adjust the self-supervised learning task between positive and negative sample pairs. By maximizing the mutual information between local and global features, the model learns features that are more similar to the original image, so as to complete the molecular classification task. The deep learning model is a graph neural network model that incorporates the idea of ​​contrastive learning. The method for building its network architecture includes the following steps: Step 1: The image shows the result of processing the original molecule. The characteristic matrix of the molecule, Let be the adjacency matrix of the molecule; for the characteristic matrix Generate using row transformations Negative samples in contrastive learning are obtained by keeping the adjacency matrix unchanged. ; Step Two: and As input to the model, a multi-layer graph neural network is used to obtain high-dimensional features of positive and negative samples. Then, graph pooling is used to transform the high-dimensional features into graph-level features. The properties of molecules are obtained through graph-level features, thereby classifying molecules. Step 3: The global features and the local features of the original network are called positive sample pairs, and the global features and the local features of the replacement network are called negative sample pairs. Then, the standard binary cross-entropy loss function is used to adjust the self-supervised learning task between the positive and negative sample pairs. By maximizing the mutual information between local and global features, the model can learn features that are more similar to the original graph. The generation performance of the graph neural network model is mainly constrained by three aspects of the loss function: 1） ； in For the first The actual properties of each molecule For the first Predicted properties of individual molecules This represents the total number of molecules used for verification. 2） ； in and These are the molecular sets of positive and negative samples, respectively. and These are the high-dimensional features of positive and negative samples, respectively. The graph-level features representing positive samples; this loss function maximizes the mutual information between the local and global features of the original graph, and uses the standard binary cross-entropy loss function between positive and negative sample pairs to adjust the self-supervised learning task, so that the model learns features that are more similar to positive samples. 3） ； Here Representing the graph-level features of the surrogate network, this loss function maximizes the mutual information between the local and global features of the surrogate graph, enhances the model's generalization ability by exploring more information from the data itself, and makes the entire model more robust by considering both the original graph network and the data-augmented surrogate graph network. The final total loss function is shown below: ； wherein is a hyperparameter representing the weight of different tasks. 2.The self-supervised molecular classification method based on a deep learning model according to claim 1, wherein, The prediction process of the molecular classification method includes a training phase and an application phase: During the training phase, data augmentation is performed by row transformation of the feature matrix of the molecules. In the application phase, the trained and predictive model is directly used to predict new molecular data outside the training set.

Images