Drug target interaction prediction method and system based on neighborhood enhancement graph contrast learning

Abstract

The invention provides a drug-target interaction prediction method and system based on neighborhood enhancement graph comparative learning, and relates to the technical field of computer-aided drug design, in the method, firstly, feature coding is performed on drug and target sequences, sequence information is converted into digital vectors, deep feature extraction is performed through a Transform encoder and a GCN network, and feature extraction is performed through the GCN network; clustering contrast learning is used to better learn features of drugs and targets. Compared with a common deep learning method, the method has the advantages that complex molecular sequences can be extracted, network information is integrated, most importantly, a novel clustering algorithm is used for comparative learning, and finally a bilinear decoder is used for drug-target interaction prediction. The method provided by the invention is helpful for improving the prediction accuracy of drug target interaction, is helpful for relieving the problem of label sparsity, and has certain value for actual disease diagnosis and treatment and subsequent drug research and development.

Description

Technical Field

[0001] The present invention relates to the technical field of computer-aided drug design, and in particular to a drug-target interaction prediction method and system based on neighborhood enhanced graph contrast learning. Background Art

[0002] Drug-target interaction validation is a key step in the drug development process, including drug repositioning, drug discovery, and side effects, and it involves identifying biomolecular targets that bind to specific drugs. This process is essential for understanding the drug's mechanism of action, guiding drug repositioning, and predicting potential side effects. The interactions between drugs and targets can exert therapeutic effects by affecting specific biological pathways, so accurately predicting these interactions can significantly improve the efficiency and safety of drug development.

[0003] However, traditional biological experiments for DTI detection are usually expensive and time-consuming. The current solution is to use 1D-CNN or RNN for feature extraction and use graph neural network combined with contrastive learning for drug-target interaction prediction.

[0004] Although 1D-CNN or RNN can perform feature extraction, they are subject to certain limitations. 1D-CNN focuses on extracting local features from sequence information and has limitations in effectively capturing global features. RNN is able to process the entire sequence through network propagation, but it has the problem of forgetting information over time. Most existing contrastive learning methods use manual graph augmentation to construct contrast conditions without considering how to construct more meaningful contrastive learning tasks for drug-target interaction prediction. Summary of the invention

[0005] In view of this, the present invention provides a drug-target interaction prediction method and system based on neighborhood enhanced graph contrast learning to overcome the limitations of existing drug-target interaction prediction schemes and improve the accuracy of drug-target interaction prediction.

[0006] To this end, the present invention adopts the following technical solutions:

[0007] In one aspect, the present invention provides a method for predicting drug-target interactions based on neighborhood enhanced graph contrast learning, comprising:

[0008] Feature encoding is performed on the drug chemistry text and target sequence respectively, and the sequence information is converted into a text vector;

[0009] The text features of the encoded drug and target text vectors are extracted through the Transformer encoder respectively;

[0010] The text features of drugs and targets extracted by the Transformer encoder are used as node features of the graph convolutional neural network (GCN). The GCN network is used to learn the interaction information between molecules and extract the network features of drugs and targets.

[0011] Using the adjacency relationship in the drug or target similarity network, find drugs or targets with dense neighborhoods, use them as clustering centers to cluster the drugs or targets, and perform neighborhood enhancement graph comparative learning on the network features of drugs and targets based on the clustering results to obtain comparative learning features of drugs and targets;

[0012] Based on the comparative learning features of drugs and targets, the interaction probability between drugs and targets is predicted by a bilinear decoder.

[0013] Furthermore, cluster comparative learning is performed on drugs and targets respectively to obtain comparative learning features of drugs and targets, including:

[0014] The Gaussian kernel similarity matrix of drugs and targets is obtained according to the drug-target interaction matrix;

[0015] Two thresholds were set to convert the similarity matrix of drugs and targets into association matrix respectively;

[0016] A threshold is obtained by the average degree and variance of the association matrix. If the degree of a node is greater than this threshold, it is marked as a strong association point. Each strong association point represents a cluster center.

[0017] The first-order neighbors and higher-order neighbors of the strongly associated points are divided into their clusters. If a node is a neighbor of multiple strongly associated points, the node is divided into the cluster of the strongly associated point with the highest similarity.

[0018] The nodes that have not been divided into clusters are divided into the clusters with the strongest correlation points with the highest similarity;

[0019] Based on the clustering results, contrastive learning is performed, and the strong connection points of each node with its cluster are positive samples, and the strong connection points with other clusters are negative samples.

[0020] Furthermore, the highest order of the high-order neighbors is sixth order.

[0021] Furthermore, for drugs, their SMILES sequences are processed to obtain chemical text information; for targets, their FASTA sequences are used as text information.

[0022] Furthermore, feature encoding is performed on the drug chemistry text and target sequence respectively, and the sequence information is converted into a text vector, including:

[0023] The k-gram method is used to segment the text sequence into phrases of length k, and a dictionary is established to record the order in which the phrases appear. The original words are replaced with the numerical word order of the dictionary, and the numbers are embedded to obtain representations.

[0024] In another aspect, the present invention further provides a drug-target interaction prediction system based on neighborhood enhanced graph contrast learning, comprising:

[0025] The initial feature module is used to encode the features of the drug chemistry text and the target sequence respectively, and convert the sequence information into a text vector;

[0026] The deep feature extraction module is used to extract text features from the encoded drug and target text vectors through the Transformer encoder. The text features of the drugs and targets extracted by the Transformer encoder are used as node features of the graph convolutional neural network (GCN). The GCN network is used to learn the interaction information between molecules and extract the network features of the drugs and targets.

[0027] The clustering comparative learning module is used to use the adjacency relationship in the drug or target similarity network to find drugs or targets with dense neighborhoods, and use them as clustering centers to cluster the drugs or targets. According to the clustering results, the network features of the drugs and targets are respectively subjected to neighborhood enhancement graph comparative learning to obtain comparative learning features of the drugs and targets.

[0028] The drug-target interaction prediction module is used to predict the interaction probability between drugs and targets through a bilinear decoder based on the comparative learning features of drugs and targets.

[0029] Advantages and positive effects of the present invention: The technical solution provided by the present invention first performs feature encoding on drug and target sequences, converts sequence information into digital vectors, performs deep feature extraction through Transformer encoder and GCN network, and uses clustering comparative learning to better learn the characteristics of drugs and targets. Compared with common deep learning methods, the present invention can not only extract complex molecular sequences, but also incorporate network information. The most important thing is that the present invention uses a novel clustering algorithm for comparative learning, and finally the bilinear decoder is used for drug-target interaction prediction. The method of the present invention helps to improve the accuracy of drug-target interaction prediction, helps to alleviate the problem of label sparsity, and has certain value for actual disease diagnosis and treatment and subsequent drug research and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1 It is a flowchart of a drug-target interaction prediction method based on neighborhood enhanced graph contrast learning in an embodiment of the present invention;

[0032] Figure 2 Schematic diagram of neighborhood enhancement graph comparative learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0035] Technical term explanation:

[0036] Rdkit tool: a powerful open source chemical informatics software toolkit, mainly used for processing molecular data and performing chemical information analysis.

[0037] SMILES sequence: The full name is Simplified Molecular Input Line Entry System, a simplified molecular linear input system, a specification that uses ASCII strings to clearly describe molecular structures.

[0038] FASTA sequence: A text format commonly used to store biological sequence information (such as DNA, RNA or protein sequences), mainly used to save sequences in databases.

[0039] k-gram method: A technique commonly used in the field of natural language processing (NLP) and information retrieval to extract a segment of characters starting at a specific position from a text string.

[0040] Transformer: The Transformer model is a deep learning architecture that is widely used in natural language processing (NLP) and other sequence-to-sequence tasks. The core of the Transformer is the self-attention mechanism, which enables the model to consider information at all positions simultaneously when processing the input sequence, thereby capturing global semantic relationships.

[0041] GCN network: Graph convolutional neural network is a deep learning model specially designed to process graph structure data. The core of GCN is that it can perform operations similar to traditional convolutional neural networks on irregular graph structures, thereby effectively extracting features.

[0042] like Figure 1 As shown, a drug-target interaction prediction method based on neighborhood enhanced graph contrast learning in an embodiment of the present invention specifically includes three stages: data preparation, model building, and model testing. Each stage is described in detail below:

[0043] (I) Data preparation:

[0044] Step 11: For drugs, use the Rdkit tool to process their SMILES sequences to obtain chemical text information. For targets, use their FASTA sequences as text information.

[0045] Step 12: Use the k-gram method to segment the text sequence into phrases of length k, and build a dictionary to record the order in which the phrases appear. Use the numerical word order of the dictionary to replace the original words, and embed these numbers to obtain representations.

[0046] (II) Model building:

[0047] Step 21: Text feature extraction.

[0048] Transformer can more fully understand the semantics and structure of text and better capture long-term dependencies in text. Therefore, Transformer is used as a feature extractor to extract complex molecular sequences. A single encoder is sufficient for this task, and more complex models require longer training time due to potential convergence difficulties, without necessarily enhancing information extraction.

[0049] Step 22: Network feature extraction.

[0050] In order to learn the information in the drug-target interaction network, the GCN network is used, and the drug or target representation vector learned in the previous step is used as the node feature of the GCN network to learn the interaction information between molecules.

[0051] Step 23: Cluster contrastive learning.

[0052] In the embodiment of the present invention, drugs and targets are clustered according to a novel clustering algorithm, and the points in the cluster are connected as closely as possible through contrastive learning, while the distance between clusters is as large as possible, so as to learn better drug and target features. Specifically, the adjacency relationship in the drug or target similarity network is used to find drugs or targets with dense neighborhoods, and the drugs or targets are clustered using the drugs or targets as clustering centers. According to the clustering results, the network features of the drugs and targets are respectively subjected to neighborhood enhancement graph contrast learning to obtain the contrastive learning features of the drugs and targets. The clustering algorithm used in the embodiment of the present invention is as follows:

[0053] Step 231: Obtain a Gaussian kernel similarity matrix of drugs and targets based on the drug-target interaction matrix.

[0054] Step 232: Two thresholds are set to convert the similarity matrix of the drug and the target into a correlation matrix respectively.

[0055] Step 233: A threshold is obtained through the average degree and variance of the association matrix. If the degree of a node is greater than the threshold, it is marked as a strong association point. Each strong association point represents a cluster center.

[0056] Step 234: The first-order neighbors and high-order neighbors (up to the sixth order) of the strongly associated point are divided into their clusters. If a node is a neighbor of multiple strongly associated points, the node is divided into the cluster of the strongly associated point with the highest similarity.

[0057] Step 235: The nodes that have not been grouped into clusters are grouped into the cluster where the strongly associated points with the highest similarity are located.

[0058] according to Figure 2 For contrastive learning, each node's strong correlation points with its cluster are positive samples, and its strong correlation points with other clusters are negative samples. Figure 2 The solid lines are positive samples, and the dotted lines are negative samples.

[0059] (III) Model testing:

[0060] Step 31: Use the Rdkit tool to convert the SMILES sequence of the drug into chemical text information, segment it and the FASTA sequence of the target through the k-gram method, replace the original words with numerical word sequences, and embed these numbers to obtain initial features. Perform deep feature extraction through the Transformer encoder and GCN network, then cluster the drugs and targets, obtain better drug and target features through comparative learning, and finally predict the interaction probability between the drug and the target through the bilinear decoder.

[0061] Step 32: Test various parameters and important modules that affect model performance. Observe the impact of the Transformer encoder and contrastive learning in the model on the model test results. Develop model variants such as replacing the Transformer encoder, removing contrastive learning, and replacing contrastive learning, and test the importance of different modules to the overall model results.

[0062] Corresponding to the drug-target interaction prediction method based on neighborhood-enhanced graph contrast learning in the present invention, the present invention also provides a drug-target interaction prediction system based on neighborhood-enhanced graph contrast learning, comprising:

[0063] The initial feature module is used to encode the features of the drug chemistry text and the target sequence respectively, and convert the sequence information into a text vector;

[0064] The deep feature extraction module is used to extract text features from the encoded drug and target text vectors through the Transformer encoder. The text features of the drugs and targets extracted by the Transformer encoder are used as node features of the graph convolutional neural network (GCN). The GCN network is used to learn the interaction information between molecules and extract the network features of the drugs and targets.

[0065] The clustering comparative learning module is used to use the adjacency relationship in the drug or target similarity network to find drugs or targets with dense neighborhoods, and use them as clustering centers to cluster the drugs or targets. According to the clustering results, the network features of the drugs and targets are respectively subjected to neighborhood enhancement graph comparative learning to obtain comparative learning features of the drugs and targets.

[0066] The drug-target interaction prediction module is used to predict the interaction probability between drugs and targets through a bilinear decoder based on the comparative learning features of drugs and targets.

[0067] As for the prediction system in the embodiment of the present invention, since it corresponds to the prediction method in the above embodiment, the description is relatively simple. For the relevant similarities, please refer to the description of the prediction method in the above embodiment, which will not be described in detail here.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A drug-target interaction prediction method based on neighborhood enhanced graph contrast learning, characterized in that: include: Feature encoding is performed on the drug chemistry text and target sequence respectively, and the sequence information is converted into a text vector; The text features of the encoded drug and target text vectors are extracted through the Transformer encoder respectively; The text features of drugs and targets extracted by the Transformer encoder are used as node features of the graph convolutional neural network (GCN). The GCN network is used to learn the interaction information between molecules and extract the network features of drugs and targets. Using the adjacency relationship in the drug or target similarity network, find drugs or targets with dense neighborhoods, use them as clustering centers to cluster the drugs or targets, and perform neighborhood enhancement graph comparative learning on the network features of drugs and targets based on the clustering results to obtain comparative learning features of drugs and targets; Based on the comparative learning features of drugs and targets, the interaction probability between drugs and targets is predicted by a bilinear decoder.

2. The drug-target interaction prediction method based on neighborhood enhanced graph contrast learning according to claim 1, characterized in that: Perform clustering comparative learning on drugs and targets respectively to obtain comparative learning features of drugs and targets, including: The Gaussian kernel similarity matrix of drugs and targets is obtained according to the drug-target interaction matrix; Two thresholds were set to convert the similarity matrix of drugs and targets into association matrix respectively; A threshold is obtained by the average degree and variance of the association matrix. If the degree of a node is greater than this threshold, it is marked as a strong association point. Each strong association point represents a cluster center. The first-order neighbors and higher-order neighbors of the strongly associated points are divided into their clusters. If a node is a neighbor of multiple strongly associated points, the node is divided into the cluster of the strongly associated point with the highest similarity. The nodes that have not been divided into clusters are divided into the clusters with the strongest correlation points with the highest similarity; Based on the clustering results, contrastive learning is performed, and the strong connection points of each node with its cluster are positive samples, and the strong connection points with other clusters are negative samples.

3. The drug-target interaction prediction method based on neighborhood enhanced graph contrast learning according to claim 2, characterized in that: The highest order of the high-order neighbors is six.

4. The drug-target interaction prediction method based on neighborhood enhanced graph contrast learning according to claim 1, characterized in that: For drugs, their SMILES sequences are processed to obtain chemical text information; for targets, their FASTA sequences are used as text information.

5. The drug-target interaction prediction method based on neighborhood enhanced graph contrast learning according to claim 4, characterized in that: Feature encoding is performed on the drug chemistry text and target sequence respectively, and the sequence information is converted into a text vector, including: The k-gram method is used to segment the text sequence into phrases of length k, and a dictionary is established to record the order in which the phrases appear. The original words are replaced with the numerical word order of the dictionary, and the numbers are embedded to obtain representations.

6. A drug-target interaction prediction system based on neighborhood enhanced graph contrast learning, characterized in that: include: The initial feature module is used to encode the features of the drug chemistry text and the target sequence respectively, and convert the sequence information into a text vector; The deep feature extraction module is used to extract text features from the encoded drug and target text vectors through the Transformer encoder. The text features of the drugs and targets extracted by the Transformer encoder are used as node features of the graph convolutional neural network (GCN). The GCN network is used to learn the interaction information between molecules and extract the network features of the drugs and targets. The clustering comparative learning module is used to use the adjacency relationship in the drug or target similarity network to find drugs or targets with dense neighborhoods, and use them as clustering centers to cluster the drugs or targets. According to the clustering results, the network features of the drugs and targets are respectively subjected to neighborhood enhancement graph comparative learning to obtain comparative learning features of the drugs and targets. The drug-target interaction prediction module is used to predict the interaction probability between drugs and targets through a bilinear decoder based on the comparative learning features of drugs and targets.

Images