Drug-disease association prediction method and system based on cross-view contrastive learning

Abstract

The present invention provides a drug-disease association prediction method and system based on cross-view contrast learning, which belongs to the field of drug discovery technology in bioinformatics. Construct semantic views and interactive views; construct and train a drug-disease association probability prediction model: integrate semantic views and interactive views to obtain fused drug vector representations and disease vector representations as fused views; use the fused drug and disease vector representations to output drug-disease association probabilities; perform cross-view contrast learning on the fused view and the semantic view; perform cross-view contrast learning on the fused view and the interactive view; and transfer implicit knowledge from contrast learning to drug-disease association probability prediction through underlying shared parameters in a multi-task learning manner. Avoid the problem of discarding necessary association information between drugs and diseases caused by using random perturbations to enhance contrast views, thereby improving the accuracy of drug-disease association prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of drug discovery in bioinformatics, and in particular to a drug-disease association prediction method and system based on cross-view contrast learning. Background Art

[0002] The traditional drug development process involves many links such as target identification, pharmacology and toxicology testing, efficacy evaluation, and clinical verification, which requires a lot of time and money, and the success rate of development is low. According to statistics, it costs an average of nearly 2.6 billion US dollars to develop a new drug approved by the FDA, which takes about 10-12 years, but the success rate of new drug development is less than 10%. Therefore, how to effectively shorten the drug development cycle, increase the success rate of development, and thus reduce the economic burden on patients is of great significance to the construction of a healthy China. In order to solve the above problems, the drug repositioning technology of "old drugs for new uses" has attracted the attention of researchers. Drug repositioning refers to a method of discovering new applicable diseases or new uses from approved clinical drugs. With the increasing abundance of multi-omics data and biomedical knowledge bases, computational drug repositioning methods provide a new idea for drug reuse. Drug repositioning can generally be classified into two categories: "disease-centric" and "target-centric", which are used to predict drug-disease associations and drug-target interactions, respectively. The present invention focuses on the former, that is, predicting the probability that a drug can be associated with a given disease.

[0003] Since the relationship between drugs and diseases can be naturally modeled as a network structure, the existing technology proposes to construct a heterogeneous biological network, in which the nodes are drugs, diseases, proteins, etc., and the edges represent the interaction between nodes. The complex topological structure of the network is modeled using graph mining methods such as graph neural networks and graph contrastive learning, and the drug and disease nodes are projected into a low-dimensional vector space. Finally, the unknown drug-disease relationship is predicted by a neural network classifier. However, there are two defects in the existing technology that make its prediction accuracy insufficient: the number of annotated drug-disease association positive samples is limited, resulting in label sparsity problems, which makes the model unable to fully capture the inherent complex semantic associations between drugs and diseases; the contrast view enhancement strategy based on random perturbations causes the necessary association information between drugs and diseases in the graph topology to be discarded.

[0004] Therefore, a drug-disease association prediction method is needed that has sparse link labels and fully captures high-order semantics and interaction information to improve prediction accuracy. Summary of the invention

[0005] In view of this, the present invention provides a drug-disease association prediction method and system based on cross-view contrastive learning, which solves the problems of label sparsity and lack of necessary association structure in enhanced views faced by existing methods by defining effective cross-view contrastive learning auxiliary tasks.

[0006] The present invention provides the following technical solutions:

[0007] Drug-disease association prediction method based on cross-view contrastive learning, including:

[0008] Constructing a drug similarity network, a disease similarity network and a heterogeneous biological interaction network; the drug similarity network and the disease similarity network are used as semantic views, and the heterogeneous biological interaction network is used as an interaction view;

[0009] Construct and train a drug-disease association probability prediction model; the model includes:

[0010] Integrate the semantic view and the interaction view to obtain the fused drug vector representation and disease vector representation;

[0011] The fused drug vector representation and disease vector representation are used as a fused view;

[0012] The fused drug and disease vector representation is used to output the drug-disease association probability through the predictor;

[0013] The model training includes:

[0014] Perform cross-view comparative learning on fusion view and semantic view;

[0015] Conduct cross-view comparative learning of fused views and interactive views;

[0016] Through a multi-task learning approach, implicit knowledge is transferred from contrastive learning to drug-disease association probability prediction via underlying shared parameters.

[0017] Furthermore, the construction of the drug similarity network, the disease similarity network and the heterogeneous biological interaction network includes:

[0018] Based on the drug fingerprint similarity and Gaussian interaction spectrum kernel similarity, the similarity between any drugs is calculated; for each drug, the top-K nearest neighbor drug nodes are screened to construct a drug similarity network

[0019] Based on the similarity of disease phenotypes and Gaussian interaction kernel similarity, the similarity between any diseases is calculated; for each disease, the top-K nearest neighbor disease nodes are selected to construct a disease similarity network

[0020] Construct heterogeneous biological interaction networks by integrating drug-disease association networks, drug-protein, and disease-protein interaction networks

[0021] Further, integrating the semantic view and the interaction view to obtain the fused drug vector representation and disease vector representation includes:

[0022] Using the Basic Graph Transformer model to generate the drug and disease vector representations under the semantic view;

[0023] Using the Heterogeneous Graph Transformer model to generate the drug and disease vector representations under the interaction view;

[0024] Performing average pooling operations on all layers to generate the fused drug and disease vector representations:

[0025]

[0026]

[0027] Among them, represents the vector representation of drug u after fusion at the l-th layer, represents the vector representation of disease v after fusion at the l-th layer; the vector fusion is completed through the multi-head self-attention mechanism.

[0028] Further, the loss function for cross-view contrastive learning between the fused view and the semantic view is:

[0029]

[0030] Among them, τ represents the temperature coefficient; the vector representation of node i on is obtained through average pooling:

[0031] Further, the loss function for cross-view contrastive learning between the fused view and the interaction view is:

[0032]

[0033] Among them, τ represents the temperature coefficient; the vector representation of node i on is obtained through average pooling:

[0034] Further, using the fused drug and disease vector representations, the drug-disease association probability is output through a predictor:

[0035]

[0036] Among them, MLP is a multi-layer perceptron, || represents concatenation, and y u，v is the drug-disease association probability.

[0037] Furthermore, the prediction of drug-disease association probability is optimized by the cross entropy loss function:

[0038]

[0039] Among them, S is a set of positive and negative training samples; is the true label.

[0040] Furthermore, the implicit knowledge is transferred from contrastive learning to drug-disease association probability prediction through underlying shared parameters by multi-task learning:

[0041]

[0042] where α 1 With α 2 is the task weight hyperparameter.

[0043] The drug-disease association prediction system based on cross-view contrastive learning includes:

[0044] A view construction module, which constructs a drug similarity network, a disease similarity network and a heterogeneous biological interaction network; the drug similarity network and the disease similarity network are used as semantic views, and the heterogeneous biological interaction network is used as an interaction view;

[0045] A fusion view representation module integrates the semantic view and the interactive view to obtain a fused drug vector representation and a disease vector representation; the fused drug vector representation and the disease vector representation are used as a fusion view;

[0046] The drug-disease association probability prediction module uses the fused drug and disease vector representation to output the drug-disease association probability through the predictor;

[0047] The model training module performs cross-view comparative learning between fusion view and semantic view; performs cross-view comparative learning between fusion view and interaction view; and transfers implicit knowledge from comparative learning to drug-disease association probability prediction through underlying shared parameters through multi-task learning.

[0048] Advantages and positive effects of the present invention:

[0049] The present invention introduces two self-supervised contrastive learning auxiliary tasks, which are jointly optimized with the main task of drug-disease association prediction through multi-task learning, so that implicit knowledge is transferred from the auxiliary tasks to the main task through underlying shared parameters, thereby alleviating the impact caused by label sparsity. The semantic view and the interactive view are integrated through a multi-head self-attention mechanism to generate a fused view. The fused view is further compared with the semantic view and the interactive view respectively, thereby avoiding the problem of discarding the necessary association information between drugs and diseases caused by the use of random perturbations for contrast view enhancement, thereby improving the accuracy of drug-disease association prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] 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.

[0051] Figure 1 It is a schematic diagram of a flow chart of an embodiment of the present invention;

[0052] Figure 2 Schematic diagram of drug-disease similarity network and heterogeneous biological interaction network in an embodiment of the present invention;

[0053] Figure 3 This is a diagram of the overall architecture of the model in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] 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.

[0055] 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.

[0056] The present invention proposes a drug-disease association prediction method based on cross-view contrast learning, the concept of which is:

[0057] From the perspective of graph structure, the semantic view and the interactive view can be regarded as a homogeneous information network and a heterogeneous information network, respectively. Compared with homogeneous networks, heterogeneous information networks contain multiple types of nodes and edges, and therefore contain richer structural and semantic information. In order to fully encode and retain the key information in the two views, the present invention uses two GraphTransformer models to encode them separately. Heterogeneous Graph Transformer can better capture the heterogeneous attributes in the interactive view.

[0058] In order to integrate the implicit high-order information (drug-drug, disease-disease) and explicit interaction relationship (drug-disease) in the original association network, a fusion view is introduced. In addition, in order to make the generated contrast-enhanced view fully capture the high-order interaction semantics, the semantic view and the interaction view are not directly compared. Instead, the constructed fusion view is compared with the semantic view and the interaction view respectively, so as to design two cross-view contrast learning auxiliary tasks. The above auxiliary tasks enable the model to effectively learn high-quality drug and disease vector representations by supplementing additional self-supervised training signals.

[0059] Combination Figure 1 The flow chart shown further illustrates the steps of the invention:

[0060] S1. Calculate the similarity between any drugs based on drug fingerprint similarity and Gaussian interaction profile kernel similarity. And for each drug, select the top-K nearest neighbor drug nodes to construct a drug similarity network. Similarly, the similarity between any diseases is calculated based on the disease phenotype similarity and Gaussian interaction kernel similarity, and a disease similarity network is constructed based on the top-K nearest neighbor disease nodes.

[0061] Based on the original drug-disease association network, considering the important role of protein molecules in life, drug-protein and disease-protein interactions were introduced to construct a heterogeneous biological interaction network. The network contains three types of nodes (drug, disease, protein) and three types of edges (drug-disease, drug-protein, disease-protein).

[0062] S2, the drug and disease similarity network constructed by S1 ( and ) as a semantic view, and use the BasicGraph Transformer model to learn the drug and disease vector representation under the semantic view.

[0063] In this implementation, the drug vector make It represents the vector representation of drug u at layer l, and d represents the vector dimension. The multi-head attention mechanism is used to evaluate the importance of each neighbor node.

[0064] Specifically, for drug u and its neighbor node j, drug u is converted into a query vector Map neighbor node j to a key vector and a vector of values Where h represents the hth attention head, and the total number of attention heads is H. as well as is a trainable parameter. The scaled dot product attention corresponding to the hth attention head is used to calculate the degree of association between the query vector and the key vector:

[0065]

[0066] in, are all neighbor nodes of drug u; the message vector from neighbor node j to drug u is defined as:

[0067]

[0068] Here, || represents a concatenation operation. The message propagation mechanism is used to aggregate information from all neighboring nodes and add residual connections to update the vector representation of drug u at the (l+1)th layer:

[0069]

[0070] Among them, σ(.) represents nonlinear activation, is a trainable parameter. After calculating the L layer, the vector representation of each layer corresponding to each drug is retained and recorded as in Indicates the total number of drugs.

[0071] Similarly, another Basic Graph Transformer model is used to train the disease similarity network Modeling is performed to generate the corresponding layer representations for each disease, recorded as in, Represents the total number of diseases.

[0072] S3: The heterogeneous biological interaction network constructed by S1 As an interactive view, the Heterogeneous GraphTransformer model is used to generate drug and disease vector representations under the interactive view. In this embodiment, Indicated in The vector representation of the lth layer of node i is given by: given an edge e from source node j to target node i, the corresponding node type and edge type are φ(j), φ(i) and Project the vector representations corresponding to the target node i and the source node j to the multi-head query vector, key vector, and value vector respectively, that is, Here h also represents the hth attention head, and the total number of attention heads is set to H. as well as is a parameter matrix specific to the node type. Calculate the h-th attention head value for each edge e:

[0073]

[0074] in, For the target node i The neighbor nodes on is a parameter specific to the edge type. Next, the message vector from source node j along edge e to target node i is defined as:

[0075]

[0076] in, This is also a parameter specific to the edge type. The update formula for the vector representation of node i at the (l+1)th layer is:

[0077]

[0078] in, are node type specific parameters. After L-layer calculation, The drug and disease vector representations obtained in each layer are respectively denoted as

[0079] S4: Take the drug and disease vector representations under the two views obtained in S2 and S3 as input, use the multi-head self-attention mechanism to fuse the vector representations obtained from the semantic view and the interaction view layer by layer, and generate the fused drug and disease vector representations.

[0080] In this embodiment, taking drug u as an example, the vector representations obtained from the semantic view and the interaction view at the lth layer are respectively as well as The above two vectors are fused using the multi-head self-attention mechanism, and the vector representation of drug u after fusion at the lth layer is recorded as Similarly, for disease v, its vector representation after fusion at layer l is recorded as Perform average pooling operations on all layers to generate the final fused drug and disease vector representations:

[0081]

[0082]

[0083] S5. Define two self-supervised contrastive learning auxiliary tasks to compare the fusion view with the semantic view and the interaction view respectively, where the positive sample pairs are the same drug (or disease) nodes in different views, and the negative sample pairs are different drug (or disease) nodes in different views. Accordingly, define two cross-view contrastive learning loss functions to bring the distance between positive sample pairs closer in the representation space and push the distance between negative sample pairs farther.

[0084] In this embodiment, for the semantic view ( and ) and interactive views The node vector representation is also obtained by average pooling. Therefore, node i is as well as The vector representations on The fusion view is compared with the semantic view. To avoid false negatives, the top-K nearest neighbor nodes corresponding to node i are removed from the negative samples. The cross-view comparison learning loss function between the fusion view and the semantic view is defined as:

[0085]

[0086] where τ is the temperature coefficient. Similarly, the cross-view contrastive learning loss function between the fusion view and the interaction view is defined as:

[0087]

[0088] S6: Given a drug-disease entity pair to be predicted, concatenate the fused drug and disease vector representations obtained in S4 and input them into the association predictor to predict the final association probability. The model is optimized using the cross-entropy loss function. This loss function is jointly optimized with the two cross-view contrastive learning loss functions defined in S5 in a multi-task learning manner to ensure that implicit knowledge is transferred from the contrastive learning auxiliary task to the main task of drug-disease association prediction through the underlying shared parameters.

[0089] In this example, given a drug-disease pair (u, v), the corresponding fused drug is represented as Disease They are connected in series and input into a multilayer perceptron to calculate the probability that drug u can be associated with disease v:

[0090]

[0091] Where MLP is a multi-layer perceptron. The drug-disease association prediction is optimized by the cross entropy loss function:

[0092]

[0093] Where S is a set of positive and negative training samples, is the true label. The above cross entropy loss function is jointly optimized with the two cross-view contrastive learning loss functions defined in S5) in a multi-task learning manner, and the overall loss function is:

[0094]

[0095] where α 1 With α 2 is the task weight hyperparameter.

[0096] The present invention also provides a drug-disease association prediction system based on cross-view contrast learning, comprising:

[0097] A view construction module, which constructs a drug similarity network, a disease similarity network and a heterogeneous biological interaction network; the drug similarity network and the disease similarity network are used as semantic views, and the heterogeneous biological interaction network is used as an interaction view;

[0098] A fusion view representation module integrates the semantic view and the interaction view layer by layer to obtain a fused drug vector representation and a disease vector representation; the fused drug vector representation and the disease vector representation are used as a fusion view;

[0099] The drug-disease association probability prediction module uses the fused drug and disease vector representation to output the drug-disease association probability through the predictor;

[0100] The model training module performs cross-view comparative learning between the fusion view and the semantic view; performs cross-view comparative learning between the fusion view and the interaction view to provide additional training signals for the main task of drug-disease association prediction; and transfers implicit knowledge from comparative learning to drug-disease association probability prediction through underlying shared parameters in a multi-task learning manner.

[0101] Among them, the fusion view representation module includes: a view semantic view encoding module and an interaction view encoding module; the semantic view encoding module learns the node vector representation from the drug-disease similarity network; the interaction view encoding module is used to capture different types of interaction patterns in heterogeneous biological networks (including drug-disease associations, drug-protein and disease-protein interactions).

[0102] The present invention has a cross-view contrast learning module, which is jointly optimized with the main task of drug-disease association prediction through multi-task learning, so that implicit knowledge is transferred from the auxiliary task to the main task through the underlying shared parameters. The multi-head self-attention mechanism is used to integrate the semantic view and the interaction view to generate a fused view, and the fused view is further compared with the semantic view and the interaction view respectively, so as to fully capture high-order semantics and interaction information, and further improve the accuracy of prediction.

[0103] 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-disease association prediction method based on cross-view contrastive learning, characterized in that: include: Construct drug similarity networks, disease similarity networks and heterogeneous biological interaction networks; The drug similarity network and the disease similarity network are used as semantic views, and the heterogeneous biological interaction network is used as an interaction view; Construct and train a drug-disease association probability prediction model; the model includes: Integrate the semantic view and the interaction view to obtain the fused drug vector representation and disease vector representation; The fused drug vector representation and disease vector representation are used as a fused view; The fused drug and disease vector representation is used to output the drug-disease association probability through the predictor; The model training includes: Perform cross-view comparative learning on fusion view and semantic view; Conduct cross-view comparative learning of fused views and interactive views; Through a multi-task learning approach, implicit knowledge is transferred from contrastive learning to drug-disease association probability prediction via underlying shared parameters.

2. The drug-disease association prediction method based on cross-view contrastive learning according to claim 1, characterized in that: The construction of drug similarity network, disease similarity network and heterogeneous biological interaction network includes: Based on the drug fingerprint similarity and Gaussian interaction spectrum kernel similarity, the similarity between any drugs is calculated; for each drug, the top-K nearest neighbor drug nodes are screened to construct a drug similarity network Based on the similarity of disease phenotypes and Gaussian interaction spectrum kernel similarity, the similarity between any diseases is calculated; for each disease, the top-K nearest neighbor disease nodes are selected to construct a disease similarity network Construct heterogeneous biological interaction networks by integrating drug-disease association networks, drug-protein, and disease-protein interaction networks 3. The drug-disease association prediction method based on cross-view contrastive learning according to claim 1, characterized in that: The integration of the semantic view and the interactive view to obtain the fused drug vector representation and disease vector representation includes: Use the Basic Graph Transformer model to generate drug and disease vector representations under the semantic view; Use the Heterogeneous Graph Transformer model to generate drug and disease vector representations in interactive views; Perform average pooling operations on all layers to generate fused drug and disease vector representations: in, represents the vector representation of drug u after fusion at layer l, Represents the vector representation of the disease v after fusion at the lth layer; vector fusion is completed through the multi-head self-attention mechanism.

4. The drug-disease association prediction method based on cross-view contrastive learning according to claim 1, characterized in that: The loss function for cross-view contrast learning of the fusion view and the semantic view is: Among them, τ represents the temperature coefficient; the node i is obtained by average pooling The vector representation on is:

5. The drug-disease association prediction method based on cross-view contrastive learning according to claim 1, characterized in that: The loss function for cross-view contrast learning of the fused view and the interactive view is: Among them, τ represents the temperature coefficient; the node i is obtained by average pooling The vector representation on is:

6. The drug-disease association prediction method based on cross-view contrastive learning according to claim 1, characterized in that: The fused drug and disease vector representation is used to output the drug-disease association probability through the predictor: Among them, MLP is a multi-layer perceptron, || represents concatenation, and y u,v is the drug-disease association probability.

7. The drug-disease association prediction method based on cross-view contrastive learning according to claim 1, characterized in that: Optimizing drug-disease association probability prediction through cross entropy loss function: Among them, S is a set of positive and negative training samples; is the true label.

8. The drug-disease association prediction method based on cross-view contrastive learning according to claim 1, characterized in that: The multi-task learning method transfers implicit knowledge from contrastive learning to drug-disease association probability prediction through underlying shared parameters: Among them, α1 and α2 are task weight hyperparameters.

9. A drug-disease association prediction system based on cross-view contrastive learning, characterized in that: include: View construction module, constructing drug similarity network, disease similarity network and heterogeneous biological interaction network; The drug similarity network and the disease similarity network are used as semantic views, and the heterogeneous biological interaction network is used as an interaction view; A fusion view representation module integrates the semantic view and the interactive view to obtain a fused drug vector representation and a disease vector representation; the fused drug vector representation and the disease vector representation are used as a fusion view; The drug-disease association probability prediction module uses the fused drug and disease vector representation to output the drug-disease association probability through the predictor; Model training module, which performs cross-view comparative learning between fusion view and semantic view; Conduct cross-view comparative learning of fused views and interactive views; Through a multi-task learning approach, implicit knowledge is transferred from contrastive learning to drug-disease association probability prediction via underlying shared parameters.

Images