A drug-drug interaction prediction method based on secure multi-party computing

Abstract

The present invention discloses a method for predicting drug-drug interactions based on secure multi-party computing, comprising the following steps: S1, obtaining a drug-drug interaction network, a drug-protein interaction network, a drug-disease association network, and a drug-side effect association network; S2, calculating Jaccard similarity based on different drug features to obtain similar features between all drugs, and using principal component analysis technology to reduce the dimensionality of all drug similarity features; S3, dividing each user's private feature data into four parts, encrypting them using secret sharing technology, and sending them to four servers using an additional secret sharing mechanism; S4, inputting the four parts of feature data into a preset private deep learning model to predict drug-drug interactions. The present invention enables high-quality collaboration between pharmaceutical companies and research institutions without leaking drug privacy information, thereby improving drug-drug interaction prediction.

Description

Technical Field

[0001] The present invention relates to the field of pharmaceutical technology, and more particularly to a method for predicting drug-drug interactions based on secure multi-party computing. Background Art

[0002] Drug-drug interactions are a significant issue in drug discovery and pharmacovigilance. When two or more medications are taken simultaneously, some can alter the activity of others, leading to adverse drug reactions that may cause illness or death. Furthermore, as the number of medications a patient takes increases, the incidence of drug-drug interactions increases rapidly. Therefore, rapidly predicting potential drug-drug interactions can effectively reduce the incidence of adverse drug reactions and disease.

[0003] In recent years, an increasing number of researchers have begun using computational techniques to improve the performance of drug-drug interaction prediction. In particular, various deep learning models have achieved excellent predictive performance and have become a key technology for drug-drug interaction prediction. However, most deep learning models overly rely on drug feature data and training data. Unfortunately, high-quality drug data is relatively scarce in real-world drug discovery. Consequently, these deep learning-based drug-drug interaction prediction models may not perform well in real-world drug development scenarios.

[0004] With the rapid development of biomedical research, various public bioinformatics and pharmacogenomics databases have been constructed. For example, DrugBank is a comprehensive database for drug discovery, providing detailed drug and target information. The Therapeutic Target Database (TTD) provides extensive drug information, including nucleic acid targets, therapeutic proteins, pathways, and targeted diseases. The SuperTarget database is a one-stop database and a resource for exploring drug-target interactions. More importantly, pharmaceutical companies and research institutions possess a vast amount of valuable private drug data. Existing research has shown that integrating multi-source data from various public databases and institutions can improve the performance of drug-drug interaction prediction. However, concerns about drug data privacy have led pharmaceutical companies to be reluctant to share their data, severely hindering the development of drug-drug interaction prediction. Furthermore, integrating privacy-preserving computing with deep learning technologies to facilitate collaboration among research institutions without leaking private information, thereby improving the performance of drug-drug interaction prediction, is crucial in the drug development process. To this end, it is necessary to develop a drug-drug interaction prediction method based on secure multi-party computation. Summary of the Invention

[0005] The purpose of the present invention is to provide a drug-drug interaction prediction method based on secure multi-party computing to overcome the defects of the existing technology.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting drug-drug interactions based on secure multi-party computing comprises the following steps:

[0008] S1. Obtain drug-drug interaction networks, drug-protein interaction networks, drug-disease association networks, and drug-side effect association networks, and convert drug SMILES sequences into drug molecular fingerprints;

[0009] S2. Calculate the Jaccard similarity features between all drugs based on the drug molecular fingerprints in step S1, and calculate the Jaccard similarity features between drugs based on the drug-protein interaction network, drug-disease interaction network, and drug-side effect association network in step S1, and use principal component analysis technology to reduce the dimension of all drug similarity features and use them as the private drug features of each user;

[0010] S3, divide each user's private feature data into four parts, encrypt them through secret sharing technology, and send them to four servers using additional secret sharing mechanism. Each server obtains part of the feature data F from four users. MoFP ,F DPI ,F DDA ,F DSA ;

[0011] S4. Obtain four partial feature data F on either side MoFP ,F DPI ,F DDA ,F DSA Then, the four parts of feature data F MoFP ,F DPI ,F DDA ,F DSA Input into a preset private deep learning model to predict drug-drug interactions.

[0012] Furthermore, in step S1, drug-drug interaction and drug-protein interaction networks are extracted from the DrugBank database, drug-disease association networks are extracted from the CTD database, and drug-side effect association networks are extracted from the CTD and SIDER databases, and then the SMILES sequences of the molecules are converted into molecular fingerprints using RDKit technology.

[0013] Furthermore, the formula for calculating the Jaccard similarity in step S2 is:

[0014]

[0015] Where A and B are binary vectors of the two drugs, |A∩B| represents the intersection of A and B, and |A| and |B| represent the total number of 1s in the A and B vectors.

[0016] Furthermore, in step S3, a 4-out-of-4 additional secret sharing mechanism is used to send the secret to four servers.

[0017] Furthermore, the private deep learning model in step S4 includes a feature projection module and a drug-drug interaction prediction module, wherein the projection module is used to map different types of drug features into the same space, and the drug-drug interaction prediction module is used to predict the output of the projection module to obtain a prediction result.

[0018] Compared with the existing technology, the advantages of the present invention are: the present invention provides a drug-drug interaction prediction method based on secure multi-party computing, which uses secret sharing technology to calculate drug-related feature data from multiple institutions, and designs a deep learning model with multimodal feature fusion to improve drug-drug interaction prediction. Moreover, all data transmission and deep learning operations are integrated in the secure multi-party computing framework, which enables high-quality collaboration between pharmaceutical companies and research institutions without leaking drug privacy information, thereby improving drug-drug interaction prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 This is a framework diagram of the drug-drug interaction prediction method based on secure multi-party computing of the present invention.

[0021] Figure 2 It is the experimental result figure of the present invention. DETAILED DESCRIPTION

[0022] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0023] See Figure 1 and Figure 2As shown, this embodiment discloses a method for predicting drug-drug interactions based on secure multi-party computing, comprising the following steps:

[0024] Step S1: Extract drug-drug interaction and drug-protein interaction networks from the DrugBank database, extract drug-disease association networks from the CTD database, and extract drug-side effect association networks from the CTD and SIDER databases. Then, convert the drug SMILES sequences into drug molecular fingerprints. These four types of data are four binary feature matrices, which are distributed to the four pharmaceutical companies as their own proprietary feature data.

[0025] Step S2, private feature preprocessing: Perform feature preprocessing in a plaintext environment based on the drug data owned by the four pharmaceutical companies.

[0026] Each pharmaceutical company calculates the Jaccard similarity between all drugs based on the characteristics of its own drugs and the drug molecular fingerprints in step S1 to obtain the similarity matrix between drugs. The formula is as follows:

[0027]

[0028] Where A and B are binary vectors of the two drugs, |A∩B| represents the intersection of A and B, and |A| and |B| represent the total number of 1s in the A and B vectors.

[0029] And according to the drug-protein interaction network, drug-disease interaction network, and drug-side effect association network in step S1, the Jaccard similarity features between the drugs are calculated respectively.

[0030] The principal component analysis technology is then used to reduce the dimension of all drug similarity features and use them as the private drug features of each user, thereby improving the quality of the input features.

[0031] Step S3: Data Sharing Based on Secure Multi-Party Computation: In this embodiment, secure multi-party computation is implemented as a client-server model, where each pharmaceutical organization with private data is considered a client, and each user possesses private drug profile data. Each server adheres to the rules, avoiding sending fabricated or incorrect information to other servers, and no two servers collude with each other. Specifically,

[0032] Each user's private feature data is divided into four parts and encrypted using secret sharing technology; it is sent to four servers using a 4-out-of-4 additional secret sharing mechanism, and each server obtains partial feature data F from four users. MoFP ,F DPI ,F DDA ,F DSA,Therefore, no server can obtain the complete data of a single pharmaceutical company.

[0033] Step S4: Obtain four partial feature data F on either side MoFP ,F DPI ,F DDA ,F DSA After that, the four parts of feature data F MoFP ,F DPI ,F DDA ,F DSA Input into a preset private deep learning model to predict drug-drug interactions.

[0034] Specifically, the private deep learning model includes a feature projection module and a drug-drug interaction prediction module. These two modules can be further decomposed into a number of basic operations, including private addition, private multiplication, comparison, and some linear and nonlinear functions. These private deep learning operations are performed through the CRYPTEN framework. Most operations, including bilinear operations and multiplication, use arithmetic secret sharing (ASIS). In contrast, binary secret sharing (BSS) is used to evaluate logical expressions, such as the ReLU activation function. The ABY sharing conversion technique is used to convert between ASIS and BSS. The four servers complete all private deep learning operations in sequence and obtain encrypted results. Finally, each server sends its own secret results to the corresponding user.

[0035] The projection module is used to map different types of drug features into the same space. The projection module is defined as:

[0036] h0=BNLayer(h MoFP ||h DPI ||h DDA ||h DSA )

[0037] Where || represents the concatenation operation, BNLayer(·) represents the batch normalization operation; h l ∈{h MoFP ,h DPI ,h DDA ,h DSA} can be calculated by formula (3):

[0038] h l =WF l +b

[0039] Where W is the learnable weight matrix of the linear transformation, b is the bias vector, and F l ∈{F MoFP ,F DPI ,F DDA ,FDSA}, in order to project different types of features into the same space, the four secret features share a weight matrix and bias vector. Finally, h0 is input into the prediction module for drug-drug interaction prediction.

[0040] The drug-drug interaction prediction module is used to predict the output of the projection module to obtain a prediction result. The drug-drug interaction prediction module uses a multilayer perceptron with three fully connected layers. The two hidden layers and the output layer use the ReLU activation function and the sigmoid activation function, respectively. This embodiment uses a private operation protocol to train the deep learning model and only transmits the predicted drug-drug interaction results to the corresponding users.

[0041] This example uses 708 drugs in the DrugBank database as an example. The above method is used to predict drug-drug interactions. The area under the receiver operating curve (AUROC) is 0.938, the area under the precision-recall curve (AUPR) is 0.942, and the recall rate (Recall) is 0.870, which is better than the other five methods. Figure 2 shown.

[0042] Although the embodiments of the present invention are described in conjunction with the accompanying drawings, the patent owner may make various changes or modifications within the scope of the appended claims. As long as they do not exceed the scope of protection described in the claims of the present invention, they should be within the scope of protection of the present invention.

Claims

1. A method for predicting drug-drug interactions based on secure multi-party computing, characterized in that: The following steps are involved: S1. Obtain drug-drug interaction networks, drug-protein interaction networks, drug-disease association networks, and drug-side effect association networks, and convert drug SMILES sequences into drug molecular fingerprints; S2. Calculate the Jaccard similarity features between all drugs based on the drug molecular fingerprints in step S1, and calculate the Jaccard similarity features between drugs based on the drug-protein interaction network, drug-disease interaction network, and drug-side effect association network in step S1, and use principal component analysis technology to reduce the dimension of all drug similarity features and use them as the private drug features of each user; S3, divide each user's private drug feature data into four parts, encrypt them through secret sharing technology, and send them to four servers using additional secret sharing mechanism. Each server obtains partial feature data F from four users. MoFP ,F DPI ,F DDA ,F DSA ; S4. Obtain four partial feature data F on either side MoFP ,F DPI ,F DDA ,F DSA After that, the four parts of feature data F MoFP ,F DPI ,F DDA ,F DSA Input into a preset private deep learning model to predict drug-drug interactions.

2. The method for predicting drug-drug interactions based on secure multi-party computing according to claim 1, characterized in that: In step S1, drug-drug interaction and drug-protein interaction networks are extracted from the DrugBank database, drug-disease association networks are extracted from the CTD database, and drug-side effect association networks are extracted from the CTD and SIDER databases, and then the SMILES sequences of the molecules are converted into molecular fingerprints using the RDKit technology.

3. The method for predicting drug-drug interactions based on secure multi-party computing according to claim 1, characterized in that: The formula for calculating the Jaccard similarity in step S2 is: Where A and B are binary vectors of the two drugs, |A∩B| represents the intersection of A and B, and |A| and |B| represent the total number of 1s in the A and B vectors.

4. The method for predicting drug-drug interactions based on secure multi-party computing according to claim 1, wherein: In step S3, the secret is sent to four servers using a 4-out-of-4 additional secret sharing mechanism.

5. The method for predicting drug-drug interactions based on secure multi-party computing according to claim 1, wherein: The private deep learning model in step S4 includes a feature projection module and a drug-drug interaction prediction module. The projection module is used to transform the four different types of drug features F in step S3 into MoFP ,F DPI ,F DDA ,F DSA Mapped into the same space, the drug-drug interaction prediction module is used to predict the output of the projection module to obtain a prediction result.

Images