A drug target interaction prediction method based on interactive inference network

Abstract

The embodiment of the application discloses a drug target interaction prediction method based on an interaction reasoning network, which comprises an embedding layer, an encoding layer, an interaction layer, a feature extraction layer and an output layer. Sequence embedding of drug and target molecules is generated in the embedding layer, the encoding layer obtains the encoded molecular representation of the drug and the target, the interaction layer simulates the interaction of the drug and the target, the feature extraction layer extracts the interaction features of the interaction matrix, and finally the prediction result is obtained. The application can be applied to the prediction of the relationship between drugs and targets. By using the application, the problem that the training of the traditional DTI prediction model is limited to the small amount of labeled data set and the unlabeled data cannot be fully utilized is solved; secondly, the interpretability of the model in the aspect of DTI prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of drug-target relationship prediction technology, and more particularly to a drug-target interaction prediction method based on interactive reasoning networks. Background Technology

[0002] New drug development is time-consuming and costly, generally divided into two phases: preclinical and clinical research. Utilizing existing drugs to treat new diseases is a feasible strategy because these "old" drugs have already undergone mechanism studies and clinical trials, which can reduce development costs and time.

[0003] Drug targets are biological macromolecules, such as proteins and nucleic acids, that can interact with drugs. Pre-identifying these targets is crucial for drug development targeting specific diseases. However, traditional drug discovery methods only consider a single target for a single disease, neglecting the complex interactions between drugs and targets, and the fact that many diseases involve multiple targets. Therefore, current research increasingly focuses on multi-target drugs and drug combination studies that simultaneously target multiple targets to improve drug efficacy and overcome drug resistance and toxic side effects. While the multipharmacological properties of drugs may lead to unexpected side effects, they may also bring new therapeutic effects, a phenomenon known as drug retargeting.

[0004] Drugs typically improve disease symptoms by interacting with proteins within cells. A vast number of compounds can serve as drug candidates, and drug targets are mostly proteins. Currently, only a small fraction of drug-target protein interactions are known, and many unknown drug-target interactions remain to be discovered.

[0005] Accurately identifying and validating the interactions between drugs and their targets is crucial for discovering new drugs or repositioning existing ones. However, due to the complex relationships between the chemical and proteomic spaces, identifying novel drugs and their targets remains a very challenging process. Multiple factors, such as chemical bonds and affinities, can influence the interactions between drugs and their targets.

[0006] Therefore, effective computational prediction methods are needed to detect complex drug target associations, improve our understanding of biological interactions and processes, and provide new potential drug target interaction candidates for biological experiments. Currently, the main computational prediction methods are molecular docking simulation and machine learning. While molecular docking simulation is widely accepted in biology, the docking process is time-consuming and requires the three-dimensional structures of both the target and the drug, which are difficult to obtain. In contrast, machine learning methods use known information about drug-target interactions to train prediction models to predict interactions between new drugs and targets. Using computers to screen potential drug-target interactions as candidates can reduce time and cost, and finally, drug target interactions can be validated through biological experiments. Summary of the Invention

[0007] This invention provides a drug target interaction prediction method based on interactive reasoning networks, which solves the technical problems of most DTI prediction models being limited to training on scarce labeled datasets, failing to fully utilize unlabeled data, and exhibiting poor interpretability in DTI predictions by machine learning-based models. The method includes the following steps:

[0008] S1: Obtain data on drugs, targets, and drug-target interactions from the BIOSNAP, BindingDB, and DAVIS databases respectively, and obtain positive sample pairs from the three datasets;

[0009] S2: Acquire and generate negative sample pair data;

[0010] S3: Based on the data of the drug, target, and drug-target interaction (i.e., positive sample pair data) and the generated negative sample pair data, the three datasets are organized and merged into a training set, a validation set, and a test set;

[0011] S4: The embedding layer generates the embedding of drug molecules and target sequences, which are then passed through the coding layer respectively;

[0012] S5: The coding layer obtains the molecular characterization of the encoded drug and target;

[0013] S6: The interaction layer simulates drug-target interactions;

[0014] S7: The feature extraction layer captures the interaction features of the interaction matrix;

[0015] S8: Finally, connect to a fully connected network to predict drug target interactions.

[0016] S1 includes the step of identifying sample pairs with Kd values ​​< 30 units as positive samples, and S2 includes the step of detecting drugs and targets contained in the BindingDB and DAVIS databases, identifying sample pairs with Kd values ​​>= 30 as negative samples, and then randomly generating negative sample pairs from the BIOSNAP database.

[0017] S3 includes the following steps: stratified sampling of positive and negative samples in the dataset, and using the hold-out method to divide the dataset into mutually exclusive subsets in a ratio of 7:1:2, which serve as the training set, validation set, and test set.

[0018] S4 includes the following steps:

[0019] Three methods—direct encoding of the entire molecular sequence, subsequence encoding, and fingerprint encoding—are used to convert molecular sequences into vector representations, integrating sequence information into feature vectors. During embedding, the SMILES expression of the drug molecule and the amino acid sequence of the target are encoded separately to generate drug and target sequence embeddings. and These are then input into the encoding layer.

[0020] The step of directly encoding the complete molecular sequence includes: directly encoding SMILES and amino acid sequences; setting the maximum length of SMILES descriptor-generated embeddings to 100 and the maximum length of amino acid sequence-generated embeddings to 1000; truncating embeddings longer than the maximum length and padding those shorter than the maximum length with zeros to generate target and drug embeddings. and ;

[0021] The steps of encoding the molecular subsequence include:

[0022] The ESPF algorithm was applied to extract drug and target substructures from the Uniprot dataset and the ChEMBL database. By setting different frequency thresholds and data sources, vocabulary sets V of different sizes were obtained. Frequent subsequences C were used as the segmentation criteria to decompose drug and target sequences into a set of subsequences.

[0023] Both drug and target sequences are decomposed into substructures and converted into corresponding embedding matrices. The corresponding target embedding is generated through the queryable content embedding matrix. and drug embedding The generation rules are as follows:

[0024]

[0025]

[0026] in, and These are vocabulary sets based on target and drug breakdown subsequences. The size and preset embedding length automatically form a queryable matrix;

[0027] The fingerprint encoding steps include:

[0028] Using molecular fingerprints As a fingerprint embedding It is compressed using a flag bit encoding method.

[0029] S5 includes the following steps:

[0030] The encoding layer employs three encoding methods: multilayer perceptron encoding, convolutional neural network encoding, and self-attention mechanism encoding. It encodes the embeddings by fusing contextual information to extract features.

[0031]

[0032]

[0033] in and These represent the encoding of the drug and target sequences after passing through the coding layer, respectively.

[0034] The multilayer perceptron encoding method includes:

[0035] A three-layer fully connected neural network is stacked, with hidden layers containing 1024, 256, and 64 neurons respectively. In each layer, the input vector... Connection weights with hidden units The components are multiplied to obtain their respective outputs, which are then passed through a non-linear activation function and three layers of neurons to obtain the fingerprint code. ;

[0036] The method of encoding via the convolutional neural network includes:

[0037] A three-layer one-dimensional convolutional neural network is used as the encoder for drug and target embeddings, directly encoding SMILES and amino acid sequences to generate sequence embeddings of drugs and targets. and Input into the CNN encoder;

[0038] A convolutional neural network (CNN) is used to encode and stack three layers of a one-dimensional CNN. Drug encoding uses convolutional kernels of sizes 4, 6, and 8 sequentially according to the layer number, while target encoding uses convolutional kernels of sizes 4, 8, and 12 sequentially according to the layer number. After the convolution operations, the results are pooled and passed through a fully connected layer to finally obtain the drug and target encodings. and ;

[0039] The method for encoding the self-attention mechanism includes:

[0040] By combining sequence information with content embedding through positional encoding, sequence embeddings with the ability to learn sequence sequence information are formed, enabling positional encoding of drugs and targets. and The location encoding is learned through a learning process. Each location learns an independent vector, which is generated using a queryable location embedding matrix. The generation rules are as follows:

[0041]

[0042]

[0043] in, and These are the unique heat codes of the i-th subsequence of the target and the j-th subsequence corresponding to the drug, respectively; and It is a queryable two-dimensional embedding matrix automatically constructed based on the maximum length of the subsequences after drug and target decomposition and the preset embedding length. The preset embedding length is consistent with the content embedding length during generation.

[0044] Input of a self-attention encoder and It is obtained by adding content embedding and location embedding.

[0045] Specifically, S6 includes the following steps:

[0046] For each target subsequence i and each drug subsequence j Generate interactive values:

[0047]

[0048] in F The aggregation function measures the interaction between drug target pairs, and a tensor matrix is ​​obtained after the interaction layer. I Using dot product as an aggregation function can produce a single scalar to measure the strength of the interaction between the smallest units of a single drug target pair.

[0049] Specifically, S7 includes the following steps: modeling the interaction between neighboring regions of individual substructures, and using a convolutional neural network to extract features on the interaction graph.

[0050] Specifically, S8 includes the following steps: the output layer decodes the collected features, flattens the extracted interactive features into a vector, and outputs the prediction result through a linear layer.

[0051] Implementing the embodiments of this invention has the following beneficial effects: This invention uses an interactive inference network for drug-target interactions, including an embedding layer, an encoding layer, an interaction layer, a feature extraction layer, and an output layer. In the embedding layer: to ensure richer feature information for drug and target embeddings, drug molecule fingerprints are used as a supplement, and the ESPF algorithm is used to segment drug and target sequences into subsequences. In the encoding layer: to capture the features of each functional group in the molecular sequences of drugs and targets, a convolutional neural network is used to encode the original molecular sequences; to capture the correlation features between functional groups, a self-attention encoder is used to encode the subsequences of the molecules. The interaction layer design simulates the interaction process, using dot product as the aggregation function to generate a single scalar to measure the interaction strength between the smallest unit pairs of drug-target interactions, providing interpretability for the model's interactive prediction results. The component-based design of the embedding and encoding layers of the interactive inference network gives it strong scalability. By selecting different embedding generation methods in conjunction with different encoders, different prediction effects and experimental significance can be obtained. Attached Figure Description

[0052] Figure 1 This is the network structure implemented in this invention;

[0053] Figure 2 These are statistical tables for the DAVIS, BindingDB, and BIOSNAP datasets;

[0054] Figure 3 It is a table for allocating training, validation, and test sets. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0056] like Figure 1 As shown, an embodiment of the present invention provides a drug target interaction prediction method based on an interactive reasoning network, which is implemented through the following steps.

[0057] 1. Data Acquisition and Preprocessing

[0058] Data on drugs, targets, and drug-target interactions were collected from the BIOSNAP, BindingDB, and DAVIS databases. The BIOSNAP database used in this invention primarily consisted of the MINER DTI dataset, which contains information on 4510 drugs and 2181 targets, and these were considered positive DTI pairs. In addition to the BIOSNAP dataset, the DAVIS and BindingDB datasets were also included. The DAVIS dataset contains experimental detection Kd values ​​for 68 drugs and 379 targets, while BindingDB contains experimental detection Kd values ​​for 7165 drugs and 1254 targets. DTI pairs with Kd values ​​<30 units were considered positive samples.

[0059] 2. Obtain negative sample pairs. This step specifically includes:

[0060] Drug and target pairs with a Kd value >= 30 were considered negative DTI pairs, and then negative DTI pairs were randomly generated from the BIOSNAP database, resulting in a total of three sets of negative DTI pairs. Figure 2 As shown.

[0061] 3. Organize and merge the data from the three datasets. This step specifically includes:

[0062] The dataset is stratified by positive and negative samples, and then divided into mutually exclusive subsets in a 7:1:2 ratio using the hold-out method. These subsets serve as the training set, validation set, and test set, respectively. The allocation is as follows: Figure 3 As shown. For each experiment, five independent runs were performed.

[0063] 4. The sequence embedding of drug and target molecules is generated in the embedding layer. This step specifically includes:

[0064] Three methods are employed to convert molecular sequences into vector representations: direct encoding of the entire molecular sequence, molecular subsequence encoding, and fingerprint encoding (Morgan fingerprint and PubChem fingerprint encoding), integrating sequence information into the feature vector. During embedding, SMILES and amino acid sequences are encoded separately to generate drug and target sequence embeddings. and These are then input into the encoding layer.

[0065] Three methods are employed to convert molecular sequences into vector representations: direct encoding of the entire molecular sequence, molecular subsequence encoding, and fingerprint encoding (Morgan fingerprint and PubChem fingerprint encoding), integrating sequence information into the feature vector. During embedding, SMILES and amino acid sequences are encoded separately to generate drug and target sequence embeddings. and The data are then input into the coding layer. The specific steps for the three methods—direct encoding of the entire molecular sequence, encoding of molecular subsequences, and fingerprint encoding (Morgan fingerprint and PubChem fingerprint encoding)—are as follows:

[0066] (1) Direct encoding of the entire molecular sequence

[0067] Since both SMILES and amino acid sequences are represented by letter symbols, they can be directly encoded. The maximum length for embeddings generated from SMILES descriptors is set to 100, and the maximum length for embeddings generated from amino acid sequences is set to 1000. Embeddings longer than the maximum length are truncated, and those shorter than the maximum length are padded with '0's to generate embeddings for targets and drugs. and .

[0068] In this experiment, SMILES are represented by 63 labels, and the amino acid sequence is represented by 21 labels. The labels are used to represent the SMILES descriptor and the amino acid sequence. For example, 'C', 'H', and 'N' are represented by '1', '2', and '3' respectively, so the SMILES sequence 'CN=C=O' is represented as [CN=C=O]=[1 3 63 1 63 5].

[0069] (2) Encoding of molecular subsequences

[0070] This invention uses the ESPF algorithm to extract drug and target substructures from the Uniprot dataset and the ChEMBL database. By setting different frequency thresholds (the minimum number of times a single subsequence appears in the dataset) and data sources, vocabulary sets V of different sizes are obtained. Frequent subsequences C are used as the segmentation criterion to decompose drug and target sequences into a set of subsequences. These are encoded as bit vectors and embeddings are generated. Since the decomposed SMILES descriptors and amino acid sequences have relatively short lengths, drugs (100) and targets (500) with lower frequency thresholds are selected to retain more original information. In this experiment, the maximum length of drug sequences is 50 subsequences, and the maximum length of target sequences is 545 subsequences. Both target and drug sequences are decomposed into substructures and converted into corresponding embedding matrices. These matrices generate corresponding target embeddings through queryable content embedding matrices. and drug embedding The generation rules are as follows:

[0071]

[0072]

[0073] in, and These are vocabulary sets based on target and drug breakdown subsequences. The size and preset embedding length automatically form a queryable matrix.

[0074] (3) Fingerprint encoding

[0075] Fingerprint encoding is further divided into two types: direct fingerprint embedding and fingerprint marker bit encoding. Due to molecular fingerprints... It is a binary vector composed of '0' and '1', so this vector can be directly used as a fingerprint embedding. However, because molecular fingerprints are long and have sparse information density, they can be compressed using marker bit encoding. This allows for the identification of molecular fingerprints. All flag bits with a value of 1 Ignore the flag bits with a value of 0, and extract the flag bits according to the original order encoding in the fingerprint. Corresponding order i Then, a shorter binary vector is generated according to the extraction order and used as the embedding. .

[0076] Because each drug corresponds to a different number of fingerprint marker bits, the vector The lengths are also inconsistent and need to be set. Maximum length. Based on Morgan's fingerprint and PubChem's fingerprint encoding. Length distribution. The maximum embedding length for Morgan fingerprint encoding was ultimately set to 76, and the maximum embedding length for PubChem encoding was set to 250. Embeddings shorter than the maximum length after encoding were padded with '0's.

[0077] 5. Obtain molecular characterization of the encoded drug and target at the coding layer. This step specifically includes:

[0078] The encoding layer employs three encoding methods: Multilayer Perceptron (MLP) encoding, Convolutional Neural Network (CNN) encoding, and Self-Attention encoding. These methods encode the embedding by fusing contextual information to extract features.

[0079]

[0080]

[0081] in and These represent the encoding of the drug and target sequences after passing through the coding layer, respectively.

[0082] The specific steps for the three encoding methods—Multilayer Perceptron (MLP) encoding, Convolutional Neural Network (CNN) encoding, and Self-Attention encoding—are as follows:

[0083] (1) Multilayer perceptron (MLP) coding

[0084] Embedded representations generated directly from fingerprints are simple, context-independent, and non-linear. Embedded representations generated directly from drug molecule fingerprints can be used... Input MLP encoder.

[0085] The MLP encoder is stacked with three fully connected neural network layers, with hidden layers containing 1024, 256, and 64 neurons respectively. In each layer, the input vector... Connection weights with hidden units The components are multiplied to obtain their respective outputs, which are then passed through a non-linear activation function. After passing through three layers of neurons, a fingerprint encoding is obtained. :

[0086]

[0087]

[0088] (2) Convolutional Neural Network (CNN) Encoding

[0089] Embedsion lengths generated using directly encoded SMILES and complete amino acid sequences can reach 1000 units. Fully connected neural networks can be used for encoding, but this approach is prone to overfitting and does not reflect the group-based interaction characteristics of drugs and targets. Convolutional neural networks, by directly connecting neurons in upper and lower layers through convolutional kernels and sharing parameters, can extract local features of molecules, reduce data processing, and retain useful information, making them more suitable for encoding.

[0090] The experiment employed a three-layer one-dimensional convolutional neural network as the encoder for drug and target embedding, directly encoding SMILES and amino acid sequences to generate drug and target sequence embeddings. and , input into the CNN encoder.

[0091] A CNN encoder is stacked with three layers of a one-dimensional convolutional neural network. For drugs, convolutional kernels of sizes 4, 6, and 8 are used sequentially according to the layer number; for targets, convolutional kernels of sizes 4, 8, and 12 are used sequentially according to the layer number. After the convolution operations, the results are pooled and passed through a fully connected layer to finally obtain the encodings for the drug and target. and :

[0092]

[0093]

[0094] (3) Self-Attention Mechanism Encoding

[0095] The computation time and storage requirements of a self-attention encoder are approximately quadratic of the input size, making it computationally infeasible to use embeddings generated from directly encoded whole molecular structures as input. Embeddings generated from molecular sub-sequence encoding, however, are relatively short. Encoders based on the self-attention mechanism can capture the chemical semantics and contextual relationships between drug and target sub-structures; therefore, embeddings generated from sub-sequence encoding are used as input to the self-attention encoder.

[0096] Since the embeddings generated from the sequentially decomposed subsequences of drugs and targets do not contain contextual information about the molecular sequence, the Self-Attention encoder uses positional encoding to combine sequence information with content embeddings, forming sequence embeddings capable of learning sequence sequence information. Positional encoding of drugs and targets. and This can be learned. The learning location encoding involves learning an independent vector for each location, generated using a queryable location embedding matrix, according to the following rules:

[0097]

[0098]

[0099] in, and These are the target numbers i The unique thermal encoding of the j-th subsequence corresponding to each subsequence and the drug; and It is a queryable two-dimensional embedding matrix automatically constructed based on the maximum length of the subsequences after drug and target decomposition and the preset embedding length. The preset embedding length is consistent with the content embedding generation.

[0100] Finally, the input to the Self-Attention encoder and The fingerprint is obtained by adding content embedding and location embedding. To maintain consistency between the embedded content and location in the fingerprint generation, Obtained by adding the content embedding to a constant:

[0101]

[0102]

[0103]

[0104] The generated embedding containing location information is input into a Self-Attention encoder, which consists of N identical layers, each layer consisting of two sub-layers. The first is a multi-head self-attention encoder. The multi-head self-attention encoder is... After several different linear transformations, the different attention results are concatenated. Using multi-head attention allows the model to focus on different positions. In this invention, a 12-head self-attention mechanism is used.

[0105]

[0106]

[0107] The results from the self-attention layer are then fed into a feed-forward neural network. Simultaneously, a sub-layer is appended to both the self-attention layer and the feed-forward neural network layer for concatenation and normalization. Finally, the codes for the drug and target are obtained. and :

[0108]

[0109] .

[0110] 6. The interaction layer simulates the drug-target interaction. This step specifically includes:

[0111] To simulate the interaction of paired drug targets, for each target subsequence i and each drug subsequence j Generate interactive values:

[0112]

[0113] in F This is an aggregation function used to measure the interaction between drug target pairs; in experiments, it is represented by a dot product. A tensor matrix is ​​obtained after the interaction layer. I Each column in this interaction table considers the interaction between individual minimum units of the target and the drug. Using the dot product as the aggregation function produces a single scalar that measures the strength of the interaction between a single pair of drug-target minimum units.

[0114] 7. The feature extraction layer extracts the interaction features of the interaction matrix. This step specifically includes:

[0115] The interaction layer outputs a two-dimensional interaction table. IThen, feature extraction is performed in the downstream layer. The higher the value in the interaction table, the higher the probability of DTI interaction. If the dot product at the corresponding position in the interaction table is high, it indicates that the interacting units do indeed interact. Visualizing the interaction table can more intuitively show which units contribute to the final result, clarifying the internal working principle of the drug-target interaction prediction model as much as possible, and thus providing strong support for biological interpretation.

[0116] The smallest adjacent units in drug and target sequences can influence each other and trigger interactions. Therefore, it is necessary to simulate interactions between neighboring regions, not just pairwise interactions. To achieve this, convolutional neural networks are used to model interactions in the interaction table. I Feature extraction is performed on the convolutional layers. By using a fixed number of order-invariant convolutional kernels, the convolutional layers can capture and aggregate the interactions between neighborhoods and output a matrix. O :

[0117]

[0118] 8. Finally, obtain the prediction results. This step specifically includes:

[0119] The output layer decodes the collected features, flattens the extracted interaction features into a vector, and outputs the prediction result after passing through a linear layer. L :

[0120]

[0121]

[0122] in, , and These represent the weight matrix and offset of the fully connected layer, respectively. This indicates the true label of the drug target interaction pair.

[0123] In this embodiment of the invention, the drug target interaction prediction model based on interactive reasoning network includes five different combinations of algorithms: (1) generating embeddings by direct encoding of the whole molecular sequence and using a CNN encoder for encoding and prediction; (2) generating embeddings by subsequence encoding and using a Self-Attention encoder for encoding and prediction; (3) generating embeddings by direct encoding of PubChem fingerprints, generating drug codes by using an MLP encoder, and then concatenating them with the drug codes generated in (1) for prediction; (4) generating embeddings by direct encoding of Morgan fingerprints, generating drug codes by using an MLP encoder, and then concatenating them with the drug codes generated in (1) for prediction; (5) generating embeddings by encoding PubChem fingerprint marker bits, generating drug codes by using a Self-Attention encoder, and then concatenating them with the drug codes generated in (2) for prediction.

[0124] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for predicting drug target interactions based on interactive reasoning networks, characterized in that, Includes the following steps: S1: Obtain data on drugs, targets, and drug-target interactions from the BIOSNAP, BindingDB, and DAVIS databases respectively, and obtain positive sample pairs from the three datasets; S2: Acquire and generate negative sample pair data; S3: Based on the data of the drug, target and drug-target interaction, as well as the generated negative sample pair data, the three datasets are integrated to construct a training set, a validation set and a test set; S4: The drug molecule and target sequence are generated in the embedding layer and passed through the coding layer respectively, and then processed by the following method: Three methods—direct encoding of the entire molecular sequence, encoding of molecular subsequences, and fingerprint encoding—are used to convert molecular sequences into vector representations, and sequence information is integrated into feature vectors. During embedding, the SMILES expression of the drug molecule and the amino acid sequence of the target are encoded separately to generate drug and target sequence embeddings. and Each is input into the encoding layer; S5: The coding layer obtains the molecular characterization of the encoded drug and target; S6: The interaction layer performs drug-target interaction simulations, including the following steps: for each target subsequence i and each drug subsequence j Generate interactive values: in F The aggregation function measures the interaction between drug-target pairs, and a tensor matrix is ​​obtained after the interaction layer. I Using dot product as an aggregation function can produce a single scalar to measure the strength of the interaction between the smallest units of a single drug target pair. S7: The feature extraction layer captures the interaction features of the interaction matrix; S8: Finally, connect to a fully connected network to predict drug target interactions.

2. The drug target interaction prediction method based on interactive reasoning networks according to claim 1, characterized in that, S1 includes the step of identifying sample pairs with Kd values ​​< 30 units as positive samples. S2 includes the step of detecting drugs and targets contained in the BindingDB and DAVIS databases, identifying sample pairs with Kd values ​​>= 30 as negative samples, and then randomly generating negative sample pairs from the BIOSNAP database.

3. The drug target interaction prediction method based on interactive reasoning networks according to claim 2, characterized in that, S3 includes the following steps: stratified sampling of positive and negative samples in the dataset, and using the hold-out method to divide the dataset into mutually exclusive subsets in a ratio of 7:1:2, which serve as the training set, validation set, and test set.

4. The drug target interaction prediction method based on interactive reasoning networks according to claim 1, characterized in that, The steps of direct encoding of the complete molecular sequence include: directly encoding SMILES and amino acid sequences; setting the maximum length of SMILES descriptor-generated embeddings to 100 and the maximum length of amino acid sequence-generated embeddings to 1000; truncating embeddings longer than the maximum length and padding those shorter than the maximum length with zeros to generate target and drug embeddings. and ; The steps of encoding the molecular subsequence include: Substructures of drugs and targets were extracted from the Uniprot dataset and the ChEMBL database. The ESPF algorithm was used to obtain vocabulary sets V of different sizes by setting different frequency thresholds and data sources. Frequent subsequences C were used as the segmentation criteria to decompose drug and target sequences into a set of subsequences. Both drug and target sequences are decomposed into substructures and converted into corresponding embedding matrices. The corresponding target embedding is generated through the queryable content embedding matrix. and drug embedding The generation rules are as follows: in, and These are vocabulary sets based on target and drug breakdown subsequences. The size and preset embedding length automatically form a queryable matrix; The fingerprint encoding steps include: Using molecular fingerprints As a fingerprint embedding It is compressed using a flag bit encoding method.

5. The drug target interaction prediction method based on interactive reasoning networks according to claim 4, characterized in that, S5 includes the following steps: The encoding layer employs three encoding methods: multilayer perceptron encoding, convolutional neural network encoding, and self-attention mechanism encoding. It encodes the embeddings by fusing contextual information to extract features. in and These represent the encoding of the drug and target sequences after passing through the coding layer, respectively.

6. The drug target interaction prediction method based on interactive reasoning networks according to claim 5, characterized in that, The multilayer perceptron encoding method includes: A three-layer fully connected neural network is stacked, with hidden layers containing 1024, 256, and 64 neurons respectively. In each layer, the input vector... Connection weights with hidden units The components are multiplied to obtain their respective outputs, which are then passed through a non-linear activation function and three layers of neurons to obtain the fingerprint code. ; The method of encoding via the convolutional neural network includes: A three-layer one-dimensional convolutional neural network is used as the encoder for drug and target embeddings, directly encoding SMILES and amino acid sequences to generate sequence embeddings of drugs and targets. and Input into the CNN encoder; A convolutional neural network (CNN) is used to encode and stack three layers of a one-dimensional CNN. Drug encoding uses convolutional kernels of sizes 4, 6, and 8 sequentially according to the layer number, while target encoding uses convolutional kernels of sizes 4, 8, and 12 sequentially according to the layer number. After the convolution operations, the results are pooled and passed through a fully connected layer to finally obtain the drug and target encodings. and ; The method for encoding the self-attention mechanism includes: By combining sequence information with content embedding through positional encoding, sequence embeddings with the ability to learn sequence sequence information are formed, enabling positional encoding of drugs and targets. and The location encoding is learned through a learning process. Each location learns an independent vector, which is generated using a queryable location embedding matrix. The generation rules are as follows: in, and These are the unique heat codes of the i-th subsequence of the target and the j-th subsequence corresponding to the drug, respectively; and It is a queryable two-dimensional embedding matrix automatically constructed based on the maximum length of the subsequences after drug and target decomposition and the preset embedding length. The preset embedding length and the content embedding length are consistent during generation. Among them, the input of the self-attention mechanism encoder and It is obtained by adding content embedding and location embedding.

7. The drug target interaction prediction method based on interactive reasoning networks according to claim 1, characterized in that, S7 specifically includes the following steps: modeling the interaction between neighboring regions of individual substructures, and using a convolutional neural network to extract features on the interaction graph.

8. The drug target interaction prediction method based on interactive reasoning networks according to claim 1, characterized in that, S8 specifically includes the following steps: the output layer decodes the collected features, flattens the extracted interactive features into a vector, and outputs the prediction result through a linear layer.

Images