Drug target affinity prediction interaction and expansion causal convolutional enhancement network and method

Abstract

The invention provides a drug target affinity prediction interaction and expansion causal convolutional enhancement network and method. The network comprises a substructure decomposition module based on knowledge inspiration for mining features of substructures; the method comprises the following steps: constructing an enhanced transform encoder module to effectively mine a semantic relationship between substructures extracted from a large amount of unmarked drug data; an interaction module is constructed to better capture drug information so as to realize more accurate and explainable DTA prediction; the sequential relationship is captured through an expansion causal convolution module, and complex feature representations are provided to handle long-term dependencies to improve prediction accuracy. The method solves the technical problems that the result is inaccurate and difficult to explain, unlabeled molecular data is ignored, and the time dependency relationship is not considered.

Description

Technical Field

[0001] The present invention relates to the field of drug target affinity prediction, and in particular to a drug target affinity prediction interaction and dilated causal convolution enhanced network and method. Background Art

[0002] Accurately predicting drug-target affinity (DTA) is crucial in the drug discovery and development process. DTA reveals the strength of the protein-ligand interaction between the drug and the target, and it provides more information about the drug-target interaction. In addition, DTA is the basis for finding new drugs (i.e., virtual screening) and new indications for existing drugs (i.e., drug repositioning) because the therapeutic effects of drug compounds are detected by examining DTA. Therefore, people need to develop effective technologies to determine the potential connection between drugs and targets.

[0003] Existing drug-target affinity prediction methods are mainly divided into two categories: predicting binary DTA and predicting drug-target affinity values. The first category is direct prediction of DTA, that is, determining whether a drug and a target interact with each other through known drug-target interactions and their feature vectors or through pairwise similarities extracted by combining the chemical or structural descriptors of the drug with the descriptors of the target. Most methods formulate the prediction of DTA as a classification problem based on the feature vectors of drugs or targets or the pairwise similarities between drugs or targets. For example, if it is known that drug-target pairs have interactions, they can be considered positive, otherwise they are negative. Therefore, models built based on methods such as deep neural networks (DNNs), deep belief networks (DBNs), and convolutional neural networks (CNNs) can be learned by training on positive and negative drug-target pairs, whose features or pairwise similarities can be constructed through the chemical or structural features of drugs or targets.

[0004] However, existing studies usually learn molecular representations and make predictions based on the overall molecular structure of drugs and proteins, ignoring that the interactions only involve relevant substructures of drugs and proteins. Full-structure molecular representations introduce noise and affect the prediction effect. In addition, the learned representations are also difficult to interpret because they do not provide a feasible way to explain which substructures of drugs and proteins contribute to the interaction. In addition, these studies focus on studying labeled drug data and limit the scope to thousands of drugs and proteins, while ignoring the large amount (such as millions) of available unlabeled biomedical data. In addition, these methods do not have a large receptive field, ignore the long-term dependencies between drug targets, and it is difficult to capture a wider range of relationships between drugs and targets, making it even more difficult to provide richer feature representations and more accurate predictions.

[0005] At present, a large number of prediction methods have emerged in the field of drug target affinity prediction, which can be mainly divided into structure-based methods and structure-free methods. Structure-based methods rely on the three-dimensional structure of drug molecules and target proteins to predict their interactions. However, when structural information is lacking, the effectiveness of this method will be limited. Therefore, structure-free methods are studied in depth, among which similarity-based methods use known drug-target similarity information to predict drug target affinity. For example, KronRLS and SimBoost. Feature-based methods refer to the use of the characteristics of drug-protein interactions to predict the strength or affinity of drug molecules binding to specific protein targets. This method usually relies on machine learning or computer simulation technology to construct a prediction model by analyzing the structure, physicochemical properties, biological activity and other characteristics between drug molecules and protein targets. In addition, deep learning-based methods have achieved significant prediction results. They usually use different networks (e.g., convolutional neural networks (CNN), recurrent neural networks (RNN)) to build models for DTA prediction. For example, DeepAffinity uses composite SMILES or protein SPS representations to perform unsupervised pre-training of seq2seq RNN models. GeneralizedDTA is designed for unknown drug discovery, combining self-supervised pre-training and multi-task learning with a dual adaptation mechanism to enhance protein and drug feature representation and improve generalization ability. FusionDTA uses a multi-head linear attention mechanism for feature aggregation and knowledge distillation to reduce parameter redundancy, thereby improving the performance and efficiency of DTA tasks. Nguyen et al. used transfer learning of chemical-chemical interaction (CCI) and protein-protein interaction (PPI) tasks to enhance drug-target affinity prediction.

[0006] The aforementioned existing similarity-based schemes have the following defects: lack of reliable data, similarity methods rely on a large number of reliable known data sets to establish the relationship between drug targets; at the same time, they are unable to accurately identify specific binding sites.

[0007] The aforementioned existing feature-based schemes have the following defects: the existing feature representation methods cannot fully capture the complex interactions between drug molecules and target proteins, which limits the accuracy of prediction.

[0008] The aforementioned existing deep learning-based solutions have the following defects: poor feature extraction and prediction effects on drug and target protein sequences.

[0009] The existing invention patent application document "A self-supervised graph neural network pre-training method based on contrastive learning" with publication number CN114038517A, the existing method includes: pre-processing the compound molecules in the public database to screen out organic molecules; structurally decomposing and extracting the screened organic molecules, using the obtained substructures as identifiers, and constructing a corpus of substructures; treating the decomposed substructures as super nodes and constructing corresponding subgraph data, the subgraph data and the original molecular graph data constitute positive sample pairs, and randomly selecting a number of subgraph data and the original molecular graph data to form negative sample pairs; constructing a graph convolutional neural network based on an attention mechanism, a multi-level gated recurrent unit and a multi-layer perceptron module to form a self-supervised learning model; inputting all positive and negative sample pair data into the self-supervised learning model for pre-training and saving, so as to facilitate fine-tuning of downstream tasks. However, in the aforementioned existing schemes: in the field of drug target prediction, it is difficult for graph neural networks to obtain high-quality data sets with labels, and at the same time, the computational complexity is high. The scheme takes a lot of time to make predictions, and has little advantage in practical applications.

[0010] In summary, the existing technologies have technical problems such as inaccurate and difficult to interpret results, ignoring unlabeled molecular data, and not considering time dependence. Summary of the invention

[0011] The technical problem to be solved by the present invention is: how to solve the technical problems in the prior art that the results are inaccurate and difficult to interpret, unlabeled molecular data are ignored, and time dependency is not considered.

[0012] The present invention adopts the following technical solutions to solve the above technical problems: the ETID2C-DTA structure of the drug target affinity prediction interaction and dilated causal convolution enhanced network includes:

[0013] The lower structure decomposition module SDM is used to take drugs, proteins and unlabeled data as input, decompose and process the molecular sequences of proteins, drugs and unlabeled data based on data-driven sequence patterns, obtain substructure sequences, and obtain decomposed substructures;

[0014] An enhanced transformer embedding module ETEM is used to receive the decomposed substructure and mine the semantic relationship between the substructures of the decomposed substructure. The enhanced transformer embedding module ETEM embeds a learnable substructure lookup dictionary and learns all decomposed substructures of the same molecule to modify the embedded input substructure. The enhanced transformer embedding module ETEM also includes: a transformer encoder to enhance the embedding context substructure information. The enhanced transformer embedding module ETEM is connected to the lower structure decomposition module SDM.

[0015] The interaction module IM is used to perform learning and modeling operations based on the interactions between substructures in the semantic relationship between substructures, simulate the interactions between paired substructures and extract neighborhood interaction information to capture high-order interactions and extract drug and target structural features. The interaction module IM is connected to the enhanced transformer embedding module ETEM;

[0016] The dilated causal convolution module DCCM is used to expand the receptive field through the dilated convolution layer, capture the temporal information between drugs and targets, and perform non-recursive parallel computing using the dilated causal convolution module DCCM to detect and extract amino acid features.

[0017] The present invention considers the influence of unlabeled data and the properties of drug substructures on DTA prediction at the same time, and for the first time in the field of DTA, expands the causal convolutional network and combines it with the existing transformer and CNN for predictive modeling. The present invention uses the DTA prediction model IEC2T-DTA to fully explore the characteristics of protein and drug substructures, and simultaneously learns the labeled and unlabeled data, which can more accurately predict drug target affinity. The present invention can quickly and extensively perform preliminary screening of compounds, minimizing resource utilization and cost. Secondly, our enhanced encoder can more effectively mine the semantic relationships in the substructures of a large amount of unlabeled drug data. The temporal relationship is captured by expanding causal convolution, and complex feature representation is provided to handle long-term dependencies, thereby improving the prediction accuracy. Compared with the above-mentioned deep learning method, the present invention proposes an expanded causal convolution module for the first time, which can more effectively extract the characteristics of drug and target protein sequences, and has better prediction effect. The present invention can simultaneously learn labeled and unlabeled data, and can more accurately predict drug target binding affinity.

[0018] Specifically, the present invention designs a knowledge-inspired substructure decomposition module to mine the features of substructures; secondly, an enhanced transformer encoder module is constructed to effectively mine the semantic relationship between substructures extracted from a large amount of unlabeled drug data; in addition, an interaction module is constructed to better capture drug information to achieve more accurate and explainable DTA prediction; finally, the temporal relationship is captured by the dilated causal convolution module, and a complex feature representation is provided to handle long-term dependencies to improve the prediction accuracy. Experimental results on three datasets show that the model adopted by the present invention has better prediction performance than the baseline model.

[0019] In a more specific technical solution, in the lower structure decomposition module SDM, a vocabulary set V is initialized; and a tokenization operation is performed on a preset drug / protein corpus according to the vocabulary set V to obtain a tokenization set W;

[0020] Scan the tokenized set W and find the most frequent consecutive tokens (A, B);

[0021] Update each most frequent consecutive token (A, B) in the tokenized set W with the new token (AB) and add the most frequent consecutive token (A, B) to the vocabulary set V;

[0022] Continue scanning, identifying, and updating until no most frequent consecutive tokens (A, B) meet the preset update stop condition, merge the frequent substructure sequences into frequent tokens, and decompose the remaining substructure sequences into tokens of preset length;

[0023] The following logic is used to express the substructure sequence of drugs and proteins:

[0024] C=(C1,C2,…,C k )

[0025] Where k is the size of the substructure sequence.

[0026] In a more specific technical solution, the vocabulary set V includes: different amino acid tokens and SMILES string characters.

[0027] In a more specific technical solution, the update stop condition includes: the most frequent consecutive tokens (A, B) are higher than a threshold θ, or the most frequent consecutive tokens (A, B) are higher than the size of the vocabulary set V.

[0028] In a more specific technical solution, in the enhanced transformer embedding module ETEM, for each input drug-target pair, the corresponding substructure sequence C p , C d Convert to matrix M p 、M d , where k and l are the total sizes of the drug and protein substructures, respectively, and Θ p and θ d is the maximum length of protein and drug substructure sequences, and each column of the matrix is a one-hot vector, corresponding to the substructure index of the i-th substructure of the protein sequence and the j-th substructure of the drug sequence, respectively;

[0029] Using the learnable search matrix Generating contextual embeddings of proteins and drugs and

[0030]

[0031] Where θ is the size of the hidden embedding of each substructure;

[0032] Using the Lookup Matrix Building positional embeddings

[0033]

[0034] In the formula, and is a one-hot vector.

[0035] Sum the context embedding and position embedding to get the independent embedding and

[0036]

[0037]

[0038] Using the Transformer encoder layer, we embed and Enhanced embedding.

[0039]

[0040] This paper proposes a data-driven substructure decomposition module to discover repeated subsequences in drug and protein data. In addition, an enhanced transformer embedding module is constructed to obtain enhanced contextual embeddings for each substructure.

[0041] In a more specific technical solution, in the interaction module IM, the interaction between a single drug and a target pair is modeled, and the following logic is used to process each subsequence i in the protein and each subsequence j in the drug to obtain a tensor interaction graph I:

[0042]

[0043] Where e is the dot product, Θ d and θ p are the lengths of drug and protein subsequences, respectively, and Φ is the output size of l;

[0044] Check the tensor interaction graph to get the final result contribution substructure;

[0045] Model the interaction between adjacent regions, use sliding convolution kernel to extract features from input data, and obtain extracted feature maps;

[0046] The extracted feature map is nonlinearly mapped to obtain the output of the input drug-target pair:

[0047] Q=CNN(I) (10)

[0048] In the formula, CNN represents the calculation of convolutional neural network.

[0049] In a more specific technical solution, the dilated causal convolution module DCCM includes: no less than 2 parallel dilated causal convolution modules to capture and output long-term dependencies.

[0050] In a more specific technical solution, in the dilated causal convolution module DCCM, the dilated causal convolution layers are stacked in increasing order:

[0051]

[0052] Where ζ1, ζ2, b and c are model parameters, g(·) is the activation function of the output, and σ(·) is the Sigmoid function, which determines the ratio of information passed to the next layer;

[0053] Will Flatten to a vector; use a linear layer parameterized by the weight matrix W0 and the bias vector b0:

[0054]

[0055] in,

[0056] This paper proposes a dilated causal convolution module that takes into account the temporal dependence of drug targets and can capture the association patterns between drugs and targets at different scales to effectively extract features in drug molecules and target protein sequences. An interaction is also constructed to pair drugs with protein substructures to capture higher-order interactions.

[0057] In a more specific technical solution, in the dilated causal convolution module DCCM, a joint optimization operation is performed through binary classification loss:

[0058] Loss = Y truth log(P)+(1-Y truth )log(1-P) (13)

[0059] Where Y truth is the true value.

[0060] In more specific technical solutions, the drug-target affinity prediction interaction and dilated causal convolution enhancement methods include:

[0061] S1. Taking drugs, proteins and unlabeled data as input, based on data-driven sequence patterns, decompose and process the molecular sequences of proteins, drugs and unlabeled data to obtain substructure sequences and decomposed substructures;

[0062] S2, receiving the decomposed substructure, mining the semantic relationship between the substructures of the decomposed substructure, the enhanced transformer embedding module ETEM embedding can learn the substructure lookup dictionary, learn all the decomposed substructures of the same molecule to modify the embedded input substructure, wherein the enhanced transformer embedding module ETEM also includes: a transformer encoder, which is used to enhance the embedded context substructure information;

[0063] S3, based on the interaction between substructures in the semantic relationship between substructures, learning and modeling operations are performed to simulate the interaction between paired substructures and extract neighborhood interaction information to capture high-order interactions and extract drug and target structural features. The interaction module IM is connected to the enhanced transformer embedding module ETEM;

[0064] S4. By expanding the convolutional layer, the receptive field is expanded to capture the temporal information between drugs and targets, and the dilated causal convolution module DCCM is used for non-recursive parallel computing to detect and extract amino acid features.

[0065] Compared with the prior art, the present invention has the following advantages:

[0066] The present invention considers the influence of unlabeled data and the properties of drug substructures on DTA prediction, and for the first time expands the causal convolutional network in the DTA field and combines it with the existing transformer and CNN for predictive modeling. The present invention uses the DTA prediction model IEC2T-DTA to fully explore the characteristics of protein and drug substructures, and simultaneously learns labeled and unlabeled data, which can more accurately predict drug target affinity.

[0067] The present invention designs a knowledge-inspired substructure decomposition module to mine the features of substructures; secondly, constructs an enhanced transformer encoder module to effectively mine the semantic relationship between substructures extracted from a large amount of unlabeled drug data; in addition, constructs an interaction module to better capture drug information to achieve more accurate and explainable DTA prediction; finally, the dilated causal convolution module is used to capture the temporal relationship and provide complex feature representation to handle long-term dependencies to improve the prediction accuracy. Experimental results on three datasets show that the model adopted by the present invention has better prediction performance than the baseline model.

[0068] This paper proposes a data-driven substructure decomposition module to discover repeated subsequences in drug and protein data. In addition, an enhanced transformer embedding module is constructed to obtain enhanced contextual embeddings for each substructure.

[0069] This paper proposes a dilated causal convolution module that takes into account the temporal dependence of drug targets and can capture the association patterns between drugs and targets at different scales to effectively extract features in drug molecules and target protein sequences. An interaction is also constructed to pair drugs with protein substructures to capture higher-order interactions.

[0070] The present invention solves the technical problems existing in the prior art of inaccurate and difficult to interpret results, neglect of unlabeled molecular data, and failure to consider time dependency. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic diagram of the structure of the drug target affinity prediction model IEC2T-DTA in the interactive and dilated causal convolution enhanced network for drug target affinity prediction according to Example 1 of the present invention;

[0072] Figure 2 This is a schematic diagram of the structure of the interaction module IM according to Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are 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 are within the scope of protection of the present invention.

[0074] Example 1

[0075] like Figure 1 As shown, the drug target affinity prediction interaction and dilated causal convolution enhanced network provided by the present invention adopts the overall structure of ETID2C-DTA to predict drug target affinity. In this embodiment, the overall structure of ETID2C-DTA includes but is not limited to: a lower structure decomposition module SDM1, an enhanced transformer embedding module ETEM2, an interaction module IM3, and a dilated causal convolution module DCCM4.

[0076] In this embodiment, the lower structure decomposition module SDM1 is used to take drugs, proteins and a large amount of unlabeled data as input, and decompose them into a set of explicit substructure sequences through the lower structure decomposition module SDM1; the decomposed substructures are fed to the enhanced transformer embedding module ETEM2, and the semantic relationship between the substructures is extracted by the enhanced transformer embedding module ETEM2; the interaction module IM3 performs more complex learning and modeling on the interaction between substructures to capture high-order interactions, which can effectively extract features from the structures of drugs and targets, and help improve the model's understanding of DTA; the dilated causal convolution module DCCM4 uses a dilated convolution layer to expand its receptive field, which can effectively capture the temporal information between drugs and targets, and then effectively capture long-term dependencies. Specifically, DCCM adopts a lightweight structure and can achieve good results when processing long sequence data. Finally, it is output through the decoder.

[0077] In this embodiment, the aforementioned lower structure decomposition module SDM1 decomposes the molecular sequences of proteins and drugs into substructures based on data-driven sequence patterns, which is used to discover recurring subsequences in drug and protein databases. The lower structure decomposition module SDM1 hierarchically decomposes each protein / drug sequence into subsequences, smaller subsequences, and single atoms or amino acid symbols.

[0078] In this embodiment, a vocabulary set V consisting of different amino acid tokens or SMILES string characters is initialized, and the entire drug / protein corpus is tokenized according to the given vocabulary set V; the tokenized set is recorded as W, and W is scanned to find the most frequent consecutive tokens (A, B); SDM updates each (A, B) in the token set W with a new token (AB) and adds this token to the vocabulary set V; then the scanning, identification, and updating process are repeated until no frequent tokens are above the threshold θ or above the size of V. Through this operation, frequent subsequences are merged into one token, and less frequent subsequences are decomposed into a set of shorter tokens; for a drug / protein, SDM will generate a substructure drug / target protein sequence C of size k = (C1, C2, ..., C k ), where C i (1≤i≤k) comes from the set V.

[0079] In this embodiment, IEC2T-DTA uses the substructure decomposition module to convert the input drugs and targets into explicit substructure sequences C d and C pSequence. Unlike previous substructure fingerprinting methods, SDM is more interpretable. Because many substructures are subsets of other substructures, it is difficult to know which substructure affects the results. In contrast, SDM decomposes each drug molecule into discrete, medium-sized substructure partitions, allowing the use of large amounts of unlabeled data for substructure mining. In addition, SDM can capture basic and meaningful biomedical semantics because the generated substructures are related to the basic units of frequently occurring drugs and proteins.

[0080] In this embodiment, an enhanced transformer embedding module ETEM2 is constructed to fully explore the chemical semantics of substructures. The enhanced transformer embedding module ETEM2 first embeds a learnable substructure lookup dictionary; the embedded context substructure information is enhanced through the transformer encoder. Transformer is built based on the self-attention mechanism. While learning the complex relationship between drugs and target sequences, it can also analyze the attention weights to determine which parts of drug molecules and target protein sequences play a key role in binding affinity prediction.

[0081] In this example, the enhanced transformer embedding module ETEM2 modifies each input substructure of the embedding by learning from all substructures of the same molecule. Specifically, for each input drug-target pair, the corresponding substructure sequence C p and C d Convert to two matrices M p and M d , where k and l are the total sizes of the drug and protein substructures, respectively, and Θ p and θ d is the maximum length of protein and drug substructure sequences, each column and is a one-hot vector corresponding to the substructure index of the i-th substructure of the protein sequence and the j-th substructure of the drug sequence, respectively. For each protein and drug context embedding and This can be done through a learnable lookup matrix and Generate, the specific operations are as follows:

[0082]

[0083]

[0084] where v is the size of the hidden embedding of each substructure.

[0085] In this embodiment, since IEC2T-DTA uses sequential subsequences, a lookup matrix is ​​used. and To construct position embedding and

[0086]

[0087] in, and is a one-hot vector.

[0088] In this embodiment, the final embedding and is generated by the sum of context embedding and position embedding:

[0089]

[0090]

[0091] In this embodiment, a set of independent substructure embeddings can be output. However, there are chemical relationships between these substructures, and in order to capture this contextual information, IEC2T-DTA uses Transformer encoder layers to further enhance the embedding.

[0092]

[0093] In this embodiment, an interaction module IM3 is constructed based on a convolutional neural network to simulate the interaction of paired substructures and extract neighborhood interaction information. In order to simulate the pairing interaction, for each subsequence i in the protein and each subsequence j in the drug, the following operations are performed:

[0094]

[0095] where l is the dot product, which is useful for measuring the strength of interactions between individual target drug substructure pairs.

[0096] Therefore, we get a tensor I, where Θ d and θ p are the lengths of the drug and protein subsequences, respectively, and Φ is the output size of l. Each column in I considers the interaction between individual substructures of the protein and the drug. Since the output of the dot product is one-dimensional for each pair, I becomes a 2D interaction graph. If the value in the graph is high, it will be activated in the downstream layer and have a higher DTA interaction possibility. Through end-to-end learning, if a pair of substructures do interact, they will have a higher interaction score at the corresponding substructure pair position in the interaction graph. Therefore, by examining this graph, it is directly possible to conclude which substructure pairs contribute to the final result.

[0097] In addition, the adjacent substructures of proteins and drugs will also affect each other. Therefore, in addition to modeling single pairwise interactions, it is also necessary to model the interactions of adjacent regions. The interaction module IM3 is built based on CNN to effectively capture local features. Feature extraction is performed on the input data through a sliding convolution kernel to form a feature map. The extracted features are then nonlinearly mapped to increase the network's expressive power and learn complex nonlinear relationships. Therefore, the output of the input drug-target pair can be expressed as follows:

[0098] Q=CNN(I) (23)

[0099] Among them, CNN represents the calculation of convolutional neural network.

[0100] The dilated causal convolution module DCCM4 can effectively extract features in drug molecules and target protein sequences at different scales. By learning these features, IEC2T-DTA can better predict the affinity between drugs and targets. In addition, DTA may be affected by long-range interactions, such as remote action sites in drug molecules or domains in target protein sequences. The dilated causal convolution module DCCM4 has an expanded receptive field that can capture longer sequence dependencies. This enables IEC2T-DTA to better model long-term dependencies between drugs and targets and improve the prediction performance of DTA.

[0101] like Figure 2 As shown, specifically, the dilated causal convolution module DCCM4 contains two parallel dilated causal convolution modules, which are mainly used to capture temporal dependencies and mine them at a higher level of temporal abstraction. As a special case of standard 1D convolution, the dilated causal convolution operation slides on the input by skipping values ​​with a specific step size.

[0102] See also Figure 2 , dilated causal convolution with kernel size 2. Using a dilation factor h, the dilated causal convolution module DCCM4 selects inputs at every h steps and applies a standard 1D convolution to the selected inputs.

[0103] The dilated causal convolutional network uses non-recursive parallel computing to detect and extract the features of amino acids, improve the learning speed, and alleviate the gradient vanishing problem. By stacking the dilated causal convolutional layers in ascending order, the DCCM can capture longer sequences with fewer layers, saving computing resources while improving the accuracy of long-term predictions. The form of DCCM is:

[0104]

[0105] where ζ1, ζ2, b, and c are model parameters, g(·) is the activation function of the output, and σ(·) is the Sigmoid function that determines the rate of information passed to the next layer.

[0106] Finally, Q is first flattened into a vector, and then a linear layer parameterized by the weight matrix W0 and the bias vector b0 is used:

[0107]

[0108] in Furthermore, we jointly optimize via a binary classification loss:

[0109] Loss = Y truth log(P)+(1-Y truth )log(1-P) (26)

[0110] where Y truth is the true value.

[0111] Example 2

[0112] In this example, the prediction performance of the IEC2T-DTA model was verified on three widely used datasets: BIOSNAP, Davis, and BindingDB. It is worth noting that in the BIOSNAP dataset, its subset MINER dataset was used as the main experimental dataset. It includes 4510 drugs and 2181 protein targets. The BIOSNAP dataset only contains positive DTA pairs. For negative pairs, sampling is performed from unseen pairs according to conventional methods to obtain a balanced dataset with equal positive and negative samples. The Davis dataset mainly contains K d BindingDB is a public database containing the measured binding affinities, which contains K values ​​for 4510 drugs and 2181 proteins. d In the experiment, K d DTA pairs with values ​​less than 30 units are considered positive. At the same time, negative DTA pairs with the same number of positive samples as the training set are subsampled. The negative ratio of the dataset is maintained in the validation and test sets. All datasets are divided into training set, validation set and test set in a ratio of 7:1:2. The details of the dataset are shown in Table 1 below:

[0113] Table 1. Statistics of the dataset

[0114]

[0115] In the parameter setting process of this embodiment, the experimental environment can be, for example, a Pytorch environment, using the Adam optimizer to optimize hyperparameters, with an initial learning rate of 0.001 and a discard rate of 0.1. The minimum number of occurrences of drug and protein substructures in the data set is set to 500 in SDM, resulting in 23,532 drug substructures and 16,693 protein substructures. In ETEM, two layers of transformer encoders are used for both drugs and proteins, with an input embedding size of 384, a self-attention head number of 12, and an intermediate dimension of 1536. In addition, the maximum sequence lengths of drugs and proteins are set to 50 and 545, respectively, to cover 95% of drugs and proteins in the data set. The parts exceeding and below the maximum length are cut and filled, respectively. The batch size is set to 64 and run for 30 epochs. The model converges between 8 and 15 epochs. All experiments were performed on a 22vCPU AMD EPYC 7T83 64-Core Processor with a RTX 4090GPU Card.

[0116] In this embodiment, in order to evaluate the performance of the method, the area under the ROC curve (AUROC) and the area under the PRC curve (AUPRC) are used as indicators to measure the performance of binary classification. In addition, sensitivity and specificity indicators are used, where the threshold is the threshold with the best F1 score in the validation set.

[0117] In this example, a baseline model is selected to compare IEC2T-DTA with the following state-of-the-art models:

[0118] LR: This method applies a logistic regression model to the concatenated drug and protein feature vectors.

[0119] DNN: This method uses a three-layer deep neural network (DNN) with a hidden size of 1024 on top of ECFP4 and PSC concatenated vectors.

[0120] GNN-CPI: This method uses a graph neural network to encode drugs and a CNN to encode proteins. The latent vectors are then connected to a neural network for compound-protein interaction prediction.

[0121] DeepDTI: This baseline was originally designed for CPI / DTI prediction and cannot be directly used for DTA tasks. Therefore, we need to change it to a regression task. Specifically, we replace the cross entropy loss function with MSE and set the dimension of the output layer to 1. The rest is consistent with the original paper.

[0122] DeepDTA: DeepDTA trains two 3-layer CNNs using labels / one-hot encodings of compounds and protein sequences to predict DTA tasks. Their CNN model consists of two separate CNN modules that learn features from SMILES compound strings and protein sequences, respectively. The representations of drugs and targets are concatenated and passed to a fully connected layer for DTA prediction.

[0123] DeepConv-DTI: This method uses CNN and global max pooling layers to extract local patterns of various lengths in protein sequences and employs fully connected layers to encode drug molecular fingerprints for DTA prediction.

[0124] In this example, in order to test the competitiveness of the proposed model, the experimental results of IEC2T-DTA and the most advanced models (including classical models and deep learning models) were compared, as shown in Table 2. It can be seen that IEC2T-DTA achieves the best prediction results in AUROC and AUPRC on all data sets. In addition, because the thresholds of different methods are different, the sensitivity and specificity may be different.

[0125] As can be seen from Table 2, the performance of classical methods such as LR is worse than that of deep learning-based methods. This is because classical methods rely heavily on handcrafted features and similarity matrices of drugs and targets. In contrast, deep learning-based methods can capture more information through automatic feature engineering of CNN and GNN. In addition, it is significantly shown that IEC2T-DTA performs better than other methods. The reasons include: compared with DeepCPI, our method comprehensively considers the topological structure and local chemical background, which is conducive to improving the model prediction performance; compared with DeepDTA, we construct a transformer encoding module based on a multi-head attention mechanism to obtain the topological information of drugs, and use advanced embedding technology to provide more contextual information for drug and target modeling. The designed dilated causal convolution can effectively process the sequence data of drugs and targets, and capture the temporal relationship and local dependency between them. Compared with traditional convolution operations, dilated causal convolution not only considers the long-range dependency of sequence data, but also has a larger receptive field and can mine a wider range of contextual information.

[0126]

[0127] Table 2. Performance comparison of different models

[0128] In this embodiment, in order to study the prediction problem of unlabeled drugs and targets, 20% of the drugs and target proteins and drug-target pairs related to these drugs and targets are randomly selected as test sets for experiments, and the results are shown in Table 3 below. It can be seen that the model adopted by the present invention also has excellent prediction performance for unlabeled drugs and proteins, but the performance of the other three baseline models is not much different. This is because the substructure decomposition of IEC2T-DTA can effectively mine substructure features and remove the influence of repeated subsequences on prediction performance. In addition, the enhanced transformer embedding module, interactive module (IM) and diluted causal convolution module also make significant contributions to the excellent performance of the model.

[0129] Table 3. Comparison of prediction performance for unlabeled drugs and proteins on the BIOSNAP dataset (average AUROC values ​​from five random runs)

[0130]

[0131] In this example, the performance of sparse data is studied, where the amount of DTA data is growing rapidly, but in some actual drug discovery, due to budget constraints, there are some new target proteins or drugs with only a few labels. Therefore, in order to simulate the real situation, we conduct experiments on 5%, 10%, 20% and 30% of the data set, and make predictions on the remaining data sets. The results are shown in Table 4 below. It can be seen that IEC2T-DTA is the most robust method. In contrast, the prediction performance of advanced baseline models such as DeepDTI and DeepConv-DTI decreases with the increase of missing data. Once again, the effectiveness of our model is verified.

[0132] Table 4. Comparison of prediction performance in the case of missing data on the BIOSNAP dataset (average AUROC values ​​from five random runs)

[0133] Percentage of missing data (%) DeepDTI DeepDTA DeepConv-DTI IEC2T-DTA 70 0.853±0.004 0.838±0.004 0.845±0.003 0.864±0.003 80 0.828±0.007 0.821±0.008 0.825±0.003 0.843±0.004 90 0.767±0.010 0.787±0.011 0.792±0.004 0.815±0.004 95 0.659±0.011 0.762±0.004 0.726±0.008 0.781±0.005

[0134] In this example, in order to further study the impact of different modules of the proposed model, the following five variants are proposed and compared with IEC2T-DTA on the BIOSNAP dataset. The experimental results are shown in Table 5. The differences between these variants and IEC2T-DTA are:

[0135] w / o CNN: We remove the CNN from the interaction module and feed the output of the interaction map into the decoder.

[0136] w / o transformer: We remove the transformer in the enhanced transformer embedding module and provide both position and content embeddings to the interaction module.

[0137] w / o IM: We remove the interaction module, and keep everything else the same.

[0138] w / o SDM: We remove the substructure decomposition step and conduct the experiments directly.

[0139] w / o DCCM: We remove the dilated causal convolutional module and replace it with a fully connected layer, while keeping everything else unchanged.

[0140] Table 5. Performance comparison of different models on the BIOSNAP dataset

[0141] Different models AUROC AUPRC w / o CNN 0.877±0.003 0.884±0.003 w / o transformer 0.876±0.004 0.873±0.003 w / o IM 0.851±0.003 0.864±0.005 w / o SDM 0.847±0.002 0.861±0.004 w / o DCCM 0.842±0.003 0.858±0.004 IEC2T-DTA 0.895±0.004 0.902±0.003

[0142] In this example, through the ablation experiment results, it can be seen that the dilated causal convolution module has the greatest impact on the prediction performance of the model, followed by the substructure decomposition and interaction modules. At the same time, it can be seen that the influence of other components on IEC2T-DTA. This shows that these components enable our model to effectively capture and utilize high-level drug information, thereby improving the overall performance of drug-target interaction prediction.

[0143] The present invention proposes a novel drug-target affinity prediction method IEC2T-DTA. The method combines substructure decomposition, transformer encoding, CNN-based interaction module and dilated causal convolution to enhance the model's understanding of molecular representation. First, substructure decomposition can fully mine drug target features and help establish associations between drugs and targets. Secondly, transformer encoding can process sequence data of molecular structures while capturing long-range dependencies and important features in molecular structures. The interaction module is used to capture higher-order interactions. Finally, dilated causal convolution can effectively capture local and global features in molecular structures, and can process input sequences of different lengths, which helps to improve the performance of the prediction model. Experiments on three public datasets demonstrate the superiority of IEC2T-DTA in dealing with DTA prediction problems, with the best affinity prediction performance.

[0144] In summary, the present invention considers the influence of unlabeled data and the properties of drug substructures on DTA prediction, and for the first time expands the causal convolutional network in the field of DTA and combines it with the existing transformer and CNN for predictive modeling. The present invention uses the DTA prediction model IEC2T-DTA to fully explore the characteristics of protein and drug substructures, and simultaneously learns labeled and unlabeled data, which can more accurately predict drug target affinity.

[0145] The present invention designs a knowledge-inspired substructure decomposition module to mine the features of substructures; secondly, constructs an enhanced transformer encoder module to effectively mine the semantic relationship between substructures extracted from a large amount of unlabeled drug data; in addition, constructs an interaction module to better capture drug information to achieve more accurate and explainable DTA prediction; finally, the dilated causal convolution module is used to capture the temporal relationship and provide complex feature representation to handle long-term dependencies to improve the prediction accuracy. Experimental results on three datasets show that the model adopted by the present invention has better prediction performance than the baseline model.

[0146] This paper proposes a data-driven substructure decomposition module to discover repeated subsequences in drug and protein data. In addition, an enhanced transformer embedding module is constructed to obtain enhanced contextual embeddings for each substructure.

[0147] This paper proposes a dilated causal convolution module that takes into account the temporal dependence of drug targets and can capture the association patterns between drugs and targets at different scales to effectively extract features in drug molecules and target protein sequences. An interaction is also constructed to pair drugs with protein substructures to capture higher-order interactions.

[0148] The present invention solves the technical problems existing in the prior art of inaccurate and difficult to interpret results, neglect of unlabeled molecular data, and failure to consider time dependency.

[0149] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. Drug-target affinity prediction interaction and dilated causal convolutional enhanced network, characterized by, The ETID2C-DTA structure of the network includes: The lower structure decomposition module SDM is used for taking the drug, protein and unlabeled data as input, decomposing and processing the molecular sequence of the protein, the drug and the unlabeled data based on the data-driven sequence pattern, obtaining the substructure sequence, and obtaining the decomposed substructure; An enhanced transformer embedding module ETEM is used to receive the decomposed substructure, mine the semantic relationship between the substructures of the decomposed substructure, and embed a learnable substructure lookup dictionary to learn all the decomposed substructures of the same molecule to modify the embedded input substructure, wherein the enhanced transformer embedding module ETEM also includes: a transformer encoder, which is used to enhance the embedded context substructure information, and the enhanced transformer embedding module ETEM is connected to the lower structure decomposition module SDM; An interaction module IM is used to perform learning and modeling operations according to the interactions between substructures in the semantic relationship between the substructures, simulate the interactions between paired substructures and extract neighborhood interaction information to capture high-order interactions and extract drug and target structural features. The interaction module IM is connected to the enhanced transformer embedding module ETEM; The dilated causal convolution module DCCM is used to expand the receptive field through the dilated convolution layer, capture the temporal information between the drug and the target, and use the dilated causal convolution module DCCM to perform non-recursive parallel calculation to detect and extract amino acid features.

2. The drug target affinity prediction interactive and dilated causal convolutional enhanced network according to claim 1, characterized in that: In the lower structure decomposition module SDM, a vocabulary set V is initialized; and a tokenization operation is performed on a preset drug / protein corpus according to the vocabulary set V to obtain a tokenization set W; Scan the tokenized set W to find the most frequent consecutive tokens (A, B); updating each of the most frequent consecutive tokens (A, B) in the tokenized set W with a new token (AB), and adding the most frequent consecutive tokens (A, B) to the vocabulary set V; Continue scanning, identifying and updating until no more of the most frequent consecutive tokens (A, B) satisfy the preset update stop condition, merge the frequent substructure sequences into frequent tokens, and decompose the remaining substructure sequences into tokens of preset length; The substructure sequence of the drug and the protein is expressed using the following logic: C=(C1,C2,...,C k ) Wherein, k is the size of the substructure sequence.

3. The drug target affinity prediction interactive and dilated causal convolutional enhanced network according to claim 2, characterized in that: The vocabulary set V includes: different amino acid tokens and SMILES string characters.

4. The drug target affinity prediction interactive and dilated causal convolutional enhanced network according to claim 2, characterized in that: The update stop condition includes: the most frequent consecutive tokens (A, B) is higher than a threshold θ, or the most frequent consecutive tokens (A, B) is higher than the size of the vocabulary set V.

5. The interactive and dilated causal convolutional enhanced network for drug target affinity prediction according to claim 1, characterized in that: In the enhanced transformer embedding module ETEM, for each input drug-target pair, the corresponding substructure sequence C p , C d Convert to matrix M p 、M d , where k and l are the total sizes of the drug and protein substructures, respectively, and Θ p and θ d is the maximum length of protein and drug substructure sequences, and each column of the matrix is a one-hot vector, corresponding to the substructure index of the i-th substructure of the protein sequence and the j-th substructure of the drug sequence, respectively; Using the learnable search matrix Generating contextual embeddings of proteins and drugs and In the formula, is the size of the hidden embedding of each substructure; Using the Lookup Matrix Building positional embeddings In the formula, and is a one-hot vector. Sum the context embedding and the position embedding to get an independent embedding and Using the Transformer encoder layer, the independent embedding and Enhanced embedding.

6. The interactive and dilated causal convolutional enhanced network for drug target affinity prediction according to claim 1, characterized in that: In the interaction module IM, the interaction between a single drug and a target pair is modeled, and each subsequence i in the protein and each subsequence j in the drug are processed using the following logic to obtain a tensor interaction graph I: In the formula, is the dot product, Θ d and θ p are the lengths of drug and protein subsequences respectively, and Φ is The output size of Checking the tensor interaction graph to obtain the final result contribution substructure; Model the interaction between adjacent regions, use sliding convolution kernel to extract features from input data, and obtain extracted feature maps; The extracted feature map is nonlinearly mapped to obtain the output of the input drug-target pair: Q=CNN(I) (36) In the formula, CNN represents the calculation of convolutional neural network.

7. The interactive and dilated causal convolutional enhanced network for drug target affinity prediction according to claim 1, characterized in that: The dilated causal convolution module DCCM includes: no less than 2 parallel dilated causal convolution modules, which are used to capture and output long-term dependencies.

8. The interactive and dilated causal convolutional enhanced network for drug target affinity prediction according to claim 7, characterized in that: In the dilated causal convolution module DCCM, the dilated causal convolution layers are stacked in increasing order: Where ζ1, ζ2, b and c are model parameters, g(·) is the activation function of the output, and σ(·) is the Sigmoid function, which determines the ratio of information passed to the next layer; Will Flatten to a vector; use a linear layer parameterized by the weight matrix W0 and the bias vector b0: in, 9. The interactive and dilated causal convolutional enhanced network for drug target affinity prediction according to claim 1, characterized in that: In the dilated causal convolution module DCCM, a joint optimization operation is performed through binary classification loss: Loss=Y truth log(P)+(1-Y truth )log(1-P) (39) Where Y truth is the true value.

10. Drug-target affinity prediction interaction and dilated causal convolution enhancement method, characterized in that, The method comprises: S1, taking drugs, proteins and unlabeled data as input, decomposing and processing the molecular sequences of the proteins, drugs and unlabeled data based on data-driven sequence patterns, obtaining substructure sequences, and obtaining decomposed substructures; S2, receiving the decomposed substructure, mining the semantic relationship between the substructures of the decomposed substructure, the enhanced transformer embedding module ETEM embeds a learnable substructure lookup dictionary, learns all the decomposed substructures of the same molecule to modify the embedded input substructure, wherein the enhanced transformer embedding module ETEM also includes: a transformer encoder for enhancing the embedded context substructure information; S3, according to the interaction between substructures in the semantic relationship between substructures, learning and modeling operations are performed to simulate the interaction between paired substructures and extract neighborhood interaction information to capture high-order interactions and extract drug and target structural features. The interaction module IM is connected to the enhanced transformer embedding module ETEM; S4. By expanding the convolution layer, the receptive field is expanded to capture the temporal information between the drug and the target, and the dilated causal convolution module DCCM is used to perform non-recursive parallel calculation to detect and extract amino acid features.

Images