Drug target affinity prediction interaction and expansion causal convolution enhanced network and method

Abstract

The application provides a drug target affinity prediction interaction and expansion causal convolution enhanced network and method, and the network comprises: a knowledge heuristic-based substructure decomposition module to mine the characteristics of the substructure; an enhanced transformer encoder module is constructed to effectively mine the semantic relationship between the substructures extracted from a large amount of unlabeled drug data; an interaction module is constructed to better capture drug information to achieve more accurate and interpretable DTA prediction; a time series relationship is captured through an expansion causal convolution module, and complex feature representation is provided to process long-term dependencies to improve prediction accuracy. The application solves the technical problems of inaccurate and difficult-to-explain results, neglecting unlabeled molecular data and not considering time dependence.

Description

Technical Field

[0001] This invention relates to the field of drug target affinity prediction, specifically to interactive and extended causal convolutional enhancement networks and methods for drug target affinity prediction. Background Technology

[0002] Accurate prediction of drug-target affinity (DTA) is crucial in drug discovery and development. DTA reveals the strength of protein-ligand interactions between a drug and its target, providing further information about drug-target interactions. Furthermore, DTA is fundamental to finding new drugs (i.e., virtual screening) and identifying new indications for existing drugs (i.e., drug retargeting), as the therapeutic efficacy of drug compounds is detected by examining DTA. Therefore, there is a need to develop effective techniques to determine the potential link between drugs and targets.

[0003] Existing methods for predicting drug-target affinity are mainly divided into two categories: predicting binary DTA and predicting drug-target affinity values. The first category is direct DTA prediction, which determines whether a drug and target interact through known drug-target interactions and their feature vectors, or by extracting pairwise similarities by combining the drug's chemical or structural descriptor with the target's descriptor. Most methods, based on the feature vectors of the drug or target, or the pairwise similarities between drugs or targets, formulate DTA prediction as a classification problem; for example, if a drug-target pair is known to interact, it can be considered positive, otherwise negative. Therefore, models built using methods such as deep neural networks (DNNs), deep belief networks (DBNs), and convolutional neural networks (CNNs) can be trained on positive and negative drug-target pairs. The features or pairwise similarities of these drug-target pairs can be constructed from the chemical or structural features of the drug or target.

[0004] However, existing research typically learns molecular characterizations and makes predictions based on the overall molecular structure of drugs and proteins, neglecting the fact that interactions only involve relevant substructures of the drug and protein. Whole-structure molecular characterization introduces noise, affecting prediction accuracy. Furthermore, the learned characterizations are difficult to interpret because they do not provide a viable pathway to explain which substructures of the drug and protein contribute to the interaction. In addition, these studies focus on labeled drug data and limit their scope to thousands of drugs and proteins, ignoring the vast amount (e.g., millions) of available label-free biomedical data. Moreover, these methods lack a large receptive field, ignore long-term dependencies between drug targets, and struggle to capture broader relationships between drugs and targets, thus making it even more difficult to provide richer feature representations and more accurate predictions.

[0005] Currently, numerous prediction methods have emerged in the field of drug target affinity prediction, mainly categorized into structure-based and structure-free methods. Structure-based methods rely on the three-dimensional structures of drug molecules and target proteins to predict their interactions. However, the effectiveness of these methods is limited when structural information is lacking. Therefore, structure-free methods have been a focus of research, with similarity-based methods utilizing known drug-target similarity information to predict drug target affinity. Examples include KronRLS and SimBoost. Feature-based methods utilize the characteristics of drug-protein interactions to predict the strength or affinity of drug molecules binding to specific protein targets. These methods typically rely on machine learning or computer simulation techniques, analyzing features such as structure, physicochemical properties, and biological activity between drug molecules and protein targets to construct predictive models. Furthermore, deep learning-based methods have achieved significant predictive results, often employing different networks (e.g., convolutional neural networks (CNNs) and recurrent neural networks (RNNs)) to build models for DTA prediction. For example, DeepAffinity uses composite SMILES or protein SPS representations for unsupervised pre-training of a seq2seq RNN model. GeneralizedDTA is designed specifically 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. 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 the DTA task. Nguyen et al. used transfer learning from chemical-chemical interaction (CCI) and protein-protein interaction (PPI) tasks to enhance drug-target affinity prediction.

[0006] The aforementioned similarity-based existing schemes have the following drawbacks: they lack reliable data, and similarity methods rely on a large number of reliable known datasets to establish relationships between drug targets; at the same time, they cannot accurately identify specific binding sites.

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

[0008] The aforementioned existing deep learning-based solutions have the following drawbacks: poor performance in feature extraction and prediction of drug and target protein sequences.

[0009] The existing invention patent application document CN114038517A, entitled "A Self-Supervised Graph Neural Network Pre-training Method Based on Contrastive Learning," includes the following steps: preprocessing compound molecules from a public database to screen organic molecules; decomposing and extracting the structures of the screened organic molecules, using the resulting substructures as identifiers, and constructing a corpus of substructures; treating the decomposed substructures as supernodes and constructing corresponding subgraph data, where the subgraph data and the original molecular graph data form positive sample pairs, and randomly selecting several subgraph data and the original molecular graph data to form negative sample pairs; constructing a self-supervised learning model based on an attention-based graph convolutional neural network, multi-level gated recurrent units, and multilayer perceptron modules; and inputting all positive and negative sample pairs into the self-supervised learning model for pre-training and saving them for fine-tuning in downstream tasks. However, in the aforementioned existing scheme, in the field of drug target prediction, graph neural networks struggle to obtain high-quality labeled datasets, and the computational complexity is high. Using this scheme requires a significant amount of time for prediction, offering little advantage in practical applications.

[0010] In summary, existing technologies suffer from technical problems such as inaccurate and difficult-to-interpret results, neglect of unlabeled molecular data, and failure to consider time-dependent relationships. Summary of the Invention

[0011] The technical problem to be solved by this invention is: how to solve the technical problems of inaccurate and difficult-to-interpret results, ignoring unlabeled molecular data, and not considering time dependence in the prior art.

[0012] This invention solves the above-mentioned technical problems by employing the following technical solution: The ETID2C-DTA structure of the drug target affinity prediction interaction and extended causal convolutional enhancement network includes:

[0013] The lower structure decomposition module (SDM) takes drugs, proteins, and unlabeled data as input and, based on a data-driven sequence mode, decomposes and processes the molecular sequences of proteins, drugs, and unlabeled data to obtain daughter structure sequences and decomposed daughter structures.

[0014] The Enhanced Transformer Embedding Module (ETEM) receives decomposed substructures, mines semantic relationships between substructures, and learns a substructure lookup dictionary to learn all decomposed substructures of the same molecule to modify the embedded input substructure. ETEM also includes a transformer encoder to enhance the embedded context substructure information. ETEM is connected to the Substructure Decomposition Module (SDM).

[0015] The interaction module IM is used to learn and model based on the interaction between substructures in the semantic relationship between substructures, simulate the interaction of paired substructures and extract neighborhood interaction information to capture higher-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 Convolutional Module (DCCM) is used to expand the receptive field by dilating the convolutional layers, thereby capturing the temporal information between the drug and the target. The DCCM is used for non-recursive parallel computation to detect and extract amino acid features.

[0017] This invention, considering both unlabeled data and the influence of drug substructure properties on DTA prediction, is the first to combine an extended causal convolutional network with existing transformers and CNNs for predictive modeling in the DTA field. Utilizing the IEC2T-DTA prediction model, this invention can fully mine the features of protein and drug substructures and learn from both labeled and unlabeled data simultaneously, resulting in more accurate predictions of drug target affinity. This invention enables rapid and extensive preliminary screening of compounds, minimizing resource utilization and cost. Furthermore, our enhanced encoder can more effectively mine semantic relationships within the substructures of large amounts of unlabeled drug data. By using extended causal convolution to capture temporal relationships and providing complex feature representations to handle long-term dependencies, prediction accuracy is improved. Compared to the aforementioned deep learning methods, this invention is the first to propose an extended causal convolution module, which can more effectively extract features from drug and target protein sequences and achieve better prediction results. This invention can learn from both labeled and unlabeled data simultaneously, resulting in more accurate predictions of drug target binding affinity.

[0018] Specifically, this invention designs a knowledge-inspired substructure decomposition module to mine substructure features; secondly, it constructs an enhanced transformer encoder module to effectively mine semantic relationships between substructures from a large amount of unlabeled drug data; furthermore, it constructs an interaction module to better capture drug information for more accurate and interpretable DTA prediction; finally, it captures temporal relationships by using an expanded causal convolution module and provides complex feature representations to handle long-term dependencies, thereby improving prediction accuracy. Experimental results on three datasets show that the model used in this invention has superior prediction performance compared to the baseline model.

[0019] In a more specific technical solution, in the lower structure decomposition module SDM, the vocabulary set V is initialized; and a tokenization operation is performed on the pre-set drug / protein corpus based on the vocabulary set V to obtain the tokenized 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 a new token (AB), and add the most frequent consecutive token (A,B) to the vocabulary set V;

[0022] Continue scanning, identification, and updating operations until no most frequent consecutive token (A,B) satisfies the preset update stop condition. Merge frequent substructure sequences into frequent tokens and decompose the remaining substructure sequences into tokens of preset length.

[0023] Using the following logic, express the substructure sequences of drugs and proteins:

[0024] C = (C1, C2, ..., C k )

[0025] In the formula, k is the size of the substructure sequence.

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

[0027] In a more specific technical solution, the update stopping conditions include: the most frequent consecutive tokens (A,B) are higher than the 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, within the enhanced transformer embedding module ETEM, for each input drug-target pair, the corresponding substructure sequence C is... p C d Transform into matrix M p M d Where k and l are the total size of the drug and protein substructure, respectively, and Θ p and Θ d It represents the maximum length of the protein and drug substructure sequence, and the matrix for each column. These are one-hot vectors, corresponding to the substructure indices of the i-th substructure of the protein sequence and the j-th substructure of the drug sequence, respectively.

[0029] Using Learnable Lookup Matrices Contextual embeddings for generating proteins and drugs and

[0030]

[0031] In the formula, θ is the size of the hidden embedding of each substructure;

[0032] Using a lookup matrix Constructing position embeddings

[0033]

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

[0035] Summing the context embeddings and position embeddings yields the independent embeddings. and

[0036]

[0037]

[0038] Using the Transformer encoder layer for independent embeddings and Enhanced embedding is performed.

[0039]

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

[0041] In a more specific technical solution, within the interaction module IM, the interaction between individual drug-target pairs is modeled. Using the following logic, each subsequence i in the protein and each subsequence j in the drug are processed to obtain the tensor interaction graph I:

[0042]

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

[0044] Examine the tensor interaction graph to obtain the final result contribution substructure;

[0045] The interaction between neighboring regions is modeled, and the input data is extracted using a sliding convolution kernel to obtain an extracted feature map;

[0046] The extracted feature map is then subjected to a nonlinear mapping to obtain the output of the input drug-target pair:

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

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

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

[0050] In a more specific technical solution, within the Dilated Causal Convolutional Module (DCCM), the dilated causal convolutional layers are stacked in ascending order:

[0051]

[0052] In the formula, 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 Flattened into vectors; using a linear layer parameterized by the weight matrix W0 and the bias vector b0:

[0054]

[0055] in,

[0056] This invention proposes an extended causal convolution module that considers the time dependence of drug targets, enabling the capture of association patterns between drugs and targets at different scales to effectively extract features from 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, the dilated causal convolutional module (DCCM) performs joint optimization operations using binary classification loss:

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

[0059] In the formula, Y truth It is the actual value.

[0060] In more specific technical solutions, drug target affinity prediction interactions and expanded causal convolution enhancement methods include:

[0061] S1. Using drugs, proteins, and unlabeled data as input, based on a data-driven sequence mode, the molecular sequences of proteins, drugs, and unlabeled data are decomposed and processed to obtain the daughter structure sequence and the decomposed daughter structure.

[0062] S2. Receive the decomposed substructures, mine the semantic relationships between the substructures of the decomposed substructures, and the enhanced transformer embedding module ETEM can learn the substructure lookup dictionary and learn all the 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 embedded context substructure information.

[0063] S3. Based on the interaction between substructures in the semantic relationship between substructures, learn and model the interaction of paired substructures and extract neighborhood interaction information to capture higher-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 extended to capture the temporal information between the drug and the target. The extended causal convolution module DCCM is used for non-recursive parallel computation to detect and extract amino acid features.

[0065] The present invention has the following advantages over the prior art:

[0066] This invention, considering both unlabeled data and the influence of drug substructure properties on DTA prediction, is the first to extend causal convolutional networks in the DTA field and combine them with existing transformers and CNNs for predictive modeling. Utilizing the DTA prediction model IEC2T-DTA, this invention can fully mine the features of protein and drug substructures and learn from both labeled and unlabeled data simultaneously, enabling more accurate prediction of drug target affinity.

[0067] This invention designs a knowledge-inspired substructure decomposition module to mine substructure features; secondly, it constructs an enhanced transformer encoder module to effectively mine semantic relationships between substructures from a large amount of unlabeled drug data; furthermore, it constructs an interaction module to better capture drug information for more accurate and interpretable DTA prediction; finally, it captures temporal relationships by using an expanded causal convolution module and provides complex feature representations to handle long-term dependencies, thereby improving prediction accuracy. Experimental results on three datasets show that the model used in this invention has superior prediction performance compared to the baseline model.

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

[0069] This invention proposes an extended causal convolution module that considers the time dependence of drug targets, enabling the capture of association patterns between drugs and targets at different scales to effectively extract features from drug molecules and target protein sequences. An interaction is also constructed to pair drugs with protein substructures to capture higher-order interactions.

[0070] This invention solves the technical problems of inaccurate and difficult-to-interpret results, neglect of unlabeled molecular data, and failure to consider time dependence in the prior art. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the structure of the drug target affinity prediction model IEC2T-DTA in the drug target affinity prediction interaction and extended causal convolutional enhancement network of Embodiment 1 of the present invention;

[0072] Figure 2 This is a schematic diagram of the interactive module IM structure in Embodiment 2 of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] Example 1

[0075] like Figure 1 As shown, the drug target affinity prediction interactive and extended causal convolutional enhancement 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 an extended causal convolutional module DCCM4.

[0076] In this embodiment, the lower structure decomposition module SDM1 takes the drug, protein, and a large amount of unlabeled data as input and decomposes them into a set of explicit substructure sequences. The decomposed substructures are then fed into the enhanced transformer embedding module ETEM2, which extracts the semantic relationships between the substructures. The interaction module IM3 performs more complex learning and modeling of the interactions between substructures to capture higher-order interactions, effectively extracting features from the structures of the drug and target, thus improving the model's understanding of DTAs. The dilated causal convolution module DCCM4 uses dilated convolutional layers to expand its receptive field, effectively capturing temporal information between the drug and target, and thus effectively capturing long-term dependencies. Specifically, DCCM adopts a lightweight structure while achieving good results when processing long sequence data. Finally, the output is processed by a decoder.

[0077] In this embodiment, the aforementioned substructure decomposition module SDM1 decomposes the molecular sequences of proteins and drugs into substructures based on data-driven sequence patterns, thereby discovering recurring subsequences in drug and protein databases. The substructure decomposition module SDM1 hierarchically decomposes each protein / drug sequence into subsequences, smaller subsequences, and individual atom 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 denoted as W, and the most frequent consecutive tokens (A, B) are found by scanning W; 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 is repeated until no frequent tokens are higher than a threshold θ or larger than the size of V. Through this operation, frequent subsequences are merged into a single token, while less frequent subsequences are broken down into a set of shorter tokens; for a drug / protein, SDM generates a substructured drug / target protein sequence C = (C1, C2, ..., C...) of size k. k ), where C i (1≤i≤k) comes from set V.

[0079] In this embodiment, IEC2T-DTA utilizes a substructure decomposition module to convert the input drug and target into explicit substructure sequences C, respectively. d and C pSequence. Unlike previous substructure fingerprinting methods, SDM is more interpretable. Because many substructures are subsets of others, it is difficult to know which substructure will affect the results. Instead, SDM breaks down each drug molecule into discrete, medium-sized substructure partitions, allowing for substructure mining using large amounts of unlabeled data. Furthermore, SDM can capture fundamental and meaningful biomedical semantics because the generated substructures are associated with frequently occurring basic units of drugs and proteins.

[0080] In this embodiment, an enhanced transformer embedding module, ETEM2, was constructed to fully mine the chemical semantics of substructures. ETEM2 first embeds a learnable substructure lookup dictionary; then, the embedding context substructure information is enhanced through a transformer encoder. The Transformer is built based on a self-attention mechanism, which learns the complex relationships between drug and target sequences while analyzing attention weights to determine which parts of the drug molecule and target protein sequence play a crucial role in predicting binding affinity.

[0081] In this embodiment, the enhanced transformer embedding module ETEM2 modifies each input substructure by learning from all substructures of the same molecule. Specifically, for each input drug-target pair, the corresponding substructure sequence C is modified. p and C d Transform into two matrices M p and M d Where k and l are the total size of the drug and protein substructure, respectively, and Θ p and Θ d It is the maximum length of the protein and drug substructure sequence, per column. and These are one-hot vectors, corresponding to the substructure indices of the i-th substructure of the protein sequence and the j-th substructure of the drug sequence, respectively. For each protein and drug, the context embedding... and A learnable lookup matrix can be used. and The specific steps for generating the generator 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 location embeddings and

[0086]

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

[0088] In this embodiment, the final embedding and It 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, these substructures have chemical relationships, and in order to capture this contextual information, IEC2T-DTA uses a Transformer encoder layer to further enhance the embeddings.

[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. To simulate pairing interactions, the following operations are performed for each subsequence i in the protein and each subsequence j in the drug:

[0094]

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

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

[0097] Furthermore, the adjacent substructures of proteins and drugs also influence each other. Therefore, in addition to modeling individual pair interactions, it is also necessary to model the interactions of neighboring regions. The interaction module IM3 is constructed based on CNN to effectively capture local features. Feature extraction is performed on the input data using sliding convolutional kernels 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 represented as follows:

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

[0099] CNN stands for Convolutional Neural Network.

[0100] The expanded causal convolution module DCCM4 can effectively extract features from drug molecule and target protein sequences at different scales. By learning these features, IEC2T-DTA can better predict the affinity between drugs and targets. Furthermore, DTA may be affected by long-range interactions, such as long-range acting sites in drug molecules or domains in target protein sequences. The expanded causal convolution module DCCM4 has an extended receptive field, which can capture longer sequence dependencies. This allows IEC2T-DTA to better model long-term dependencies between drugs and targets, improving the predictive 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 over the input by skipping values ​​with a specific stride.

[0102] See Figure 2 The dilated causal convolution has a kernel size of 2. Using a dilation factor h, the dilated causal convolution module DCCM4 selects the input every h steps and applies a standard 1D convolution to the selected input.

[0103] Dilated Causal Convolutional Networks (DCCMs) use non-recursive parallel computation to detect and extract amino acid features, improving learning speed and mitigating the vanishing gradient problem. By stacking the dilated causal convolutional layers in ascending order, DCCMs can capture longer sequences with fewer layers, saving computational resources while improving long-term prediction accuracy. The form of DCCM is as follows:

[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, which determines the ratio 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, joint optimization is achieved through binary classification loss:

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

[0110] Where Y truth It is the actual value.

[0111] Example 2

[0112] In this embodiment, the predictive performance of the IEC2T-DTA model was validated on three widely used datasets: BIOSNAP, Davis, and BindingDB. Notably, in the BIOSNAP dataset, a subset of the MINER dataset was used as the primary experimental dataset. It includes 4510 drugs and 2181 protein targets. The BIOSNAP dataset contains only positive DTA pairs. For negative pairs, samples were drawn from unseen pairs using conventional methods to obtain a balanced dataset with equal positive and negative samples. The Davis dataset primarily contains K... d Bioactivity data in (dissociation constant) format, including 379 proteins and 68 drugs. BindingDB is a public database containing determined binding affinity data for 4510 drugs and 2181 proteins. d Value. In the experiment, K was... d DTA pairs with values ​​less than 30 units were considered positive. Simultaneously, negative DTA pairs were subsampled in the same number as the positive samples in the training set. The negative ratio of the datasets was maintained in the validation and test sets. All datasets were divided into training, validation, and test sets in a 7:1:2 ratio. Detailed information about the datasets is shown in Table 1 below.

[0113] Table 1. Statistics of the dataset

[0114]

[0115] In the parameter settings 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 dropout rate of 0.1. In SDM, the minimum occurrence count of drug and protein substructures in the dataset is set to 500, resulting in 23,532 drug substructures and 16,693 protein substructures. In ETEM, a two-layer transformer encoder is used for both drugs and proteins, with an input embedding size of 384, 12 self-attention heads, and a median dimension of 1536. Furthermore, the maximum sequence lengths for drugs and proteins are set to 50 and 545, respectively, to cover 95% of the drugs and proteins in the dataset. Parts exceeding and falling below the maximum length are cut and padded. The batch size is set to 64, and the model is run for 30 epochs. The model converges between 8 and 15 epochs. All experiments are performed on a 22vCPU AMD EPYC 7T83 64-Core Processor with an RTX 4090 GPU Card.

[0116] In this embodiment, 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 metrics to measure the performance of binary classification. Furthermore, sensitivity and specificity metrics are used, where the threshold is the threshold that yields the best F1 score on the validation set.

[0117] In this embodiment, a baseline model was selected, and the IEC2T-DTA was compared with the following state-of-the-art models:

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

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

[0120] GNN-CPI: This method uses a graph neural network to encode drugs and a CNN to encode proteins. The latent vectors are then concatenated into a neural network to predict compound-protein interactions.

[0121] DeepDTI: This baseline was originally designed for CPI / DTI prediction and cannot be directly used for DTA tasks. Therefore, we need to modify it for regression tasks. 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 label / one-hot encodings of compound and protein sequences to predict DTA tasks. Their CNN model consists of two independent CNN modules that learn features from SMILES compound strings and protein sequences, respectively. Drug and target representations are concatenated and passed to fully connected layers 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 molecule fingerprints for DTA prediction.

[0124] In this embodiment, to test the competitiveness of the proposed model, experimental results of IEC2T-DTA were compared with state-of-the-art models (including classical and deep learning models), as shown in Table 2 below. It can be clearly seen that IEC2T-DTA achieves the best prediction performance in AUROC and AUPRC across all datasets. Furthermore, because different methods have different thresholds, sensitivity and specificity may vary.

[0125] As shown in Table 2, classical methods such as LR outperform deep learning-based methods. This is because classical methods heavily rely on handcrafted features and the similarity matrix between drugs and targets. In contrast, deep learning-based methods capture more information through automatic feature engineering using CNNs and GNNs. Furthermore, IEC2T-DTA significantly outperforms other methods. Reasons include: compared to DeepCPI, our method comprehensively considers topological structure and local chemical background, which improves model prediction performance; compared to DeepDTA, we construct a transformer encoding module based on a multi-head attention mechanism to obtain topological information about the drug and employ advanced embedding techniques to provide more contextual information for drug and target modeling. The designed dilated causal convolution can effectively process sequence data of drugs and targets, capturing temporal relationships and local dependencies between them. Compared to traditional convolution operations, dilated causal convolution not only considers the long-range dependencies 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, to investigate the prediction of unlabeled drugs and targets, 20% of the drugs and target proteins, along with drug-target pairs associated with these drugs and targets, were randomly selected as the test set for experiments. The results are shown in Table 3 below. It can be seen that the model used in this invention also exhibits excellent prediction performance for unlabeled drugs and proteins; however, the performance of the other three baseline models is not significantly different. This is because the substructure decomposition of IEC2T-DTA can effectively mine substructure features and remove the influence of repetitive subsequences on prediction performance. Furthermore, the enhanced transformer embedding module, the interactive module (IM), and the diluted causal convolution module also significantly contribute to the model's excellent performance.

[0129] Table 3. Performance comparison of unlabeled drug and protein prediction on the BIOSNAP dataset (mean AUROC value of five random runs)

[0130]

[0131] In this embodiment, we investigate the performance under sparse data conditions. While DTA data is rapidly increasing, in some real-world drug discovery, due to budget constraints, some novel target proteins or drugs have only a few tags. Therefore, to simulate real-world scenarios, we conducted experiments on 5%, 10%, 20%, and 30% of the dataset and performed predictions on the remaining datasets. 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 state-of-the-art baseline models such as DeepDTI and DeepConv-DTI decreases with increasing missing data. This further validates the effectiveness of our model.

[0132] Table 4. Comparison of prediction performance in the case of missing data in the BIOSNAP dataset (average AUROC value of 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 embodiment, to further investigate the impact of different modules of the proposed model, the following five variants were proposed and compared with IEC2T-DTA on the BIOSNAP dataset. The experimental results are shown in Table 5 below. The differences between these variants and IEC2T-DTA are as follows:

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

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

[0137] w / o IM: We removed the interaction module, everything else remains the same.

[0138] w / o SDM: We remove the substructure decomposition step and proceed directly to the experiment.

[0139] w / o DCCM: We remove the dilated causal convolution 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 embodiment, the ablation experiment results show that the expanded causal convolution module has the greatest impact on the model's predictive performance, followed by the substructure decomposition and interaction modules. The influence of other components on IEC2T-DTA is also evident. This demonstrates that these components enable our model to effectively capture and utilize advanced drug information, thereby improving the overall performance of drug-target interaction prediction.

[0143] This invention proposes a novel drug-target affinity prediction method, IEC2T-DTA. This method combines substructure decomposition, transformer encoding, a CNN-based interaction module, and dilated causal convolution to enhance the model's understanding of molecular representations. First, substructure decomposition fully mines drug-target features, facilitating the establishment of associations between drugs and targets. Second, transformer encoding can process sequence data of molecular structures, capturing long-range dependencies and important features within the molecular structure. The interaction module captures higher-order interactions. Finally, dilated causal convolution effectively captures both local and global features of the molecular structure and can handle input sequences of varying lengths, contributing to improved prediction model performance. Experiments on three public datasets demonstrate the superiority of IEC2T-DTA in handling DTA prediction problems, exhibiting the best affinity prediction performance.

[0144] In summary, this invention, considering both unlabeled data and the influence of drug substructure properties on DTA prediction, is the first to extend causal convolutional networks in the DTA field and combine them with existing transformers and CNNs for predictive modeling. Utilizing the DTA prediction model IEC2T-DTA, this invention can fully mine the features of protein and drug substructures and learn from both labeled and unlabeled data simultaneously, enabling more accurate prediction of drug target affinity.

[0145] This invention designs a knowledge-inspired substructure decomposition module to mine substructure features; secondly, it constructs an enhanced transformer encoder module to effectively mine semantic relationships between substructures from a large amount of unlabeled drug data; furthermore, it constructs an interaction module to better capture drug information for more accurate and interpretable DTA prediction; finally, it captures temporal relationships by using an expanded causal convolution module and provides complex feature representations to handle long-term dependencies, thereby improving prediction accuracy. Experimental results on three datasets show that the model used in this invention has superior prediction performance compared to the baseline model.

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

[0147] This invention proposes an extended causal convolution module that considers the time dependence of drug targets, enabling the capture of association patterns between drugs and targets at different scales to effectively extract features from drug molecules and target protein sequences. An interaction is also constructed to pair drugs with protein substructures to capture higher-order interactions.

[0148] This invention solves the technical problems of inaccurate and difficult-to-interpret results, neglect of unlabeled molecular data, and failure to consider time dependence in the prior art.

[0149] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A drug target affinity prediction interaction and extended causal convolution enhancement system, characterized in that, The ETID2C-DTA structure of the network includes: The lower structure decomposition module (SDM) takes drugs, proteins, and unlabeled data as input and, based on a data-driven sequence mode, decomposes the molecular sequences of the proteins, drugs, and unlabeled data to obtain substructure sequences and decomposed substructures. In the lower structure decomposition module (SDM), the vocabulary set is initialized. ; and according to the vocabulary set Tokenization is performed on a pre-built drug / protein corpus to obtain a tokenized set. ; Scan the tokenized set Find the most frequent consecutive tokens ; Use new token Update the tokenized set In, each of the most frequent consecutive tokens and the most frequent consecutive token Add to the vocabulary set middle; Continue scanning, identifying, and updating until the most frequent consecutive token is no longer available. If the preset update stop condition is met, the frequent substructure sequences are merged into frequent tokens, and the remaining substructure sequences are decomposed into tokens of preset length. The substructure sequence of the drug and the protein is expressed using the following logic: In the formula, The size of the substructure sequence; An enhanced transformer embedding module (ETEM) is used to receive the decomposed substructures, mine the semantic relationships between the substructures of the decomposed substructures, and learn a substructure lookup dictionary to learn all the 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 embedded context substructure information. The enhanced transformer embedding module (ETEM) is connected to the lower structure decomposition module (SDM). In the enhanced transformer embedding module ETEM, for each input drug-target pair, the corresponding substructure sequence is... , Convert to matrix , , respectively corresponding to the first digit of the protein sequence The first substructure and drug sequence Substructure index of a substructure; Using Learnable Lookup Matrices , Generate contextual embeddings for proteins and drugs and : (1) (2)； Using a lookup matrix , Constructing position embeddings , : (3) (4) In the formula, and It is a one-hot vector; Summing the context embedding and the position embedding yields the independent embedding. and : (5) (6) Using the Transformer encoder layer, the independent embeddings are... and Enhanced embedding: (7) (8) The interaction module IM is used to learn and model based on the interaction between substructures in the semantic relationship between substructures, simulate the interaction of paired substructures and extract neighborhood interaction information to capture higher-order interactions and extract drug and target structural features. The interaction module IM is connected to the enhanced transformer embedding module ETEM. The extended causal convolution module (DCCM) is used to expand the receptive field by expanding the convolutional layers, thereby capturing the temporal information between the drug and the target. The DCCM is used to perform non-recursive parallel computation to detect and extract amino acid features.

2. The drug target affinity prediction interaction and extended causal convolution enhancement system according to claim 1, characterized in that, The vocabulary set Includes: different amino acid tokens, SMILES string characters.

3. The drug target affinity prediction interaction and extended causal convolution enhancement system according to claim 1, characterized in that, The update stopping condition includes: the most frequent consecutive token. Above the threshold , or the most frequent consecutive token Higher than the vocabulary set Size.

4. The drug target affinity prediction interaction and extended causal convolution enhancement system according to claim 1, characterized in that, In the interaction module IM, the interaction between individual drug-target pairs is modeled, and each subsequence in the protein is processed using the following logic. and each subsequence in the drug The tensor interaction graph is obtained. : (9) In the formula, It is a dot product; Inspect the tensor interaction graph to obtain the final result contribution substructure; The interaction between neighboring regions is modeled, and the input data is extracted using a sliding convolution kernel to obtain an extracted feature map; The extracted feature map is then subjected to a nonlinear mapping to obtain the output of the input drug-target pair: (10) In the formula, This represents the computation of a convolutional neural network.

5. The drug target affinity prediction interaction and extended causal convolution enhancement system according to claim 1, characterized in that, The dilated causal convolution module DCCM includes at least two parallel dilated causal convolution modules for capturing and outputting long-term dependencies.

6. The drug target affinity prediction interaction and expanded causal convolution enhancement system according to claim 5, characterized in that, In the Dilated Causal Convolutional Module (DCCM), the dilated causal convolutional layers are stacked in ascending order: (11) In the formula, , , and These are model parameters. It is the output activation function. It is the Sigmoid function, which determines the ratio of information passed to the next layer; Will Flattened into a vector; using the weight matrix and bias vector Parameterized linear layers: (12) in, .

7. The drug target affinity prediction interaction and extended causal convolution enhancement system according to claim 1, characterized in that, In the Dilated Causal Convolutional Module (DCCM), a joint optimization operation is performed using binary classification loss: (13) In the formula, It is the actual value.

8. A method for enhancing drug target affinity prediction through interaction and expanded causal convolution, used to control the drug target affinity prediction through interaction and expanded causal convolution enhancement system according to any one of claims 1 to 7, characterized in that, The method includes: S1. Using drugs, proteins, and unlabeled data as input, and based on a data-driven sequence mode, decompose and process the molecular sequences of the proteins, drugs, and unlabeled data to obtain substructure sequences and decomposed substructures. S2. Receive the decomposed substructure, mine the semantic relationship between the substructures of the decomposed substructure, the enhanced transformer embedding module ETEM can learn the substructure lookup dictionary, learn all the decomposed substructures of the same molecule, so as to modify the embedding input substructure, wherein the enhanced transformer embedding module ETEM also includes: transformer encoder, used to enhance the embedding context substructure information; S3. Based on the interaction between substructures in the semantic relationship between substructures, perform learning and modeling operations, simulate the interaction of paired substructures and extract neighborhood interaction information to capture higher-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 convolutional layer, the receptive field is expanded to capture the temporal information between the drug and the target. The expanded causal convolution module DCCM is used for non-recursive parallel computation to detect and extract amino acid features.

Images