A Drug-Target Interaction Prediction Method Based on Reparameterized Convolution

Abstract

This invention discloses a drug-target interaction prediction method based on reparameterized convolution. This method is mainly aimed at virtual screening technology for drug lead compounds. It addresses the problem that many existing deep learning methods only learn high-dimensional representations of ligands and targets at a small scale, making it difficult for the model to obtain the potential patterns of binding pockets or multiple binding sites at a large scale. The invention introduces the reparameterized convolution method to improve the model's feature learning ability for protein targets and compound ligands. At the same time, a gated attention mechanism is designed to more effectively interact the characterization information of drugs and targets. The potential of this invention as a drug screening tool is verified through drug screening experiments on the target of Cystathionine-β-synthase (CBS).

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of biomedicine and artificial intelligence, specifically a method for improving convolutional neural networks using reparameterization to predict drug-target interactions. Background Technology

[0002] Defining the interaction between target proteins and ligands is a crucial step in small molecule drug development. However, due to the vast search space of compounds, finding compound ligands with strong interaction capabilities against specific targets often requires enormous costs. Even with the availability of computational simulation methods with a certain degree of reliability, finding a suitable lead compound can still take months or even years.

[0003] In recent years, artificial intelligence technology has developed rapidly, with deep learning models achieving remarkable success in fields such as image processing, text translation, and speech recognition. Simultaneously, thanks to the application of high-throughput methods, massive amounts of bioinformatics data have made it possible to construct data-driven virtual drug screening models. By formalizing virtual screening tasks into binary classification tasks (DTI drug-target interaction) or regression tasks (DTA drug-target affinity), many deep learning modeling methods can be applied to drug-target interaction prediction. Examples include HyperAttentionDTI, MGraphDTA, ZeroBind, and Al-Bind.

[0004] While deep learning has demonstrated superior predictive performance compared to molecular docking and traditional machine learning methods in drug-target interaction prediction, most current deep learning prediction techniques remain limited to extracting the overall characterization of target proteins from local amino acid sequences. In reality, the binding pocket of a target is often composed of specific regions on the peptide chain, which may involve the synergistic interaction of one or more peptide chains. Although this local sequence information can reveal the characteristics of some binding sites, it is difficult to comprehensively capture the global structural characteristics of the target, nor can it deeply reveal the complex interactions between the entire binding pocket and the ligand. Therefore, an excellent drug-target interaction prediction model should be able to simultaneously grasp the overall structural information of the target macroscopically and meticulously consider the details of local binding sites. Summary of the Invention

[0005] In view of the aforementioned deficiencies in existing technologies, the purpose of this invention is to overcome the shortcomings of existing methods, improve the encoding ability of deep learning models for protein targets and compound ligands, and thus enhance the model's predictive ability for drug-target interactions. To this end, this invention introduces a reparameterization method to improve the convolutional network model and combines it with an ensemble learning method to predict drug-target interaction results. This method is expected to overcome the bottlenecks of existing technologies, effectively reduce prediction bias, and decrease experimental costs.

[0006] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0007] A method for predicting drug-target interactions based on reparameterized convolution, the method comprising the following steps:

[0008] Step S1: Collect and preprocess data from the public datasets DAVIS, KIBA, and DUDE for drug-target interactions, and use two methods, "target protein cold start" and "target protein hot start," to obtain the training set, validation set, and test set required for model training and evaluation.

[0009] Step S2: Encode the target proteins and ligands in all samples.

[0010] Step S3: Use a reparameterized convolutional neural network to learn the high-dimensional features of the target protein and compound ligands.

[0011] Step S4: Using a gated attention mechanism, the high-dimensional feature vector of the target and the high-dimensional feature vector of the ligand are interacted to model the potential drug-target interaction site information, and the two features are merged to obtain the final mixed feature vector v. f .

[0012] Step S5: Use the ensemble learning XGBoost model to decode the final feature vector and obtain the model prediction result.

[0013] Step S6: Use the model to screen for bindable ligands for CBS targets.

[0014] The establishment of the dataset in step S1 includes:

[0015] Step S11: To eliminate the label dissociation parameter K in the DAVIS dataset d To address the label sparsity problem caused by excessively large numerical ranges, all Kd labels in the DAVIS dataset are converted to pKd as shown in Equation (1), and all samples with pKd label values ​​less than 5 are labeled as positive samples, while the others are labeled as negative samples.

[0016]

[0017] Step S12: The interaction labels in the KIBA dataset are KIBA scores. Drug-target sample pairs with KIBA scores less than 12.1 are generally considered to have almost no affinity. Therefore, all samples in the KIBA dataset with KIBA score labels greater than 12.1 are labeled as positive samples, and the others are labeled as negative samples.

[0018] Step S13: Randomly sample all negative samples contained in the DAVIS, KIBA, and DUDE datasets, remove redundant negative samples, and maintain the ratio of positive to negative samples in the datasets to approximately 1:1.

[0019] Step S14: Taking all samples as a whole, use five-fold cross-validation to divide the dataset into a training set (64% of the dataset), a validation set (16% of the dataset), and a test set (20% of the dataset). Since all target proteins in the test set appear in the training set, it is used as the "target protein hot start" dataset.

[0020] Step S15: Taking all target proteins as a whole, use five-fold cross-validation to split the dataset into a training set (accounting for approximately 65% ​​of the dataset), a validation set (accounting for approximately 15% of the dataset), and a test set (accounting for approximately 20% of the dataset). Since all target proteins in the test set do not appear in the training set, it is used as the "target protein cold start" dataset.

[0021] In step S2, to enable the computer to better understand proteins and compounds, all target proteins and compound ligands are represented as amino acid sequences and normalized SMILES strings, respectively. The input is encoded into a one-hot representation by creating a corresponding character sequence dictionary. Finally, the embedding layer transforms these representations into an embedding matrix. Similarly, the amino acid sequence of each target site can be represented as: Where L d and L p V represents the length of the ligand string and the length of the amino acid sequence, respectively. D and V P These are the dimensions of ligand and target embedding features, respectively.

[0022] In step S3, this invention introduces a reparameterization method to improve the modeling ability of one-dimensional convolutional neural networks, addressing the problem of insufficient learning of sequence features in existing deep learning models. This method is similar to a reparameterization method proposed by Ding et al. of Tsinghua University in 2022 to improve the performance of two-dimensional large convolutional kernels in the field of machine vision. It improves the ability of large convolutional kernel networks to capture local information by training multiple small convolutional kernels in parallel, thereby improving model performance. The main steps include:

[0023] Step S31: The features of the target and ligand will be further encoded in the reparameterized convolutional layer. In this convolutional layer, the present invention designs a 1D convolutional block similar to the 2D convolutional blocks commonly used in visual tasks: for the input feature map x∈R L×V Low-dimensional features of C channels are extracted through reparameterized convolution of size k, and after processing, novel feature matrices x for the target and ligand are obtained respectively.′ ∈R L×C .

[0024] Step S32: Perform layer normalization on all feature matrices as shown in Equation 2. This aims to accelerate the training speed of the model, reduce the impact of gradient vanishing and gradient exploding, and enable the model to converge faster.

[0025]

[0026] Where ∈ is the denominator correction term, u b and The mean and variance are respectively, calculated as shown in equations (3) and (4):

[0027]

[0028] Step S33: After standardizing the feature values ​​of each output channel at the layer, compression and extraction blocks are used to dynamically change the influence of different channel features on the output, thereby increasing the contribution of channels containing combined information to the results. The calculation method is as shown in equations (5) and (6):

[0029]

[0030] in, The features of the target or ligand after compression and extraction block processing, where L is the sequence length. The layer normalization result for the i-th input target or ligand, where σ and δ are the activation functions sigmoid and ReLU, respectively. and Here, r represents the learnable weight parameters, and r is the dimensionality reduction ratio.

[0031] Step S34: Return to step S32 and perform layer normalization again to finally obtain the high-dimensional features X of the target protein and compound ligands. D and X P .

[0032] In step S4, inspired by the gating mechanism in recurrent neural networks, this invention designs a gated attention mechanism that forgets information unrelated to the interaction in the drug and ligand characterization and updates the corresponding feature matrix, thereby uncovering the interaction region of the local structure between the ligand and the target. Its computational steps include:

[0033] Step S41: Given the high-dimensional feature matrix of the ligand and target high-dimensional feature matrix First, obtain the hidden layer representations of each layer through the fully connected layer, as shown in equations (7) and (8):

[0034]

[0035] in, and Let b1 be the weight matrix and b2 be the bias vectors.

[0036] Step S42: Calculate the update gate output results according to equations (9)(10)(11):

[0037]

[0038]

[0039] in, and Each is after dimensional expansion and F stands for encoder-decoder network layer. Given the activation function Tanh, MEAN(Input, dim) returns the result of taking the mean over the dimension.

[0040] Step S43: Integrate the information interaction results with the original features, as shown in equations (12) and (13):

[0041]

[0042] Where α is a hyperparameter.

[0043] Step S44, the fusion characteristics of the target protein and the compound ligand, and The vectors are fed into a global max pooling layer and concatenated, then a fully connected layer is used to convert the concatenated vectors into a mixed feature v. f As input features for the decoding layer.

[0044] In step S5, the input feature v of sample i f The prediction results in k iterations are shown in equation (14):

[0045]

[0046] in, This is the prediction result of the t-1 round model, f t (v f ) is the output of the t-th tree.

[0047] In step S6, the present invention screened three ligands with strong interactions from 22 candidate lead compounds, and proved the reliability of the screening results through wet experiments.

[0048] The present invention has the following beneficial effects and advantages:

[0049] 1. This invention introduces a novel reparameterization method, which improves the feature extraction capability of convolutional neural networks for target proteins and compound ligands.

[0050] 2. This invention designs a gated attention mechanism, which is beneficial for the model to further explore potential drug-target interaction site information and improve the interpretability of the model.

[0051] 3. This invention combines biological big data and artificial intelligence to construct a drug-target interaction prediction model based on deep learning and evaluates it using wet experimental data, verifying that the model has the potential to become a new generation of drug screening tools. Attached Figure Description

[0052] Figure 1 This is a flowchart of the present invention;

[0053] Figure 2 This is a schematic diagram of the model structure;

[0054] Figure 3 This is a schematic diagram of a large convolutional kernel block structure;

[0055] Figure 4 A schematic diagram of a reparameterized one-dimensional convolution structure;

[0056] Figure 5 This is a schematic diagram of a gating attention mechanism. Detailed Implementation

[0057] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0059] This invention uses drug-target interaction as the prediction object. The training platform configuration is as follows: Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz, 512GB RAM, NVIDIA A800 GPU with 80GB VRAM, Linux system, and Python version 3.11.3. The method of this invention is implemented on the deep learning framework PyTorch version 2.4.0, CUDA version 11.8, and cuDNN version 9.0.1.

[0060] like Figure 1-5 As shown, a drug-target interaction prediction method based on reparameterized convolution includes the following parts:

[0061] (1) Download the DAVIS, KIBA, and DUDE datasets and write a Python script to convert all Kd labels in the DAVIS dataset to pKd as shown in equation (1), and label all samples with pKd label values ​​less than 5 as positive samples and the rest as negative samples. Label all samples in the KIBA dataset with kiba score labels greater than 12.1 as positive samples and the rest as negative samples. The DUDE dataset itself is divided into positive and negative samples.

[0062]

[0063] We then randomly sampled all negative samples from the DAVIS, KIBA, and DUDE datasets, removing redundant negative samples to maintain an approximately 1:1 ratio of positive to negative samples. The final sample size is shown in Table 1.

[0064] Table 1 Summary of Datasets

[0065]

[0066]

[0067] Using all protein targets as a whole, a five-fold cross-validation was employed to divide the dataset into a training set (approximately 65% ​​of the dataset), a validation set (approximately 15% of the dataset), and a test set (approximately 20% of the dataset). Since none of the protein targets in the test set appeared in the training set, it was used as the "target protein cold start" dataset.

[0068] (2) Establish a framework for a drug-target interaction prediction model based on reparameterized convolution, such as... Figure 2 As shown.

[0069] All target proteins and compound ligands are represented as amino acid sequences and normalized SMILES strings, respectively. The input encodes each ligand into a one-hot representation by creating a corresponding character dictionary. Finally, the embedding layer transforms these representations into an embedding matrix. Similarly, each target point can be represented as: Where L d and L p V represents the length of the ligand string and the length of the amino acid sequence, respectively. D and V P These are the dimensions of ligand and target embedding features, respectively.

[0070] The ligands and targets, represented as embedding vectors, are further encoded in the reparameterized convolutional layers. These reparameterized convolutional layers consist of several large convolutional blocks (LGCNN Blocks) and small convolutional blocks (SMCNN Blocks), as shown in Table 2.

[0071] Table 2. Structure of Reparameterized Convolutional Layers

[0072]

[0073] The structure of the large convolutional block (LGCNN Block) is as follows: Figure 3 As shown, the difference between large and small convolutional blocks lies only in the fact that in small convolutional blocks, the reparameterized one-dimensional convolutions are replaced with ordinary convolutions. In large convolutional blocks, the reparameterized one-dimensional convolutional network improves the original large convolutional network's ability to capture local features by training multiple dilated small convolutions in parallel. Figure 4 As shown. Then, for all characteristic matrices x... ′ Layer standardization, as shown in Equation 2, aims to accelerate model training, reduce the impact of gradient vanishing and gradient exploding, and enable the model to converge faster.

[0074]

[0075] in, For the layer normalization results of the target or ligand, x i A novel feature matrix x for the target or ligand ′ ∈R L×C ∈ is the denominator correction term, u b and The mean and variance are respectively, calculated as shown in equations (3) and (4):

[0076]

[0077] After standardizing the feature values ​​of each output channel at each layer, compression and extraction blocks are used to dynamically change the influence of different channel features on the output, thereby increasing the contribution of channels containing combined information to the results. The calculation method is shown in equations (5) and (6):

[0078]

[0079] in, The features of the target or ligand after compression and extraction block processing, where L is the sequence length. The layer normalization result for the i-th input target or ligand, where σ and δ are the activation functions sigmoid and ReLU, respectively. and Here, r represents the learnable weight parameters, and r is the dimensionality reduction ratio.

[0080] After being further encoded by reparameterized convolutional layers, the drug and target representation vectors will interact and combine information in a gated attention layer. The interaction process is as follows: Figure 5 As shown. Given the ligand feature matrix. and target feature matrix First, obtain the hidden layer representations of each layer through the fully connected layer, as shown in equations (7) and (8):

[0081]

[0082]

[0083] in, and Let b1 be the weight matrix and b2 be the bias vectors.

[0084] The update gate output is calculated according to equations (9), (10), and (11):

[0085]

[0086] in, and Each is after dimensional expansion and F stands for encoder-decoder network layer. Given the activation function Tanh, MEAN(Input, dim) returns the result of taking the mean over the dimension.

[0087] The information interaction results are integrated with the original features, as shown in equations (12) and (13):

[0088]

[0089] Where α is a hyperparameter. Then, and The vectors are fed into a global max pooling layer and concatenated, then a fully connected layer is used to convert the concatenated vectors into a mixed feature v. f As input features for the decoding layer.

[0090] Finally, the fused features are fed into XGBoost for decoding to obtain the input features v of sample i. f The prediction results in k iterations are shown in equation (14):

[0091]

[0092] in, This is the prediction result of the t-1 round model, f t (v f ) is the output of the t-th tree.

[0093] (3) Using the cross-entropy loss as shown in Equation (15), the model is trained using the dataset obtained in (1):

[0094]

[0095] Where y is the actual label value. Predict labels for the model,

[0096] (4) Predictive analysis was performed on 22 ligands identified in the Docker molecular docking calculation software as having significant interaction potential with the CBS target, and the accuracy of the predictions was verified by wet experiments. The results are shown in Table 3. Several of the predicted results from this invention are consistent with the experimental results, demonstrating the potential of this invention as a drug screening tool, which is expected to help researchers select potentially active ligands.

[0097] Table 3 Model Selection Results

[0098]

[0099] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting drug-target interactions based on reparameterized convolution, characterized in that, The following steps are performed to obtain an ideal model for interaction prediction, which is used to screen bindable ligands for CBS targets. The method includes the following steps: Step S1: Collect and preprocess data on drug-target interaction, and use two methods, "target protein cold start" and "target protein hot start", to obtain the training set, validation set and test set required for model training and evaluation. Step S2: Encode the target proteins and ligands in all samples; the encoding in step S2 includes: All pretreated target proteins and compound ligands are represented as amino acid sequences and normalized SMILES strings, respectively. By creating a corresponding character sequence dictionary, each target protein and ligand is encoded into a one-hot characterization; The embedding layer transforms the two types of representations into embedding matrices: Where L d and L p V represents the length of the ligand string and the amino acid sequence, respectively. D and V P These are the dimensions of ligand and target embedding features, respectively; Step S3: Construct a reparameterized convolutional neural network to learn the high-dimensional features of the target protein and compound ligands; Step S3 includes the following steps: Step S31: The feature information of the target and ligand will be further encoded in the reparameterized convolutional layer; wherein, the 1D convolutional block is used to encode the input feature map x∈R L×V Low-dimensional features of C channels are extracted through reparameterized convolution of size k, resulting in novel feature matrices x for the target and ligand, respectively. ′ ∈R L×C Where L is the length of the ligand string or amino acid sequence, V is the dimension of the target or ligand embedding feature, and C is the number of channels; Step S32: Perform layer normalization on all feature matrices to accelerate model training, reduce the impact of gradient vanishing and gradient exploding, and enable the model to converge faster; the calculation method is as shown in equation (2): in, For the layer normalization result of the target or ligand, E is the denominator correction term, u b and These are the mean and variance, respectively, x ′ A novel feature matrix for the target or ligand; Step S33: After standardizing the feature values ​​of each output channel at the layer, the influence of different channel features on the output is dynamically changed by compression and extraction blocks to improve the contribution of channels containing combined information to the results; the calculation method is as shown in equations (5) and (6): in, The features of the target or ligand after compression and extraction block processing, where L is the sequence length. The layer normalization result for the i-th input target or ligand, where σ and δ are the activation functions sigmoid and ReLU, respectively. and Here, F represents the learnable weight parameters, r is the dimensionality reduction ratio, and F is the dimensionality reduction ratio. sq for The compression result of each feature channel in the image, where C is the number of channels; Step S34: Return to step S32 and perform layer normalization again to finally obtain the high-dimensional features X of the target protein and compound ligands. D and X P ; Step S4: Using a gated attention mechanism, the high-dimensional feature vector of the target and the high-dimensional feature vector of the ligand are interacted to form potential site information for drug-target interaction, and then concatenated to obtain the final mixed feature vector v. f Step S4 includes the following steps: Step S41: Given the high-dimensional feature matrix of the ligand and target high-dimensional feature matrix First, obtain the hidden layer representations of each layer through the fully connected layer, as shown in equations (7) and (8): in, and Here, b1 and b2 are the weight matrix and the bias vectors, respectively. This represents the hidden layer representation of the high-dimensional feature matrix of the ligand. This represents the hidden layer representation of the high-dimensional feature matrix of the target. Step S42: Calculate the information interaction result z of the update gate output according to equations (9)(10)(11). D z P ; in, and Each is after dimensional expansion and F stands for encoder-decoder network layer. The activation function is Tanh, and MEAN(Input, dim) returns the result of taking the mean over the dimension; σ and δ are the activation functions sigmoid and ReLU, respectively; Step S43: Integrate the information interaction results with the original features, as shown in equations (12) and (13): Where α is a hyperparameter; Step S44, fusion characteristics of the target protein and compound ligand. and The data is fed into a global max pooling layer and concatenated to output a mixed feature v. f ; Step S5: Use the ensemble learning XGBoost model to decode the final mixed feature vector to obtain the model prediction result; Step S6: Use the model to screen for bindable ligands for CBS targets.

2. The method for predicting drug-target interactions based on reparameterized convolution according to claim 1, characterized in that, The data on drug-target interaction is either a public dataset or pre-collected data; the public dataset uses at least one of DAVIS, KIBA, and DUDE.

3. The method for predicting drug-target interactions based on reparameterized convolution according to claim 1, characterized in that, The preprocessing includes: Data normalization, random sampling, and maintaining an approximate ratio of positive to negative samples in the dataset; Five-fold cross-validation was used to split the dataset into training, validation, and test sets for "target protein cold start" and "target protein hot start". In the "target protein hot start" test set, all target proteins appeared in the training set, while in the "target protein cold start" test set, none of the target proteins appeared in the training set.

4. The drug-target interaction prediction method based on reparameterized convolution according to claim 1, characterized in that, In step S5, the input feature v of sample i f The prediction results in k iterations are shown in equation (14): in, This is the prediction result of the k-1 round model, f k (v f ) is the output of the k-th tree.

5. The method for predicting drug-target interactions based on reparameterized convolution according to claim 1, characterized in that, Step S6 is used to actually screen ligands with strong interactions from a variety of candidate lead compounds.

Images