Drug and target interaction prediction method based on re-parameterized convolution

Abstract

The invention discloses a drug and target interaction prediction method based on re-parameterized convolution, which is mainly oriented to a drug lead compound virtual screening technology and aims to solve the problem that many existing deep learning methods only learn high-dimensional representation of ligands and targets from a small scale. In order to solve the problem that the model is difficult to obtain a potential rule of a binding pocket or a plurality of binding sites on a large scale, a re-parametric convolution method is introduced to improve the feature learning ability of the model to protein targets and compound ligands, and meanwhile, a gated attention mechanism is designed to more effectively interact characterization information of drugs and targets, so that the accuracy of drug delivery is improved. And a drug screening experiment on a Cystathionine-beta-synthase (CBS) target spot verifies the potential of the CBS-beta-synthase aptamer as a drug screening tool. The CBS-beta-synthase aptamer disclosed by the invention is used as a drug screening tool.

Description

Technical Field

[0001] The present invention relates to the intersection of biomedicine and artificial intelligence, and specifically is a method for improving a convolutional neural network using a reparameterization method to predict the interaction between a drug and a target. Background Art

[0002] Identifying the interaction between target proteins and ligands is a key step in the development of small molecule drugs. However, due to the problem of searching a huge compound space, it often costs a lot to find compound ligands with strong interaction ability against specific targets. Even though we now have computational simulation methods with a certain degree of credibility, it still takes months or even years to find a suitable lead compound.

[0003] In recent years, artificial intelligence technology has developed rapidly, among which deep learning models have achieved remarkable success in image processing, text translation, speech recognition and other fields. At the same time, thanks to the application of high-throughput methods, massive bioinformatics data makes it possible to build data-driven virtual drug screening models. By formalizing virtual screening tasks as binary classification tasks (DTI drug-target interaction) or regression (DTA drug-target affinity) tasks, many deep learning modeling methods can be applied to drug-target interaction prediction work. For example, HyperAttentionDTI, MGraphDTA, ZeroBind and Al-Bind.

[0004] Although deep learning has shown better prediction performance in drug-target interaction prediction than molecular docking and traditional machine learning methods, most current deep learning prediction technologies are still limited to extracting the overall representation of the target protein from the local amino acid sequence. In fact, the binding pocket of the target is often composed of specific regions on the peptide chain, which may involve the synergistic effect of one or more peptide chains. Although these local sequence information can reveal the characteristics of some binding sites, it is difficult to fully capture the global structural characteristics of the target, nor can it deeply reveal the complex laws of the interaction between the entire binding pocket and the ligand. Therefore, an excellent drug-target interaction prediction model should have the ability to simultaneously grasp the overall structural information of the target at a macro level and carefully consider the details of the local binding site. Summary of the invention

[0005] In view of the above-mentioned defects of the prior art, the purpose of the present invention is to overcome the shortcomings of the existing methods, improve the encoding ability of the deep learning model for protein targets and compound ligands, and thus enhance the model's ability to predict drug-target interactions. To this end, the present invention introduces a reparameterization method to improve the convolutional network model, and combines the ensemble learning method to predict the results of drug-target interactions. This method is expected to break through the bottleneck of the prior art, effectively reduce prediction bias, and reduce experimental costs.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

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

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

[0009] Step S2: Encode the target protein and ligand in all samples.

[0010] Step S3: using a re-parameterized convolutional neural network to learn the high-dimensional features of the target protein and the compound ligand.

[0011] Step S4: Use the gated attention mechanism to interact the target high-dimensional feature vector with the ligand high-dimensional feature vector information, model the potential drug-target interaction site information, and merge the features of the two to obtain the final mixed feature vector v f .

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

[0013] Step S6: Use the model to screen the ligands that can bind to the CBS target.

[0014] The step S1 of establishing a data set includes:

[0015] Step S11: To eliminate the label dissociation parameter K in the DAVIS data set d To solve the label sparseness problem caused by a large value range, all Kd labels in the DAVIS dataset are converted to pKd as shown in formula (1), and all samples with pKd label values ​​less than 5 are marked as positive samples, and the others are marked as negative samples.

[0016]

[0017] Step S12: In the KIBA dataset, the interaction label is the kiba score, and drug-target sample pairs with a kiba score less than 12.1 are generally considered to have almost no affinity. Therefore, all samples in the KIBA dataset with a kiba score label greater than 12.1 are marked as positive samples, and the others are marked as negative samples.

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

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

[0020] Step S15, taking all target proteins as a whole, five-fold cross validation is used to divide the data set into a training set (accounting for about 65% of the data set), a validation set (accounting for about 15% of the data set) and a test set (accounting for about 20% of the data set). Since all target proteins in the test set do not appear in the training set, they are used as the "target protein cold start" data set.

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

[0022] In step S3, the present invention introduces a reparameterization method to improve the modeling ability of one-dimensional convolutional neural networks in order to make up for the problem that existing deep learning models are insufficient in learning sequence features. 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 convolution kernels in the field of machine vision. It improves the model performance by training multiple small convolution kernels in parallel to improve the large convolution kernel network's ability to capture local information. It mainly includes the following steps:

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

[0024] Step S32: perform layer norm on all feature matrices as shown in Formula 2, which aims to speed up the training of the model, reduce the impact of gradient vanishing and gradient exploding, and make the model converge faster.

[0025]

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

[0027]

[0028] Step S33: After the characteristic values ​​of each output channel are normalized, the compression and extraction blocks are used to dynamically change the influence of different channel characteristics on the output, thereby increasing the contribution of the channel containing the combined information to the result. The calculation method is as follows:

[0029]

[0030] in, is the feature of the target or ligand after compression and extraction block processing, L is the sequence length, is the layer normalization result of the i-th input target or ligand, σ and δ are the activation functions sigmoid and ReLU respectively, and is a learnable weight parameter, r is the dimension 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 the compound ligand. D and X P .

[0032] In step S4, the present invention is inspired by the gating mechanism in the recurrent neural network to design a gated attention mechanism, which forgets the information irrelevant to the interaction in the drug and ligand representation, and updates the corresponding feature matrix, thereby mining the interaction area of ​​the local structure between the ligand and the target. The calculation steps include:

[0033] Step S41: Given a ligand high-dimensional feature matrix And the target high-dimensional feature matrix First, obtain the respective hidden layer representations through the fully connected layer, as shown in formula (7) (8):

[0034]

[0035] in, and is the weight matrix, b1 and b2 are the bias vectors.

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

[0037]

[0038]

[0039] in, and After dimension expansion, and F is the encoding-decoding network layer, For the activation function Tanh, MEAN(Input, dim) returns the result of taking the mean of the dimension.

[0040] Step S43: merge the information interaction result with the original feature, as shown in formula (12) (13):

[0041]

[0042] Among them, α is a hyperparameter.

[0043] Step S44: fusion characteristics of the target protein and the compound ligand, and It is sent to the global maximum pooling layer and concatenated, and then the concatenated vector is converted into a mixed feature v using a fully connected layer f As the input features of the decoding layer.

[0044] The input feature v of sample i in step S5 f The prediction results in the k-round iteration are shown in formula (14):

[0045]

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

[0047] In step S6, the present invention screened 3 ligands with strong interactions from 22 candidate lead compounds, and verified 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 new reparameterization method to improve the feature extraction capability of convolutional neural network for target proteins and compound ligands.

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

[0051] 3. The present invention combines biological big data with artificial intelligence to construct a deep learning-based drug-target interaction prediction model and uses wet experimental data for evaluation, verifying that the model has the potential to become a new generation of drug screening tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of the present invention;

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

[0054] Figure 3 Schematic diagram of the large convolution kernel block structure;

[0055] Figure 4 Schematic diagram of the re-parameterized one-dimensional convolution structure;

[0056] Figure 5 Schematic diagram of the gated attention mechanism. DETAILED DESCRIPTION

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation method of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full 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 improvements without violating the connotation of the invention, so the present invention is not limited by the specific implementation disclosed below.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0059] The present invention uses the interaction between drugs and targets as the prediction object, and the configuration of the selected training platform is as follows: CPU is Intel (R) Xeon (R) Platinum 8358P CPU @ 2.60GHz, memory is 512GB, GPU is NVIDIA A800, video memory is 80GB video memory, operating environment is Linux system, Python version is 3.11.3. The method of the present invention is implemented on the deep learning framework Pytorch version 2.4.0, CUDA version is 11.8, and cuDNN version is 9.0.1.

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

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

[0062]

[0063] All negative samples in the DAVIS, KIBA, and DUDE datasets are randomly sampled, and redundant negative samples are removed to keep the ratio of positive and negative samples in the dataset approximately 1:1. The final number of samples is shown in Table 1:

[0064] Table 1 Dataset summary

[0065]

[0066]

[0067] Taking all protein targets as a whole, five-fold cross validation was used to divide the dataset into a training set (accounting for about 65% of the dataset), a validation set (accounting for about 15% of the dataset) and a test set (accounting for about 20% of the dataset). Since all protein targets in the test set did not appear in the training set, it was used as a "target protein cold start" dataset.

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

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

[0070] After the ligand and target are represented as Embedding vectors, they will be further encoded in the re-parameterized convolutional layer. The re-parameterized convolutional layer consists of several large convolutional blocks (LGCNN Block) and small convolutional blocks (SMCNN Block), as shown in Table 2.

[0071] Table 2 Structure of re-parameterized convolutional layers

[0072]

[0073] The structure of the large convolution block (LGCNN Block) is as follows Figure 3 As shown in the figure, the only difference between the large convolution block and the small convolution block is that the re-parameterized one-dimensional convolution in the small convolution block will be replaced by a normal convolution. The re-parameterized one-dimensional convolution network in the large convolution block can improve the original large convolution network's ability to capture local features by training multiple small dilated convolutions in parallel. Figure 4 As shown. Then for all feature matrices x ′ Layer normalization is performed as shown in Formula 2, which aims to speed up the training of the model, reduce the impact of gradient disappearance and gradient explosion, and make the model converge faster.

[0074]

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

[0076]

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

[0078]

[0079] in, is the feature of the target or ligand after compression and extraction block processing, L is the sequence length, is the layer normalization result of the i-th input target or ligand, σ and δ are the activation functions sigmoid and ReLU respectively, and is a learnable weight parameter, r is the dimension reduction ratio;

[0080] The drug and target representation vectors are further encoded by the re-parameterized convolutional layer and then interactively combined in the gated attention layer. The interactive process is as follows: Figure 5 As shown. Given a ligand feature matrix and target feature matrix First, obtain the respective hidden layer representations through the fully connected layer, as shown in formula (7) (8):

[0081]

[0082]

[0083] in, and is the weight matrix, b1 and b2 are the bias vectors.

[0084] According to equations (9), (10), and (11), the update gate output is calculated as follows:

[0085]

[0086] in, and After dimension expansion, and F is the encoding-decoding network layer, For the activation function Tanh, MEAN(Input, dim) returns the result of taking the mean of the dimension.

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

[0088]

[0089] Among them, α is a hyperparameter. After that, and It is sent to the global maximum pooling layer and concatenated, and then the concatenated vector is converted into a mixed feature v using a fully connected layer f As the input features of the decoding layer.

[0090] Finally, the fused features are sent to XGBoost for decoding, and the input feature v of sample i can be obtained. f The prediction results in the k-round iteration are shown in formula (14):

[0091]

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

[0093] (3) Use the cross entropy loss as shown in formula (15) and use the data set obtained in (1) to train the model:

[0094]

[0095] Among them, y is the true label value, Predict labels for the model,

[0096] (4) Prediction analysis was performed on 22 ligands with significant interaction potential with CBS targets identified in the Docker molecular docking calculation software, and the prediction accuracy was verified by wet experiments, and the results are shown in Table 3. In the results shown, many of the predicted results of the present invention are consistent with the experimental results, which illustrates the potential of the present invention as a drug screening tool and is expected to help researchers select potential active ligands.

[0097] Table 3 Model screening results

[0098]

[0099] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should be regarded as within the scope of protection of the present invention.

Claims

1. A drug-target interaction prediction method 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 the binding ligands of the CBS target. The method comprises the following steps: Step S1, collecting the data of drug-target interaction for preprocessing, and using the "target protein cold start" and "target protein hot start" methods to obtain the training set, validation set and test set required for model training and evaluation; Step S2, encoding the target protein and ligand in all samples; Step S3, constructing a re-parameterized convolutional neural network to learn the high-dimensional features of the target protein and the compound ligand; Step S4: Use the gated attention mechanism to interact the target high-dimensional feature vector with the ligand high-dimensional feature vector information to form potential site information of drug-target interaction, and concatenate them to obtain the final mixed feature vector v f ; Step S5: Use the integrated learning XGBoost model to decode the final feature vector to obtain the model prediction result. Step S6: Use the model to screen the ligands that can bind to the CBS target.

2. A drug-target interaction prediction method based on reparameterized convolution according to claim 1, characterized in that: The data on the interaction between the drug and the target is a public data set or data actually collected in advance; the public data set adopts at least one of DAVIS, KIBA and DUDE.

3. A drug-target interaction prediction method based on reparameterized convolution according to claim 1, characterized in that: The pre-processing comprises: Data normalization, random sampling and keeping the ratio of positive and negative samples in the dataset similar; Using five-fold cross validation, the data set was divided into training set, validation set and test set for "target protein cold start" and "target protein hot start"; all target proteins in the "target protein hot start" test set appeared in the training set, and all target proteins in the "target protein cold start" test set did not appear in the training set.

4. The method for predicting drug-target interaction based on reparameterized convolution according to claim 1, characterized in that: The encoding in step S2 includes: All pre-processed target proteins and compound ligands are expressed in the form of amino acid sequences and standardized SMILES strings, respectively; Each target protein and ligand is encoded into a one-hot representation by creating a dictionary of corresponding character sequences; The embedding layer converts the two types of representations into embedding matrices: Where L d and L p Represent the length of the ligand string and the amino acid sequence, respectively, V D and V P are the dimensions of ligand and target embedding features, respectively.

5. The method for predicting drug-target interaction based on reparameterized convolution according to claim 1, characterized in that: The step S3 comprises the following steps: Step S31, the characteristic information of the target and the ligand will be further encoded in the re-parameterized convolutional layer; wherein the 1D convolutional block is used to encode the input feature map x∈R L×V , through the reparameterized convolution of size k, the low-dimensional features of C channels are extracted, and the new feature matrices x of the target and ligand are obtained respectively. ′ ∈R L×C ; Step S32: perform layer normalization on all feature matrices to speed up the model training, reduce the impact of gradient vanishing and gradient exploding, and make the model converge faster; the calculation method is as shown in formula (2): in, is the target or ligand layer normalization result, ∈ is the denominator correction term, u b and are the mean and variance respectively, x ′ A completely new feature matrix for a target or ligand; Step S33, after the characteristic values ​​of each output channel are normalized by the layer, the compression and extraction blocks are used to dynamically change the influence of different channel characteristics on the output, so as to improve the contribution of the channel containing the combined information to the result; the calculation method is as shown in formulas (5) and (6): in, is the feature of the target or ligand after compression and extraction block processing, L is the sequence length, is the layer normalization result of the i-th input target or ligand, σ and δ are the activation functions sigmoid and ReLU respectively, and is a learnable weight parameter, r is the dimension reduction ratio; Step S34: Return to step S32 and perform layer normalization again to finally obtain the high-dimensional features X of the target protein and the compound ligand. D and X P .

6. A drug-target interaction prediction method based on reparameterized convolution according to claim 1, characterized in that: The step S4 comprises the following steps: Step S41: Given a ligand high-dimensional feature matrix And the target high-dimensional feature matrix First, obtain the respective hidden layer representations through the fully connected layer, as shown in formula (7) (8): in, and is the weight matrix, b1 and b2 are the bias vectors; Step S42: Calculate the information interaction result z output by the update gate according to equations (9), (10), and (11). D 、z P ; in, and After dimension expansion, and F is the encoding-decoding network layer, For the activation function Tanh, MEAN(Input, dim) returns the result of taking the mean of the dimension. Step S43: merge the information interaction result with the original feature, as shown in formula (12) (13): Among them, α is a hyperparameter; Step S44: Fusion characteristics of target protein and compound ligand and It is sent to the global maximum pooling layer and concatenated to output the mixed feature v f .

7. The method for predicting drug-target interaction based on reparameterized convolution according to claim 1, characterized in that: The input feature v of sample i in step S5 f The prediction results in the k-round iteration are shown in formula (14): in, is the prediction result of the t-1 round model, f t (v f ) is the output result of the tth tree.

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

Images