A method for detecting protein-ligand binding sites based on a deep bidirectional neural network

Abstract

The present invention provides a method for detecting protein-ligand binding sites based on a deep bidirectional neural network, including: extracting protein feature information; inputting the protein feature information into a recurrent deep bidirectional neural network introducing a feedback connection mechanism to obtain final features; inputting the final features into a convolutional neural network for detection to obtain the probability distribution of protein-ligand binding sites. By introducing a feedback connection mechanism, the present invention improves a unidirectional deep convolutional neural network into a recurrent deep bidirectional neural network, enhancing the performance of protein-ligand binding site detection. Among them, the feedback connection can dynamically improve the representation learning of the protein 3D structure. It helps the encoder to focus more on possible sites rather than the complex structure of the protein itself. And the binary voxel-level masking close to the prediction contour can remove the noise patterns that may be caused by the feedback connection, further enhancing the detection effect.

Description

Technical Field

[0001] The present invention relates to the technical field of using computers to detect protein ligand binding sites. Specifically, it relates to a method for detecting protein ligand binding sites based on a deep bidirectional neural network. Background Art

[0002] Proteins are important mediators for maintaining genetic information in biological systems. In drug design, potential drug molecules or ligands are developed for a given protein receptor to form a stable complex at its binding site. Therefore, detecting the drug binding site on the 3D structure of a protein is the first challenging problem to be solved. Existing methods can be divided into two categories: sequence-based methods and structure-based methods. The former relies on the evolutionary information preserved in the amino acid sequence, while the latter directly uses the 3D structure of the protein to identify binding site residues.

[0003] With the progress of experimental techniques and recent breakthroughs in computational protein structure prediction, a large amount of protein 3D structure data has become available, and structure-based methods have received increasing attention. The main difficulty of this method lies in mining the complex structural patterns of small protein binding pockets, where the appearance of protein sites is irregular cavities or tunnels in the 3D structure. Traditional structure-based methods predict binding sites by using geometric features, interaction energies, or existing templates. These methods usually rely on expert knowledge and have a very low prediction accuracy, about 30% to 40%.

[0004] With the wide application of deep learning, especially the development of deep convolutional neural networks, end-to-end technology applications have become increasingly widespread. Currently, there have been some methods using deep learning to solve the binding site prediction task and achieved breakthrough progress. For example, DeepSite uses a four-layer convolutional neural network to extract the structural features of proteins, and Kalasanty follows the popular U-Net model in biomedical image segmentation and constructs a 3D U-Net to enhance representation learning. Compared with traditional models, the convolutional neural network model has improved the prediction performance to an unprecedented level.

[0005] However, although existing convolutional neural network models are effective for binding site prediction, they still belong to the feedforward computing architecture and are different from the human brain in many aspects. In the brain, visual information not only travels from a lower cognitive level to a higher cognitive level through a bottom-up path, but also selectively processes information in the opposite direction of the attention mechanism. This bidirectional process enables humans to perform a wide range of visual tasks in complex environments.

[0006] An early work on a self-organizing network called Least Mean Square Error Reconstruction (Lmser) was developed by folding an autoencoder along the central hidden layer. This folding equivalently established forward skip connections and feedback connections between the paired layers of the encoder and decoder. Recently, it was revisited and developed into an Lmser based on a deep convolutional neural network, and was proven to be superior to the U-Net model in image inpainting and biomedical image segmentation. In view of the limitations of the feed-forward calculation of existing protein ligand binding site prediction methods, the present invention proposes a recurrent Lmser network based on a deep bidirectional neural network to solve the difficulties in protein 3D structure representation learning and improve the prediction performance of ligand binding sites.

[0007] Currently, no description or report of similar technologies to the present invention has been found, and no similar materials at home and abroad have been collected. Summary of the Invention

[0008] In view of the defects in the prior art, the purpose of the present invention is to provide a method for detecting protein ligand binding sites based on a deep bidirectional neural network.

[0009] According to one aspect of the present invention, there is provided a method for detecting protein ligand binding sites based on a deep bidirectional neural network, including:

[0010] Extracting protein feature information;

[0011] Inputting the protein feature information into a recurrent deep bidirectional neural network introducing a feedback connection mechanism to obtain final features;

[0012] Inputting the final features into a convolutional neural network for detection to obtain the probability distribution of protein ligand binding sites.

[0013] Preferably, the extracting protein feature information includes: extracting the overall structure information of the protein or the local structure information after pre-sequence screening processing through a convolutional neural network to obtain protein feature information.

[0014] Preferably, the inputting the protein feature information into a recurrent deep bidirectional neural network introducing a feedback connection mechanism to obtain final features includes:

[0015] Encoder path: Using the encoder path of the deep bidirectional neural network for feature extraction to obtain features;

[0016] Decoder path: Inputting the features into the decoder path of the deep bidirectional neural network and fusing them with the skip connections transmitted from the encoder path for feature dimensionality reduction;

[0017] Repeat the process of the encoder path and the decoder path, and introduce and fuse the filtered feedback connection transmitted from the decoder path in the previous cycle in the encoder path.

[0018] Preferably, the deep bidirectional neural network includes: 4 encoder modules, 1 intermediate transformation module, and 4 decoder modules, and the 4 encoder modules and 4 decoder modules correspond to each other respectively.

[0019] Preferably, each encoder module includes 2 layers of 3D convolutional layers and 1 layer of average pooling layer, where:

[0020] The execution of each encoder module includes:

[0021] Input: Receive the input from the previous module, double the number of channels through the first layer of 3D convolutional layer, and rectify through the relu function to obtain the input features;

[0022] Fusion: Fuse the feedback connection from the corresponding decoder module, and perform weighted summation of the input features and the feedback connection to obtain the fused features;

[0023] Extraction: Input the fused features into the second layer of 3D convolutional layer, keep the number of channels unchanged, rectify through the relu function to obtain the extracted features, and use the extracted features as the skip connection output to the corresponding decoder module;

[0024] Output: Input the extracted features into the average pooling layer, obtain the features with reduced size and output them as the input to the next module;

[0025] Preferably, the intermediate transformation module includes 2 layers of 3D convolutional layers, and the relu function is used for rectification after each layer of convolutional layer.

[0026] Preferably, each decoder module includes 1 layer of upsampling layer and 2 layers of 3D convolutional layers;

[0027] The execution of each decoder module includes:

[0028] Input: Receive the input from the previous module and enlarge the size of the features through the upsampling layer;

[0029] Fusion: Fuse the enlarged features and the skip connection from the corresponding encoder module using the method of weighted summation to obtain the fused features;

[0030] Extraction: Input the fused features into the first layer of 3D convolutional layer to halve the number of channels, and rectify through the relu function to obtain the extracted features;

[0031] Output: Input the extracted features into the second 3D convolutional layer with the number of channels unchanged, rectify them through the relu function, output the obtained features to the next module, and at the same time output them as feedback connections to the corresponding encoder modules.

[0032] Preferably, the feedback connections are passed to the corresponding encoder modules after being filtered by voxel-level masking, including:

[0033] Prediction: Self-copy and splice the final output of the decoder path, perform a virtual protein-ligand binding site prediction, and binarize the result with a probability threshold of 0.5 as the prediction result;

[0034] Filtering: Downsample the prediction result and use it as a mask to filter each feedback connection passed from the decoder path to the encoder path.

[0035] Preferably, the final features are input into a convolutional neural network for detection to obtain the probability distribution of protein-ligand binding sites; wherein, the convolutional neural network includes:

[0036] Concatenate the final features obtained in each loop as the input, and use a convolutional neural network to perform the detection of protein-ligand binding sites;

[0037] Use the sigmoid function to calculate the probability distribution of protein-ligand binding sites.

[0038] Preferably, minimize the cross-entropy between the calculated result and the label to reduce the difference between the obtained probability distribution and the label.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] In the protein-ligand binding site detection method based on a deep bidirectional neural network in the embodiments of the present invention, by introducing a feedback connection mechanism, the unidirectional deep convolutional neural network is improved into a cyclic deep bidirectional neural network, which improves the performance of protein-ligand binding site detection. Among them, the feedback connection can dynamically improve the representation learning of the protein 3D structure. It helps the encoder to focus more on possible sites rather than the complex structure of the protein itself. And the binary voxel-level masking close to the prediction contour can remove the noise patterns that may be caused by the feedback connection, further improving the detection effect.

[0041] The protein-ligand binding site detection method based on a deep bidirectional neural network in the embodiments of the present invention only requires about 4 GPUs and a CPU with ordinary configuration, and can run and reproduce the results within 7 days. The relatively common numpy scientific computing library and the pytorch deep learning framework are used in this embodiment.

[0042] In the protein ligand binding site detection method based on a deep bidirectional neural network in the embodiments of the present invention, there is a significant improvement in accuracy compared to existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] By reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0044] Figure 1 The overall structure diagram of the deep bidirectional neural network;

[0045] Among them: on the left is the unfolded diagram of the encoder-decoder structure, on the upper right is the schematic diagram of the corresponding encoder module and decoder module, and on the lower right is the voxel-level masking mechanism of the feedback connection. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can be made. These all belong to the protection scope of the present invention.

[0047] In view of the limitations of the feedforward calculation architecture in the prior art, an embodiment of the present invention provides a protein ligand binding site detection method based on a deep bidirectional neural network, including:

[0048] S0 Extract protein feature information;

[0049] S1 Input the protein feature information into a recurrent deep bidirectional neural network with a feedback connection mechanism to obtain the final features;

[0050] S2 Input the final features into a convolutional neural network for detection to obtain the probability distribution of the protein ligand binding site.

[0051] Among them, the feedback connection can dynamically improve the representation learning of the protein 3D structure, and the trained network can provide a significant improvement in accuracy based on existing protein ligand binding site detection technologies.

[0052] In an embodiment of the invention, in S0, protein feature information is extracted, and this process is obtained by extracting protein feature information through a convolutional neural network. The specific implementation is related to the experimental settings. For example, the structural information of the entire protein macromolecule can be input into the network, or the structural information of the preprocessed protein partial slices can be input into the network.

[0053] In a preferred embodiment of the present invention, performing S1 includes:

[0054] S101 Using the protein feature information calculated by S0, perform feature extraction using the encoder path of a deep bidirectional neural network;

[0055] S102 Then input the features into the decoder path of the deep bidirectional neural network and fuse them with the skip connections transmitted from the encoder path to detect the ligand binding site;

[0056] S103 Subsequently, repeat the process of the encoder path - decoder path, but introduce and fuse the filtered feedback connection transmitted from the previous decoder path during the encoder path.

[0057] Among them, the detection of the ligand binding site mentioned in S102, even if it is to perform dimensionality reduction of features. The simulated detection performed in each cycle shares the same convolution kernel with the final detection, but the inputs are different.

[0058] In another preferred embodiment of the present invention, S101 - S103 are performed. Specifically, the deep bidirectional neural network includes an encoder path, an intermediate transformation module, and a decoder path. Among them, the encoder path includes 4 encoder modules, the decoder path includes 4 decoder modules, and each encoder module corresponds to a decoder module.

[0059] The encoder path contains 4 encoder modules. Each encoder module includes 2 layers of 3D convolutional layers and one layer of average pooling layer. After each layer of convolutional layer, there is a relu rectification layer. The convolution kernel size of the convolutional layer is 3×3×3, and the stride is 1. The relu rectification layer makes the output of some neurons be 0, resulting in the sparsity of the network, and reducing the interdependence of parameters, alleviating the occurrence of overfitting problems.

[0060] The encoder module performs the following 4 steps:

[0061] The first step (input): Receive the input from the previous module (referring to the encoder module, or the features obtained in S0 (for the first encoder module in each cycle)), double the number of channels through a layer of 3D convolutional layer, and then perform rectification through the relu function;

[0062] The second step (fusion): Fuse the feedback connection from the corresponding decoder module, and fuse the features obtained in the previous step with the feedback connection by weighted summation for subsequent calculations. If there is no feedback connection received due to being in the first cycle, keep the original features unchanged;

[0063] Step 3 (Extraction): Input the features fused in the previous step into the second 3D convolutional layer with the number of channels unchanged, then rectify through the relu function, and the obtained features are output as skip connections to the corresponding decoder modules;

[0064] Step 4 (Output): Input the features obtained in the previous step into the average pooling layer to reduce the size of the features for output to the next module (which can be an encoder module or an intermediate transformation module (for the last encoder module in each loop)).

[0065] The intermediate transformation module includes 2 layers of 3D convolutional layers. After each convolutional layer, the relu function is used for rectification. The convolutional kernel size of the convolutional layer is 3×3×3, and the stride is 1.

[0066] The decoder path contains 4 decoder modules. Each decoder module includes 1 layer of upsampling layer and 2 layers of 3D convolutional layers. The kernel size of the upsampling layer is equal to the pooling layer of the corresponding encoder module. After each convolutional layer, there is a relu rectification layer. The convolutional kernel size of the convolutional layer is 3×3×3, and the stride is 1.

[0067] The decoder module performs the following 4 steps:

[0068] Step 1 (Input): Receive the input from the previous module (which can be an intermediate transformation module (for the first decoder module in each loop) or a decoder module), and enlarge the size of the features through the upsampling layer for subsequent calculations;

[0069] Step 2 (Fusion): Use the weighted sum method to fuse the enlarged features with the skip connection from the corresponding encoder module for subsequent calculations;

[0070] Step 3 (Extraction): Input the features fused in the previous step into the first 3D convolutional layer to halve the number of channels, and then rectify through the relu function;

[0071] Step 4 (Output): Input the features obtained in the previous step into the second 3D convolutional layer with the number of channels unchanged, then rectify through the relu function. The obtained features are output to the next module (the next module can be a decoder module or S2 (for the last decoder module in each loop)), and at the same time, they are output as feedback connections to the corresponding encoder modules.

[0072] Furthermore, the feedback connection is passed to the corresponding encoder module after being filtered by voxel-level masking, which includes the following 2 steps:

[0073] Prediction: Self-copy and splice the final output of the decoder path, that is, copy the features of the original 32 channels three times to become 96 channels to meet the input requirements of the convolutional kernel; then input it into the S2 module for a virtual prediction of protein-ligand binding sites, and binarize the result with a probability threshold of 0.5.

[0074] Filtering: Appropriately downsample the prediction result of the previous step to use it as a mask to filter each feedback connection from the decoder path to the encoder path.

[0075] In a preferred embodiment of the present invention, the final features obtained by executing S2 using S1 in multiple cycles are used to detect protein-ligand binding sites via a convolutional neural network, and the probability distribution of protein-ligand binding sites is obtained. Among them, the convolutional neural network adopted includes the following two parts:

[0076] The first part: Splice the final features obtained in each cycle in S1 and use them as inputs, and use a convolutional layer with a convolutional kernel of 1×1×1 and a stride of 1 to reduce the number of channels to 1.

[0077] The second part: Use the sigmoid function to convert the feature map obtained in the first part into the probability distribution of protein-ligand binding sites for output.

[0078] To improve the accuracy of the detection results, the present invention also provides an embodiment to update the parameters in the foregoing module including the deep bidirectional neural network by reducing the difference between the calculation result and the label.

[0079] The update is performed according to the following formula:

[0080]

[0081] Where y is the predicted probability after being processed by the softmax function, t is the label, that is, the true value of the protein-ligand binding site, and N is the number of samples. The formula calculates the cross-entropy loss between the predicted probability and the label.

[0082] All the training data of the protein-ligand binding site detection method based on the deep bidirectional neural network provided by the present invention can be obtained on the Internet.

[0083] Using the protein-ligand binding site detection method based on the deep bidirectional neural network provided by the present invention, there is an obvious improvement in accuracy compared with the existing methods.

[0084] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various deformations or modifications within the scope of the claims, which do not affect the essence of the present invention. The above preferred features can be used in any combination without conflict.

Claims

1. A method for detecting protein-ligand binding sites based on a deep bidirectional neural network, characterized in that, it includes: extracting protein feature information; inputting the protein feature information into a recurrent deep bidirectional neural network introducing a feedback connection mechanism to obtain final features, including: Encoder path: Using the encoder path of the deep bidirectional neural network for feature extraction to obtain features; Decoder path: Inputting the features into the decoder path of the deep bidirectional neural network and fusing them with the skip connections transmitted from the encoder path for dimensionality reduction of the features; Repeating the process of the encoder path and the decoder path, and introducing and fusing the filtered feedback connection transmitted from the decoder path in the previous cycle in the encoder path; inputting the final features into a convolutional neural network for detection to obtain the probability distribution of protein-ligand binding sites; wherein: the feedback connection is transmitted to the corresponding encoder module after being filtered at the voxel level, including: Prediction: Self-replicating and splicing the final output of the decoder path along the channel dimension, performing a virtual prediction of protein-ligand binding sites, and binarizing the result with a probability threshold of 0.5 as the prediction result; Filtering: Downsampling the prediction result as a mask to filter each feedback connection transmitted from the decoder path to the encoder path.

2. The method for detecting protein-ligand binding sites based on a deep bidirectional neural network according to claim 1, characterized in that, the extracting of protein feature information includes: extracting the overall structure information of the protein or the local structure information after previous screening processing via a convolutional neural network to obtain protein feature information.

3. The method for detecting protein-ligand binding sites based on a deep bidirectional neural network according to claim 1, characterized in that, the deep bidirectional neural network includes: 4 encoder modules, 1 intermediate transformation module and 4 decoder modules, and the 4 encoder modules and 4 decoder modules correspond respectively.

4. The method for detecting protein-ligand binding sites based on a deep bidirectional neural network according to claim 3, characterized in that, each encoder module includes 2 layers of 3D convolutional layers and 1 layer of average pooling layer, wherein: the execution of each encoder module includes: Input: Receiving the input from the previous module, doubling the number of channels through the first layer of 3D convolutional layer, and rectifying through the relu function to obtain input features; Fusion: Fusing the feedback connection from the corresponding decoder module, and performing weighted summation of the input features and the feedback connection to obtain fused features; Extraction: Inputting the fused features into the second layer of 3D convolutional layer with the number of channels unchanged, rectifying through the relu function to obtain extracted features, and using the extracted features as the skip connection output to the corresponding decoder module; Output: Inputting the extracted features into the average pooling layer to obtain size-reduced features and output them as the input to the next module.

5. The method for detecting protein-ligand binding sites based on a deep bidirectional neural network according to claim 3, It is characterized in that the intermediate transformation module includes two layers of 3D convolutional layers, and the relu function is used for rectification after each convolutional layer.

6. A method for detecting protein-ligand binding sites based on a deep bidirectional neural network according to claim 3, It is characterized in that each decoder module includes one layer of upsampling layer and two layers of 3D convolutional layers; The execution of each decoder module includes: Input: Receive the input from the previous module and enlarge the size of the feature through the upsampling layer; Fusion: Fuse the enlarged feature and the skip connection from the corresponding encoder module by using weighted summation to obtain a fused feature; Extraction: Input the fused feature into the first layer of 3D convolutional layer to halve the number of channels, and rectify it through the relu function to obtain an extracted feature; Output: Input the extracted feature into the second layer of 3D convolutional layer, keep the number of channels unchanged, rectify it through the relu function, output the obtained feature to the next module, and at the same time output it as a feedback connection to the corresponding encoder module.

7. A method for detecting protein-ligand binding sites based on a deep bidirectional neural network according to claim 1, It is characterized in that the final feature is input into a convolutional neural network for detection to obtain the probability distribution of protein-ligand binding sites; wherein, the convolutional neural network includes: Concatenate the final features obtained in each loop as the input, and use a convolutional neural network to perform the detection of protein-ligand binding sites; Use the sigmoid function to calculate the probability distribution of protein-ligand binding sites.

8. A method for detecting protein-ligand binding sites based on a deep bidirectional neural network according to claim 1, It is characterized in that The difference between the obtained probability distribution and the label is reduced by minimizing the cross-entropy between the calculation result and the label.

Images