Drug binding energy prediction method based on quantum convolutional neural network

Abstract

The present invention provides a method for predicting drug binding energy based on quantum convolutional neural network, which belongs to the field of quantum computing technology. The method encodes drug molecule sequence data, protein sequence data and protein binding pocket sequence data into corresponding first quantum state, second quantum state and third quantum state, and inputs the first quantum state, second quantum state and third quantum state into quantum convolutional neural network for feature extraction and fusion, and measures the fusion feature, so as to predict the optimal value corresponding to the binding energy. Because the protein binding pocket containing the local characteristics of the protein molecule is added to the input of the present invention, in the process of feature extraction of quantum convolutional neural network, according to the important physical and chemical characteristics of the protein binding pocket, the feature can be extracted more accurately, so that the predicted value of the drug binding energy is more accurate.

Description

Technical Field

[0001] The present invention belongs to the field of quantum computing technology, and specifically relates to a drug binding energy prediction method based on quantum convolutional neural network. Background Art

[0002] Biomolecular recognition between proteins and ligands (drug molecules) plays a vital role in drug discovery and drug development. However, it takes a lot of time and resources to experimentally determine the binding energy between proteins and ligands. Currently, there are many computational methods for predicting binding energy, most of which require the three-dimensional structure of the protein, which is not easy to obtain. Therefore, new methods are needed to predict the binding affinity between proteins and ligands from sequence-level features to accelerate the drug discovery process.

[0003] Convolutional Neural Network (CNN) is an important model in classical machine learning technology, especially in image processing and speech recognition tasks. Convolutional neural network has two special network layers: convolution layer and pooling layer. These two network layers are stacked to form a special hierarchical structure, which enables convolutional neural network to efficiently extract structured features from input data. In addition, the local connection and weight sharing characteristics of convolutional neural network reduce the complexity of the network model and can significantly improve the computational efficiency of the model.

[0004] However, traditional convolutional neural networks require training of a large number of parameters when dealing with complex protein molecules and ligands with large molecular weights, and the accuracy of the predicted binding energy is low. Summary of the invention

[0005] The present invention provides a drug binding energy prediction method based on quantum convolutional neural network, thereby greatly reducing the parameters of model training and improving the accuracy of binding energy prediction values.

[0006] The present invention provides a method for predicting drug binding energy based on quantum convolutional neural network, comprising the following steps:

[0007] Encoding drug molecule sequence data, protein sequence data and protein binding pocket sequence data into corresponding first quantum state, second quantum state and third quantum state;

[0008] The first quantum state, the second quantum state, and the third quantum state are respectively input into the quantum convolutional neural network to extract features and then fuse them, and the fusion features are measured to obtain the fusion features; and

[0009] The fused features are input into the classic fully connected neural network to obtain the predicted value corresponding to the drug binding energy.

[0010] The above method, wherein:

[0011] The quantum convolutional neural network includes multiple quantum convolution layers and multiple quantum pooling layers;

[0012] Each quantum convolution layer includes single-qubit gates or multi-qubit gates;

[0013] Each quantum pooling layer includes a two-qubit controlled rotation gate.

[0014] The above method, wherein:

[0015] Single-qubit gates include single-qubit rotation gates, and multi-qubit gates include controlled-NOT gates.

[0016] The above method, wherein:

[0017] The encoding methods of the first quantum state include:

[0018] Performing dimensionality reduction and transposition on the drug molecule sequence data of a first predetermined length to obtain a corresponding first semi-positive definite matrix; and

[0019] The first semi-positive definite matrix is ​​normalized to obtain a first density matrix corresponding to the first quantum state.

[0020] The above method, wherein:

[0021] The encoding methods of the second quantum state include:

[0022] Performing dimensionality reduction and transposition on the protein sequence data of a second predetermined length to obtain a corresponding second positive semi-definite matrix; and

[0023] The second semi-positive definite matrix is ​​normalized to obtain a second density matrix corresponding to the second quantum state.

[0024] The above method, wherein:

[0025] Protein sequence data includes protein secondary structure sequence data or residue sequence data.

[0026] The above method, wherein:

[0027] The encoding methods of the third quantum state include:

[0028] Performing dimensionality reduction and transposition on the protein binding pocket sequence data of a third predetermined length to obtain a corresponding third semi-positive definite matrix; and

[0029] The third positive semi-definite matrix is ​​normalized to obtain a third density matrix corresponding to the third quantum state.

[0030] The above method, wherein:

[0031] Among them, the number of quantum bits of the quantum circuit in the quantum convolutional neural network is greater than or equal to log 2 d, d is the dimension of any one of the first density matrix, the second density matrix or the third density matrix.

[0032] The above method, wherein:

[0033] It also includes using the gradient descent method to update the parameters in the quantum convolutional neural network and using the loss function to optimize the difference between the predicted value and the actual value corresponding to the binding energy.

[0034] The present invention also provides another method for predicting drug binding energy based on quantum convolutional neural network, comprising the following steps:

[0035] Encoding drug molecule sequence data, protein sequence data and protein binding pocket sequence data into corresponding first quantum state, second quantum state and third quantum state;

[0036] The first quantum state, the second quantum state, and the third quantum state are respectively input into the quantum convolutional neural network for the first feature extraction and then fused to obtain the fused features;

[0037] Performing a second feature extraction on the fused features and measuring the extracted features; and

[0038] The extracted features are input into the classic fully connected neural network to obtain the predicted value corresponding to the drug binding energy.

[0039] Functions and Effects of the Invention

[0040] According to a drug binding energy prediction method based on a quantum convolutional neural network provided by the present invention, the method encodes drug molecule sequence data, protein sequence data and protein binding pocket sequence data into corresponding first quantum states, second quantum states and third quantum states, and inputs the above-mentioned first quantum states, second quantum states and third quantum states into a quantum convolutional neural network for feature extraction and fusion, and measures the fusion features, so as to predict the optimal value corresponding to the binding energy. Because the protein binding pocket containing the local features of the protein molecule is added to the input of the present invention, in the process of feature extraction of the quantum convolutional neural network, according to the important physical and chemical features of the protein binding pocket, the features can be extracted more accurately, so that the predicted value of the drug binding energy is more accurate. In addition, since the number of quantum bits in the quantum convolutional neural network in the present invention is much smaller than the parameters in the traditional convolutional neural network, and with the help of quantum superposition characteristics and parallel computing, the training parameters can be greatly reduced and the calculation speed can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1It is a flow chart of a method for predicting drug binding energy based on a quantum convolutional neural network according to an embodiment of the present invention.

[0042] Figure 2 Schematic diagram of the structure of the quantum circuit of the quantum convolution layer of an embodiment of the present invention.

[0043] Figure 3 Schematic diagram of the structure of the quantum circuit of the quantum pooling layer of an embodiment of the present invention.

[0044] Figure 4 It is a schematic diagram of the processing process of the drug binding energy prediction method based on quantum convolutional neural network according to an embodiment of the present invention.

[0045] Figure 5 It is a flowchart of another method for predicting drug binding energy based on quantum convolutional neural network according to an embodiment of the present invention.

[0046] Figure 6 yes Figure 5 Corresponding schematic diagram of the processing process of another drug binding energy prediction method based on quantum convolutional neural network. DETAILED DESCRIPTION

[0047] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the drug binding energy prediction method based on quantum convolutional neural network provided by the present invention is specifically described below in combination with embodiments and drawings.

[0048] <Example>

[0049] Figure 1 It is a flow chart of a method for predicting drug binding energy based on a quantum convolutional neural network according to an embodiment of the present invention.

[0050] like Figure 1 As shown, the drug binding energy prediction method based on quantum convolutional neural network provided in this embodiment includes the following steps:

[0051] S1, encoding drug molecule sequence data, protein sequence data and protein binding pocket sequence data into corresponding first quantum state, second quantum state and third quantum state;

[0052] S2, respectively inputting the first quantum state, the second quantum state, and the third quantum state into the quantum convolutional neural network to extract features and then fuse them, and measure and obtain the fusion features;

[0053] S3. Input the fused features into the classic fully connected neural network to obtain the predicted value corresponding to the binding energy.

[0054] The drug molecule sequence data in this embodiment is first represented by a one-dimensional sequence of drug molecules in the form of Simplified Molecular Input Line Entry System (SMILES) characters, and then each SMILES character corresponding to the drug molecule is integer-encoded, for example, the C element is encoded as 1, the H element is encoded as 2, the N element is encoded as 3, the O element is encoded as 4, and so on, so that the one-dimensional drug molecule sequence can be encoded as a one-hot vector with a length of N1 and a dimension of 64.

[0055] The 64-dimensional one-hot vector corresponding to the above drug molecule sequence is encoded into the first quantum state. The specific encoding method includes the following steps:

[0056] S101, performing dimension reduction and transposition on drug molecule sequence data of a first predetermined length N1 to obtain a corresponding first semi-positive definite matrix;

[0057] S102. Normalize the first semi-positive definite matrix to obtain a first density matrix corresponding to the first quantum state.

[0058] In this embodiment, the (N1,64) unique hot vector corresponding to the drug molecule sequence data is first reduced in dimension, for example, from 64 dimensions to d dimensions, to obtain a (N1,d)-dimensional matrix X1, and the matrix X1 is transposed and multiplied with the matrix X1 to obtain a (d,d)-dimensional first semi-positive definite matrix. The first semi-positive definite matrix is ​​then normalized so that the trace of the processed matrix is ​​1, thereby obtaining a first density matrix corresponding to the drug molecule sequence data, which can be used as the first quantum state of the quantum circuit in the quantum convolutional neural network.

[0059] The protein sequence data in this embodiment includes protein secondary structure sequence data or residue sequence data. For example, 21 different types of residues in a protein sequence can be encoded using a 21-dimensional one-hot vector, and 8-dimensional one-hot vectors can be used to encode the secondary structure sequence data of a protein, the secondary structure sequence data is from the SSPro program, and the secondary structure can include α-helix (H), residue in isolated β-bridge (B), extended strand, participates in β ladder (E), hydrogen bonded turn (T), 310 helix (G), π-helix (I), bend (S), and coil (C). For each residue structure, there are four properties based on the side chain structure: non-polar, polar, acidic and basic. According to its dipole and side chain volume, it can be divided into 7 groups. Therefore, an 11-dimensional unique hot vector can also be used to encode the physicochemical properties of each residue. In summary, this embodiment uses a total of 40-dimensional unique hot vectors to represent each residue in the protein sequence. The embodiment of this application uses the secondary structure of the protein and the physicochemical properties of the residues to enrich the structural information of the protein sequence and enhance the binding affinity with the drug molecules.

[0060] The 40-dimensional one-hot vector corresponding to the protein sequence data is encoded into the second quantum state. The specific encoding method includes the following steps:

[0061] S103, performing dimension reduction and transposition on the protein sequence data of the second predetermined length N2 to obtain a corresponding second positive semi-definite matrix;

[0062] S104. Normalize the second semi-positive definite matrix to obtain a second density matrix corresponding to the second quantum state.

[0063] In this embodiment, the (N2,40) unique hot vector corresponding to the protein sequence data is first reduced in dimension, for example, from 40 dimensions to d dimensions, to obtain a (N2,d)-dimensional matrix X2, and the matrix X2 is transposed and multiplied with the matrix X2 to obtain a (d,d)-dimensional second semi-positive definite matrix. The second semi-positive definite matrix is ​​then normalized so that the trace of the processed matrix is ​​1, thereby obtaining a second density matrix corresponding to the protein sequence data, which can be used as the second quantum state of the quantum circuit in the quantum convolutional neural network.

[0064] According to the secondary structure sequence and residue sequence representation in the protein sequence data, the protein binding pocket sequence data can be encoded using a 40-dimensional unique hot vector of the same dimension as the protein sequence data. The protein binding pocket sequence data in this embodiment includes local features of the protein sequence data, such as the local features may include the physicochemical properties of key amino acids of the protein, and the physicochemical properties have a direct and close relationship with the drug binding affinity, so it can provide a reference for the subsequent protein and drug molecule binding affinity, making the predicted value of the drug molecule and protein binding affinity more accurate.

[0065] In this embodiment, the 40-dimensional one-hot vector corresponding to the protein binding pocket sequence data is encoded into a third quantum state. The specific encoding method includes the following steps:

[0066] S105, performing dimensionality reduction and transposition on the protein binding pocket sequence data of the third predetermined length N3 to obtain a corresponding third semi-positive definite matrix;

[0067] S106. Normalize the third semi-positive definite matrix to obtain a third density matrix corresponding to the third quantum state.

[0068] In this embodiment, the (N3,40) unique hot vector corresponding to the protein binding pocket sequence data is first reduced in dimension, for example, from 40 dimensions to d dimensions, to obtain a (N3,d)-dimensional matrix X3, and the matrix X3 is transposed and multiplied with the matrix X3 to obtain a (d,d)-dimensional third semi-positive definite matrix. The third semi-positive definite matrix is ​​then normalized so that the trace of the processed matrix is ​​1, thereby obtaining a third density matrix corresponding to the protein binding pocket sequence data, which can be used as the third quantum state of the quantum circuit in the quantum convolutional neural network.

[0069] Figure 2 is a schematic diagram of the structure of a quantum circuit of a quantum convolution layer of an embodiment of the present invention, Figure 3 is a schematic diagram of the structure of the quantum circuit of the quantum pooling layer of an embodiment of the present invention, Figure 4 It is a schematic diagram of the processing process of the drug binding energy prediction method based on quantum convolutional neural network according to an embodiment of the present invention.

[0070] like Figure 2-4 As shown, the first quantum state, the second quantum state, and the third quantum state prepared above are input into the constructed quantum convolutional neural network ( Figure 4 In the dotted part), since the quantum convolutional neural network needs to perform multiple convolutions and multiple pooling operations, the dotted lines in this embodiment can represent multiple executions of convolution and pooling operations.

[0071] The quantum convolutional neural network in this embodiment may include multiple quantum convolution layers and multiple quantum pooling layers. Each quantum convolution layer includes a single-qubit gate or a multi-qubit gate; the single-qubit gate includes a single-qubit rotation gate, and the multi-qubit gate includes a controlled NOT gate. Each quantum pooling layer includes a dual-qubit controlled rotation gate.

[0072] The number of qubits in the quantum circuit of the quantum convolutional neural network is greater than or equal to log 2 d, d is the dimension of any one of the first density matrix, the second density matrix, or the third density matrix. In this embodiment, the first density matrix, the second density matrix, and the third density matrix are all density matrices of (d, d). Because the quantum bits required in this embodiment are log 2 d, so compared with traditional convolutional neural networks, it can greatly reduce training parameters and speed up model convergence.

[0073] like Figure 2 As shown, the quantum convolution operation is described by taking the first quantum state corresponding to the drug molecule sequence data as an example. According to the aforementioned determined quantum bits, the first quantum state is encoded onto the determined quantum bits. Figure 2 Only two quantum bits are shown as an example for illustration. The first quantum state is encoded into the above quantum bit (such as Figure 2 After passing through two unitary matrices U, the two unitary matrices U act on the two quantum bits respectively, and after tensor calculation, they are merged into a new quantum state. Figure 2 The unitary matrix U is a universal single-qubit gate. The unitary matrix U has three adjustable parameters. The expression of the unitary matrix U is as follows:

[0074]

[0075] In the above formula, γ, θ, Represent three rotation angles respectively. The new quantum state after the U matrix passes through the CNOT gate again, and the quantum bit where the black circle above is located is the control bit, and the quantum bit containing the cross in the white circle below is the controlled bit, and the quantum bit above is superimposed on the quantum bit below by binary addition, and then the quantum bit above is rotated around the Y axis through the Ry rotating gate, and the quantum bit below is superimposed on the quantum bit above by binary addition and then rotated around the Z axis through the Rz rotating gate. The quantum bit after the rotation operation passes through the CNOT gate, the rotating Ry gate, the CNOT gate, and the single quantum bit U gate to obtain the quantum state after convolution. Both the Ry rotating gate and the Rz rotating gate have adjustable parameters θ (rotation angle), and the parameters in the single quantum bit U gate, the Ry rotating gate, and the Rz rotating gate are continuously adjusted to extract the quantum state corresponding to the drug molecule. The quantum gate of the quantum convolution circuit in this embodiment can also select other quantum gates, and the placement method includes but is not limited to the placement method provided in this embodiment.

[0076] The quantum state of the drug molecule sequence data after convolution is then Figure 3 The quantum pooling circuit shown can output the quantum state corresponding to the pooled drug molecule sequence data. Figure 3 The quantum pooling circuit shown may, in other embodiments, also be used with other quantum circuits capable of performing the pooling operation. Figure 3 The double-qubit controlled rotating gate acts between two adjacent qubits and merges the information of the two qubits into one qubit, thereby reducing the dimension of the quantum circuit. The black circle is the control bit 1, and the white circle is the control bit 0. In this embodiment, the quantum state corresponding to the pooled drug molecule sequence data is measured multiple times. For example, Figure 3 In , only the qubit where the revolving gate Rz is located is measured. That is to say, only half of the number of input qubits is measured multiple times to obtain the first pooled feature vector corresponding to the drug molecule sequence data.

[0077] The quantum convolution circuit and quantum pooling circuit of protein sequence data and protein binding pocket sequence data are the same as the quantum circuit in drug molecule sequence data, and will not be repeated here. After the protein sequence data and protein binding pocket sequence data are input into the quantum convolution circuit and quantum pooling circuit, the pooling feature vectors corresponding to the protein sequence data and protein binding pocket sequence data are output, and are recorded as the second pooling feature vector and the third pooling feature vector, respectively. After fusing the above-mentioned first pooling feature vector, the second pooling feature vector, and the third pooling feature vector, a fusion feature containing drug molecules, protein molecules, and protein binding pockets is obtained.

[0078] The above fusion features are input into a fully connected neural network, and after passing through at least one fully connected layer, a nonlinear transformation is performed using the ReLU activation function to output the predicted value corresponding to the drug binding energy.

[0079] Finally, the parameters in the quantum convolutional neural network (i.e., the rotation angle of the quantum gate) are updated using the gradient descent algorithm, and the difference between the predicted binding energy and the actual value is optimized by the loss function. After repeated multiple times to reach the preset number of training times, the optimal predicted value corresponding to the drug binding energy is obtained. In this embodiment, the mean square error loss function is used for difference optimization. In other embodiments, other loss functions can also be used for difference optimization. The expression of the mean square error loss function used in this embodiment is as follows: Where E is the error, is the actual value of drug binding energy, and Y is the predicted value of drug binding energy. After completing the above training process, the final model is obtained.

[0080] Figure 5 It is a flowchart of another method for predicting drug binding energy based on quantum convolutional neural network in an embodiment of the present invention.

[0081] like Figure 5 As shown, the other drug binding energy prediction method based on quantum convolutional neural network includes the following steps:

[0082] S1, encoding drug molecule sequence data, protein sequence data and protein binding pocket sequence data into corresponding first quantum state, second quantum state and third quantum state;

[0083] S2, respectively inputting the first quantum state, the second quantum state, and the third quantum state into the quantum convolutional neural network for the first feature extraction and then fusing to obtain fused features;

[0084] S3, performing a second feature extraction on the fused features, and measuring the extracted features;

[0085] S4. Input the extracted features into the classic fully connected neural network to obtain the predicted value corresponding to the drug binding energy.

[0086] Figure 6 yes Figure 5 Schematic diagram of the corresponding processing process of the drug binding energy prediction method based on quantum convolutional neural network.

[0087] See also Figure 5 , Figure 6 , the flowchart and processing diagram of the other method are Figure 1 , Figure 4 The corresponding flow charts are similar to the processing process schematic diagrams. Figure 5 , Figure 6The fusion features obtained after the quantum convolutional neural network are added with a quantum convolution layer and a quantum pooling layer, that is, the second feature extraction is performed. The pooled feature vector after the second feature extraction is processed in a fully connected neural network to obtain the drug binding energy prediction value. The parameter optimization process behind this other method is the same as the difference optimization process, which will not be repeated here.

[0088] The above description and drawings provide typical embodiments of specific structures of specific embodiments, and the above invention provides existing preferred embodiments, but these contents are not intended to be limiting. For those skilled in the art, various changes and modifications will undoubtedly be obvious after reading the above description. Therefore, the attached claims should be regarded as covering all changes and modifications to the true intent and scope of the present invention. Any and all equivalent scopes and contents within the scope of the claims should be considered to still be within the intent and scope of the present invention.

Claims

1. A method for predicting drug binding energy based on quantum convolutional neural network, characterized in that: The steps include: Encoding drug molecule sequence data, protein sequence data and protein binding pocket sequence data into corresponding first quantum state, second quantum state and third quantum state; The first quantum state, the second quantum state, and the third quantum state are respectively input into a quantum convolutional neural network to extract features and then fuse them, and the fused features are measured to obtain the fused features; as well as The fusion features are input into a classic fully connected neural network to obtain the predicted value corresponding to the drug binding energy.

2. The method according to claim 1, characterized in that: The quantum convolutional neural network includes multiple quantum convolution layers and multiple quantum pooling layers; Each of the quantum convolution layers includes a single-qubit gate or a multi-qubit gate; Each of the quantum pooling layers includes a dual-qubit controlled rotation gate.

3. The method according to claim 2, characterized in that: The single-qubit gate comprises a single-qubit rotation gate, and the multi-qubit gate comprises a controlled NOT gate.

4. The method according to claim 1, characterized in that: The encoding method of the first quantum state includes: Performing dimension reduction and transposition on the drug molecule sequence data of a first predetermined length to obtain a corresponding first semi-positive definite matrix; and The first positive semi-definite matrix is ​​normalized to obtain a first density matrix corresponding to the first quantum state.

5. The method according to claim 1, characterized in that: The encoding method of the second quantum state includes: Performing dimensionality reduction and transposition on the protein sequence data of a second predetermined length to obtain a corresponding second positive semi-definite matrix; and The second positive semi-definite matrix is ​​normalized to obtain a second density matrix corresponding to the second quantum state.

6. The method according to claim 5, characterized in that: The protein sequence data includes protein secondary structure sequence data or residue sequence data.

7. The method according to claim 1, characterized in that: The encoding method of the third quantum state includes: Performing dimensionality reduction and transposition on the protein binding pocket sequence data of a third predetermined length to obtain a corresponding third positive semi-definite matrix; and The third semi-positive definite matrix is ​​normalized to obtain a third density matrix corresponding to the third quantum state.

8. The method according to any one of claims 4, 5 or 7, characterized in that: in, The number of quantum bits of the quantum circuit in the quantum convolutional neural network is greater than or equal to log2d. When citing claim 4, d is the dimension of the first density matrix; when citing claim 5, d is the dimension of the second density matrix; when citing claim 7, d is the dimension of the third density matrix.

9. The method according to claim 1, characterized in that: It also includes using a gradient descent method to update the parameters in the quantum convolutional neural network and using a loss function to optimize the difference between the predicted value and the actual value corresponding to the drug binding energy.

10. A method for predicting drug binding energy based on quantum convolutional neural network, characterized in that: The steps include: Encoding drug molecule sequence data, protein sequence data and protein binding pocket sequence data into corresponding first quantum state, second quantum state and third quantum state; The first quantum state, the second quantum state, and the third quantum state are respectively input into a quantum convolutional neural network for first feature extraction and then fused to obtain fused features; Performing a second feature extraction on the fused features, and measuring to obtain the extracted features; as well as The extracted features are input into a classic fully connected neural network to obtain the predicted value corresponding to the drug binding energy.

Images