阅读说明：本技术 一种基于轻量级深度卷积网络的蛋白质扭转角预测方法 (Protein torsion angle prediction method based on lightweight deep convolutional network ) 是由 杨伟 文云光 李艳萍 葛文庚 于 2021-07-05 设计创作，主要内容包括：本发明公开一种基于轻量级深度卷积网络的蛋白质扭转角预测方法,包括：基于PISCES服务器构建蛋白质扭转角数据集；从RCSB PDB数据库中抽取各蛋白质序列中每个氨基酸残基对应的扭转角Phi和Psi,并加入数据集中；将uniref90数据库与各蛋白质序列进行多序列比对,生成对应蛋白质序列的PSSM谱矩阵,基于PSSM谱矩阵及氨基酸的物理化学特性,构造蛋白质序列特征；设计残差模块,并基于该模块构建预测蛋白质扭转角的深度卷积网络模型；构建训练网络模型的损失函数；基于构建的损失函数训练网络模型；基于训练后的网络模型进行蛋白质扭转角的预测。本发明不仅能精确地预测蛋白质扭转角,还具有模型小,预测速度快的优点。(The invention discloses a protein torsion angle prediction method based on a lightweight deep convolutional network, which comprises the following steps: constructing a protein torsion angle data set based on a PISCES server; extracting torsion angles Phi and Psi corresponding to each amino acid residue in each protein sequence from an RCSB PDB database, and adding the torsion angles Phi and Psi into a data set; performing multi-sequence comparison on the uniref90 database and each protein sequence to generate a PSSM (phosphosilicate-associated syndrome) spectrum matrix corresponding to the protein sequence, and constructing protein sequence characteristics based on the PSSM spectrum matrix and the physicochemical characteristics of amino acids; designing a residual error module, and constructing a depth convolution network model for predicting the protein torsion angle based on the residual error module; constructing a loss function of a training network model; training a network model based on the constructed loss function; and predicting the protein torsion angle based on the trained network model. The method can accurately predict the protein torsion angle, and has the advantages of small model and high prediction speed.)

1. A protein torsion angle prediction method based on a lightweight deep convolutional network is characterized by comprising the following steps:

step 1: constructing a protein torsion angle dataset based on a PISCES server, the protein torsion angle dataset comprising a plurality of protein sequences;

step 2: extracting torsion angles Phi and Psi corresponding to each amino acid residue in each protein sequence in the protein torsion angle data set from an RCSB PDB database by adopting BioPython, and adding the torsion angles Phi and Psi into the protein torsion angle data set;

and step 3: performing multi-sequence alignment on the uniref90 database and each protein sequence in the protein torsion angle dataset to generate a PSSM (phosphosilicate-like syndrome) spectrum matrix corresponding to the protein sequence, and constructing protein sequence characteristics based on the PSSM spectrum matrix of the protein sequence and the physicochemical characteristics of amino acids;

and 4, step 4: designing a residual error module, and constructing a depth convolution network model for predicting the protein torsion angle based on the designed residual error module; the residual module involves computational operations including: one-dimensional convolution, one-dimensional batch processing normalization, Hard-Swish activation function, average pooling, full connection and residual connection;

and 5: constructing a loss function of a deep convolution network model for training and predicting the protein torsion angle;

step 6: training a deep convolution network model for predicting the protein torsion angle based on the constructed loss function;

and 7: and predicting the protein torsion angle based on the trained deep convolution network model for predicting the protein torsion angle.

2. The method for predicting the torsion angle of the protein based on the lightweight deep convolutional network as claimed in claim 1, wherein the step 1 comprises:

selecting the protein structure database with resolution less than that of the protein structure database based on the PISCES serverAnd m protein sequences with sequence identity less than 25% were added to the protein torsion angle dataset.

3. The method for predicting the torsion angle of the protein based on the lightweight deep convolutional network as claimed in claim 2, further comprising, after the step 2:

the protein torsion angle data set is divided into a training set, a validation set and a test set.

4. The method for predicting the torsion angle of the protein based on the lightweight deep convolutional network as claimed in claim 1, wherein the step 3 comprises:

for each protein sequence in the protein torsion angle dataset, performing multiple sequence alignment with protein sequences in uniref90 database by calling PSI-BLAST program to generate PSSM spectrum matrix corresponding to the protein sequence, wherein each amino acid residue corresponds to a 20-dimensional numerical vector in the PSSM spectrum matrix; the vector is spliced with a 7-dimensional numerical vector representing the physicochemical properties of the amino acids corresponding to the protein sequence to obtain the final property representation of the protein sequence.

5. The method for predicting the protein torsion angle based on the lightweight deep convolutional network as claimed in claim 1, wherein the expression of the one-dimensional convolution operation in the residual error module is as follows:

wherein the content of the first and second substances,is a convolution operation, x_inAnd x_outThe convolved input tensor and output tensor, respectively, k the convolution kernel and b the offset vector.

6. The method of claim 1, wherein the expression of the Hard-Swish activation function used by the residual error module is as follows:

wherein x is input data; relu6(x +3) indicates that changing x +3 less than 0 to 0, greater than 6 takes 6.

7. The method of claim 1, wherein the one-dimensional convolution used by the residual module is a depth separable one-dimensional convolution.

8. The method for predicting the protein torsion angle based on the lightweight deep convolutional network as claimed in claim 1, wherein in the constructed deep convolutional network model for predicting the protein torsion angle, the convolutional network outputs four values for each amino acid residue for a given protein sequence: sin (φ), cos (φ), sin (ψ) and cos (ψ); the values of the torsion angles Phi and Psi are calculated by Phi arctan (sin (Phi)/cos (Phi)) and Psi arctan (sin (Psi)/cos (Psi)), respectively.

9. The method for predicting the torsion angle of the protein based on the lightweight deep convolutional network as claimed in claim 1, wherein the loss function in the step 5 is as follows:

wherein N is the number of protein sequences in the mini-batch, L is the length of the protein sequences, 4 represents the vector length formed by the sine value and the cosine value of the torsion angles Psi and Phi, lambda is the regularization parameter of weight attenuation, W is the weight parameter of the convolution network,is the output of the prediction network and,is composed ofThe corresponding true tensor.

Technical Field

The invention belongs to the technical field of bioinformatics, and particularly relates to a protein torsion angle prediction method based on a lightweight deep convolutional network.

Background

Predicting the torsion angle based on the amino acid sequence of a protein is an important task in computational molecular biology. The function of a protein is determined by its structure. However, the determination of protein structure using experimental methods such as X-ray crystallography and nuclear magnetic resonance is extremely expensive and time consuming. Therefore, it is necessary to determine the structure of a protein by calculation. For a protein chain consisting of L amino acid residues, the protein backbone is a repeating sequence consisting of nitrogen, α -carbon and carbon atoms: n is a radical of⁽¹⁾、C⁽¹⁾、N⁽²⁾、C⁽²⁾、…、N^(L)、C⁽ _α ^L)、C^(L). In particular, the torsion angle Psi is defined by N⁽ⁱ⁾,And C⁽ⁱ⁾A determined plane andC⁽ⁱ⁾and N⁽ⁱ⁺¹⁾A dihedral angle between the determined planes. The torsion angle Phi is formed by C⁽ⁱ⁾、N^{(i +1)}Anddetermined plane and N⁽ⁱ⁺¹⁾,And C⁽ⁱ⁺¹⁾A dihedral angle between the determined planes. Due to phaseThe bond length and angle between adjacent skeleton atoms are fixed, and only the torsion angle is uncertain, so that the torsion angle is determined, and the skeleton structure is determined. Therefore, it is very important to accurately predict the torsion angle of a protein, and the prediction result of the torsion angle can be used not only for template-based tertiary structure prediction and fold recognition, but also for determining the class of the protein structure.

Currently, some Deep learning methods such as stacked sparse self-encoders, Deep cyclic limited boltzmann machine [ h.li, j.hou, b.adhikari, q.lyu, and j.cheng, "Deep learning methods for protein conversion prediction," BMC bioinf, vol.18, No.1, p.417,2017 ], and bidirectional recurrent neural networks [ Heffernan, r., Yang, y., Paliwal, K. & Zhou, y.capturen-local interactions by local distribution-local networking for enhancing the prediction of protein conversion, etc. (2842, and 2842, protein conversion angle prediction and others have been successfully applied to protein prediction. However, the models have the disadvantages of more parameters, large network model and long prediction time.

Disclosure of Invention

The invention provides a protein torsion angle prediction method based on a lightweight deep convolution network, aiming at the problems of more parameters, large network model and long prediction time of the conventional protein torsion angle prediction model.

In order to achieve the purpose, the invention adopts the following technical scheme:

a protein torsion angle prediction method based on a lightweight deep convolutional network comprises the following steps:

step 1: constructing a protein torsion angle dataset based on a PISCES server, the protein torsion angle dataset comprising a plurality of protein sequences;

step 2: extracting torsion angles Phi and Psi corresponding to each amino acid residue in each protein sequence in the protein torsion angle data set from an RCSB PDB database by adopting BioPython, and adding the torsion angles Phi and Psi into the protein torsion angle data set;

and step 3: performing multi-sequence alignment on the uniref90 database and each protein sequence in the protein torsion angle dataset to generate a PSSM (phosphosilicate-like syndrome) spectrum matrix corresponding to the protein sequence, and constructing protein sequence characteristics based on the PSSM spectrum matrix of the protein sequence and the physicochemical characteristics of amino acids;

and 4, step 4: designing a residual error module, and constructing a depth convolution network model for predicting the protein torsion angle based on the designed residual error module; the residual module involves computational operations including: one-dimensional convolution, one-dimensional batch processing normalization, Hard-Swish activation function, average pooling, full connection and residual connection;

and 5: constructing a loss function of a deep convolution network model for training and predicting the protein torsion angle;

step 6: training a deep convolution network model for predicting the protein torsion angle based on the constructed loss function;

and 7: and predicting the protein torsion angle based on the trained deep convolution network model for predicting the protein torsion angle.

Further, the step 1 comprises:

selecting the protein structure database with resolution less than that of the protein structure database based on the PISCES serverAnd m protein sequences with sequence identity less than 25% were added to the protein torsion angle dataset.

Further, after the step 2, the method further comprises the following steps:

the protein torsion angle data set is divided into a training set, a validation set and a test set.

Further, the step 3 comprises:

for each protein sequence in the protein torsion angle dataset, performing multiple sequence alignment with protein sequences in uniref90 database by calling PSI-BLAST program to generate PSSM spectrum matrix corresponding to the protein sequence, wherein each amino acid residue corresponds to a 20-dimensional numerical vector in the PSSM spectrum matrix; the vector is spliced with a 7-dimensional numerical vector representing the physicochemical properties of the amino acids corresponding to the protein sequence to obtain the final property representation of the protein sequence.

Further, the expression of the one-dimensional convolution operation in the residual error module is as follows:

wherein the content of the first and second substances,is a convolution operation, x_inAnd x_outThe convolved input tensor and output tensor, respectively, k the convolution kernel and b the offset vector.

Further, the expression of the Hard-Swish activation function used by the residual module is:

wherein x is input data; relu6(x +3) indicates that changing x +3 less than 0 to 0, greater than 6 takes 6.

Further, the one-dimensional convolution used by the residual module is a depth separable one-dimensional convolution.

Further, in the constructed deep convolutional network model for predicting protein torsion angles, the convolutional network outputs four values for each amino acid residue for a given protein sequence: sin (φ), cos (φ), sin (ψ) and cos (ψ); the values of the torsion angles Phi and Psi are calculated by Phi arctan (sin (Phi)/cos (Phi)) and Psi arctan (sin (Psi)/cos (Psi)), respectively.

Further, the loss function in step 5 is:

wherein N is the number of protein sequences in the mini-batch, L is the length of the protein sequences, 4 represents the vector length formed by the sine value and the cosine value of the torsion angles Psi and Phi, lambda is the regularization parameter of weight attenuation, W is the weight parameter of the convolution network,is the output of the prediction network, Y ∈ R^N×L×4Is composed ofThe corresponding true tensor.

Compared with the prior art, the invention has the following beneficial effects:

the method adopts the physicochemical properties of protein amino acids and the PSSM spectrum matrix to represent the protein sequence characteristics, realizes the prediction of the protein torsion angle based on the lightweight convolution network designed by deep separable convolution, can accurately predict the protein torsion angle, and has the advantages of small model and high prediction speed.

Drawings

FIG. 1 is a basic flowchart of a method for predicting a torsion angle of a protein based on a lightweight deep convolutional network according to an embodiment of the present invention;

fig. 2 is a schematic structural diagram of a residual error module constructed in a protein torsion angle prediction method based on a lightweight deep convolutional network according to an embodiment of the present invention;

fig. 3 is a schematic structural diagram of a deep convolution network model for predicting a protein torsion angle, which is constructed in a protein torsion angle prediction method based on a lightweight deep convolution network according to an embodiment of the present invention.

Detailed Description

The invention is further illustrated by the following examples in conjunction with the accompanying drawings:

as shown in fig. 1, a method for predicting a protein torsion angle based on a lightweight deep convolutional network includes:

step S101: constructing a protein torsion angle dataset based on a PISCES server, the protein torsion angle dataset comprising a plurality of protein sequences; specifically, the PISCES server is a protein sequence selection server that can select a data set satisfying criteria from a protein structure database (PDB) according to a user-specified structural quality and maximum sequence identity;

step S102: extracting torsion angles Phi and Psi corresponding to each amino acid residue in each protein sequence in the protein torsion angle data set from an RCSB PDB database by adopting BioPython, and adding the torsion angles Phi and Psi into the protein torsion angle data set;

step S103: performing multi-sequence alignment on the uniref90 database and each protein sequence in the protein torsion angle dataset to generate a PSSM (phosphosilicate-like syndrome) spectrum matrix corresponding to the protein sequence, and constructing protein sequence characteristics based on the PSSM spectrum matrix of the protein sequence and the physicochemical characteristics of amino acids;

step S104: designing a residual error module, and constructing a depth convolution network model for predicting the protein torsion angle based on the designed residual error module; the residual module involves computational operations including: one-dimensional convolution, one-dimensional batch processing normalization, Hard-Swish activation function, average pooling, full connection and residual connection;

step S105: constructing a loss function of a deep convolution network model for training and predicting the protein torsion angle;

step S106: training a deep convolution network model for predicting the protein torsion angle based on the constructed loss function;

step S107: and predicting the protein torsion angle based on the trained deep convolution network model for predicting the protein torsion angle.

Further, the step S101 includes:

selecting the protein structure database with resolution less than that of the protein structure database based on the PISCES serverIs identical with the sequenceProtein sequences with a potency of less than 25% were added to the protein torsion angle dataset. As an implementation mode, the resolution smaller than the resolution selected from the protein structure database based on the PISCES server10701 protein sequences with sequence identity less than 25%.

Further, the step S102 includes:

based on the PDBID in the constructed data set, firstly downloading a PDB file from an RCSB PDB database; then the pdb file is processed using PDBParser in the BioPython toolkit; the torsion angle value in the corresponding protein sequence can be obtained by calling the function get _ phi _ psi _ list () in Polypeptide class. One for each amino acid residue, Phi and Psi.

Specifically, after the step S102, the method further includes:

the protein torsion angle data set is divided into a training set, a validation set and a test set. As an embodiment, 9677 protein sequences and their corresponding torsion angle data are randomly selected as a training set, 512 protein sequences and their corresponding torsion angle data are selected as a verification set, and the remaining 512 protein sequences and their corresponding torsion angle data are selected as a test set for 10701 protein sequences in the protein torsion angle data set and the torsion angle data corresponding to each protein sequence.

Further, the step S103 includes:

first download uniref90 database from link ftp:// ftp. ebi. ac. uk/pub/databases/uniprot/current _ release/uniref; then, for each protein sequence in the protein torsion angle dataset, performing multiple sequence alignment with the protein sequences in uniref90 database by calling PSI-BLAST program, generating PSSM spectrum matrix corresponding to the protein sequence, in which each amino acid residue corresponds to a 20-dimensional numerical vector; and splicing the vector with a 7-dimensional numerical vector representing the physicochemical characteristics (including hydrophilicity, hydrophobicity, chargeability, molecular weight, accessibility, volume and specific volume) of the amino acid corresponding to the protein sequence to obtain the final characteristic representation of the protein sequence. For example, for a protein sequence of length N, the corresponding feature matrix size is N × 27.

Further, in step S104:

the expression of the convolution operation in the residual error module is as follows:

wherein the content of the first and second substances,is a convolution operation, x_inAnd x_outThe convolved input tensor and output tensor, respectively, k the convolution kernel and b the offset vector.

The purpose of batch normalization is to normalize the features in each channel to a standard normal distribution. By introducing batch normalization, not only can the training speed of the network be increased, but also the overfitting of the network to a certain specific sample can be prevented, and therefore the generalization capability of the network is improved.

The expression of the Hard-Swish activation function used by the residual module is:

where x is input data, relu6(x +3) indicates that x +3 is less than 0 and becomes 0, and 6 is taken if x +3 is greater than 6, i.e., relu6(x +3) ═ min (max (0, x +3), 6). The Hard-Swish activation function can maintain a lower computational burden than the Swish activation function with unchanged performance. In particular, the Hard-Swish function is still capable of gradient updates in intervals where the input value is less than zero. Specifically, the input data to the first residual module in the deep convolutional network model that predicts the protein torsion angle is the protein sequence feature vector.

To reduce the amount of computation, the present invention uses a depth separable one-dimensional convolution. Deep separable convolution significantly reduces the number of parameters required for convolution computation and hence the amount of computation by splitting the normal convolution into a channel-by-channel convolution (i.e., a group convolution, the number of groups being the same as the number of input channels, one convolution kernel being responsible for one channel, and one channel being convolved by only one convolution kernel) and a point-by-point convolution (a convolution with a convolution kernel size of 1).

When input data enters the module, the operation amount is reduced by using the deep separable convolution, so that the channel number is increased firstly, and then batch processing normalization, Hard-Swish nonlinear activation and channel-by-channel convolution operations are carried out. This is because information in a part of the channels is inevitably lost when data passes through the active function layer. If there are more channels, then the information may still be stored in other channels. And the separable convolution is used for reducing the operation amount, so that the operation amount can be kept not to be increased under the condition that the number of channels is increased, and the subsequent processing can be carried out after the number of channels is increased. The invention chooses to increase the number of input channels by k times the number of input channels.

In addition, in order to excavate the relevance between channels, the method introduces channel attention operation for a residual error module, namely sequentially carrying out average pooling, full connection, Hard-Swish nonlinear activation, full connection and Sigmoid nonlinear activation, and multiplying the output result serving as the result of weight and depth convolution channel by channel in sequence. This operation may improve the representational capacity of the network by modeling the channel dependencies. In particular, by adjusting features on a channel-by-channel basis, important channel features can be made more interesting and those less important channel features can be suppressed.

Based on the designed residual module, the invention constructs a deep convolution network model for torsion angle prediction (as shown in FIG. 3). In particular, the present invention employs n consecutive residual modules to capture local and non-local interactions between amino acid residues in a protein sequence. For a given protein sequence, the convolutional network will output four values for each amino acid residue: namely sin (Phi), cos (Phi), sin (Psi) and cos (Psi), where Phi is the twist angle Phi and Psi is the twist angle Psi. The values of the torsion angles Phi and Psi can be calculated by Phi ═ a rc t a n (Phi s i n () Phi/and Psi ═ arctan (sin (Psi)/cos (Psi)) respectively.

Further, in step S105:

order toAnd the prediction network outputs, wherein N is the number of protein sequences in the mini-batch, L is the length of the protein sequences, and 4 represents the length of a vector consisting of sine values and cosine values of the torsion angles Psi and Phi. In addition, order andthe corresponding real tensor is Y ∈ R^N×L×4. Then based on the mean square error, the following loss function may be defined:

wherein λ is the regularization parameter of weight attenuation, and W is the weight parameter of the convolutional network.

Further, in step S106:

in order to train a deep convolution network model for predicting a protein torsion angle, firstly, initializing weight parameters in a convolution layer by adopting a kaiming _ normal method, and simultaneously setting a bias parameter to be 0; then training a deep convolution network for predicting the protein torsion angle based on the loss function in the step S105 by adopting an Adam optimizer; as an implementation mode, the mini-batch size (N) during training is 32, and the learning rate of Adam is 0.001. In particular, to prevent the network from over-fitting the training data, we introduce Dropout with a p-value of 0.1 in each convolutional layer and terminate the network training with an early stop method based on the validation set, specifically, the tolerance value (probability) of the early stop method is 8.

In conclusion, the protein sequence characteristics are represented by the physicochemical properties of protein amino acids and the PSSM spectrum matrix, and the prediction of the protein torsion angle is realized based on the lightweight convolution network designed by the deep separable convolution.

The above shows only the preferred embodiments of the present invention, and it should be noted that it is obvious to those skilled in the art that various modifications and improvements can be made without departing from the principle of the present invention, and these modifications and improvements should also be considered as the protection scope of the present invention.