CN109817276B - A Method for Protein Secondary Structure Prediction Based on Deep Neural Network - Google Patents

Abstract

The invention relates to a protein secondary structure prediction method based on a deep neural network. According to the method, the interdependent characteristics of protein sequences can be fused through a plurality of different layers of convolutional neural networks, single amino acid characteristics are extracted at the same time, then the characteristics are input into the convolutional neural network for further fusion, and a mapping relation between the characteristics and 8 categories of proteins is established through a full-connection layer. Finally, the RMSProp optimizer is used for training the deep neural network based on the cross entropy error between labels and logits, so that the accuracy of protein secondary structure prediction is effectively improved.

Description

A Method for Protein Secondary Structure Prediction Based on Deep Neural Network

technical field

The invention relates to a protein secondary structure prediction method based on a deep neural network, which includes technologies such as a convolutional neural network, a recurrent neural network, and protein structure prediction.

Background technique

Proteins have four levels of structure, among which the secondary structure refers to the first level of folding of the protein, which is a general structure formed by the folding of part of the protein chain. The structure formed by a protein chain depends entirely on its amino acid sequence, but so far we have not fully understood the folding rules of protein sequences. The structure of protein plays an important role in the analysis of its function and pharmacology, so how to predict the structure of protein based on amino acid sequence is one of the challenges faced by bioinformatics. Accurate protein structure and function prediction, partly based on the accuracy of protein secondary structure prediction.

Protein secondary structure prediction methods have been extensively studied, and related methods such as neural network, hidden Markov model, and support vector machine have been successfully applied to predict three states of secondary structure, and a few methods predict eight states. Realizing the prediction of the secondary structure of the eight states of the protein can provide more detailed protein structure information, but it also increases the difficulty of prediction.

In recent years, deep learning methods based on hierarchical concepts to construct multi-layer architectures to learn complex concepts have made breakthroughs in many fields such as computer vision, speech processing, natural language processing, and bioinformatics. Among them, the convolutional neural network, as a neural network that specializes in processing data with a similar network structure, performs well on time series data and image data. Recurrent neural network, as a neural network that specializes in processing sequence data, performs well on time series data.

Contents of the invention

The technical problem solved by the present invention is: there are relatively few existing methods for predicting the eight states of proteins, and the prediction accuracy is low, which cannot meet the daily application requirements.

The technical solution adopted by the present invention to solve the problems existing in the prior art is to provide a protein secondary structure prediction method based on a deep neural network, which can fuse protein sequence interdependence features through a convolutional neural network and extract single amino acid features at the same time , and then input these features into the recurrent neural network for further fusion, and finally establish the mapping relationship with the 8 categories of proteins through the fully connected layer. Concrete technical scheme of the present invention comprises the following steps:

Extract protein sequence features: the input features of the network are amino acid sequence and structural information, and the data is downloaded from PDB and marked with the DSSP system; among them, the number of features of amino acid sequence information is 21, and the number of features of amino acid structure information is also 21, each 42 features of each amino acid were used to predict its corresponding secondary structure;

Multi-layer combined convolutional neural network feature extraction: Use multiple convolutional neural networks at different levels to extract features separately, and combine them with the original features and pass them to the next layer;

Bidirectional long-term short-term memory neural network feature extraction: use bidirectional long-term short-term memory neural network to extract features and pass them to the next layer; fully connected layer to establish feature mapping: use fully connected layer to map features to 8 different categories of proteins;

Train the deep neural network: Use the RMSProp optimizer to train the deep neural network built in the previous step based on the cross-entropy error between labels and logits, and use the L2 regularization method to prevent overfitting.

The further technical solution of the present invention is: in the feature extraction of multi-layer combined convolutional neural network, use a plurality of one-dimensional convolution kernels of different lengths to simulate windows of different sizes to extract amino acid sequence features, and specifically combine the construction of convolutional neural network method, including the following steps:

One-layer one-dimensional convolutional neural network construction: use one-dimensional convolutional networks with convolution kernel nodes of 3, 5, 7, 9, and 11 to perform convolution operations on amino acid sequences in proteins to extract features. The number of input channels is 42, corresponding to 42 features of an amino acid, the number of output channels is 50, and then the output of the convolutional network is activated with the ReLU function, and finally the batch normalization function is used to prevent the model from overfitting;

K-layer one-dimensional convolutional neural network construction: run the above-mentioned one-layer one-dimensional convolutional neural network construction method iteratively k times to construct a k-layer one-dimensional convolutional neural network, wherein, starting from the second layer, The number of input channels and the number of output channels of the convolutional network are both 50;

Combined output of original features and multiple convolutional neural network output features: In the present invention, the features of the amino acid sequence are extracted with 1-layer, 2-layer and 3-layer one-dimensional convolutional neural networks, and combined with the original features for output, then A total of 42+50×3=792 features are output to the bidirectional long-short-term memory neural network for further processing.

The technical effect of the present invention is: the present invention relates to a method for predicting protein secondary structure based on deep neural network, through the method of combining multi-layer combined convolutional neural network and bidirectional long-term and short-term memory neural network, the system can automatically complete protein secondary structure The hierarchical structure prediction solves the problem of low accuracy of traditional prediction methods. In the multi-layer combined convolutional neural network, multiple one-dimensional convolution kernels of different lengths are used to simulate windows of different sizes to extract amino acid sequence features, which avoids the problem that the classical method cannot effectively extract effective information between amino acid sequences, and enables the system to Simultaneously extract features between amino acid sequences and within sequences, further improving the feature extraction capabilities of deep architectures.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Fig. 2 is a flow chart of feature extraction of multi-layer combined convolutional neural network of the present invention.

Fig. 3 is a flow chart of the one-layer one-dimensional convolutional neural network of the present invention.

Detailed ways

The technical solutions of the present invention will be further described below in conjunction with specific embodiments.

As shown in Figure 1, the specific embodiment of the present invention is: provide a kind of protein secondary structure prediction method based on deep neural network, comprise the steps:

Step 100: Extract protein sequence features. Among the input features of the network, the number of features of amino acid sequence information is 21, the number of features of amino acid structure information is also 21, and 42 features of each amino acid are used to predict its corresponding secondary structure. Each protein sequence contains a maximum of 700 amino acids, so each protein can be represented by a 700×42 matrix. If a protein has less than 700 amino acids, fill the amino acid sequence features with 0. A protein sequence can be represented as:

中的一个向量，其中第j个坐标对应第j个特征。在本发明中，L＝700，D＝42。where L is the length of the protein sequence and D is the number of features per amino acid. Each column of X is an amino acid x. An amino acid with all the features can be seen as the space

A vector in , where the jth coordinate corresponds to the jth feature. In the present invention, L=700, D=42.

Y is a label dataset corresponding to L amino acids, which can be expressed as:

中的向量，其中第j个坐标对应第j个类别。where C is the number of categories in the dataset. In the present invention, C=8. Each column of Y is a space

A vector in , where the jth coordinate corresponds to the jth category.

The present invention will use the deep framework to use L amino acid training to construct the mapping function of X→Y. After training, when a new amino acid x is input, the deep architecture can use the mapping function to determine the label y corresponding to x.

Step 200: Multi-layer combined convolutional neural network feature extraction, using multiple convolutional neural networks at different levels to extract features respectively, and combining with the original features to pass to the next layer.

As shown in Figure 2, in the multi-layer combined convolutional neural network feature extraction step, the following steps are included:

Step 210: 1-layer one-dimensional convolutional neural network feature extraction, as shown in Figure 3, respectively use the one-dimensional convolutional network with the number of convolution kernel nodes as 3, 5, 7, 9, 11 to carry out the amino acid sequence in the protein The convolution operation extracts features. The number of input channels of the convolutional network is 42, corresponding to 42 features of an amino acid, and the number of output channels is 50.

Take a one-dimensional convolutional network with 3 convolution kernel nodes as an example. Its function is to extract features from the input data. It contains 50 convolution kernels inside, and each element that makes up the convolution kernel corresponds to 3 weight coefficients. and a bias, similar to the neurons of a feed-forward neural network. For each row feature of the input X, the convolution kernel will perform matrix element multiplication and summation on each amino acid and its adjacent amino acids in turn; after the convolution operation, the weighted summation based on each column feature obtained and superimposed deviation:

i∈{0,1,…,L ₁ }

等于输入X的第k个氨基酸。L₁是输出通道数，K₀是特征图的通道数，f是卷积核大小。在这里L₁＝50，K₀＝42，f＝3。where b ⁰ is the amount of bias. h ⁰ and h ¹ represent the input and output of the convolution,

Equal to the kth amino acid of input X. L ₁ is the number of output channels, K ₀ is the number of channels of the feature map, and f is the size of the convolution kernel. Here L ₁ =50, K ₀ =42, f=3.

Then activate the output of the convolutional network with the ReLU function:

Finally, use the batch normalization function to prevent the model from overfitting:

The one-layer one-dimensional convolutional network with the number of convolution kernel nodes of 5, 7, 9, and 11 is the same as the above structure, and the value of f is set to 5, 7, 9, and 11 respectively. Then the features extracted by the 1-layer one-dimensional convolutional neural network are

Step 220: 2-layer one-dimensional convolutional neural network feature extraction, one more operation of convolution, activation and batch normalization than the one-layer one-dimensional convolutional neural network feature extraction. Taking a one-dimensional convolutional network with a convolution kernel node number of 3 as an example, two convolution, activation, and batch normalization operations are performed on the input X in sequence:

i∈{0,1,…,L ₁ }

i∈{0,1,…,L ₂ }

where ^b1 is the amount of bias. d ¹ is the output of 1 convolution, activation and batch normalization operation, and h ² represents the input and output of convolution. L ₂ is the number of output channels, K ₁ is the number of channels of the feature map, and f is the size of the convolution kernel. Here L ₂ =50, K ₁ =50, f=3.

Then activate the output of the convolutional network with the ReLU function:

Finally, use the batch normalization function to prevent the model from overfitting:

The 2-layer one-dimensional convolutional network with convolution kernel nodes of 5, 7, 9, and 11 is the same as the above structure, and the value of f is set to 5, 7, 9, and 11 respectively. Then the features extracted by the 2-layer one-dimensional convolutional neural network are

Step 230: 3-layer one-dimensional convolutional neural network feature extraction, one more operation of convolution, activation and batch normalization than 2-layer one-dimensional convolutional neural network feature extraction. For the input X, three convolution, activation and batch normalization operations are performed sequentially, and the number of output channels of the third convolution operation is L ₃ =50. The features extracted by the three-layer one-dimensional convolutional neural network are

Step 240: Combining the original features with multiple convolutional neural network output features and outputting the input X with the features extracted in step 210, step 220, and step 230

combined and output to step 300. Step 210, step 220 and step 230 each output 50×5 features, then a total of 42+50×5×3=792 features are output to the bidirectional long-short-term memory neural network for further processing.

Step 300: Bidirectional long-short-term memory neural network feature extraction, respectively extracting features from the forward long-term short-term memory neural network and the backward long-term short-term memory neural network and combining them.

表示长短期记忆神经网络基于前t-1个氨基酸特征的在第t个位置提取到的特征表示。/>

表示长短期记忆神经网络基于后L-t个氨基酸特征的在第t个位置提取到的特征表示。长短期记忆神经网络隐藏层的结点数为800，则前向和后向长短期记忆神经网络整合后输出的特征个数为1600。in,

Indicates the feature representation extracted by the long short-term memory neural network based on the first t-1 amino acid features at the tth position. />

Indicates the feature representation extracted by the long-short-term memory neural network based on the last Lt amino acid features at the tth position. The number of nodes in the hidden layer of the long-term short-term memory neural network is 800, and the number of features output after the integration of the forward and backward long-term short-term memory neural networks is 1600.

Step 400: The fully connected layer establishes a feature map, and uses the fully connected layer to map the feature g _t extracted in step 300 to 8 different categories of proteins.

where w _st is the weight connecting the sth feature to the tth category, and _bt is the bias of the tth category. g _st is the t-th feature of the s-th amino acid output in step 300 . N is the number of features of g _st . Here N=1600.

Step 500: Train the deep neural network, use L labeled data to optimize the parameter space W, so that the deep architecture has better discrimination ability. This task can be transformed into an optimization problem:

in

T represents the loss function. In the present invention, T is a cross-entropy error function, and the RMSProp optimizer is used to train the deep neural network constructed in the previous steps based on T, and an L2 regularization method is used to prevent overfitting.

The present invention proposes a protein secondary structure prediction method based on a deep neural network, using the feature extraction of a multi-layer convolutional neural network, the combination of multiple different-level features and the feature extraction of a bidirectional long-short-term memory neural network to achieve a combination of Based on the deep structure of protein secondary structure prediction, it effectively improves the accuracy of protein secondary structure prediction.

The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be assumed that the specific implementation of the present invention is limited to these descriptions. For those of ordinary skill in the technical field of the present invention, without departing from the concept of the present invention, some simple deduction or replacement can be made, which should be regarded as belonging to the protection scope of the present invention.

Claims (2)

1. A protein secondary structure prediction method based on a deep neural network comprises the following steps:

extracting protein sequence characteristics: the input features of the network are the sequence and structure information of amino acids, and the data is downloaded from PDB and marked by DSSP system; wherein the number of the characteristics of the amino acid sequence information is 21, the number of the characteristics of the amino acid structure information is 21, and 42 characteristics of each amino acid are used for predicting the corresponding secondary structure;

multilayer combined convolutional neural network feature extraction: respectively extracting features by using a plurality of convolutional neural networks with different layers, combining the features with the original features, and transmitting the combined features to the next layer;

two-way long-short-term memory neural network feature extraction: extracting features by using a two-way long-short-term memory neural network and transmitting the features to the next layer; the number of nodes of the hidden layer of the long-short-period memory neural network is 800, and the number of features output after the integration of the forward and backward long-short-period memory neural networks is 1600;

the full connection layer establishes a feature map: mapping features to 8 different classes of proteins using the full-ligation layer;

training the deep neural network: the deep neural network constructed in the previous step was trained based on cross entropy errors between labels and logits using an RMSProp optimizer, and an L2 regularization method was used to prevent overfitting.

2. The protein secondary structure prediction method based on deep neural network according to claim 1, wherein in the multilayer combined convolutional neural network feature extraction, a plurality of one-dimensional convolution kernels with different lengths are used for simulating windows with different sizes to extract amino acid sequence features, and the construction mode of the combined convolutional neural network specifically comprises the following steps:

1 layer one-dimensional convolution neural network construction: the method comprises the steps of performing convolution operation on amino acid sequences in proteins by using one-dimensional convolution networks with the number of convolution kernel nodes of 3,5,7,9 and 11 respectively to extract characteristics, wherein the number of input channels of the convolution network is 42, the number of output channels of the convolution network corresponds to 42 characteristics of one amino acid, the number of output channels of the convolution network is 50, activating output of the convolution network by using a ReLU function, and finally preventing model overfitting by using a batch normalization function;

2-layer one-dimensional convolutional neural network construction: the characteristic extraction of the one-dimensional convolution neural network is more than that of the 1-layer one-dimensional convolution neural network by 1 time, and the convolution, activation and batch normalization operations are performed; sequentially performing convolution, activation and batch normalization operation on the input X for 2 times;

3-layer one-dimensional convolution neural network construction: the characteristic extraction of the one-dimensional convolution neural network is more than that of the 2-layer one-dimensional convolution neural network by 1 convolution, activation and batch normalization operation; sequentially carrying out convolution, activation and batch normalization operation on the input X for 3 times;

the original features are combined with a plurality of convolutional neural network output features to output: extracting amino acid sequence features by using 1 layer, 2 layer and 3 layer one-dimensional convolutional neural networks respectively, and combining with original features to output, so as to share

The characteristic is output to the two-way long-short-term memory neural network for continuous processing.

Images