CN113743301B - Solid-state nanopore sequencing electric signal noise reduction processing method based on residual self-encoder convolutional neural network - Google Patents

Abstract

The application provides a solid-state nanopore sequencing electric signal noise reduction processing method based on a residual error self-encoder convolutional neural network, which is characterized in that a self-encoder, a residual error bottleneck module, a shortcut connection, a convolutional layer, a batch regularization layer, an activation layer and other structures are introduced to construct a deep neural network model, a solid-state nanopore sequencing electric signal data set is utilized to train the model, so that the model accurately learns the characteristic mode of sequencing electric signal noise, a mapping from a noise signal to a clean signal is established, and finally the clean signal corresponding to the noise signal is predicted and estimated by using the learned mapping. The signal noise reduction method based on the residual error self-encoder convolutional neural network, disclosed by the application, enhances the recognition capability of the neural network on the noise part of the electric signal, establishes the accurate mapping from the noise electric signal to the clean signal, and realizes real-time denoising.

Description

A solid-state nanopore sequencing telecommunications based on residual autoencoder convolutional neural network Signal noise reduction processing method

Technical field

The invention belongs to the field of digital signal processing technology, and specifically relates to a solid-state nanopore sequencing electrical signal denoising processing method based on a residual autoencoder convolutional neural network, which can be applied to machine learning and digital signal processing technology.

Background technique

In recent years, solid-state nanopores have received widespread attention as an emerging nanostructured material. Compared with existing gene sequencing systems, the gene sequencing system developed based on solid-state nanopores is stable, has the advantages of high throughput and low cost, and has therefore attracted widespread attention from researchers. However, there are still a lot of problems that need to be solved in the development of solid-state nanopore sequencing systems, among which the noise in the solid-state nanopore sequencing signals seriously affects the accuracy of the final sequencing results. Further research shows that the noise in the solid-state nanopore sequencing signal is mainly composed of flicker noise in the low-frequency part and thermalnoise, dielectric noise, and capacitive noise in the high-frequency part, but essentially: the noise sequencing electrical signal Y is composed of a clean sequencing electrical signal X and the additive sequencing system noise N are superimposed, which can be expressed as a mathematical expression: Y=X+N. In order to improve the sequencing accuracy, it is necessary to perform noise reduction processing on the noisy electrical signal, remove the superimposed noise as much as possible, and restore the signal to the state of a clean signal.

At present, the noise reduction processing method for nanopore sequencing electrical signals mainly uses various low-pass filters such as Bessel filter to remove the high-frequency noise part to improve the signal-to-noise ratio. However, low-pass filtering can only process the noisy electrical signals to a level that is acceptable. The extent to which signal analysis is performed, and at the expense of the high-frequency characteristic portion of the sequencing signal. At the same time, this type of signal processing method cannot solve the impact of low-frequency flicker noise in the noise signal, and the processing effect is unsatisfactory, which brings great inconvenience to subsequent sequencing signal analysis.

Contents of the invention

In view of the problems existing in the existing technology and based on the needs of reality and production practice, the applicant invested a large amount of money and after long-term research, provided a solid-state nanopore sequencing electrical signal reduction method based on a residual autoencoder convolutional neural network. Noise processing method, which enhances the neural network's ability to identify the noisy part of electrical signals and establishes an accurate mapping of noisy electrical signals to clean signals to achieve real-time denoising of solid-state nanopore electrical signals.

According to the technical solution of the patent of the present invention, a solid-state nanopore electrical signal denoising processing method based on residual autoencoder convolutional neural network is provided. The method introduces the autoencoder structure, residual bottleneck module, shortcut connection, convolution Create a deep neural network model using stacked layers, batch regularization layers, activation layers and other structures. Combined with the created solid-state nanopore sequencing electrical signal training data set, the training model accurately learns the characteristics of noise in the signal and establishes a process from noise signal to clean signal. Mapping, so that the learned mapping can be used to predict and estimate the corresponding clean signal based on the noise signal.

Further, the solid-state nanopore electrical signal denoising processing method based on residual autoencoder convolutional neural network includes the following steps:

Step S1, build a residual autoencoder convolutional neural network model;

Step S2, select and create a training data set, and set the training parameters of the residual autoencoder convolutional neural network model;

Step S3, according to the set model training parameters, train the residual autoencoder convolutional neural network model with the goal of minimizing the loss function, and complete the construction of the neural network model for solid-state nanopore sequencing electrical signal denoising processing;

Step S4: Input the noise electrical signal to be processed into the solid-state nanopore sequencing electrical signal noise reduction processing neural network model, and output the noise-reduced signal.

Preferably, the residual autoencoder convolutional neural network model in step S1 includes an encoder part and a decoder part. The encoder part includes multiple residual coding modules, and the residual coding module includes multiple convolutional layers, multiple Activation layer and multiple batch regularization layers; the decoder part consists of multiple inverse convolution decoding modules, a convolution layer and a sigmoid activation layer. The inverse convolution decoding module consists of multiple inverse convolution layers and multiple activation layers. and multiple batch regularization layers.

Preferably, in step S2, the training data set in step S2 consists of clean sequencing electrical signals and corresponding noise signals. Further, blank solid-state nanopore sequencing system noise is collected, a simulated via signal data set is constructed based on the DNA base signal standard library and superimposed with the collected solid-state nanopore sequencing system noise to create a training data set, and the residual self-encoding is set training parameters of the convolutional neural network model.

More preferably, in step S3, the loss function in step S3 is the mean square error function:

_{Among them} , Mapping to clean signals.

The initial value of the weight of the neural network model for the solid-state nanopore sequencing electrical signal denoising process described in step S3 is generated by a Gaussian random function, and the weight parameter of the last batch regularization layer of the residual module is initially set to zero.

Compared with the existing technology, the beneficial effect of the present invention is that: the present invention trains the residual autoencoder convolutional neural network model for a variety of different noise variances to form a corresponding signal denoising processing neural network model under the noise variance, and The base via signal to be processed is denoised through the signal denoising neural network model under the noise variance corresponding to the image to be processed, and the processing speed is fast.

Description of the drawings

Figure 1 is a flow chart of a solid-state nanopore sequencing electrical signal denoising processing method based on a residual autoencoder convolutional neural network according to the present invention;

Figure 2 is a schematic diagram of the internal structure of the residual autoencoder convolutional neural network model according to the present invention;

Figure 3 is a schematic diagram of the internal structure of the residual coding module according to the present invention.

Detailed ways

The technical solutions in the patent embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the patent embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the patent of the present invention, not all of them. Example. Based on the embodiments in the patent of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the patent of the present invention.

The solid-state nanopore sequencing electrical signal denoising processing method based on the residual autoencoder convolutional neural network of the present invention introduces the autoencoder structure, the residual bottleneck module, the shortcut connection, the convolution layer, the batch regularization layer, and the activation layer. Create a deep neural network model with other structures, and combine it with the created solid-state nanopore sequencing electrical signal training data set to train the model to accurately learn the characteristics of the noise in the signal, and establish a mapping from the noise signal to the clean signal, so that the learned mapping can be used Noisy signals are predicted and estimated from their corresponding clean signals.

Specifically, the solid-state nanopore sequencing electrical signal denoising processing method based on the residual autoencoder convolutional neural network of the present invention mainly includes the following steps: building a residual autoencoder convolutional neural network model, and the residual autoencoder convolutional neural network model is constructed. The encoder convolutional neural network model includes two main parts: an encoder and a decoder, each part including multiple convolutional layers and an activation layer after each of the convolutional layers; a solid-state nanopore electrical signal training set is created, and Setting the training parameters of the residual autoencoder convolutional neural network model; training the residual autoencoder neural network with the goal of minimizing the loss function according to the residual autoencoder neural network model and its training parameters The model forms a solid-state nanopore electrical signal noise reduction processing model; the solid-state nanopore electrical signal to be processed is input into the solid-state nanopore electrical signal noise reduction processing model, and the noise-reduced electrical signal is output.

The present invention will be further described below with reference to the accompanying drawings and preferred embodiments. As shown in Figure 1, a solid-state nanopore electrical signal denoising processing method based on residual autoencoder convolutional neural network includes the following steps:

Step S1, build a residual autoencoder convolutional neural network model;

Step S2, select to create a training data set and set the initial parameters of the model;

Step S3, according to the set model training parameters, train the residual autoencoder convolutional neural network model with the goal of minimizing the loss function, and complete the construction of the solid-state nanopore sequencing electrical signal noise reduction processing neural network model;

Step S4: Input the noise electrical signal to be processed into the solid-state nanopore sequencing electrical signal noise reduction processing neural network model, and output the noise-reduced signal.

Among them, in step S1, a residual autoencoder neural network model is built. The residual neural network model includes an encoder part and a decoder part; the encoder part includes multiple residual coding modules, and the residual coding module includes multiple convolutions. Product layer, multiple activation layers and multiple batch regularization layers; the decoder part includes multiple deconvolution layers, multiple activation layers and multiple batch regularization layers. Furthermore, the decoder part consists of multiple inverse convolution decoding modules, a convolution layer and a sigmoid activation layer. The inverse convolution decoding module consists of multiple inverse convolution layers, multiple activation layers and multiple batch regularization layer composition.

The residual coding module of the encoder part of the residual neural network model includes a plurality of convolution layers with a convolution kernel larger than 1×1 and a plurality of convolution layers with a convolution kernel of 1×1. Furthermore, the residual coding module includes one convolution layer with a convolution kernel size of 3×3 and two convolution layers with a convolution kernel size of 1×1; and the first and third layers of the residual coding module are The size of the three-layer convolution kernel is 1×1, and the size of the second-layer convolution kernel is 3×3. The constructed residual coding module is a bottleneck neural network structure. In a further embodiment, the residual coding module includes a shortcut connection for directly performing identity mapping on the input and output, bypassing the bottleneck convolutional network structure.

In step S2, select and create a training data set and set initial parameters of the model. The training data set in step S2 consists of clean sequencing electrical signals and corresponding noise signals. To further collect the noise of the blank solid-state nanopore sequencing system, a simulated via signal data set is constructed based on the DNA base signal standard library and superimposed with the collected solid-state nanopore sequencing system noise to create a training data set, and the residual autoencoder is set Training parameters for the convolutional neural network model.

Step S3 further trains the residual autoencoder convolutional neural network model based on the residual autoencoder convolutional neural network model and its training parameters with the goal of minimizing the loss function to form a signal noise reduction processing neural network model; The loss function in step S3 is the mean square error function:

_{Among them} , Mapping to clean signals. Furthermore, the initial value of the weight of the neural network model for the solid-state nanopore sequencing electrical signal denoising process described in step S3 is generated by a Gaussian random function, and the weight parameter of the last batch regularization layer of the residual module is initially set to zero. .

Step S4 further preprocesses the noise electrical signal to be processed and inputs it into the electrical signal noise reduction processing neural network model, and outputs the noise reduction processed signal.

As shown in Figure 2, further explanation of the residual autoencoder convolutional neural network model built in step S1 is added here. In the preferred embodiment of the present invention, the encoder part of the residual autoencoder convolutional neural network consists of eight residual coding modules. Each residual coding module includes a three-layer bottleneck convolution network structure and a shortcut connect. The decoder part consists of four inverse convolution decoding modules, a convolution layer and a sigmoid activation layer, where each inverse convolution module consists of an inverse convolution layer, an activation layer and a batch regularization layer.

As shown in Figure 3, the residual encoding module in the residual autoencoder convolutional neural network model is further explained. The residual coding module includes a total of three convolutional layers. The convolutional kernel size of the first convolutional layer is 1×1 and the number of channels is the module channel number N _C ; the convolutional kernel of the second convolutional layer is 3 ×3, the number of channels is the number of module channels N _C ; the convolution kernel of the third convolution layer is 1 × 1, and the number of channels is four times the number of module channels N _C , that is, 4N _C . The residual coding module constructed in this way belongs to the bottleneck neural network structure, and the corresponding functions implemented by the three convolutional layers are dimension compression, feature extraction, and dimension recovery. The shortcut connection is used to bypass the bottleneck convolutional network structure and directly perform identity mapping on the input and output.

Therefore, the residual coding module can be expressed as the following function:

Among them, X and Y represent the input and output matrices respectively, and W is the weight parameter variable in the bottleneck neural network in the residual coding module. represents the mapping transformation learned by the bottleneck neural network structure, and I(·) represents the identity mapping of shortcut connections. The bottleneck neural network structure in the residual coding module is/> Part of what is learned is the mapping of inputs to the differences between inputs and outputs.

Furthermore, the output after each convolutional layer is processed by the ReLU activation layer and the regularization layer before being input to the next convolutional layer for calculation. Among them, the ReLU activation layer in the network structure can eliminate neurons whose values are lower than 0 after convolution calculation to screen out more meaningful features, improve the stability of the network, and effectively avoid the problem of gradient explosion.

In the method of the present invention, the data of the method comes from the blocking current signal generated when the DNA collected by the solid-state nanopore sequencing system passes through the solid-state nanopore. Among them, the training data set is constructed by carrying out the designed DNA standard sequence and conducting the solid-state nanopore sequencing experiment to collect the standard sequence via signal data. In addition, a standard signal database can also be constructed based on standard sequence via signal data, and then a large amount of simulated signal data can be generated based on the standard signal database to be used as training data. At the same time, the simulated signal data can be superimposed with different levels of system noise to construct signal data sets affected by different system noises to train neural network models that can handle different noise effects. The regularization layer uses the BatchNormalization method to standardize the calculation results of the convolution layer so that they obey the same distribution. The introduction of BatchNormalization can make the function represented by the entire network smoother, which is beneficial to stable network training and subsequent optimization.

Step S2 of the embodiment of the present invention also includes: resampling and dimensionally converting the noise base sequence electrical signal and the clean base sequence electrical signal into 20×10 two-dimensional signal data. This is because the joint action of dimension conversion and convolution calculation is more conducive to the model mining relevant features in the time dimension of current data, speeding up training to a certain extent, and improving noise processing performance. Since the input data is only current amplitude data, the number of input data channels is 1. After that, the channel number of the first convolutional layer was set to 64, with the purpose of increasing the dimension of the input data and improving the neural network model's ability to learn data features. The number of module channels N _C of the eight residual coding modules is set to 64, 64, 32, 32, 16, 16, 8, and 8 in sequence, which is used to mine and extract features in the signal. Such processing can effectively improve the accuracy of model learning. Correspondingly, the channel numbers of the eight decoding modules are set to 32, 64, 128, and 256 in sequence for signal reconstruction. The learning rate of the neural network is set to 0.001 (in other embodiments, it can be set to any value from 0.1 to 0.001, and the optimization selection is made according to the actual situation).

In this embodiment, the initial value of the weight of the neural network model is generated by a Gaussian random function, but the weight parameter of the last batch regularization layer of each residual module is initially set to zero to ensure the residual score of the residual module. The initial generation result of the branch is zero, and the residual coding module will initially behave as an identity mapping from input to output.

In step S3, according to the residual autoencoder convolutional neural network model and its training parameters, the residual autoencoder convolutional neural network model is trained with the goal of minimizing the loss function to complete the signal denoising processing neural network model. of construction.

The optimizer for neural network training chooses the Adam optimizer. When optimizing the network parameters, the Adam optimizer combines the first-order momentum of the gradient (the exponential moving average of the gradient) and the second-order momentum (the exponential moving average of the squared gradient) when calculating the iteration step size. value) is taken into consideration in the calculation, which can ensure that the first-order momentum and second-order momentum of the gradient are adaptively adjusted during iterative optimization at each time step, effectively preventing the neural network model from converging to the local optimal solution during optimization. while accelerating the overall efficiency of optimization.

In step S4, the noise electrical signal to be processed is preprocessed and then input into the electrical signal noise reduction processing neural network model, and a noise reduction processed signal is output.

In step S2, you can choose to create solid-state nanopore sequencing electrical signal data sets affected by different noises, and train a residual autoencoder neural network model that specifically handles the corresponding noise as described in step S3. In step S4, the noise electrical signal to be processed is input into the neural network model corresponding to the influence of noise, and the corresponding clean signal can be predicted and the noise-processed sequencing electrical signal can be output.

According to the signal denoising processing method of the present invention, the trained residual autoencoder neural network model can process noise signals very quickly, and can complete signal denoising in less than 0.1 seconds under most hardware conditions. Processing has great advantages in practical applications, especially in situations where real-time noise reduction processing is required.

Compared with the existing technology, the solid-state nanopore sequencing electrical signal denoising processing method provided by the present invention combines the autoencoder and the convolutional neural network structure. In the encoder part, the bottleneck module in the residual network is used to construct the residual autoencoder. The machine neural network uses the learning ability of the convolutional layer and the screening ability of the ReLU activation layer to perform deep learning on the characteristics of sequencing electrical signal noise. It learns the most representative noise characteristics from the training set, and then uses the deconvolution layer and The conversion of the sigmoid activation layer establishes an accurate mapping of the noise signal to the clean signal, which can achieve real-time denoising.

In a further solution, the present invention can also have the following beneficial effects: the bottleneck design of the residual bottleneck module can effectively increase the depth of the model without sacrificing the time complexity of the model. The shortcut connection of identity mapping can effectively avoid the problem of gradient disappearance or gradient explosion caused by the increase in the number of network structure layers without adding additional parameters. It ensures the stability of the model while enjoying the benefits of deep networks. The accuracy rate is improved. At the same time, the residual bottleneck module uses convolution kernels of appropriate size, which can achieve excellent noise processing effects without introducing additional pooling layers, effectively avoiding the introduction of pooling layers that reduce model parameters and lead to model accuracy. The problem is that it is not enough and the noise processing effect becomes poor.

The above are only preferred specific implementations of the patent of the present invention, but the scope of protection of the patent of the present invention is not limited thereto. Any person familiar with the technical field can, within the technical scope disclosed in the embodiments of the patent of the present invention, Any easily imagined changes or substitutions should be covered by the protection scope of the patent of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the protection scope of the claims.

Claims (5)

1. A solid-state nanopore electrical signal noise reduction processing method based on a residual self-encoder convolutional neural network is characterized by comprising the following steps of: the method comprises the steps of introducing a self-encoder, a residual bottleneck module, a shortcut connection, a convolution layer, a batch regularization layer and an activation layer structure to construct a deep neural network model, training the model by utilizing a solid-state nanopore sequencing electric signal data set, accurately learning a characteristic mode of sequencing electric signal noise, establishing a mapping from a noise signal to a clean signal, and finally predicting and estimating the clean signal corresponding to the noise signal by utilizing the learned mapping;

which comprises the following steps:

step S1, constructing a residual error self-encoder convolutional neural network model;

s2, selecting and creating a training data set, and setting training parameters of the residual error self-encoder convolutional neural network model;

step S3, training the residual error self-encoder convolutional neural network model by taking a minimized loss function as a target according to set model training parameters, and completing construction of a neural network model for noise reduction processing of the solid-state nanopore sequencing electric signal;

s4, inputting the noise electric signal to be processed into a solid-state nanopore sequencing electric signal noise reduction processing neural network model, and outputting a noise reduction processed signal;

the residual self-encoder convolutional neural network model in the step S1 comprises an encoder part and a decoder part, wherein the encoder part comprises a plurality of residual coding modules, and the residual coding modules comprise a plurality of convolutional layers, a plurality of activating layers and a plurality of batch regularization layers; the decoder part consists of a plurality of deconvolution decoding modules, a convolution layer and a sigmoid activation layer, wherein the deconvolution decoding modules consist of a plurality of deconvolution layers, a plurality of activation layers and a plurality of batch regularization layers;

wherein the residual coding module comprises a plurality of convolution layers with convolution kernels larger than 1×1 and a plurality of convolution layers with convolution kernels of 1×1;

wherein, the residual coding module comprises a convolution layer with a convolution kernel size of 3×3 and two convolution layers with convolution kernels of 1×1; the convolution kernel sizes of the first convolution layer and the third convolution layer of the residual error coding module are 1 multiplied by 1, the convolution kernel size of the second convolution layer is 3 multiplied by 3, and the constructed residual error coding module is of a bottleneck neural network structure;

wherein, the loss function in step S3 is a mean square error function:

wherein X is _i 、Y _i And respectively creating a noise signal and a clean signal in the training set, wherein θ is a weight, n represents the number of data points in the signal, and F (-) is a mapping from the noise signal to the clean signal, which is obtained after training.

2. The method for noise reduction of solid state nanopore electrical signals based on residual self encoder convolutional neural network of claim 1, wherein the residual encoding module comprises a shortcut connection for directly mapping identities of inputs and outputs bypassing a bottleneck convolutional network structure.

3. The method for noise reduction processing of solid state nanopore electrical signals based on residual self encoder convolutional neural network of claim 1, wherein the deconvolution decoding module of the decoder portion consists of an deconvolution layer, an activation layer and a batch regularization layer.

4. The residual self-encoder convolutional neural network-based solid state nanopore electrical signal noise reduction processing method of claim 1, wherein the training dataset in step S2 consists of clean sequencing electrical signals and corresponding noise signals.

5. The solid state nanopore electrical signal noise reduction processing method based on the residual self encoder convolutional neural network according to claim 1, wherein in step S3, an initial value of a weight of a neural network model of the solid state nanopore sequencing electrical signal noise reduction processing is generated by a gaussian random function, and a weight parameter of a last batch regularization layer of a residual module is initially set to zero.