Deep convolutional neural network-based promoter interpretable prediction method for nannochloropsis

Abstract

The application discloses a kind of based on deep convolutional neural network's micro-nanqiqiao promoter explainable prediction method, comprising:1, the collection and pre-processing of micro-nanqiqiao promoter sequence data;2, based on dense connection convolution and attention mechanism establishes micro-nanqiqiao promoter classification network;3, using binary classification cross entropy loss function training micro-nanqiqiao promoter classification network, obtains optimal classification model;4, based on residual connection and hollow convolution establishes micro-nanqiqiao promoter importance score generation network;5, using prediction loss and saving quantity loss training micro-nanqiqiao promoter importance score generation network, obtains optimal generation model.The application can obtain relatively accurate micro-nanqiqiao promoter classification prediction result on different strain data, and based on the importance score of promoter nucleotide, the prediction result is faithfully explained, so as to provide tool and train of thought for subsequent micro-nanqiqiao research.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of bioinformatics, and particularly relates to a micro-nannochloropsis promoter interpretable prediction method based on a deep convolutional neural network. BACKGROUND

[0002] Micro-nannochloropsis is a kind of marine microalgae, which has attracted extensive attention and research due to its industrial application value in oil, pigment and food. Compared with other terrestrial plants, micro-nannochloropsis has the advantages of rich species, fast growth, high productivity and low resource consumption, which is attributed to its complex metabolic capacity. However, the wild type strain often cannot achieve the optimal yield in practical application, so it is a worthwhile attempt to improve the strain through synthetic biology methods. The promoter is a series of DNA sequences near the transcription start site. At the beginning of gene transcription, the promoter binds to the RNA polymerase and plays a key regulatory role. Therefore, in the field of synthetic biology, the micro-nannochloropsis promoter can be studied to control the expression of related genes and improve the yield of target products, so as to improve the economic feasibility of the production process.

[0003] In previous related research, a variety of machine learning models for promoter prediction have been developed. However, the training and testing of most models are based on promoter data of common experimental species such as humans and mice, and the performance of the models has not been verified for micro-nannochloropsis promoter data. The deep learning model developed for micro-nannochloropsis promoter sequence recognition still has some shortcomings in some aspects: first, the prediction performance of the existing deep learning model on some specific strain data sets still has a large room for improvement; second, the existing deep learning model can only give a classification judgment of whether the input sequence is a micro-nannochloropsis promoter, but cannot explain the specific basis for making the judgment, so the model has poor interpretability. SUMMARY

[0004] The present application is proposed to solve the above-mentioned problems in the prior art, and provides a micro-nannochloropsis promoter interpretable prediction method based on a deep convolutional neural network, so as to improve the accuracy of micro-nannochloropsis promoter prediction and improve the interpretability of the classification result through the importance evaluation of nucleotides.

[0005] In order to achieve the above-mentioned application purposes, the present application adopts the following technical solutions:

[0006] The micro-nannochloropsis promoter interpretable prediction method based on a deep convolutional neural network has the following steps:

[0007] Step 1, collection and preprocessing of micro-nannochloropsis promoter sequence data:

[0008] Step 1.1, screen out sequences with length W from the obtained microalgae promoter data, and use the extended CD-HIT-EST toolkit to remove duplicates from the screened sequences to obtain a positive sample dataset wherein, represents the mth positive sample in the positive sample dataset S + m,i represents the one-hot encoding vector of the ith nucleotide in the mth positive sample, and M is the total number of positive samples;

[0009] Step 1.2, divide the mth positive sample in the positive sample dataset S + into a plurality of subsequence fragments {{b m，1 , b m，2},..., {b m，i , b m，i+1},..., {b m，L-1 , b m，W}}, wherein {b m，i , b m，i+1} represents the i-th subsequence fragment of the mth positive sample;

[0010] randomly shuffle the arrangement order of nucleotides in each subsequence fragment in the mth positive sample , to obtain the mth negative sample corresponding to , thereby obtaining a negative sample dataset

[0011] merge S + and S - to obtain a preprocessed microalgae promoter dataset S = {s1,..., s n ,..., s N}, wherein s n represents the nth sequence sample, N = 2M is the total number of samples, and the label of the nth sequence sample s n is flag n ∈{0,1}.

[0012] Step 2, establish a microalgae promoter classification network, which successively includes an initial convolution module, a backbone convolution module, and a weight output module, and is used to process the nth sequence sample s n to obtain a normalized microalgae promoter binary classification prediction label p n .

[0013] Step 3, construct a binary classification cross-entropy loss function L using formula (1):

[0014] ​​​

[0015] Based on the microalgae promoter dataset S = {s1,..., s n ,..., s 2N} input into the microalgae promoter classification network for iterative training, and the binary cross-entropy loss function L is calculated to update the network parameters, when the number of training iterations reaches the set number of times or the binary cross-entropy loss function L converges, the training stops, thereby obtaining the optimal microalgae promoter classification model;

[0016] Step 4, establish a microalgae promoter importance score generation network, which sequentially includes a backbone module and an output module, and is used to process the mth positive sample to obtain the importance score c corresponding to the mth positive sample m ;

[0017] Step 5, training of the microalgae promoter importance score generation network:

[0018] Step 5.1, obtain the background distribution bg of the positive sample data using formula (3):

[0019]

[0020] Step 5.2, match the dimension of c m with the dimension of to obtain the importance score after dimension processing , thereby obtaining the perturbed distribution

[0021]

[0022] Step 5.3, input the mth positive sample to the optimal microalgae promoter classification model for processing to obtain the original predicted label p m ;

[0023] Step 5.4, use the Softmax Straight-Through sampling method to sample K perturbed samples from and input them into the optimal microalgae promoter classification model for processing to obtain K perturbed prediction results corresponding to p m , wherein , wherein represents the kth sample sampled from , and represents the perturbed prediction label output by the optimal microalgae promoter classification model after processing ;

[0024] Step 5.5, calculating the prediction loss L using formula (5) p :

[0025]

[0026] Step 5.6, calculating the preservation loss L using formula (6) s :

[0027]

[0028] In formula (6), t bit is a hyperparameter of the target preservation amount, and ∑ axis · represents summing up the elements in all dimensions of the tensor;

[0029] Step 5.7, constructing the overall loss function L using formula (7) total :

[0030] L total =L p +λ·L s (7)

[0031] In formula (7), λ is a weight hyperparameter balancing the two losses;

[0032] Step 5.8, inputting the positive sample dataset into the micro-nanoball algae promoter importance score generation network for training, and calculating the overall loss function L total to update the network parameters, and when the number of training iterations reaches a set number or the overall loss function L total converges, the training stops, thereby obtaining an optimal micro-nanoball algae promoter importance score generation model for evaluating the importance of the predicted labels output by the optimal micro-nanoball algae promoter classification model.

[0033] The micro-nanoball algae promoter interpretable prediction method based on a deep convolutional neural network according to the present application also has the following characteristics:

[0034] Step 2.1, the initial convolution module is sequentially composed of a two-dimensional convolution layer Conv2D and a convolution block attention module CBAM;

[0035] The nth sequence sample s n is input into the initial convolution module and sequentially processed by Conv2D and CBAM to obtain the upgraded feature initConv n ;

[0036] Step 2.2, the backbone convolution module is based on a DenseNet network and sequentially includes: a first dense convolution block, a second dimension reduction conversion block, a third dense convolution block, a fourth dimension reduction conversion block and a fifth dense convolution block;

[0037] Each dense convolution block is sequentially composed of D densely connected two-dimensional convolution blocks and a convolution block attention module CBAM, and each two-dimensional convolution block is sequentially composed of an activation function layer ReLU, a two-dimensional convolution layer Conv2D and a random inactivation layer Dropout;

[0038] Each dimension reduction conversion block is sequentially composed of a two-dimensional convolution layer with a convolution kernel of 1x1, an average pooling layer AP, a batch normalization layer BN and a convolution block attention module CBAM;

[0039] The dimension-increased feature initConv n is input into the backbone convolution module for processing, and a convolution feature denseConv n is obtained.

[0040] Step 2.3, the weight output module is sequentially composed of an average pooling layer AP, a first full connection layer DC1, an activation function layer ReLU, a batch normalization layer BN, a second full connection layer DC2 and an activation function layer Sigmoid;

[0041] The convolution feature denseConv n is input into the weight output module for processing, and a normalized micro-nanobubble algae promoter binary classification prediction label p n is obtained.

[0042] The step 4 includes:

[0043] Step 4.1, the backbone module includes: an initial convolution layer, a first residual group, a second residual group, a third residual group, a fourth residual group, a fifth residual group, a sixth convolution layer and a convolution jump connection layer;

[0044] Each residual group is composed of R residual blocks; each residual block is sequentially composed of a batch normalization layer BN, an activation function layer ReLU, a first dilated convolution layer dilatedConv1, a batch normalization layer BN, an activation function layer ReLU and a second dilated convolution layer dilatedConv2, and the input feature of the residual block is added to the output feature of the second dilated convolution layer dilatedConv2 through residual connection to obtain the output of the residual block; all dilated convolution layers in each residual group share a uniform dilated factor;

[0045] The mth positive sample is input into the backbone module of the micro-nanobubble algae promoter importance score generation network, and a first input feature is obtained after being processed by the initial convolution layer. the first input feature After processing by the first residual group, a second input feature is obtained the second input feature After processing by the second residual group, a third input feature is obtained the third input feature After processing by the third residual group, a fourth input feature is obtained the fourth input feature After processing by the fourth residual group, a fifth input feature is obtained the fifth input feature After processing by the fifth residual group, a residual group output feature ResGroupOutm is obtained; the residual group output feature ResGroupOut m After processing by the sixth convolutional layer, a backbone output feature BackboneOut is obtained m ;

[0046] The are respectively input into convolutional skip connection layers for processing, and five cross-layer features are sequentially obtained Thus, the convolutional feature BackboneOut' finally output by the backbone module is obtained according to formula (2) m :

[0047]

[0048] In formula (2), denotes the jth cross-layer feature;

[0049] Step 4.2, the output module is sequentially composed of a convolutional layer with a convolutional kernel number of 1 and an activation function layer Softplus;

[0050] After BackboneOut ′ m is input into the output module for processing, the mth positive sample is obtained. The corresponding importance score c m ={cb m,1 ,..., cb m，i ,..., cb m，W}, wherein cb m，i >0 represents the real value importance score of the ith nucleotide in the mth positive sample .

[0051] The electronic equipment of the present application comprises a memory and a processor, and is characterized in that the memory is used to store a program supporting the processor to execute the interpretable prediction method of the Nannochloropsis sp. promoter, and the processor is configured to execute the program stored in the memory.

[0052] The computer readable storage medium of the present application has a computer program stored thereon, and is characterized in that the computer program, when executed by a processor, performs the steps of the interpretable prediction method of the Nannochloropsis sp. promoter.

[0053] Compared with the prior art, the present application has the following beneficial effects:

[0054] 1. The present application directly maps the DNA sequence data of Nannochloropsis sp. to the binary classification probability of whether it is a promoter through an end-to-end neural network, which can save a lot of time and labor cost compared with traditional biological experimental methods.

[0055] 2. In the present application, the convolution block attention mechanism is combined in the densely connected convolutional neural network when constructing the Nannochloropsis sp. promoter classification network, which not only strengthens the repeated use of sequence features, but also enables the model to fully focus on the motif existence, interaction and position information in the sequence, providing effective help for training a better model, thereby improving the accuracy of Nannochloropsis sp. promoter prediction.

[0056] 3. The present application uses a deep generative model based on residual connection and cavity convolution to output the importance score of nucleotides in the Nannochloropsis sp. promoter sequence, so that the classification model has good interpretability, and compared with other interpretation methods based on perturbation or back propagation, it can more effectively capture the nonlinear interaction between biological sequence features, rather than individually evaluating the importance of each feature, thereby improving the quality of interpretability. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The flowchart of the method of the present application;

[0058] Figure 2 The overall network framework diagram of the present application;

[0059] Figure 3 The Nannochloropsis sp. promoter classification network structure diagram of the present application;

[0060] Figure 4 The comparison diagram of the five-fold cross-validation accuracy Acc of the classification network of the present application;

[0061] Figure 5 The Nannochloropsis sp. promoter importance score generation network structure diagram of the present application. DETAILED DESCRIPTION

[0062] In this embodiment, a microalgae promoter explainable prediction method based on deep convolutional neural network is provided, and the specific process is as shown in the figure Figure 1 The overall network framework is as shown in the figure Figure 2 The specific process comprises the following steps:

[0063] Step 1, collection and pretreatment of microalgae promoter sequence data:

[0064] Step 1.1, obtaining microalgae promoter data, according to the different types of sequenced strains, the obtained promoter data can be divided into 6 data sets, which are N. oceanica (IMET1), N. oculata (CCMP525), N. granulata (CCMP529), N. oceanica (CCMP531), N. salina (CCMP537) and N. gaditana (CCMP526); Since the length of the obtained sequence data is not fixed, it cannot be used as the input feature of the convolutional neural network, so it is necessary to screen the sequence with a length of W from the obtained microalgae promoter data, considering that the proportion of promoter sequences with a length of 1000bp is the highest in the 6 strain data sets, therefore, the screening length W is set to 1000 in this example; Since the data is redundant, the screened sequence is de-duplicated using the extended CD-HIT-EST toolkit, and the similarity threshold is set to 0.8, to obtain a positive sample data set Wherein, represents the mth positive sample in the positive sample data set S + m,i represents the one-hot encoding vector of the ith nucleotide in the mth positive sample, and M is the total number of positive samples;

[0065] Step 1.2, dividing the mth positive sample in the positive sample data set S+ into a plurality of subsequence fragments {{b m,1 , b m,2},..., {b m，i , b m,i+1},..., {b m，L-1 , b m,W}}, and the length of each subsequence fragment is 2bp, wherein {b m，i , b m,i+1} represents the i th subsequence fragment of the m th positive sample;

[0066] randomly rearranging the nucleotide arrangement order in each subsequence fragment of the mth positive sample to obtain the mth negative sample corresponding to , thereby obtaining a negative sample data set​​ According to the processing method, the frequency of nucleotide occurrence in each sub-sequence fragment in the mth negative sample is kept unchanged compared with the mth positive sample ;

[0067] S + is merged - to obtain a pre-processed microcystis promoter data set S = {s1,..., s n ,..., s N}, wherein s n represents the nth sequence sample, N = 2M is the total number of samples, and the label of the nth sequence sample is flag n ∈{0,1};

[0068] Step 2, a microcystis promoter classification network is established, which sequentially includes an initial convolution module, a backbone convolution module, and a weight output module, as shown in Figure 3 ;

[0069] Step 2.1, the initial convolution module is sequentially composed of a two-dimensional convolution layer Conv2D and a convolution block attention module CBAM;

[0070] The main function of the two-dimensional convolution layer Conv2D is to increase the channel dimension of the input sequence feature, and in this example, the number of convolution kernels of the convolution layer is set to 64;

[0071] The convolution block attention module CBAM mainly includes a channel attention module and a spatial attention module connected in series; in this example, the channel dimension in the convolution layer represents the sequence features recognized by the convolution kernel, that is, some base sequences recognized by the shallow convolution and the complex interaction between the base sequences recognized by the deep convolution, and the channel attention module enables the model to learn to pay attention to the feature types highly related to promoter prediction during the training process; the spatial dimension corresponds to the position and type of the base in the sequence, so the function of the spatial attention module is to enable the model to learn which positions in the abstract sequence feature should be paid attention to during the training process;

[0072] The nth sequence sample s n is input into the initial convolution module, and after being processed by Conv2D and CBAM in turn, the dimension-increased feature initConv n is obtained;

[0073] Step 2.2, the backbone convolution module is based on the DenseNet network, and sequentially includes a first dense convolution block, a second dimension-reducing conversion block, a third dense convolution block, a fourth dimension-reducing conversion block, and a fifth dense convolution block;

[0074] The role of the dense convolution block is to extract sequence features, and the dense connection can reduce the risk of gradient disappearance while increasing the number of convolution layers to improve the feature extraction capability, and at the same time, strengthen the propagation ability of sequence features in the network; each dense convolution block is sequentially composed of D dense connection two-dimensional convolution blocks and a convolution block attention module CBAM, in this example, the number of two-dimensional convolution blocks D is set to 10, and each two-dimensional convolution block is sequentially composed of an activation function layer ReLU, a two-dimensional convolution layer Conv2D and a random inactivation layer Dropout, wherein the convolution kernel size of the two-dimensional convolution layer is set to 4, the number of convolution kernels is 12, the padding is "same", and the inactivation rate of the random inactivation layer is 0.25;

[0075] The main role of the dimension reduction conversion block is to reduce the number of dimensions of the intermediate features in the network, thereby reducing the memory occupation during training and prediction, improving the receptive field of the network, and enhancing the robustness and generalization ability of the network; each dimension reduction conversion block is sequentially composed of a two-dimensional convolution layer with a convolution kernel of 1x1, an average pooling layer AP, a batch normalization layer BN, and a convolution block attention module CBAM, wherein the number of convolution kernels of the 1x1 convolution is half of the number of input feature channels of the layer, and the window size of the average pooling layer is 2 and the step is 2;

[0076] The dimension-reduced features initConv n are input into the main convolution module for processing, and the convolution features denseConv n are obtained.

[0077] Step 2.3, the weight output module is sequentially composed of an average pooling layer AP, a first fully connected layer DC1, an activation function layer ReLU, a batch normalization layer BN, a second fully connected layer DC2, and an activation function layer Sigmoid; wherein the window size of the average pooling layer is 25, the step is 25, the number of neurons of the first fully connected layer is 64, and the number of neurons of the second fully connected layer is 1;

[0078] The convolution features denseConv n are input into the weight output module for processing, and the normalized micro-nanoblast promoter binary classification prediction label p n is obtained.

[0079] Step 3, a binary classification cross-entropy loss function L is constructed using formula (1):

[0080]

[0081] Based on the micro-nanoblast promoter data set S = {s1,..., s n ,..., s 2NThe micro-nanoball algae promoter classification network is iteratively trained, and a binary classification cross-entropy loss function L is calculated to update the network parameters. In this example, the optimizer of the classification network uses SGD, the learning rate is set to 0.001, the momentum is 0.9, the batch size is set to 32, and the training stops when the number of training iterations reaches the set number of times or the binary classification cross-entropy loss function L converges, thereby obtaining an optimal micro-nanoball algae promoter classification model.

[0082] As shown in Figure 4 , two other promoter prediction deep learning algorithms reported in recent journal articles are used for performance comparison with the above-mentioned micro-nanoball algae promoter classification network of the present application. Five five-fold cross-validations are performed on six micro-nanoball algae strain data sets, and the average accuracy Acc is used as the comparison index. The results show that the micro-nanoball algae promoter classification network of the present application achieves the relatively best prediction performance.

[0083] Step 4, a micro-nanoball algae promoter importance score generation network is established, which sequentially includes a backbone module and an output module, as shown in Figure 5 .

[0084] Step 4.1, the backbone module includes an initial convolutional layer, a first residual group, a second residual group, a third residual group, a fourth residual group, a fifth residual group, a sixth convolutional layer, and a convolutional jump connection layer.

[0085] In this example, the convolutional kernel size of the initial convolutional layer and the sixth convolutional layer is set to 1, the number of convolutional kernels is 32, and the padding is "same"; each residual group is composed of R residual blocks, and R is set to 4 in this example; each residual block is composed of a batch normalization layer BN, an activation function layer ReLU, a first dilated convolutional layer dilatedConv1, a batch normalization layer BN, an activation function layer ReLU, and a second dilated convolutional layer dilatedConv2, and the input features of the residual block are added to the output features of the second dilated convolutional layer dilatedConv2 through a residual connection to obtain the output of the residual block; the convolutional kernel size of all dilated convolutional layers is set to 3, the number of convolutional kernels is 32, and the padding is "same"; all dilated convolutional layers in each residual group share a uniform dilation factor, and in this example, the dilation factors of the five residual groups are 1x, 2x, 4x, 2x, and 1x. The use of dilated convolution can replace the pooling layer to increase the receptive field of the model, allowing the model to better identify long-range interactions between sequence features that may exist in the promoter.

[0086] The mth positive sample is input into the backbone module of the micro-nanoball algae promoter importance score generation network and processed by the initial convolutional layer to obtain the first input feature The first input feature After the processing of the first residual group, a second input feature is obtained Second input feature After the processing of the second residual group, a third input feature is obtained Third input feature After the processing of the third residual group, a fourth input feature is obtained Fourth input feature After the processing of the fourth residual group, a fifth input feature is obtained Fifth input feature After the processing of the fifth residual group, a residual group output feature ResGroupOut is obtained m Residual group output feature ResGroupOut m After the processing of the sixth convolutional layer, a backbone output feature BackboneOut is obtained m .

[0087] Residual group output feature ResGroupOut is input into the convolutional skip connection layer respectively for processing, the convolutional kernel size of all convolutional skip connection layers is set to 1, the number of convolutional kernels is 32, and the padding is "same", and five cross-layer features According to formula (2), the convolutional feature BackboneOut' output by the backbone module is obtained m :

[0088]

[0089] In formula (2), denotes the jth cross-layer feature;

[0090] Step 4.2, the output module is sequentially composed of a convolutional layer with a number of convolutional kernels of 1 and an activation function layer Softplus;

[0091] After BackboneOut' m is input into the output module for processing, the mth positive sample is obtained m The corresponding importance score c m,1 ={cb m，i ,..., cb m,W}, wherein cb m,i >0 represents the real value importance score of the ith nucleotide in the mth positive sample If cb m,i tends to 0, it indicates that the nucleotide b m,i has little influence on the decision of the micro-nanoball algae promoter classification network on the sample , on the contrary, if cb m，iThe larger the value, the more it indicates b m，i The greater the impact on classification decisions.

[0092] Step 5: Training the network for generating importance scores for the Microcystis globulus promoter:

[0093] Step 5.1: Calculate the average distribution of all positive samples using equation (3) to obtain the background distribution bg of the positive sample data:

[0094]

[0095] Step 5.2, place c m Dimensions and Dimensional matching yields importance scores after dimensional processing. Thus, the perturbed distribution can be obtained using equation (4).

[0096]

[0097] Among them, importance score The closer it gets to 0, the better the perturbed distribution becomes. The closer it is to the background distribution (bg), the higher the importance score. The closer it gets to ∞, the more the distribution after the perturbation... The closer to the original sample

[0098] Step 5.3: Select the m-th positive sample. The input is processed by the optimal Micrococcus pluvialis promoter classification model to obtain the original predicted label p. m ;

[0099] Step 5.4: Use the Softmax Straight-Through sampling method to sample from... K perturbation samples were sampled from the middle. The data is then input into the optimal Microcystis globulus promoter classification model for processing, resulting in a comparison with p. m The corresponding K perturbation prediction results in, Indicates from The kth sample selected from the sample, express The perturbation prediction label output after processing by the optimal microcystis promoter classification model is set to 16 samples in this example.

[0100] Step 5.5: Calculate the predicted loss L using equation (5). p :

[0101]

[0102] Step 5.6, calculate the preservation loss L using formula (6) s :

[0103]

[0104] In formula (6), t bit is a hyperparameter of the target preservation, and ∑ axis denotes summing up the elements in all dimensions of the tensor. Step 5.7, construct the overall loss function L using formula (7) total :

[0105] L total = L p + λ·L s (7)

[0106] In formula (7), λ is a weight hyperparameter balancing the two losses.

[0107] Step 5.8, input the positive sample dataset into the microcystis promoter importance score generation network for training, and calculate the overall loss function L total to update the network parameters. In this example, the optimizer of the importance score generation network uses Adam, the learning rate is set to 0.001, and the batch size is set to 8. When the number of training iterations reaches the set number of times or the overall loss function L total converges, the training stops, thereby obtaining the optimal microcystis promoter importance score generation model for evaluating the importance of the predicted labels output by the optimal microcystis promoter classification model.

[0108] In this embodiment, an electronic device includes a memory for storing a program supporting the processor to execute the above method, and a processor configured to execute the program stored in the memory.

[0109] In this embodiment, a computer readable storage medium has a computer program stored thereon, and the computer program is run by a processor to perform the steps of the above method.

Claims

1. A method for interpretable prediction of microcystis promoter based on deep convolutional neural network, characterized in that, is performed as follows: Step 1, collection and preprocessing of microalgae Nannochloropsis sp. promoter sequence data: Step 1.1, screening sequences with length W from the obtained microalgae Nannochloropsis sp. promoter data, and using the extended CD-HIT-EST toolkit to remove the screened sequences, to obtain a positive sample data set wherein, represents the mth positive sample in the positive sample data set S + m,i represents the one-hot encoding vector of the ith nucleotide in the mth positive sample, and M is the total number of positive samples.​ Step 1.2: Transfer the positive sample dataset S + The m-th positive sample Divided into multiple subsequence segments {{b m,1 ,b m,2 },…,{b m,i ,b m,i+1 },…,{b m,L-1 ,b m,W }}, where {b m,i ,b m,i+1 } represents the m-th positive sample. Subsequence fragments; the mth positive sample is obtained the nucleotide sequence in each sub-sequence fragment is randomly shuffled, to obtain a corresponding mth negative sample the mth negative sample is obtained thereby obtaining a negative sample data set Merge S + With S - Obtain the pre-processed microalgae promoter dataset S = {s1,..., s n ..., s N}, where s n represents the nth sequence sample, N = 2M is the total number of samples, and let the label of the nth sequence sample s n be flag n ∈{0, 1}. Step 2, establishing a microalgae promoter classification network, which sequentially comprises an initial convolution module, a backbone convolution module and a weight output module, and is used for classifying an nth sequence sample s n Processing to obtain a normalized microalgae promoter binary classification prediction label p n ; Step 3, constructing a binary cross-entropy loss function L using formula (1): Based on the microalgae promoter dataset S = {s1,..., s n ,..., s 2N}, input the microalgae promoter classification network for iterative training, and calculate the binary cross-entropy loss function L to update the network parameters, when the number of training iterations reaches the set number of times or the binary cross-entropy loss function L converges, the training stops, thereby obtaining the optimal microalgae promoter classification model; Step 4, establishing a micro-nanqiqiao algae promoter importance score generation network, in turn comprising: a main module and an output module, and being used for the mth positive sample is processed to obtain the mth positive sample The corresponding importance score c m ; Step 5, training of the microalgae Nannochloropsis sp. promoter importance score generation network: Step 5.1, obtaining the background distribution bg of positive sample data using formula (3): Step 5.2, c m matches the dimension of , resulting in dimensionally processed importance score , resulting in perturbed distribution Step 5.3, obtaining the mth positive sample Input into the optimal microcystis promoter classification model for processing to obtain the original predicted label p m ; Step 5.

4. Sampling K perturbation samples from using the Softmax Straight-Through sampling method and inputting them into the optimal micromonas startup promoter classification model for processing to obtain K perturbation prediction results corresponding to p m , where represents the kth sample sampled from , and represents the perturbation prediction label output after the optimal micromonas startup promoter classification model is processed . ​ Step 5.5, compute the predicted loss L using formula (5) p : Step 5.6, Calculate the loss of reserve L using formula (6) s : In formula (6), t bit is a hyperparameter for the target reserve amount, ∑ axis • represents summing over all dimensions in the tensor; Step 5.7, constructing the overall loss function L with formula (7) total : L total = L p + λ · L s (7) In formula (7), λ is a weight hyperparameter balancing the two losses. Step 5.8, generating the positive sample dataset The input N. salina promoter importance score generation network is trained and the overall loss function L is calculated total The network parameters are updated when the number of training iterations reaches a set number or the overall loss function L total When the training converges, the training stops, and the optimal N. salina promoter importance score generation model is obtained, which is used to evaluate the importance of the predicted labels output by the optimal N. salina promoter classification model.

2. The method of claim 1, wherein the method is based on a deep convolutional neural network. The step 2 comprises: Step 2.1, the initial convolution module is sequentially composed of a two-dimensional convolution layer Conv2D and a convolution block attention module CBAM; The nth sequence sample s n is input into an initial convolution module and sequentially processed by Conv2D and CBAM to obtain the high-dimensional feature initConv n ; Step 2.2, the backbone convolution module is based on a DenseNet network and sequentially comprises: a first dense convolution block, a second dimension reduction conversion block, a third dense convolution block, a fourth dimension reduction conversion block, and a fifth dense convolution block; Each dense convolution block is sequentially composed of D densely connected two-dimensional convolution blocks and a convolution block attention module CBAM, and each two-dimensional convolution block is sequentially composed of an activation function layer ReLU, a two-dimensional convolution layer Conv2D, and a random inactivation layer Dropout; Each dimension reduction conversion block is sequentially composed of a two-dimensional convolution layer with a convolution kernel of 1x1, an average pooling layer AP, a batch normalization layer BN, and a convolution block attention module CBAM; The dimension-increased feature initConv is input into the main convolution module for processing, and a convolution feature denseConv is obtained n The dimension-increased feature initConv is input into the main convolution module for processing, and a convolution feature denseConv is obtained n ; Step 2.3, the weight output module is sequentially composed of an average pooling layer AP, a first fully connected layer DC1, an activation function layer ReLU, a batch normalization layer BN, a second fully connected layer DC2, and an activation function layer Sigmoid; The convolutional features denseConv n The normalized microcystis promoter binary classification prediction label p is obtained by processing in the input weight output module n .

3. The method of claim 1, wherein the method is based on a deep convolutional neural network. The step 4 comprises: Step 4.1, the backbone module comprises: an initial convolution layer, a first residual group, a second residual group, a third residual group, a fourth residual group, a fifth residual group, a sixth convolution layer, and a convolution jump connection layer; Each residual group is composed of R residual blocks; each residual block is sequentially composed of a batch normalization layer BN, an activation function layer ReLU, a first dilated convolution layer dilatedConv1, a batch normalization layer BN, an activation function layer ReLU, and a second dilated convolution layer dilatedConv2, and the input features of the residual block are added to the output features of the second dilated convolution layer dilatedConv2 through a residual connection to obtain the output of the residual block; all dilated convolution layers in each residual group share a uniform dilation factor; the mth positive sample The first input feature is obtained after the main module of the micro-nanoglobe promoter importance score generation network is input and processed by an initial convolutional layer The first input feature is obtained after the main module of the micro-nanoglobe promoter importance score generation network is input and processed by an initial convolutional layer The second input feature is obtained after processing by the first residual group The second input feature is obtained after processing by the first residual group The third input feature is obtained after processing by the second residual group The third input feature is obtained after processing by the second residual group The fourth input feature is obtained after processing by the third residual group The fourth input feature is obtained after processing by the third residual group The fifth input feature is obtained after processing by the fourth residual group The fifth input feature is obtained after processing by the fourth residual group The residual group output feature ResGroupOut is obtained after processing by the fifth residual group m The residual group output feature ResGroupOut is obtained after processing by the fifth residual group m The backbone output feature BackboneOut is obtained after processing by the sixth convolutional layer m The backbone output feature BackboneOut is obtained after processing by the sixth convolutional layer Will The features are processed by inputting them into the convolutional skip connection layer, resulting in 5 cross-layer features. Thus, according to equation (2), the final convolutional feature BackboneOut' of the backbone module is obtained. m : In formula (2), denotes the jth cross-layer feature; Step 4.2, the output module is sequentially composed of a convolution layer with a convolution kernel number of 1 and an activation function layer Softplus; BackboneOut m is input into the output module for processing to obtain the mth positive sample The corresponding importance score c m = {cb m,1 ,…,cb m,i ,…,cb m,W}, wherein cb m,i > 0 represents the real value importance score of the ith nucleotide in the mth positive sample .

4. An electronic device comprising a memory and a processor, characterized in that The memory is used to store a program supporting the processor to execute the microalgae Nannochloropsis sp. promoter interpretable prediction method of any one of claims 1-3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to perform the steps of the microalgae Nannochloropsis sp. promoter interpretable prediction method of any one of claims 1-3.

Images