Plant transcription factor binding site prediction method fusing fully dynamic convolution and cubemlp

Abstract

The application provides a plant transcription factor binding site prediction method fusing full dynamic convolution and CubeMLP, DNA sequence data of a plant transcription factor binding site region obtained through sequencing is subjected to One-hot coding processing and DNA double helix spatial structure feature extraction; mode feature of different categories of TFBS sequences is dynamically extracted by using an ODConv network, DNA spatial structure feature is extracted by using a convolutional neural network; efficient fusion of DNA sequence feature mode and DNA spatial structure mode feature is realized from three dimensions of sequence, mode and channel by using CubeMLP; the result of the fused feature is mapped into transcription factor binding sites of different categories by using a full connection layer and a Sigmoid function, and then the recognition of transcription factor binding sites of different categories is realized; the function of predicting transcription factor binding sites is realized. The application has high prediction accuracy, provides a new reference for the research of the recognition site of a plant transcription factor, and has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biological information, and particularly relates to a plant transcription factor binding site prediction method fusing complete dynamic convolution and CubeMLP. BACKGROUND

[0002] The transcription process is closely related to gene expression, and transcription factors (TF) play a crucial role in the growth and development of plants and environmental adaptability. TF is closely related to the morphogenesis process such as embryonic development and organ differentiation of plants, and can regulate the biosynthesis of various metabolites in plants, and has an important influence on the disease resistance, insect resistance and stress resistance of plants. TF affects the transcription and gene expression of plants by binding to specific DNA sequences (i.e., TFBS). Therefore, accurately predicting TFBS has important value for studying TF-DNA binding mechanism and understanding the transcriptional regulation mechanism of plants.

[0003] In recent years, with the continuous progress of machine learning and deep learning technology, it has been successfully applied in the field of computer vision, natural language processing and other fields. Deep learning method can automatically learn features from large-scale raw data, and can mine high-order complex nonlinear information. Therefore, applying deep learning to the prediction of plant TFBS is expected to improve the accuracy and efficiency of prediction.

[0004] However, the current deep learning-based plant TFBS prediction method is still in its infancy, and a perfect technical system has not yet been formed. The existing method uses only the sequence information of DNA to predict TFBS, and the representative deep learning method is TSPTFBS1.0 and TSPTFBS2.0, which shows strong robustness in single-class TFBS identification, but still has challenges in multi-class TFBS identification. On the other hand, the complex binding process of plant TFBS needs to be considered, and the sequence information of DNA and the spatial structure characteristics of DNA double helix are used to improve the accuracy of predicting TFBS. For example, the PlantBind method can realize the identification of multi-class TFBS, but this method does not fully fuse the features of different modalities, resulting in certain limitations in recognition accuracy. Therefore, it is of great significance to develop a plant transcription factor binding site prediction method with high accuracy. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a plant transcription factor binding site prediction method fusing complete dynamic convolution and CubeMLP, which is used for predicting transcription factor binding sites.

[0006] The technical solution adopted by the present application to solve the above technical problem is: a plant transcription factor binding site prediction method fusing complete dynamic convolution and CubeMLP, comprising the following steps:

[0007] S1: data partitioning and One-hot encoding are performed on the DNA sequence data to obtain a One-hot encoding matrix; Monte Carlo simulation method is used to extract DNA spatial structure features to obtain a DNA spatial structure encoding matrix;

[0008] S2: the One-hot encoding matrix is subjected to multi-dimensional attention mechanism of a fully dynamic convolutional network to adaptively adjust the convolution kernel weight, and DNA sequence features are extracted according to the characteristics of each type of transcription factor binding site sequence;

[0009] S3: the DNA spatial structure encoding matrix is subjected to the weight sharing characteristics of the convolution operation of a convolutional neural network to extract DNA spatial structure features;

[0010] S4: CubeMLP is used to capture the complex spatial interaction relationship of cross-modal features through the MLP mechanism of three dimensions of sequence, mode and channel, and the multi-modal features of DNA sequence features and DNA spatial structure features are fused;

[0011] S5: the result of the fused features is mapped to different types of transcription factor binding sites to realize classification of different transcription factor binding site features.

[0012] According to the above scheme, the following steps are further included before step S1:

[0013] S0: DNA sequence data of the transcription factor binding site region of the plant is obtained by sequencing the biological material.

[0014] According to the above scheme, in step S1, the specific steps are:

[0015] The DNA sequence data is divided according to a fixed length and subjected to One-hot encoding to obtain a One-hot encoding matrix;

[0016] The Monte Carlo simulation method is used to extract 14 different structure features of 6 base features within each base, 6 base pair features in double helix structure, 1 minor groove width and 1 minor groove potential in the DNA sequence to obtain a DNA spatial structure encoding matrix.

[0017] According to the above scheme, in step S2, the specific steps are:

[0018] S21: the One-hot encoding matrix is subjected to an Avgpooling layer and a 1x1 convolution to respectively compress the overall features of the TF sequence in the sequence dimension and the channel dimension, and a compressed and converged feature vector is obtained;

[0019] S22: Extract different types of attention through the multi-head attention module, including convolution kernel attention, channel attention, filter attention, and spatial attention;

[0020] S23: Weight the convolution layer weights one by one in the order of space, channel, filter, and convolution kernel to obtain the full-dimension weighted convolution layer weights;

[0021] S24: Perform convolution operation on the One-hot encoding matrix using the weighted dynamic convolution layer to gradually extract local features;

[0022] S25: Perform fast down-sampling on the local features through the down-sampling layer, and map them to high-order features of a specified dimension using the linear layer.

[0023] According to the above scheme, in step S3, the specific steps are:

[0024] S31: Gradually extract local features from the DNA spatial structure encoding matrix through the convolution block;

[0025] S32: Perform fast down-sampling on the local features through the down-sampling layer, and map them to high-order features of a specified dimension using the linear layer.

[0026] According to the above scheme, in step S4, the specific steps are:

[0027] S41: Concatenate the DNA sequence feature and the DNA spatial structure feature to obtain multi-modal fusion data;

[0028] S42: Process the multi-modal fusion data using CubeMLP to obtain fully fused features.

[0029] According to the above scheme, in step S5, the specific steps are:

[0030] S51: Synthetically fuse the features through the fully connected layer;

[0031] S52: Use the activation function Sigmoid to obtain the probability of transcription factor binding site features, and realize the classification of different transcription factor binding site features.

[0032] According to the above scheme, the following steps are also included:

[0033] S6: Calculate the network loss using the binary cross-entropy loss function BCELoss, which measures the difference between the target label value and the predicted probability value; update the parameters using the optimized gradient algorithm AdamW, and add the weight decay term to the loss function to adjust the parameters in the adaptive learning rate update process.

[0034] A DeepTFBS model comprises an input layer, a feature extraction layer, a cubic multi-layer perception layer and an output layer;

[0035] The input layer comprises a one-hot encoding layer and a spatial structure encoding layer; the one-hot encoding layer is used for data division and one-hot encoding of DNA sequence data to obtain a one-hot encoding matrix; and the spatial structure encoding layer is used for extracting DNA spatial structure features by using a Monte Carlo simulation method to obtain a DNA spatial structure encoding matrix;

[0036] The feature extraction layer comprises a fully dynamic convolution network and a convolutional neural network; the fully dynamic convolution network is used for adaptively adjusting convolution kernel weights of the one-hot encoding matrix through a multi-dimensional attention mechanism, and extracting DNA sequence features according to the characteristics of each transcription factor binding site sequence; and the convolutional neural network is used for extracting DNA spatial structure features by using the weight sharing characteristics of convolution operation on the DNA spatial structure encoding matrix;

[0037] The cubic multi-layer perception layer comprises a sequence dimension multi-layer perception machine, a modal dimension multi-layer perception machine and a channel dimension multi-layer perception machine, and is used for capturing complex spatial interaction relationships of cross-modal features through MLP mechanisms in three dimensions of sequence, mode and channel, and fusing multi-modal features of DNA sequence features and DNA spatial structure features;

[0038] The output layer comprises a plurality of multi-layer perception machines, and is used for mapping the results of fused features into different categories of transcription factor binding sites, so as to realize classification of different transcription factor binding site features.

[0039] Further, the fully dynamic convolution network comprises two dynamic convolution blocks, a down-sampling layer and a linear layer; the dynamic convolution block comprises a one-dimensional dynamic convolution, a RELU activation function and a Dropout layer; the dynamic convolution comprises a convergence-excitation multi-head attention module for adaptively adjusting convolution kernel weights;

[0040] The convolutional neural network comprises two convolution blocks, a down-sampling layer and a linear layer; the convolution block comprises a one-dimensional convolution, a RELU activation function and a Dropout layer; the convolution has the characteristics of weight sharing and local receptive field, and is used for extracting features of DNA spatial structure by using fewer parameters; and the RELU activation function and the Dropout layer enhance the generalization ability of the model by introducing a nonlinear feature.

[0041] The DeepTFBS model has the following advantages:

[0042] 1. The plant transcription factor binding site prediction method of the present application fuses complete dynamic convolution and CubeMLP, by One-hot encoding processing of the DNA sequence data of the plant transcription factor binding site region obtained by sequencing and extraction of the spatial structure features of the DNA double helix; the mode features of TFBS sequences of different categories are dynamically extracted by the ODConv network, the DNA spatial structure features are extracted by the convolutional neural network, and the nonlinear features are introduced through the RELU activation function and the Dropout layer, thereby enhancing the generalization ability of the model; the multi-axis MLP idea of CubeMLP is used to realize the fusion of multi-axis complex features through three dimensions of sequence, mode and channel, and efficient fusion of DNA sequence feature mode and DNA spatial structure mode features is realized; the results of the fused features are mapped to transcription factor binding sites of different categories by using the full connection layer and the Sigmoid function, and then the recognition of different categories of transcription factor binding sites is realized; the function of predicting transcription factor binding sites is realized.

[0043] 2. The present application recognizes different categories of transcription factor binding sites by learning the features of plant DNA sequences and spatial structures, has high prediction accuracy, provides a new reference for the recognition of plant transcription factor binding sites, and has a wide application prospect.

[0044] 3. The experimental results of the present application in predicting 315 TFBS of Arabidopsis thaliana show that, compared with existing models, the DeepTFBS method of the present application has the highest prediction accuracy; in terms of predicting TFBS by chromosome, the prediction accuracy of DeepTFBS is better than that of existing methods, indicating that DeepTFBS has strong stability; the ablation experiment results of the ODConv and CubeMLP modules in the DeepTFBS method show that ODConv and CubeMLP both have great contribution in prediction.

[0045] 4. The DeepTFBS method proposed in the present application can predict all potential TFBS at the plant genome level, can supplement the data amount of TFBS and supplement the TFBS obtained by CHIP-seq experiment, and has portability and generalizability.

[0046] Of course, implementing any product of the present application does not necessarily need to achieve all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0048] Figure 1 This is a flowchart of an embodiment of the present invention.

[0049] Figure 2 This is a principle block diagram of an embodiment of the present invention.

[0050] Figure 3 This is a diagram showing the binary and multi-class classification results of the DeepTFBS method in Arabidopsis thaliana 315 classes predicted by an embodiment of the present invention.

[0051] Figure 4 This is a graph showing the multi-class prediction results of different methods in Arabidopsis thaliana TFBS 315 classes.

[0052] Figure 5 This figure shows the performance and stability results of DeepTFBS chromosome 5-fold crossover validation in Arabidopsis thaliana 315-class TFBS multi-class prediction, according to an embodiment of the present invention.

[0053] Figure 6 This is an image showing the overlap between the five potential TFBSs obtained from DeepTFBS prediction of Arabidopsis thaliana RAP211, BIM2, SPL5, CBF4, SPL9, and SPL5 in an embodiment of the present invention and the TFBSs obtained from CHIP-seq sequencing. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0055] Example 1

[0056] See Figure 1 The specific steps of the plant transcription factor binding site prediction method that integrates fully dynamic convolution and CubeMLP are as follows:

[0057] S1: The DNA sequence data is divided and one-hot encoded to obtain the one-hot encoding matrix; the spatial structure features of DNA are extracted using the Monte Carlo simulation method to obtain the DNA spatial structure encoding matrix;

[0058] S2: For the one-hot encoding matrix, the weights of the convolution kernel are adaptively adjusted through the multi-dimensional attention mechanism of a fully dynamic convolutional network, and DNA sequence features are extracted based on the characteristics of the binding site sequences of each type of transcription factor.

[0059] S3: Extract DNA spatial structure features from the DNA spatial structure encoding matrix by utilizing the weight-sharing characteristic of convolution operations in convolutional neural networks;

[0060] S4: Capture the complex spatial interaction relationship of cross-modal features by using CubeMLP through the MLP mechanism of three dimensions of sequence, modality and channel, and fuse the multi-modal features of DNA sequence features and DNA spatial structure features;

[0061] S5: Map the results of fused features to different categories of transcription factor binding sites to realize the classification of different transcription factor binding site features.

[0062] Before step S1, the following steps are included:

[0063] S0: Obtain the DNA sequence data of the transcription factor binding site region of the plant by sequencing the biological material.

[0064] According to the above scheme, in step S1, the specific steps are:

[0065] Divide the DNA sequence data according to the fixed length and perform One-hot encoding to obtain the One-hot encoding matrix;

[0066] Use the Monte Carlo simulation method to extract 14 different structural features of 6 base intrachain features, 6 base pair features in double helix structure, 1 minor groove width and 1 minor groove potential in DNA sequence to obtain the DNA spatial structure encoding matrix.

[0067] Further, in step S2, the specific steps are:

[0068] S21: Perform overall feature compression of TF sequence in sequence dimension and channel dimension on One-hot encoding matrix through Avgpooling layer and 1x1 convolution to obtain compressed and converged feature vector;

[0069] S22: Extract different types of attention through multi-head attention module, including convolution kernel attention, channel attention, filter attention and spatial attention;

[0070] S23: Weight the convolution layer weights one by one in the order of space, channel, filter and convolution kernel to obtain the full-dimension weighted convolution layer weights;

[0071] S24: Use the weighted dynamic convolution layer to perform convolution operation on the One-hot encoding matrix to gradually extract local features;

[0072] S25: Perform fast down-sampling on local features through the down-sampling layer, and map them to high-order features of specified dimensions using the linear layer.

[0073] In step S3, the specific steps are:

[0074] S31: Local features are extracted from the DNA spatial structure encoding matrix step by step through the convolution block;

[0075] S32: The local features are quickly down-sampled through the down-sampling layer, and are mapped to high-order features of a specified dimension using a linear layer.

[0076] In step S4, the specific steps are as follows:

[0077] S41: The DNA sequence features and the DNA spatial structure features are spliced to obtain multi-modal fusion data.

[0078] S42: The multi-modal fusion data is processed using CubeMLP to obtain features after sufficient fusion.

[0079] In step S5, the specific steps are as follows:

[0080] S51: The features are comprehensively fused through a fully connected layer.

[0081] S52: The probability of transcription factor binding site features is obtained using the activation function Sigmoid to realize the classification of different transcription factor binding site features.

[0082] The following steps are also included:

[0083] S6: The network loss is calculated using the binary cross-entropy loss function BCELoss which measures the difference between the target label value and the predicted probability value; the parameters are updated using the optimized gradient algorithm AdamW, and a weight decay term is added to the loss function to adjust the parameters in the adaptive learning rate update process.

[0084] In this embodiment, the DNA sequence data of the plant transcription factor binding site region obtained by sequencing is processed by One-hot encoding and the spatial structure features of the DNA double helix are extracted; the ODConv network is used to dynamically extract the pattern features of different categories of TFBS sequences, the convolutional neural network is used to extract the DNA spatial structure features, and the non-linear features are introduced through the RELU activation function and the Dropout layer, thereby enhancing the generalization ability of the model; the multi-axis MLP idea of CubeMLP is used to realize the fusion of multi-axis complex features through three dimensions of sequence, mode and channel, and the efficient fusion of DNA sequence feature mode and DNA spatial structure mode features is realized; the fully connected layer and the Sigmoid function are used to map the fusion feature results to different categories of transcription factor binding sites, and then the recognition of different categories of transcription factor binding sites is realized; the function of predicting transcription factor binding sites is realized.

[0085] Embodiment 2

[0086] The steps of this embodiment are the same as those of Example 1, except that each step is applied to a specific example. Specifically, the following steps are included:

[0087] S0: sequencing of biological materials to obtain DNA sequence data of plant transcription factor binding site regions;

[0088] S1: data partitioning and One-hot encoding of the DNA sequence data to obtain a One-hot encoding matrix; DNA spatial structure features are extracted using the Monte Carlo simulation method to obtain a DNA spatial structure encoding matrix;

[0089] The plant transcription factor binding site (TFBS) region data obtained by sequencing is subjected to data partitioning and other processing, and the original DNA sequence with a length of 101 bp S ={ s 1， s 2... s i}, s i is taken as the input of the model. Then the sequence is One-hot encoded, converting base A to [1, 0, 0, 0], base T to [0, 1, 0, 0], base C to [0, 0, 1, 0], and base G to [0, 0, 0, 1], thereby obtaining a 4*101 One-hot encoding matrix Seq ( Seq ∈ R 4×101 ) which is input into the fully dynamic convolution network (ODConv) for further processing.

[0090] Then, 14 different structural features are extracted using the Monte Carlo simulation method, including 6 intrabase features (Buckled, Sheared, Stretched, Propeller Twisted (ProT), Opened, and Staggered) of each base, 6 base pair features (Tilted, Shifted, Slid, Rolled, Risen, and Helix Twisted) in the double helix structure, 1 minor groove width, and 1 minor groove potential. Finally, a DNA spatial structure encoding matrix Shape ( Shape ∈R 14×101 ) is obtained, which is input into the convolutional neural network for further processing.

[0091] S2: Extract the unique patterns and features of multi-class transcription factor binding site (TFBS) sequences using the ODConv network. Dynamic convolution enhances the feature extraction capability of the convolution layer by introducing a multi-dimensional attention mechanism. This operation allows the convolution kernel weight to adaptively adjust, so that each class of TFBS sequence can be processed with different convolution kernels, and then the model can extract the best feature representation according to the characteristics of each class of TFBS sequence.

[0092] The ODConv network is composed of two dynamic convolution blocks, one down-sampling layer and one linear layer. Among them, the dynamic convolution block contains a one-dimensional dynamic convolution, a RELU activation function and a Dropout layer. The adaptive weight of dynamic convolution is realized by a convergence-excitation multi-head attention module to adaptively adjust the convolution kernel weight.

[0093] S21: Input sequence Seq Through the Avgpooling layer and a 1x1 convolution, this operation realizes the overall feature compression of the TF sequence in the sequence dimension and the channel dimension, respectively, to obtain the compressed and converged feature vector Seq G ( Seq G ∈R C×1 ). Specifically as shown in equation (1).

[0094] (1)

[0095] S22: Extract different types of attention through the multi-head attention module, including convolution kernel attention a k ( a k ∈R K num), channel attention a c ( a c ∈R C in), filter attention a f ( a f ∈R C out), spatial attention a s ( a s ∈R K size), the specific implementation process is shown in equation (1)-(5).

[0096] (2)

[0097] (3)

[0098] (4)

[0099] (5)

[0100] S23: Weight the convolution layer weights one by one in the order of space, channel, filter, and convolution kernel to obtain full-dimension weighted convolution layer weights W ODconv ( W ODconv ∈R C out ×C in ×K size), ensuring that the convolution operation has differences in all dimensions. The specific implementation is shown in formula (6).

[0101] (6)

[0102] where, K num is the number of convolution kernels, C out is the number of filters in each convolution kernel, Wi j ( Wi j ∈R Cin ×Ksize ) represents the i th filter of the j th convolution kernel. a s , a c respectively, are the attention assigned to the convolution kernel in the spatial dimension and the channel dimension, ai f is the attention assigned to the different filters of the convolution kernel, a k and the attention assigned to the different convolution kernels.

[0103] S24: Perform convolution operation on Seq using the weighted dynamic convolution layer, and also use the RELU activation function and Dropout to introduce nonlinear features. The complete implementation process of the dynamic convolution block is shown in formula (7).

[0104] (7)

[0105] The ODConv network uses ODConv 1 and ODConv 2 convolution blocks to gradually extract local features Seq '' from the encoded Seq matrix. The specific implementation is shown in formula (8).

[0106] (8)

[0107] S25: Local features are down-sampled by a down-sampling layer Seq '' Fast down-sampling and mapping to high-order features of specified dimensions using a linear layer Seq * The specific implementation is shown in equation (9).

[0108] (9)

[0109] S3: The weight sharing strategy of the convolutional neural network is used to effectively extract the spatial structure features of DNA with fewer parameters. In addition, the RELU activation function and the Dropout layer introduce nonlinear features, thereby enhancing the generalization ability of the model.

[0110] The convolutional neural network layer is composed of two convolutional blocks, a down-sampling layer, and a linear layer. The convolutional block contains a one-dimensional convolution, a RELU activation function, and a Dropout layer. The characteristics of weight sharing and local receptive field in the convolution operation enable effective extraction of the spatial structure features of DNA while using fewer parameters. The RELU activation function and the Dropout layer introduce nonlinear features, thereby enhancing the generalization ability of the model. The specific implementation of the convolutional block is shown in equation (10).

[0111] (10)

[0112] S31: The convolutional module uses Conv 1 and Conv 2 two convolutional blocks to gradually extract local features from the encoded Shape matrix Shape . The specific implementation is shown in equation (11).

[0113] (11)

[0114] S32: Local features are down-sampled by a down-sampling layer Shape , and then mapped to high-order features of specified dimensions using a linear layer Shape* . The specific implementation is shown in equation (12).

[0115] (12)

[0116] S4: In order to effectively fuse the extracted DNA sequence features and different modal features of DNA spatial structure, CubeMLP is used to capture the complex spatial interaction relationship of cross-modal features, so as to realize the fusion between multi-modal features. The CubeMLP structure uses MLP units of three dimensions of sequence (L), mode (M) and channel (C) to realize the fusion of multi-axis complex features, ensures the effective sharing of modal information in multiple dimensions, and significantly reduces the calculation cost.

[0117] S41: splice sequence features Seq* and spatial structure features Shape* to obtain multi-modal fusion data S ( S ∈R L ×M×C ), as shown in equation (13).

[0118] (13)

[0119] S42: use CubeMLP to process multi-modal fusion data S Each MLP unit is composed of two fully connected layers, a GELU activation function and a layer normalization. The GELU activation function based on probability activation is introduced to capture more complex patterns. At the same time, in the process of feature fusion, residual connection is used to prevent information loss of deep network and speed up model convergence. Specifically, the feature fusion of sequence (L) dimension, mode (M) dimension and channel (C) dimension is implemented as shown in equations (14)-(16).

[0120] (14)

[0121] (15)

[0122] (16)

[0123] CubeMLP layer uses two layers of CubeMLP stacking, and finally obtains the features after sufficient fusion S * ( S* ∈R L'×M'×C' ) as shown in equation (17).

[0124] (17)

[0125] S5: the output layer is used to predict the category of TFBS, and realize the classification of plant transcription factor binding site features.

[0126] S51: the output feature of CubeMLP S *Input and output layers; combined through two fully connected layers, the fully connected neural network is used to further extract the fused features, capturing the hierarchical features of different transcription factors, and the fully connected layers are further prevented from overfitting through the ReLU activation function and Dropout layer;

[0127] S52: The probability of obtaining transcription factor binding site characteristics using the sigmoid activation function, with a value between 0 and 1. The specific implementation is shown in formula (18).

[0128] (18)

[0129] S6: The network loss is calculated using the binary cross-entropy loss function BCELoss, which measures the difference between the target label value and the predicted probability value. The parameters are updated using the optimized gradient algorithm AdamW, which adds the weight decay term directly to the loss function to ensure that the parameters can be adjusted more accurately during the adaptive learning rate update process.

[0130] When calculating the network loss, the difference between the target label value and the predicted probability value is used. binary The cross-entropy loss function BCELoss is specifically implemented as shown in formula (19).

[0131] (19)

[0132] p For theoretical labels, take either 0 or 1. q The model outputs predicted values, ranging from [0, 1]. w For weights.

[0133] The optimized gradient algorithm AdamW is used to update the parameters. It adds the weight decay term directly to the loss function to ensure that the parameters can be adjusted more accurately during the adaptive learning rate update process. The specific implementation is shown in formula (20).

[0134] (20)

[0135] in, weight_decay It is the weight decay coefficient. lr It's the learning rate. m It is a first-moment estimate of the gradient. v It is the second moment estimate of the gradient. eps It is a small constant used to increase numerical stability.

[0136] This invention identifies binding sites for different types of transcription factors by learning the characteristics of plant DNA sequences and spatial structures, with high prediction accuracy. It provides a new reference for the study of transcription factor recognition sites in plants and has broad application prospects.

[0137] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0138] Embodiment 3

[0139] The following is an example result of applying the method of the present application to TFBS prediction in Arabidopsis 315 class TFBS data. The superiority of the method of the present application in TFBS prediction is explained in detail through related experiments, thereby providing a reference for plant TFBS prediction.

[0140] First, the model was trained and tested on the Arabidopsis dataset to evaluate the ability of DeepTFBS to predict plant TFBS. The specific experimental results are shown in Figure 3 Seven evaluation indicators including SN, SP, ACC, AUC, AP, and MCC were used to evaluate the results of DeepTFBS prediction of binary classification and multi-classification. In addition, the evaluation indicator names were not distinguished between binary classification and multi-classification, and the binary classification indicators were named: SN_b, SP_b, ACC_b, AUC_b, AP_b, and MCC_b; the multi-classification indicators were named: SN_m, SP_m, ACC_m, AUC_m, AP_m, and MCC_m; In the multi-classification task, AP and MCC indicators were mainly used for method comparison. Because AP and MCC indicators can more accurately reflect the multi-classification ability of the model in the case of data imbalance. Specifically, the average AUC_b of DeepTFBS is 0.9714, the average AUC_m is 0.9755, the average AP_b is 0.9736, the average AP_m is 0.5017, the average MCC_b is 0.9310, and the average MCC_m is 0.3139. In the multi-classification case, the AP_m and MCC_m indicators have declined. This may be related to the serious class imbalance problem in the Arabidopsis dataset. In addition, there are similar motif sequences within the family between the TFBS of different TFs in Arabidopsis, which further leads to misjudgment of negative samples. The distribution of each indicator is shown in Figure 3 (F), it can be seen that the indicator has less abnormal value and is concentrated. The above results show that DeepTFBS has stable classification performance and can effectively distinguish positive and negative samples.

[0141] Further, in order to more deeply evaluate the multi-classification ability of DeepTFBS in the Arabidopsis dataset, four state-of-the-art methods for predicting plant TFBS were trained and tested, including PlantBind, DeepSTF, D_SSCA, and DenseNet. The specific experimental results are shown in Figure 4The results of DeepTFBS are 0.9755, which is improved by 0.07% compared with PlantBind, 0.51% compared with DeepSTF, 19.03% compared with D_SSCA, and 47.2% compared with DenseNet. For the class-imbalance dataset, the average MCC_m index is used for the experiment. The results show that the MCC_m index of DeepTFBS is 0.3139, which is improved by 0.41% compared with PlantBind, 1.46% compared with DeepSTF, 20.06% compared with D_SSCA, and 31.08% compared with DenseNet. At the same time, the distribution of each index of multi-classification is analyzed, such as Figure 4 As shown in the ridge plots B, C, and D, it is found that the result distribution of DeepTFBS is more concentrated than other methods. In summary, DeepTFBS exhibits the best multi-classification performance on the Arabidopsis dataset.

[0142] In order to further verify the stability of DeepTFBS, 5 times of cross-validation experiments of each chromosome are performed. Specifically, for the 5 chromosomes of Arabidopsis, all TFBS data on one chromosome are selected as the test set each time, and the rest are used as the training set. The results are shown in Figure 5 It can be seen that each multi-classification index decreases slightly, the average AUC_m is 0.9719, the average AP_m is 0.4818, and the average MCC_m is 0.3039. The ridge plots B, C, and D reveal that there is no significant distribution difference between the results of each cross-validation. Based on these results, it can be concluded that DeepTFBS has the ability to efficiently and stably predict plant TFBS based on a lightweight model architecture.

[0143] Further, in order to evaluate the influence of ODconv and CubeMLP on the prediction ability of DeepTFBS model, ablation experiments are performed to analyze the different modules in DeepTFBS method. The results of the ablation experiments are shown in Table 1.

[0144] Table 1 Ablation experiment results of DeepTFBS

[0145]

[0146] In each experiment, the hyperparameters are consistent with DeepTFBS. In experiment A, only ODconv is used for feature extraction, and no feature fusion is performed. In experiment B, the same parameters are used to replace Conv with ODconv. The results show that the average AUC_m drops by 3.84% and the average AP_m drops by 11.847% after removing CubeMLP. This shows that the fusion of CubeMLP in three dimensions helps to deeply fuse sequence features and structure features, and the fused representation is more conducive to model classification. However, although the performance of model B is slightly better than DeepTFBS, the static structure of Conv limits the robustness of model B, and its accuracy in subsequent cross-species prediction is not high. The dynamic structure of ODconv enables it to adapt to different inputs, thereby providing DeepTFBS with greater robustness.

[0147] Based on the good prediction accuracy of DeepTFBS on the Arabidopsis dataset, all potential TFBSs on the Arabidopsis genome level were further predicted. First, for the Arabidopsis whole genome dataset, a sliding window method was used to extract sequences. The sliding window was set to 101 bp, and the sliding step was set to 10 bp, thereby obtaining 11894312 DNA sequences. Based on this dataset, the DNA sequence fragments were input into the trained DeepTFBS model for prediction. With a threshold of 0.5, sequences with a prediction value greater than 0.5 are potential TFBSs. Finally, the number of potential TFBSs predicted is much larger than the number of TFBSs actually sequenced. This difference may be due to the fact that the actual sequencing data is obtained in a specific time and a single cell type, while DeepTFBS predicts all potential TFBSs at the genome level, in different time periods and different tissues. The specific prediction results are shown in Table 2. Figure 6

[0148] The experimental results of the present application in predicting 315 types of TFBSs in Arabidopsis show that, compared with existing models, the DeepTFBS method of the present application has the highest prediction accuracy; in terms of predicting TFBSs by chromosome, the prediction accuracy of DeepTFBS is better than that of existing methods, indicating that DeepTFBS has strong stability; the ablation experiment results of ODConv and CubeMLP modules in the DeepTFBS method show that ODConv and CubeMLP both have a great contribution to prediction.

[0149] The DeepTFBS method proposed by the present application can predict all potential TFBSs at the genome level of plants, can supplement the amount of TFBS data and the TFBS obtained by CHIP-seq experiment, and has portability and generalizability.

[0150] Example 4​

[0151] The embodiment is used to implement the principle of the above-mentioned method embodiment to construct a DeepTFBS model, including an input layer, a feature extraction layer, a cubic multi-layer perception layer and an output layer.

[0152] The input layer includes a one-hot encoding layer and a spatial structure encoding layer; the one-hot encoding layer is used for data division and one-hot encoding of DNA sequence data to obtain a one-hot encoding matrix; and the spatial structure encoding layer is used for extracting DNA spatial structure features by using a Monte Carlo simulation method to obtain a DNA spatial structure encoding matrix.

[0153] The feature extraction layer includes a fully dynamic convolution network and a convolutional neural network; the fully dynamic convolution network is used for adaptively adjusting convolution kernel weights of the one-hot encoding matrix through a multi-dimensional attention mechanism, and extracting DNA sequence features according to the characteristics of each transcription factor binding site sequence; and the convolutional neural network is used for extracting DNA spatial structure features by using the weight sharing characteristics of convolution operation on the DNA spatial structure encoding matrix.

[0154] The cubic multi-layer perception layer includes a sequence dimension multi-layer perception machine, a modal dimension multi-layer perception machine and a channel dimension multi-layer perception machine, which are used for capturing complex spatial interaction relationships of cross-modal features through MLP mechanisms in three dimensions of sequence, mode and channel, and fusing multi-modal features of DNA sequence features and DNA spatial structure features.

[0155] The output layer includes a plurality of multi-layer perception machines, which are used for mapping the results of fused features into different categories of transcription factor binding sites, so as to realize classification of different transcription factor binding site features.

[0156] The fully dynamic convolution network includes two dynamic convolution blocks, a down-sampling layer and a linear layer; the dynamic convolution block includes a one-dimensional dynamic convolution, a RELU activation function and a Dropout layer; the dynamic convolution includes a convergence-excitation multi-head attention module for adaptively adjusting convolution kernel weights.

[0157] The convolutional neural network includes two convolution blocks, a down-sampling layer and a linear layer; the convolution block includes a one-dimensional convolution, a RELU activation function and a Dropout layer; the convolution has the characteristics of weight sharing and local receptive field, and is used for extracting features of DNA spatial structure using fewer parameters; the RELU activation function and the Dropout layer enhance the generalization ability of the model by introducing a nonlinear feature.

[0158] It should be noted that, according to the needs of implementation, each step / component described in the present application can be split into more steps / components, or two or more steps / components or partial operations of the steps / components can be combined into a new step / component, to achieve the purpose of the present application.

[0159] The embodiment also includes a processor, a communication interface, a memory, and a communication bus; wherein the processor, the communication interface, and the memory complete the communication among each other through the communication bus; the memory stores a computer program; when the program is executed by the processor, the processor executes the steps of the plant transcription factor binding site prediction method fusing full dynamic convolution and CubeMLP.

[0160] The embodiment also provides a computer readable storage medium, which stores executable instructions, and the instructions make the processor implement the plant transcription factor binding site prediction method fusing full dynamic convolution and CubeMLP when executed by the processor.

[0161] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0162] Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0163] The present application is described with reference to the flowcharts of the method and the computer program product according to the embodiment 1 of the present application and the block diagrams of the apparatus (system) of the embodiment 4. It should be understood that each flow or block in the flowcharts or block diagrams, and the combination of the flows or blocks in the flowcharts or block diagrams can be implemented by computer program instructions.

[0164] These computer program instructions can be provided to a processor of a general purpose computer, a special purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus generate one or more functions specified in the flowcharts or block diagrams. Figure 1 one or more flows or blocks Figure 1 a DeepTFBS model of the function specified in one or more flows or blocks.

[0165] These computer program instructions can also be stored in a computer readable storage medium capable of guiding the computer or other programmable data processing apparatus to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction devices, which implement the functions specified in the flowcharts or block diagrams. Figure 1one or more processes or blocks Figure 1 the function specified in the one or more blocks.

[0166] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operation steps are performed on the computer or other programmable data processing devices to generate computer implemented processes, so that the instructions executed on the computer or other programmable data processing devices provide processes for implementing the flow Figure 1 one or more processes or blocks Figure 1 the steps of the plant transcription factor binding site prediction method fusing full dynamic convolution and CubeMLP specified in the one or more blocks.

[0167] The above examples are only used to illustrate the design ideas and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and to implement it, and the protection scope of the present application is not limited to the above examples. Therefore, any equivalent changes or modifications made in accordance with the principles and design ideas disclosed by the present application are within the protection scope of the present application.

Claims

1. A method for predicting plant transcription factor binding sites by integrating fully dynamic convolution and CubeMLP, characterized by: Includes the following steps: S1: The DNA sequence data is divided and one-hot encoded to obtain the one-hot encoding matrix; the spatial structure features of DNA are extracted using the Monte Carlo simulation method to obtain the DNA spatial structure encoding matrix; S2: For the one-hot encoding matrix, the weights of the convolution kernel are adaptively adjusted through the multi-dimensional attention mechanism of a fully dynamic convolutional network, and DNA sequence features are extracted based on the characteristics of the binding site sequences of each type of transcription factor. S3: Extract DNA spatial structure features from the DNA spatial structure encoding matrix by utilizing the weight-sharing characteristic of convolution operations in convolutional neural networks; S4: Utilizing CubeMLP to capture complex spatial interactions of cross-modal features through MLP mechanisms in three dimensions: sequence, modality, and channel, thus fusing multimodal features of DNA sequence features and DNA spatial structure features; S5: Map the results of the fusion features to different categories of transcription factor binding sites to achieve classification of different transcription factor binding site features.

2. The method for predicting plant transcription factor binding sites by fusing fully dynamic convolution and CubeMLP according to claim 1, characterized in that: The steps preceding step S1 include the following: S0: Sequencing biological materials to obtain DNA sequence data of transcription factor binding site regions in plants.

3. The method for predicting plant transcription factor binding sites by fusing fully dynamic convolution and CubeMLP according to claim 1, characterized in that: The specific steps in step S1 are as follows: DNA sequence data is divided into fixed lengths and one-hot encoded to obtain a one-hot encoding matrix; The Monte Carlo simulation method was used to extract 14 different structural features, consisting of 6 intrabase features of each base in the DNA sequence, 6 base pair features in the double helix structure, 1 minor groove width, and 1 minor groove potential, to obtain the DNA spatial structure coding matrix.

4. The method for predicting plant transcription factor binding sites by fusing fully dynamic convolution and CubeMLP according to claim 1, characterized in that: The specific steps in step S2 are as follows: S21: The One-hot encoding matrix is ​​compressed into the overall features of the TF sequence in both the sequence dimension and the channel dimension by using an Avgpooling layer and a 1×1 convolution to obtain the compressed and converged feature vector. S22: Extract different types of attention through a multi-head attention module, including kernel attention, channel attention, filter attention, and spatial attention; S23: Weight the convolutional layer weights one by one in the order of space, channel, filter and convolutional kernel to obtain the full-dimensional weighted convolutional layer weights; S24: Use a weighted dynamic convolutional layer to perform convolution operations on the One-hot encoding matrix to gradually extract local features; S25: Local features are quickly downsampled using a downsampling layer, and then mapped to higher-order features of a specified dimension using a linear layer.

5. The method for predicting plant transcription factor binding sites by fusing fully dynamic convolution and CubeMLP according to claim 1, characterized in that: The specific steps in step S3 are as follows: S31: Local features are extracted step by step from the DNA spatial structure coding matrix through convolutional blocks; S32: Local features are quickly downsampled using a downsampling layer, and then mapped to higher-order features of a specified dimension using a linear layer.

6. The method for predicting plant transcription factor binding sites by fusing fully dynamic convolution and CubeMLP according to claim 1, characterized in that: The specific steps in step S4 are as follows: S41: Splice DNA sequence features and DNA spatial structure features to obtain multimodal fusion data; S42: Use CubeMLP to process multimodal fusion data to obtain fully fused features.

7. The method for predicting plant transcription factor binding sites by fusing fully dynamic convolution and CubeMLP according to claim 1, characterized in that: The specific steps in step S5 are as follows: S51: Features are integrated and fused through a fully connected layer; S52: Use the sigmoid activation function to obtain the probability of transcription factor binding site features, and classify different transcription factor binding site features.

8. The method for predicting plant transcription factor binding sites by fusing fully dynamic convolution and CubeMLP according to claim 1, characterized in that: It also includes the following steps: S6: Calculate the network loss using the binary cross-entropy loss function BCELoss, which measures the difference between the target label value and the predicted probability value; update the parameters using the optimized gradient algorithm AdamW, and add a weight decay term to the loss function to adjust the parameters during the adaptive learning rate update process.

Images