Method for predicting maize chromatin open regions based on multi-scale convolutional network and gMLP

Abstract

The application discloses a method for predicting maize chromatin open regions based on a multi-scale convolutional network and a gMLP, which is realized through a DeepOCR model constructed and comprises the following steps: performing One-hot coding processing on maize chromatin open region data obtained through sequencing after processing; inputting the coded sequence into a multi-scale convolutional network, utilizing different sizes of convolution kernels to pay attention to and fuse multi-scale features and capture important motifs of different lengths, and using Softpool to reduce information loss in a feature map; preventing overfitting through Batch Normalization and Dropout; capturing global dependence in a DNA sequence through gMLP, and extracting more discriminative feature representations to improve network performance and enhance the accuracy of model prediction; and using a full connection function and a Sigmoid function to complete the prediction of the maize chromatin open region in an output layer. The method has high prediction accuracy, can provide a new reference for the study on the sequence open conservation and specificity of maize and different tissues of different species, and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP. Background Art

[0002] As one of the most important food crops in the world, the yield and quality of maize are of great significance to global food security. However, the genetic mechanism of maize is complex, and many gene regulatory mechanisms related to key traits such as yield, quality, and stress resistance have not yet been fully revealed. As an important part of gene expression regulation, the accurate prediction of chromatin open regions is of great significance for understanding the genetic mechanism of maize and field breeding. Chromatin open regions refer to relatively loose regions in chromatin that are easy for transcription factors to bind. These regions are usually associated with gene expression levels. Accurately identifying maize chromatin open regions will help understand the molecular mechanism of gene expression regulation and provide a theoretical basis for analyzing the genetic basis of complex maize traits.

[0003] In recent years, deep learning has been widely used in bioinformatics. Its powerful feature learning and representation capabilities provide new approaches for processing and analyzing complex biological data. Deep learning models can automatically learn features from large amounts of data and exploit deep insights within the data through multi-layered nonlinear transformations. Therefore, applying deep learning to the prediction of accessible chromatin regions in maize is expected to improve both accuracy and efficiency, thereby facilitating the analysis of conservation and specificity across different maize tissues and varieties.

[0004] However, deep learning-based methods for predicting maize chromatin open regions are still in their infancy and lack a comprehensive technical framework. Existing methods rely on traditional machine learning algorithms, such as support vector machines and random forests. These methods have limitations in feature extraction and model building, resulting in incomplete sequence information mining. Furthermore, existing methods utilize relatively simple deep learning algorithms, such as CharPlant, Basset, and SMOC. These methods cannot effectively integrate sequence information, thus affecting prediction accuracy. Therefore, developing highly accurate methods for predicting maize chromatin open regions is of great research significance. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for predicting open regions of corn chromatin based on multi-scale convolutional networks and gMLP in response to the defects in the existing technology.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] The present invention provides a method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP. The method is used to construct a DeepOCR model, which includes an input layer, a multi-scale convolutional network layer, a gMLP layer, and an output layer. The method specifically includes the following steps:

[0008] Step 1: In the input layer, ATAC-seq data of maize chromatin is obtained and processed using one-hot encoding. The resulting matrix is ​​then input into the multi-scale convolutional network layer.

[0009] Step 2: Use convolution kernels of different sizes in the multi-scale convolutional network layer to focus on and fuse multi-scale features and capture important motifs of different lengths. SoftPool is used to remove a large amount of redundant information while reducing information loss. The output of the multi-scale convolutional network layer is used as the input of the gMLP layer.

[0010] Step 3: The gMLP layer takes the output of the multi-scale convolutional network layer as its input to capture complex spatial interactions across tokens, thereby obtaining long-range dependencies between features in the sequence. The gMLP layer includes layer normalization, three layers of gMLP blocks, and a residual-like structure.

[0011] Step 4: The output features of the gMLP layer are used as the input of the output layer. Then, through a layer normalization and two fully connected layers, a complex combination of gMLP features can be learned, which enhances the expressive power of the model and realizes the classification of the open region features of maize chromatin.

[0012] In step 5, 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, and the parameters are updated using the optimized gradient algorithm AdamW, which directly adds the weight decay term to the loss function to ensure more accurate parameter adjustment during the adaptive learning rate update process.

[0013] Furthermore, the method in step 1 of the present invention includes:

[0014] The maize chromatin open region data obtained by sequencing was processed by data partitioning and other processing, and a sequence length of 1000bp was used as the input of the model. The sequence was then 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]. This resulted in a 1000*4 one-hot encoded matrix, which was input into the multi-scale convolutional layer for the next step.

[0015] Furthermore, the method in step 2 of the present invention includes:

[0016] The multi-scale convolutional network layer includes: multiple convolution kernels of different sizes, activation functions, SoftPool, BatchNormalization and Dropout; among them: each convolution kernel of different sizes is connected to an activation function and a SoftPool respectively; then the output ends of multiple SoftPools are added and fused and connected to Batch Normalization and Dropout in sequence; the convolution operation in the multi-scale convolutional network layer extracts key data features through the convolution kernel, realizing local perception and dimensionality reduction functions.

[0017] Furthermore, the method in step 2 of the present invention includes:

[0018] The specific implementation of the convolution operation is as follows:

[0019] ConV(x) i，j =∑ m ∑ n X(m*s,n*s)*W k (im,jn)

[0020] Where i, j are the output position indexes, k is the kernel index, and m and n are the index variables in the convolution kernel. The convolution kernel slides on the feature map with a step size of s. At each position (i, j), the feature map and the convolution kernel are multiplied element by element, and then all products are summed to output the value at that position in the feature map. In order to simultaneously perceive and extract features of different scales, convolution kernels of different scales are used to implement multi-scale convolution.

[0021] The activation function ReLU is used to learn the nonlinear relationship between features. The expression is as follows:

[0022] ReLU(X)=max(0,X)

[0023] Based on the Softmax function, Softpool makes the output obey a certain probability distribution by highlighting the activation values ​​with great influence rather than directly selecting the maximum value or average value. The expression of Softpool is as follows:

[0024]

[0025] Among them, i is the original feature map position index, j is the pooling weight index; the input sequence is subjected to convolution, ReLU and Softpool operations of different scales, and then the important features extracted are added and fused;

[0026] In order to improve the training speed and model stability, a batch normalization operation is added in the middle of the multi-scale convolutional layer. Its expression is as follows:

[0027]

[0028] Among them, μ, σ 2 is the global mean and variance, γ and β are learnable parameters used for feature scaling and translation operations, and ε is a small constant used to stabilize the value and avoid division by zero;

[0029] The Dropout operation changes the activation values ​​of some neural network units to 0 with probability p, that is, the weights are not updated to reduce the number of intermediate features, thereby increasing the orthogonality between the features of each layer and preventing overfitting.

[0030] Furthermore, the gMLPblocks in step 3 of the present invention includes a micro Attention module for capturing long-distance dependencies between features; the output of the gMLPblocks is merged with the input of the gMLP layer as the output of the gMLP layer, thereby realizing a residual structure that maintains network performance.

[0031] Furthermore, the method in step 3 of the present invention includes:

[0032] The gMLP layer consists of several blocks of the same size and structure, based on gated MLPs. In the spatial dimension rather than the channel dimension, the gMLP layer captures complex spatial interactions across tokens. Patches Embedding is performed on the module input X. The expression of the entire gMLP layer is as follows:

[0033] Y = norm(X)

[0034] Y = gMLPblocks(Y)

[0035] Z=Y+X

[0036] Among them, norm refers to Layer Normalization;

[0037] To enhance the learning effect of the network, a total of three layers of gMLPblocks are used. Blocks contain a micro Attention module to capture long-range dependencies between features. The expression of the Attention mechanism is as follows:

[0038]

[0039] By performing a linear transformation on the input sequence, we can obtain the query vector Query, key vector Key, and value vector Value; ij Used to calculate the similarity between Q and K, and then calculate each attention weight, and then perform weighted summation with V;

[0040] The part that runs in parallel with Attention, called Proj_in, is implemented by a fully connected layer and the activation function GELU, which is used to perform channel projection mapping operations on the input. GELU introduces random regularization in activation and is a probabilistic description of neuron input. It is considered to be a comprehensive functional implementation of Dropout, Zoneout, and ReLU. The GELU expression is as follows:

[0041] GELU(X)=xP(X<=x)=xΦ(x)

[0042] Among them, Φ(x) is the probability function of normal distribution, using the standard normal distribution N(0,1);

[0043] The result is input into the core Spatial Gating Unit of the gMLP layer. The Spatial Gating Unit only operates on the spatial dimension without position embedding, thereby capturing the interaction of spatial dimension features. Its expression is as follows:

[0044] O=x1⊙(W(x2+x3)+b)

[0045] Among them, x1 and x2 are the separation results of Proj_in in the channel dimension, x3 is the result of the Attention module, W is the spatial projection matrix, b represents the specific Token bias, and ⊙ represents element-wise multiplication. Finally, the result of the Spatial Gating Unit is mapped to the channel dimension to obtain the output of the gMLPblock;

[0046] Finally, the output of gMLPblocks is merged with the input of gMLP as the output of gMLP.

[0047] Furthermore, the method in step 4 of the present invention includes:

[0048] The output feature Z of the gMLP layer is used as the input of the output layer. Then, through a layer normalization and two fully connected layers, a complex combination of gMLP features can be learned to enhance the expressive power of the model. The specific implementation is as follows:

[0049] Z′=norm(W1*reduce_dim(Z)+b1)

[0050] Output=sigmoid(ReLu(W2*Z′+b2))

[0051] Among them, W1 and W2 in the fully connected layer are trainable weight matrices, and b1 and b2 are trainable bias vectors;

[0052] Finally, the activation function Sigmoid is used to obtain the probability of the chromatin open region feature, with a value range between 0 and 1; the specific implementation is shown below:

[0053]

[0054] Furthermore, the method in step 5 of the present invention includes:

[0055] When calculating the network loss, the binary cross entropy loss function BCELoss is used to measure the difference between the target label value and the predicted probability value, as shown below:

[0056] BCELoss=-w*[p*log(q)+(1-p)*log(1-q)]

[0057] Where p is the theoretical label, which can be 0 or 1; q is the model output prediction value, which ranges from [0, 1]; and w is the weight. For each sample, when p is 1, the loss value is only w*log(q); when p is 0, the loss value is w*log(1-q). The final loss value is obtained by summing and negating the loss values ​​corresponding to the above two cases.

[0058] The optimized gradient algorithm AdamW is used to update the parameters. The weight decay term is directly added to the loss function to ensure that the parameters can be adjusted more accurately during the adaptive learning rate update process. The specific expression is as follows:

[0059] L2_regularization=weigt_decay*w

[0060] w=w-lr*(m / (sqrt(v)+eps)+L2_regularization)

[0061] Among them, weight_decay is the weight decay coefficient, lr is the learning rate, m is the first-order moment estimate of the gradient, v is the second-order moment estimate of the gradient; eps is a small constant used to increase numerical stability.

[0062] The present invention provides a system for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP. The system constructs a DeepOCR model, including a data input module, a multi-scale convolutional network module, a gMLP module, and a data output module; specifically, the system includes:

[0063] The data input module is used to obtain maize chromatin ATAC-seq data, process the ATAC-seq data using one-hot encoding, and input the resulting matrix into the multi-scale convolutional network layer;

[0064] The multi-scale convolutional network module is used to focus on and fuse multi-scale features and capture important motifs of different lengths using convolution kernels of different sizes in the multi-scale convolutional network layer. SoftPool is used to remove a large amount of redundant information while reducing information loss. The output of the multi-scale convolutional network layer is used as the input of the gMLP layer.

[0065] The gMLP module takes the output of the multi-scale convolutional network layer as its input and captures complex spatial interactions across tokens, thereby obtaining long-range dependencies between features in the sequence. The gMLP module includes layer normalization, three layers of gMLP blocks, and a residual-like structure.

[0066] The data output module is used to convert the gMLP layer output features into inputs. Then, through a layer normalization and two fully connected layers, a complex combination of gMLP features can be learned, which enhances the model's expressiveness and enables the classification of maize chromatin open region features.

[0067] The model parameter update module is used to 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, and update the parameters using the optimized gradient algorithm AdamW, which directly adds the weight decay term to the loss function to ensure more accurate parameter adjustment during the adaptive learning rate update process.

[0068] Furthermore, the multi-scale convolutional network module of the present invention includes: multiple convolution kernels of different sizes, activation functions, SoftPool, Batch Normalization and Dropout; wherein: each convolution kernel of different sizes is connected to an activation function and a SoftPool respectively; then the output ends of multiple SoftPools are added and fused and then connected to BatchNormalization and Dropout in sequence.

[0069] The beneficial effects produced by the present invention are:

[0070] The present invention provides a deep learning method that integrates multi-scale convolution and gMLP to learn the characteristics of corn DNA sequences to predict corn chromatin open regions. The technical solution provided by the present invention is as follows Figure 1 shown.

[0071] The present invention first performs one-hot encoding on the processed ATAC-seq data. A multi-scale convolutional network is then used to extract and fuse local features of the sequence using convolution kernels of different sizes. Softpool is used to reduce information loss in the feature map, and batch normalization and dropout are used to prevent overfitting. Convolution kernels of lengths 8 and 13 are capable of capturing important motifs of varying lengths. A gating mechanism based on spatial projection, gMLP, is then used to capture global dependencies between features, with the attention mechanism helping to enhance gating capabilities. Finally, a fully connected layer and a sigmoid function are used to predict open regions of maize chromatin.

[0072] The present invention proposes a method DeepOCR for predicting corn chromatin open regions based on multi-scale convolutional networks and gMLP. The corn B73 variety contains five tissues: Ear, Leaf, Flower, Root, and Tassel. Compared with existing models, DeepOCR performs best in the AV_AUROC and AV_AUPR values ​​of the five tissues and as a whole. At the same time, the standard deviation of AUROC of DeepOCR is the lowest, indicating that DeepOCR has higher accuracy and stronger stability in predicting corn chromatin open regions. In addition, the model trained for each tissue was used to predict the open regions corresponding to the other four tissues. It was found that the accuracy of predicting the open regions of the other four tissues using the model of the Flower tissue was the highest, and the performance was most significant in the Tassel tissue. For the Ear tissue, the prediction results of the models corresponding to the Leaf and Root were relatively similar. For the Leaf tissue, the prediction results of the models corresponding to the Root and Tassel were similar. The performance of the models corresponding to the Ear and Leaf in the other three tissues was not as good as other models, which provides a reference for the conservation and specificity among corn B73 tissues. Furthermore, DeepOCR was used to explore the relatedness and genetic diversity among different varieties, as well as the similarities and differences between the same tissues of different species. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0074] Figure 1 is a schematic diagram of a flow chart of an embodiment of the present invention;

[0075] Figure 2 The ROC and PR results of different methods for predicting the characteristics of chromatin open regions in maize B73;

[0076] Figure 3 The box plots and histograms of the prediction results of chromatin open region features in maize B73 using different methods are shown;

[0077] Figure 4 This is the prediction result of chromatin open regions in five tissues corresponding to maize B73 by DeepOCR;

[0078] Figure 5 This is the result of DeepOCR predicting chromatin open regions across tissues in maize B73;

[0079] Figure 6 This is the prediction result of chromatin open regions in Leaf tissues of different maize varieties by DeepOCR;

[0080] Figure 7 This is the prediction result of chromatin open regions by DeepOCR in Leaf tissues of different species. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 only used to explain the present invention and are not intended to limit the present invention.

[0082] Example 1

[0083] The method for predicting maize chromatin open regions based on a multi-scale convolutional network and gMLP in the embodiment of the present invention is a method for predicting maize chromatin open regions by integrating a multi-scale convolutional network and gMLP. The method constructs a DeepOCR model through the following steps. The steps of the DeepOCR model are as follows:

[0084] Step 1: One-hot encoding is used to process ATAC-seq data, and the resulting matrix is ​​input into a multi-scale convolutional network.

[0085] In a preferred embodiment of the present invention, the sequenced maize chromatin open region data was processed by data segmentation and other methods, and a 1000-bp sequence length was used as the model input. The sequence was then 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]. This resulted in a 1000*4 one-hot encoded matrix, which was then fed into a multi-scale convolutional layer for further processing.

[0086] Step 2: Utilize convolution kernels of varying sizes within a multi-scale convolutional network to focus on and fuse multi-scale features and capture important motifs of varying lengths. Smaller convolution kernels can capture fine-grained features, while larger kernels can capture features across a wider range, thereby enhancing the network's ability to capture multi-dimensional features of DNA sequence data. Furthermore, a more refined pooling method, Softpool, compared to maximum pooling and average pooling, is used to remove redundant information while minimizing information loss, ensuring efficient computation and reducing memory consumption.

[0087] In a preferred embodiment of the present invention, convolution kernels of varying sizes within a multi-scale convolutional network are used to focus on and fuse multi-scale features and capture motifs of varying lengths. The convolution operation extracts key data features through the convolution kernels, achieving local perception and dimensionality reduction. Its parameter sharing feature effectively reduces the number of network parameters, thereby simplifying computation and improving network performance. The specific implementation of the convolution operation is shown below:

[0088] ConV(x) i,j =∑ m ∑ n X(m*s,n*s)*W k (im,jn) (1)

[0089] Here, i and j are the output position indices, and k is the kernel index. m and n are the index variables in the convolution kernel. The convolution kernel slides across the feature map with a stride of s. At each position (i, j), the value at that position in the feature map is output by element-by-element multiplication of the feature map and the convolution kernel, and then all products are summed. To simultaneously perceive and extract features at different scales and more comprehensively capture information in the data, convolution kernels of different scales are used to implement multi-scale convolution operations. Smaller convolution kernels can capture fine-grained features, while larger convolution kernels can capture features over a wider range, thereby enhancing the network's ability to capture multi-dimensional features of DNA sequence data.

[0090] Then, the activation function ReLU is used to learn the nonlinear relationship between features. The expression is as follows:

[0091] ReLU(X)=max(0,X) (2)

[0092] At the same time, the pooling method Softpool is used to remove a large amount of redundant information while reducing information loss, thereby ensuring high computational efficiency and reducing memory consumption. The expression of Softpool is as follows:

[0093]

[0094] Where i is the original feature map position index, and j is the pooling weight index. After the input sequence undergoes convolution, ReLU, and Softpool operations of different scales, it is summed up and fused to extract the important features.

[0095] In order to improve training speed and model stability, a batch normalization operation is added in the middle of the multi-scale convolution layer. Its expression is as follows:

[0096]

[0097] Among them, μ, σ 2 are the global mean and variance. γ and β are learnable parameters used for feature scaling and translation. ε is a small constant used to stabilize the value and avoid division by zero. This function aims to normalize the data in each batch to reduce the range of input data variation, thereby improving model stability and convergence speed. BatchNormalization also introduces some noise to provide regularization, effectively preventing model overfitting and reducing internal covariate shift in each layer of the network, thereby improving network training performance.

[0098] At the same time, in order to prevent overfitting, the Dropout operation is added, which changes the activation values ​​of some neural network units to 0 with probability p, that is, the weights are not updated to reduce the number of intermediate features.

[0099] In step 3, gMLP is used to capture complex spatial interactions across tokens, thereby identifying long-range dependencies between features in the sequence. The gMLP consists of layer normalization, three layers of gMLP blocks, and a residual-like structure. The gMLP blocks include a miniature Attention module, which helps capture long-range dependencies between features. Finally, the output of the gMLP blocks is merged with the input of the gMLP as the output of the gMLP. This creates a residual structure that maintains network performance, ensuring good information fitting when training deep networks and minimizing performance degradation.

[0100] In a preferred embodiment of the present invention, gMLP is used to capture complex spatial interactions across tokens, thereby obtaining long-range dependencies between features in a sequence. gMLP includes layer normalization, three layers of gMLP blocks, and a residual-like structure. First, Patches Embedding is performed on the module input X. The expression of the entire gMLP module is as follows:

[0101] Y=norm(X) (5)

[0102] Y=gMLPblocks(Y) (6)

[0103] Z=Y+X (7)

[0104] Here norm refers to Layer Normalization.

[0105] Blocks contains a micro Attention module that helps capture long-distance dependencies between features. The Attention mechanism expression is as follows:

[0106]

[0107] By performing a linear transformation on the input sequence, we can obtain the query vector (Query), key vector (Key), and value vector (Value). ij It is used to calculate the similarity between Q and K, and then calculate each attention weight, and then perform weighted summation with V.

[0108] The parallel part of Attention, called Proj_in, is implemented by a fully connected layer and the activation function GELU, which is used to perform channel projection mapping on the input. GELU introduces random regularization in the activation, which is a probabilistic description of neuron input. It can be considered a comprehensive implementation of Dropout, Zoneout, and ReLU. The GELU expression is as follows:

[0109] GELU(X)=xP(X<=x)=xΦ(x) (10)

[0110] Where Φ(x) is the probability function of the normal distribution, using the standard normal distribution N(0,1).

[0111] The above two results are input to the Spatial Gating Unit, the core of gMLP. The Spatial Gating Unit operates only on the spatial dimension without the need for position embedding, thereby capturing the interaction of spatial dimension features. Its expression is as follows:

[0112] O=x1⊙(W(x2+x3)+b) (11)

[0113] Here, x1 and x2 represent the channel-wise separation results of Proj_in, x3 represents the result of the Attention module, W represents the spatial projection matrix, b represents the specific token bias, and ⊙ represents element-wise multiplication. Finally, the results of the Spatial Gating Unit are mapped to the channel dimension to form the output of the gMLP block.

[0114] Finally, the results of the gMLP blocks are combined with the gMLP input as the gMLP output. This operation implements a residual structure that maintains network performance. Through the identity mapping, it can still ensure good information fitting ability when training deep networks, thereby reducing network performance degradation.

[0115] In step 4, the output features of the gMLP are used as the input to the output layer. Then, through a layer normalization and two fully connected layers, a complex combination of gMLP features can be learned, which enhances the model's expressiveness and enables the classification of maize chromatin open region features. The specific implementation is as follows:

[0116] Z′=norm(W1*reduce_dim(Z)+b1) (12)

[0117] Output=sigmoid(ReLu(W2*Z′+b2)) (13)

[0118] In the fully connected layer, W1 and W2 are trainable weight matrices, and b1 and b2 are trainable bias vectors.

[0119] Finally, the activation function Sigmoid is used to obtain the probability of the chromatin open region feature, with a value range between 0 and 1. The specific implementation is shown below:

[0120]

[0121] In step 5, 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, and the parameters are updated using the optimized gradient algorithm AdamW, which directly adds the weight decay term to the loss function to ensure that the parameters can be adjusted more accurately during the adaptive learning rate update process.

[0122] In a preferred embodiment of the present invention, the network loss is calculated using a binary cross entropy loss function BCELoss that measures the difference between the target label value and the predicted probability value, as shown below:

[0123] BCELoss=-w*[p*log(q)+(1-p)*log(1-q)] (15)

[0124] Where p is the theoretical label, which can be 0 or 1. q is the model's predicted output value, which ranges from [0 to 1]. w is the weight. For each example, when p is 1, the loss is simply w*log(q). When p is 0, the loss is w*log(1-q). The final loss is obtained by summing and inverting the corresponding loss values ​​for the two cases above.

[0125] The optimized gradient algorithm AdamW is used to update the parameters. The weight decay term is directly added to the loss function to ensure that the parameters can be adjusted more accurately during the adaptive learning rate update process. The specific expression is as follows:

[0126] L2_regularization=weight_decay*w (16)

[0127] w=w-lr*(m / (sqrt(v)+eps)+L2_regularization) (17)

[0128] Where weight_decay is the weight decay coefficient, lr is the learning rate, m is the first moment estimate of the gradient, and v is the second moment estimate of the gradient. eps is a small constant used to increase numerical stability.

[0129] Example 2

[0130] The following are example results of using the method of the present invention to predict chromatin open regions in maize B73 (five tissues: Ear, Leaf, Flower, Root, Tassel), Leaf tissue from 24 parents in the NAM, and six additional species (Tair, ZS97, MH63, Bdistachyon, Sitalica, and Sbicolor). These experiments demonstrate the superiority of the method of the present invention in predicting chromatin open regions, providing a reference for tissue conservation and specificity across different maize tissues, varieties, and species.

[0131] In the prediction of chromatin open regions in maize B73, Figure 2It can be seen that DeepOCR has high prediction accuracy for all 19 types of chromatin open region features. In the ROC curve results, the area under the curve of CharPlant and SMOC is significantly smaller, and the curves of Baaset and DanQ have a small downward deviation. There is little difference between DeepOCR and PlantDeepSEA. In the PR curve results, the curves of Basset, CharPlant, and SMOC are significantly concave, and the area under the curve of DeepOCR is significantly larger than that of DanQ and PlantDeepSEA. This shows that DeepOCR has a strong ability to distinguish between positive and negative examples and can achieve a good balance between precision and recall. In addition, for Ear tissue, DeepOCR's AV_AUROC reached 0.9859 and AV_AUPR reached 0.5905, which are at least 0.11% and 1.89% better than the best results of the other five methods, respectively. For the Leaf tissue, DeepOCR achieved an AV_AUROC of 0.9880 and an AV_AUPR of 0.5782, outperforming the best results of the other five methods by at least 0.09% and 2.46%, respectively. For the Flower tissue, DeepOCR achieved an AV_AUROC of 0.9877 and an AV_AUPR of 0.3550, outperforming the best results of the other five methods by at least 0.01% and 0.09%, respectively. For the Root tissue, DeepOCR achieved an AV_AUROC of 0.9878 and an AV_AUPR of 0.5788, outperforming the best results of the other five methods by at least 0.07% and 2.45%, respectively. For the Tassel tissue, DeepOCR achieved an AV_AUROC of 0.9879 and an AV_AUPR of 0.5647, outperforming the best results of the other five methods by at least 0.03% and 1.97%, respectively.

[0132] Furthermore, in order to compare the performance of different methods as a whole, the results are compared through box plots and histograms, such as Figure 3 shown. Figure 3 (A-B) show the AUROC and AUPR results for each open chromatin region feature predicted by different methods. DeepOCR achieves the highest median values ​​of 0.9871 and 0.5859, respectively. Furthermore, DeepOCR's predictions are more concentrated. DeepOCR's standard deviation of AUROC is 0.0019, the lowest among all methods. Figure 3(CD) shows the average AUROC and AUPR results of the six methods. The DeepOCR model achieved the best AV_AUROC value of 0.9870. DeepOCR also achieved the best AV_AUPR value of 0.5581. Therefore, the DeepOCR method of the present invention has higher accuracy and greater stability in predicting open chromatin regions in maize.

[0133] In the task of predicting chromatin open regions in independent tissues of corn B73, DeepOCR was used to train different models on different tissues. Among the ACC indicators, the prediction accuracy of Flower tissue was the highest, reaching 0.9173. At the same time, the prediction accuracy of Ear tissue and Leaf tissue was similar, and the prediction accuracy of Root tissue and Tassel tissue was also close, with all four around 0.9. Among the AUROC and AUPR indicators, the AUC curve and PR curve of Flower tissue were the most prominent, with AUROC values ​​and AUPR values ​​of 0.9708 and 0.9697, respectively. The AUC curve and PR curve of Leaf tissue were both located in the innermost part, with AUROC values ​​and AUPR values ​​of 0.9486 and 0.9502, respectively. The AUROC and AUPR values ​​of Root tissue and Tassel tissue were similar, and both were better than the results of Ear tissue. In short, the prediction of the five tissues was roughly the same under different indicators. The above results are as follows. Figure 4 shown.

[0134] In the cross-tissue prediction task of corn B73, the model trained in each tissue was used to predict the chromatin open regions of the other four tissues, and then the conservation and specificity between tissues were explored. The results are as follows: Figure 5 (AC) is shown. It can be seen that the prediction trends of chromatin open regions among tissues are very similar. Among them, the chromatin open regions of Flower tissue were predicted with the highest accuracy by the models corresponding to the other four tissues, and the performance was most significant in the model corresponding to Tassel tissue, with ACC of 0.9570, AUC of 0.9886, and AP of 0.9900. The chromatin open regions of Ear tissue were predicted with the lowest accuracy by the models corresponding to the other four tissues. At the same time, the model corresponding to Flower tissue had the lowest accuracy in predicting the other four tissues. In addition, the models corresponding to other tissues had similar accuracy in predicting Ear and Leaf tissues, and the models corresponding to other tissues had similar accuracy in predicting Root and Tassel tissues. Overall analysis, the results are as follows Figure 5As shown in Figures 5 and 6 (D-F), the model corresponding to Flower tissue had the lowest prediction accuracy for the other tissues, with an AV_ACC of only 0.8876. Using the AUC and AP metrics, the prediction results for the models corresponding to Ear and Leaf tissues were more concentrated for the other four tissues, with standard deviations of 0.0107 and 0.0120, respectively. In contrast, the prediction results for the model corresponding to Tassel tissue were more dispersed, with a standard deviation of 0.0187. Furthermore, the model corresponding to Ear tissue had the highest median prediction accuracy for the other tissues, followed by the models corresponding to Leaf and Tassel tissues.

[0135] Based on the method of the present invention, the Leaf tissue data of 24 parents in NAM were selected to explore the influence of kinship and genetic diversity between different varieties on the prediction results. The model of Leaf tissue of B73 was used to predict the chromatin open regions of Leaf tissue of other varieties. The results are as follows Figure 6 As shown in the figure. With the exception of varieties Il14H, M37W, and Mo18W, the ACC for the other 21 varieties all exceeded 0.96, with AUC and AP reaching 0.99. Among them, NC358, CML333, and NC350, all tropical varieties, had the highest prediction accuracy across various metrics, indicating that their leaf tissue is more similar to that of B73. In contrast, the prediction results for varieties Il14H, M37W, and Mo18W were similar, differing significantly from those of the other 21 varieties, with an average ACC of 0.5726. This suggests that these three varieties may have significant genetic differences from B73.

[0136] The method of the present invention was applied to the prediction of chromatin open regions in leaf tissues of different species. The three species with the highest prediction accuracy were Sbicolor, Sitalica, and ZS97. Among them, the AUC value of Sbicolor reached 0.9065, indicating that the chromatin open region characteristics of its leaf tissue were more similar to those of B73. The prediction results of MH63 and Bd were similar, with AUC reaching 0.83. Tair had the lowest prediction accuracy, with an ACC value of only 0.5763, indicating that there were significant differences in the chromatin open region characteristics of its leaf tissue from B73. The above results are as follows Figure 7 shown.

[0137] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional improvements or changes can be easily achieved. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

[0138] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

Claims

1. A method for predicting corn chromatin open regions based on multi-scale convolutional networks and gMLP, characterized in that: The DeepOCR model is constructed by this method, including an input layer, a multi-scale convolutional network layer, a gMLP layer, and an output layer. The method specifically includes the following steps: Step 1: In the input layer, ATAC-seq data of maize chromatin is obtained and processed using one-hot encoding. The resulting matrix is ​​then input into the multi-scale convolutional network layer. Step 2: Use convolution kernels of different sizes in the multi-scale convolutional network layer to focus on and fuse multi-scale features and capture important motifs of different lengths. SoftPool is used to remove a large amount of redundant information while reducing information loss. The output of the multi-scale convolutional network layer is used as the input of the gMLP layer. Step 3: The gMLP layer takes the output of the multi-scale convolutional network layer as its input to capture complex spatial interactions across tokens, thereby obtaining long-range dependencies between features in the sequence. The gMLP layer includes layer normalization, three layers of gMLP blocks, and a residual-like structure. Step 4: The output features of the gMLP layer are used as the input of the output layer. Then, a layer normalization and two fully connected layers are used to learn a complex combination of gMLP features, which enhances the model's expressive power and enables the classification of maize chromatin open region features. In step 5, 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, and the parameters are updated using the optimized gradient algorithm AdamW, which directly adds the weight decay term to the loss function to ensure more accurate parameter adjustment during the adaptive learning rate update process.

2. The method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP according to claim 1, characterized in that: The method in step 1 includes: The maize chromatin open region data obtained by sequencing was processed by data partitioning and other processing, and a sequence length of 1000bp was used as the input of the model. The sequence was then 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]. This resulted in a 1000*4 one-hot encoded matrix, which was input into the multi-scale convolutional layer for the next step.

3. The method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP according to claim 1, characterized in that: The method in step 2 includes: The multi-scale convolutional network layer includes: multiple convolution kernels of different sizes, activation functions, SoftPool, BatchNormalization and Dropout; among them: each convolution kernel of different sizes is connected to an activation function and a SoftPool respectively; then the output ends of multiple SoftPools are added and fused and connected to Batch Normalization and Dropout in sequence; the convolution operation in the multi-scale convolutional network layer extracts key data features through the convolution kernel, realizing local perception and dimensionality reduction functions.

4. The method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP according to claim 3, characterized in that: The method in step 2 includes: The specific implementation of the convolution operation is as follows: in, i , j is the output position index, k is the kernel index, m and n Is the index variable in the convolution kernel; the convolution kernel is in steps of s Slide on the feature map, at each position ( i , j ), by multiplying the feature map and the convolution kernel element by element, and then summing all the products, the value at that position in the feature map is output; in order to simultaneously perceive and extract features of different scales, convolution kernels of different scales are used to implement multi-scale convolution; The activation function ReLU is used to learn the nonlinear relationship between features. The expression is as follows: Based on the Softmax function, Softpool makes the output obey a certain probability distribution by highlighting the activation values ​​with great influence rather than directly selecting the maximum value or average value. The expression of Softpool is as follows: in, i is the original feature map position index, j is the pooling weight index; after the input sequence undergoes convolution, ReLU and Softpool operations of different scales, it is added together to fuse the important features extracted; In order to improve the training speed and model stability, a batch normalization operation BatchNormalization is added in the middle of the multi-scale convolution layer. Its expression is as follows: in, , are the global mean and variance, and are learnable parameters for feature scaling and translation operations, It is a small constant used to stabilize the value and avoid division by zero; The Dropout operation is based on the probability p The activation values ​​of some neural network units are changed to 0, that is, the weights are not updated, so as to reduce the number of intermediate features, thereby increasing the orthogonality between the features of each layer and preventing overfitting.

5. The method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP according to claim 1, wherein: The gMLPblocks in step 3 contain a micro Attention module for capturing long-distance dependencies between features. The output of the gMLPblocks is merged with the input of the gMLP layer as the output of the gMLP layer, thereby realizing a residual structure that maintains network performance.

6. The method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP according to claim 5, characterized in that: The method in step 3 includes: The gMLP layer consists of several blocks of the same size and structure, based on gated MLPs. In terms of space rather than channel dimensions, the gMLP layer captures complex spatial interaction information across tokens. X Perform the Patches Embedding operation. The expression of the entire gMLP layer is as follows: in, norm Refers to Layer Normalization; To enhance the learning effect of the network, a total of three layers of gMLPblocks are used. Blocks contain a micro Attention module to capture long-range dependencies between features. The expression of the Attention mechanism is as follows: By performing a linear transformation on the input sequence, we can obtain the query vector Query, key vector Key, and value vector Value; s ij Used for calculation Q , K The similarity between them is calculated, and then each attention weight is calculated and then compared with V Perform weighted summation; The parallel part of Attention is called Proj_in, which is implemented by a fully connected layer and the activation function GELU, and is used to perform channel projection mapping operations on the input. GELU introduces random regularization in the activation, which is a probabilistic description of neuron input and is considered to be a comprehensive functional implementation of Dropout, Zoneout, and ReLU. The GELU expression is as follows: in, Φ ( x ) is the probability function of the normal distribution, using the standard normal distribution N(0, 1); The result is input into the core Spatial Gating Unit of the gMLP layer. The Spatial Gating Unit only operates on the spatial dimension without position embedding, thereby capturing the interaction of spatial dimension features. Its expression is as follows: in, x 1, x 2 is the separation result of Proj_in in the channel dimension, x 3 is the result of the Attention module, W is the spatial projection matrix, b Indicates a specific Token deviation, represents element-wise multiplication; finally, the result of SpatialGating Unit is mapped to the channel dimension to obtain the output of gMLPblock; Finally, the output of gMLPblocks is merged with the input of gMLP as the output of gMLP.

7. The method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP according to claim 1, characterized in that: The method in step 4 includes: Output features of the gMLP layer Z The input of the output layer is then passed through a layer normalization and two fully connected layers to learn the complex combination of gMLP features and enhance the expressive power of the model. The specific implementation is as follows: Among them, in the fully connected layer W 1 and W 2 is the trainable weight matrix, b 1 and b 2 is the trainable bias vector; Finally, the activation function Sigmoid is used to obtain the probability of the chromatin open region feature, with a value range between 0 and 1; the specific implementation is shown below: 。 8. The method for predicting corn chromatin open regions based on a multi-scale convolutional network and gMLP according to claim 1, wherein: The method in step 5 includes: When calculating the network loss, the binary cross entropy loss function BCELoss is used to measure the difference between the target label value and the predicted probability value, as shown below: in, p is the theoretical label, which takes 0 or 1; q Output prediction value for the model, the value range is [0, 1], w is the weight; for each sample, when p When it is 1, the loss value is only w *log( q ),when p When it is 0, the loss value is w *log(1- q ); The final loss value is obtained by summing and negating the loss values ​​corresponding to the above two cases; The optimized gradient algorithm AdamW is used to update the parameters. The weight decay term is directly added to the loss function to ensure that the parameters can be adjusted more accurately during the adaptive learning rate update process. The specific expression is as follows: in, weight_decay is the weight decay coefficient, lr is the learning rate, m is the first moment estimate of the gradient, v is the second moment estimate of the gradient; eps is a small constant used to increase numerical stability.

9. A system for predicting open regions of maize chromatin based on multi-scale convolutional networks and gMLP, characterized in that: The DeepOCR model is constructed in this system, including a data input module, a multi-scale convolutional network module, a gMLP module, and a data output module; specifically, it includes: The data input module is used to obtain maize chromatin ATAC-seq data, process the ATAC-seq data using one-hot encoding, and input the resulting matrix into the multi-scale convolutional network layer; The multi-scale convolutional network module is used to focus on and fuse multi-scale features and capture important motifs of different lengths using convolution kernels of different sizes in the multi-scale convolutional network layer. SoftPool is used to remove a large amount of redundant information while reducing information loss. The output of the multi-scale convolutional network layer is used as the input of the gMLP layer. The gMLP module takes the output of the multi-scale convolutional network layer as its input and captures complex spatial interactions across tokens, thereby obtaining long-range dependencies between features in the sequence. The gMLP module includes layer normalization, three layers of gMLP blocks, and a residual-like structure. The data output module is used to convert the gMLP layer output features into inputs. Then, through a layer normalization and two fully connected layers, the complex combination of gMLP features can be learned, which enhances the model's expressiveness and enables the classification of maize chromatin open region features. The model parameter update module is used to 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, and update the parameters using the optimized gradient algorithm AdamW, which directly adds the weight decay term to the loss function to ensure more accurate parameter adjustment during the adaptive learning rate update process.

10. The system for predicting corn chromatin open regions based on multi-scale convolutional networks and gMLP according to claim 9, characterized in that: The multi-scale convolutional network module includes: multiple convolution kernels of different sizes, activation functions, SoftPool, Batch Normalization and Dropout; wherein: each convolution kernel of different sizes is connected to an activation function and a SoftPool respectively; then the output ends of multiple SoftPools are added and fused and then connected to Batch Normalization and Dropout in sequence.

Images