DNA methylation prediction method and related equipment

Abstract

The invention discloses a DNA methylation prediction method and related equipment, and the method comprises the steps: obtaining DNA methylation data of a plurality of species, and carrying out the preprocessing to obtain a data sample set; performing iterative optimization on a preset deep learning model by using the data sample set to obtain an initial prediction model; performing parameter fine tuning on the initial prediction model by using the data sample set of the target species to obtain a target prediction model; and performing DNA methylation prediction on the target species by using the target prediction model. According to the method, the multi-species public methylation data is integrated through a multi-species pre-training mechanism, and the initial model with generalization ability is constructed, so that the high-precision model can be finely adjusted by the target species only through a small number of samples, and the dependence on large-scale entity tissue data is remarkably reduced. The method can be widely applied to the fields of DNA methylation scientific research, disease risk genetic variation recognition, genetic counseling and the like; the related prediction model technology can be widely applied to the technical field of deep learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a DNA methylation prediction method and related equipment. BACKGROUND

[0002] DNA methylation, as a key epigenetic marker, is directly related to a variety of disease mechanisms. By locating DNA methylation quantitative trait loci (mQTL) to study the regulation of genomic variation on methylation, it is the core path to analyze the causal relationship of diseases. However, it is difficult to obtain samples of solid organ tissues (such as lung, kidney, and intestinal tract) on a large scale, which leads to the fact that existing mQTL research relies heavily on blood tissue data, and restricts the understanding of the real environment of diseases.

[0003] Current machine learning methods rely on artificial design features (such as GC content, transcription factor binding sites, etc.), which have two major defects: 1. Feature engineering bias: it is difficult to fully capture the deep regulation rules of DNA sequences by manually extracting features; 2. Insufficient generalization ability: the model trained based on limited samples of a single species (especially human) has a significant decrease in prediction accuracy in solid organs where data is scarce. SUMMARY

[0004] The main purpose of the embodiments of the present application is to propose a DNA methylation prediction method, device, electronic equipment and storage medium, aiming at solving at least one problem of the prior art.

[0005] To achieve the above-mentioned purpose, one aspect of the embodiments of the present application proposes a DNA methylation prediction method, which comprises:

[0006] Obtaining DNA methylation data of multiple species, and preprocessing to obtain a data sample set;

[0007] Among them, the sample data of the data sample set includes the DNA sequence of the target site and the DNA methylation level;

[0008] Using the data sample set to iteratively optimize a preset deep learning model to obtain an initial prediction model;

[0009] Using the data sample set of the target species to fine-tune the parameters of the initial prediction model to obtain a target prediction model;

[0010] Using the target prediction model to predict DNA methylation of the target species.

[0011] In some embodiments, the data sample set is preprocessed, including the following steps:

[0012] Data preprocessing is performed on the DNA methylation data;

[0013] The data preprocessing includes data cleaning and smoothing processing.

[0014] A data sample set is obtained according to the result of data preprocessing.

[0015] The data sample set includes a training set, a validation set and a test set divided based on a preset ratio.

[0016] In some embodiments, the method further comprises the following steps:

[0017] The deep learning model is configured according to the sequentially connected multi-scale convolution kernel, the species-specific batch normalization layer, the GELU activation function, the DNA sequence feature extraction module, the multi-scale feature storage and the multi-tissue prediction head.

[0018] The species-specific batch normalization layer includes a batch normalization layer corresponding to each species, the DNA sequence feature extraction module includes a multi-scale CNN block stacked a preset number of times, the multi-scale feature storage is used to store the feature map output by each multi-scale CNN block, and the multi-tissue prediction head is used to predict the output methylation level according to the feature map.

[0019] In some embodiments, the method further comprises the following steps:

[0020] The multi-scale CNN block is configured according to the sequentially connected first convolutional layer, the species-specific batch normalization layer, the GELU activation function, the second convolutional layer, the species-specific batch normalization layer and the GELU activation function.

[0021] The input data of the first convolutional layer and the output data of the second species-specific batch normalization layer in the multi-scale CNN block are connected by a residual connection as the input data of the second GELU activation function in the multi-scale CNN block.

[0022] The multi-scale CNN block is stacked a preset number of times to configure the DNA sequence feature extraction module.

[0023] In some embodiments, the method further comprises the following steps:

[0024] The multi-tissue prediction head is configured according to the sequentially connected global average pooling layer, the first linear layer, the ELU activation function, the second linear layer and the Sigmoid activation function.

[0025] In some embodiments, the data sample set includes a training set and a validation set, and the preset deep learning model is iteratively optimized using the data sample set to obtain an initial prediction model, including the following steps:

[0026] One-hot encoding is performed on the DNA sequences in the training set;

[0027] The DNA sequence encoded by one-hot encoding is used as input data into the deep learning model so that the deep learning model outputs the first predicted methylation level.

[0028] The mean squared error is constructed based on the first predicted methylation level and the DNA methylation level as the loss function, and the model parameters of the deep learning model are optimized and adjusted based on the loss function using a preset optimizer.

[0029] The deep learning model is evaluated for accuracy and its hyperparameters are tuned using a validation set. The process then returns to the step of inputting the one-hot encoded DNA sequence as input data into the deep learning model until the preset training conditions are met, and the initial prediction model is obtained.

[0030] The deep learning model is configured with a species-specific batch normalization layer, which includes a batch normalization layer corresponding to each species. The deep learning model calls the corresponding batch normalization layer of the species based on the DNA sequence of the input data.

[0031] In some embodiments, the data sample set includes a training set, a validation set, and a test set. The initial prediction model is fine-tuned using the data sample set of the target species to obtain the target prediction model, including the following steps:

[0032] One-hot encoding of the DNA sequences in the training set of the target species;

[0033] The DNA sequence encoded by one-hot encoding is used as input data into the initial prediction model so that the initial prediction model outputs a second predicted methylation level.

[0034] The mean squared error is constructed based on the second predicted methylation level and the DNA methylation level as the loss function, and the model parameters of the initial prediction model are fine-tuned based on the loss function using a preset optimizer.

[0035] The initial prediction model is evaluated for accuracy and its hyperparameters are adjusted using the validation set. The process is then repeated until the preset adjustment conditions are met. Finally, the prediction accuracy of the initial prediction model is verified using the test set.

[0036] If the prediction accuracy meets the accuracy requirements, the initial prediction model is used as the target prediction model. If the prediction accuracy does not meet the accuracy requirements, the process returns to the step of inputting the one-hot encoded DNA sequence as input data into the initial prediction model until the prediction accuracy meets the accuracy requirements.

[0037] The deep learning model is configured with a species-specific batch normalization layer, and the species-specific batch normalization layer includes a batch normalization layer corresponding to each species.

[0038] To achieve the above-mentioned purpose, another aspect of the embodiment of the present application proposes a DNA methylation prediction device, the device comprises:

[0039] The data processing module is configured to obtain DNA methylation data of multiple species, and preprocess the data to obtain a data sample set.

[0040] The sample data of the data sample set includes the DNA sequence of the target site and the DNA methylation level.

[0041] The iterative optimization module is configured to iteratively optimize a preset deep learning model using the data sample set to obtain an initial prediction model.

[0042] The parameter fine-tuning module is configured to fine-tune the initial prediction model using the data sample set of the target species to obtain a target prediction model.

[0043] The model application module is configured to use the target prediction model to predict the DNA methylation of the target species.

[0044] In some embodiments, the device further comprises a model configuration module configured to perform the following operations:

[0045] The deep learning model is configured according to the sequentially connected multi-scale convolution kernel, the species-specific batch normalization layer, the GELU activation function, the DNA sequence feature extraction module, the multi-scale feature storage, and the multi-tissue prediction head.

[0046] The species-specific batch normalization layer includes a batch normalization layer corresponding to each species, the DNA sequence feature extraction module includes a multi-scale CNN block stacked a preset number of times, the multi-scale feature storage is configured to store the feature map output by each multi-scale CNN block, and the multi-tissue prediction head is configured to predict the output methylation level according to the feature map.

[0047] In some embodiments, the device further comprises a configuration module of the DNA sequence feature extraction module configured to perform the following operations:

[0048] The multi-scale CNN block is configured according to the sequentially connected first convolution layer, the species-specific batch normalization layer, the GELU activation function, the second convolution layer, the species-specific batch normalization layer, and the GELU activation function.

[0049] The input data of the first convolutional layer and the output data of the second species-specific batch normalization layer in the multi-scale CNN block are connected through a residual connection as input data of a second GELU activation function in the multi-scale CNN block.

[0050] The multi-scale CNN block is repeatedly stacked for a preset number of times to configure a DNA sequence feature extraction module.

[0051] In some embodiments, the device further comprises a multi-tissue prediction head configuration module configured to perform the following operations:

[0052] The multi-tissue prediction head is configured according to a global average pooling layer, a first linear layer, an ELU activation function, a second linear layer, and a Sigmoid activation function connected in sequence.

[0053] To achieve the above object, another aspect of the embodiments of the present application provides an electronic device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the foregoing method when executing the computer program.

[0054] To achieve the above object, another aspect of the embodiments of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the foregoing method.

[0055] To achieve the above object, another aspect of the embodiments of the present application provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the foregoing method.

[0056] The embodiments of the present application at least have the following beneficial effects: the present application provides a DNA methylation prediction method, device, electronic device, storage medium and program product, the scheme obtains DNA methylation data of multiple species, and pre-processes to obtain a data sample set; wherein the sample data of the data sample set comprises a DNA sequence of a target site and a DNA methylation level; the data sample set is used to iteratively optimize a preset deep learning model to obtain an initial prediction model; the data sample set of the target species is used to fine-tune the parameters of the initial prediction model to obtain a target prediction model; and the target prediction model is used to predict DNA methylation of the target species. The present application integrates multiple species public methylation data through a multi-species pre-training mechanism, constructs an initial model with generalization ability, solves the core pain point of insufficient single species entity organ sample, makes the target species (such as a specific tissue of human) only need a small amount of sample to fine-tune a high-precision model, and significantly reduces the dependence on large-scale entity tissue data. Moreover, the present application uses a two-stage optimization strategy (pre-training + fine-tuning): the initial model provides universal representation ability, and the target species fine-tuning focuses on learning tissue-specific patterns, providing a reliable basis for disease-related mQTL causal inference. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is an implementation environment schematic diagram of a DNA methylation prediction method provided by an embodiment of the present application;

[0058] Figure 2 is a flow schematic diagram of a DNA methylation prediction method provided by an embodiment of the present application;

[0059] Figure 3 is a structure schematic diagram of a deep learning model applied by a DNA methylation prediction method provided by an embodiment of the present application;

[0060] Figure 4 is a flow application example schematic diagram of a DNA methylation prediction method provided by an embodiment of the present application;

[0061] Figure 5 is a flow schematic diagram of a DNA methylation prediction device provided by an embodiment of the present application;

[0062] Figure 6 is a structure schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementation described in the following exemplary embodiments does not represent all the implementations consistent with the embodiments of the present application. They are only examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the appended claims.

[0064] It can be understood that the terms "first", "second", and the like used in the present application can be used herein to describe various concepts, but unless specifically stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of the present application, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information. Depending on the context, the word "if" as used herein can be interpreted as "when" or "when" or "in response to determining".

[0065] The terms "at least one", "multiple", "each", "any", and the like used in the present application include one, two or more than two, multiple includes two or more than two, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing only embodiments of this application and is not intended to be limiting of this application.

[0067] In the related art, the current machine learning method relies on artificial design features (such as GC content, transcription factor binding sites, etc.), which has two major defects: 1. Feature engineering bias: it is difficult to fully capture the deep regulation rules of DNA sequences by manually extracting features; 2. Insufficient generalization ability: the model trained based on limited samples of a single species (especially humans) has a significant decrease in prediction accuracy in data-scarce solid organs.

[0068] Therefore, in the embodiments of the present application, a DNA methylation prediction method and related equipment are provided. The scheme obtains DNA methylation data of multiple species, and pre-processes to obtain a data sample set. The sample data of the data sample set includes a DNA sequence of a target site and a DNA methylation level. The data sample set is used to iteratively optimize a preset deep learning model to obtain an initial prediction model. The data sample set of the target species is used to fine-tune the parameters of the initial prediction model to obtain a target prediction model. The target prediction model is used to predict the DNA methylation of the target species. The present application integrates multiple species of public methylation data through a multi-species pre-training mechanism, constructs an initial model with generalization ability, solves the core pain point of insufficient solid organ samples of a single species, and makes the target species (such as a specific tissue of a human) only need a small amount of samples to fine-tune a high-precision model, thereby significantly reducing the dependence on large-scale solid tissue data. Moreover, the present application uses a two-stage optimization strategy (pre-training + fine-tuning): the initial model provides universal representation ability, and the target species fine-tuning focuses on learning tissue-specific patterns, thereby providing a reliable basis for disease-related mQTL causal inference.

[0069] It can be understood that the DNA methylation prediction method provided by the present application can be applied to any computer device with data processing and computing capability, and the computer device can be various terminals or servers. When the computer device in the embodiments is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms. Optionally, the terminal is a smart phone, a tablet computer, a notebook computer, and a desktop computer, but is not limited thereto.

[0070] As shown in Figure 1 , it is an implementation environment schematic diagram provided by an embodiment of the present application. Referring to Figure 1 , the implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected through wireless or wired network to complete data transmission and exchange.

[0071] The server 101 can be a stand-alone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network, content distribution network), and basic cloud computing services such as big data and artificial intelligence platforms.

[0072] In addition, the server 101 can also be a node server in a blockchain network. The blockchain is a new application mode of distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, and other computer technologies.

[0073] The terminal 102 can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal 102 and the server 101 can be directly or indirectly connected through wired or wireless communication, and the embodiments of the present application do not limit this.

[0074] For example, based on the implementation environment shown in Figure 1 , the present application provides a DNA methylation prediction method, which is applied to the server 101 as an example below. It can be understood that the DNA methylation prediction method can also be applied to the terminal 102.

[0075] Referring to Figure 2 , Figure 2 , it is an optional flowchart of the DNA methylation prediction method provided by an embodiment of the present application. The execution subject of the DNA methylation prediction method can be any one of the aforementioned computer devices (including servers or terminals). Figure 2 The method in may include but is not limited to steps S100 to S400.

[0076] Step S100, obtaining DNA methylation data of multiple species, and preprocessing to obtain a data sample set;

[0077] The sample data of the data sample set includes the DNA sequence of the target site and the DNA methylation level.

[0078] It should be noted that in some embodiments, the pre-processing of the data sample set can include the following steps: data pre-processing of the DNA methylation data; wherein the data pre-processing includes data cleaning and smoothing processing; and arranging the data sample set according to the results of the data pre-processing; wherein the data sample set includes a training set, a validation set and a test set divided based on a preset ratio.

[0079] Exemplarily, in some specific embodiments, data collection and processing can be implemented as follows:

[0080] First, a large amount of DNA methylation data of human samples was collected from public databases in total; in addition, DNA methylation data of mouse, non-human primates (Macaca fascicularis / Macaca mulatta), anthropoid (Gorilla gorilla / Chimpanzee) and domestic animals (cattle / sheep / dog) and other mammalian species were also collected. The steps are as follows: download DNA methylation data based on WGBS-seq technology from public databases.

[0081] The steps of data processing are as follows:

[0082] Data cleaning: removing low-quality sequences, removing ambiguous bases such as N bases, and removing CpG sites with poor data quality.

[0083] The DNA methylation data is smoothed using the BSmooth algorithm.

[0084] Data segmentation: the data set is divided into training set, validation set and test set in the ratio of 80%:10%:10% (the specific ratio can be adjusted according to actual needs), the training set is used for training of MethylAI model, the validation set is used to evaluate whether the hyperparameter design of the model is reasonable, and the test set is used to evaluate the generalization ability of the model.

[0085] In some embodiments, the method can further include the following steps: configuring a deep learning model according to a plurality of scale convolution kernels connected in sequence, a species-specific batch normalization layer, a GELU activation function, a DNA sequence feature extraction module, a multi-scale feature storage, and a multi-tissue prediction head; wherein the species-specific batch normalization layer includes a batch normalization layer corresponding to each species, the DNA sequence feature extraction module includes a plurality of scale CNN blocks stacked repeatedly for a preset number of times, the multi-scale feature storage is used to store feature maps output by each multi-scale CNN block, and the multi-tissue prediction head is used to predict and output the methylation level according to the feature maps.

[0086] In some embodiments, the method can further comprise the following steps: configuring a multi-scale CNN block according to a first convolutional layer, a species-specific batch normalization layer, a GELU activation function, a second convolutional layer, a species-specific batch normalization layer, and a GELU activation function connected in sequence; wherein the input data of the first convolutional layer and the output data of the second species-specific batch normalization layer in the multi-scale CNN block are connected through a residual connection as the input data of the second GELU activation function in the multi-scale CNN block; and repeating the stacking of the multi-scale CNN block for a preset number of times to configure a DNA sequence feature extraction module.

[0087] In some embodiments, the method can further comprise the following steps: configuring a multi-tissue prediction head according to a global average pooling layer, a first linear layer, an ELU activation function, a second linear layer, and a Sigmoid activation function connected in sequence.

[0088] By way of example, in some specific embodiments, the present application designs and develops a deep learning model MethylAI based on a convolutional neural network (CNN), which can input the DNA sequence on both sides of a CpG in the human genome and output the DNA methylation level of the CpG site in different tissues of the human body. As shown in Figure 3 The specific structure and parameters of the MethylAI model are as follows (the set values of the parameters, specifications, and scales in the following model structure are only illustrative, and in actual applications, they can be set and adjusted according to actual needs):

[0089] Multi-scale Conv: multi-scale convolution kernel, which gives MethylAI the ability to recognize DNA feature sequences of different lengths.

[0090] Species specific BatchNorm: species-specific batch normalization layer, which calls the species-specific batch normalization layer according to the species of the DNA methylation data input into the model, solves the problem of data heterogeneity when training the model with multiple species data, and enables the loss value (an index for measuring the difference between the predicted results and the true results of the model, i.e., the loss value) to steadily decrease when training the model with multiple species data.

[0091] GELU, ELU, Sigmoid: all belong to activation functions, which can introduce a non-linear relationship into the model, enabling it to learn the complex relationship between DNA sequences and DNA methylation levels.

[0092] Conv: convolutional layer, MethylAI uses a convolutional layer with a stride of 4 at a specific position, which optimizes the operation efficiency of the convolutional layer, enables MethylAI to efficiently compress DNA sequence information, saves GPU memory resources, and can read longer DNA sequences under the same hardware.

[0093] Multi-scale CNN block: a multi-scale CNN block composed of Conv, Species specific BatchNorm, GELU, and a residual connection (indicated by a "+" symbol with an outer circle) that adds the input of the module to the output, thereby providing better gradient flow during network training and effectively preventing gradient vanishing and model degradation during model training. The module is repeated 12 times to form a DNA sequence feature extraction module.

[0094] Multi-scale Feature Storage: a multi-scale feature storage for storing the feature maps obtained by the intermediate operation of the Mult-scale CNN block, so that MethylAI can extract and store DNA features of different scales to improve prediction accuracy.

[0095] Adaptive Average Pooling: a global average pooling layer that compresses DNA sequence feature dimension information.

[0096] Multi-tissue Prediction Head: a multi-tissue prediction head that can simultaneously predict the DNA methylation level of the input DNA sequence in multiple tissues. The multi-tissue prediction head is composed of two linear layers (Linear Layer), an ELU activation function, and a Sigmoid activation function. The ELU (Exponential Linear Unit) activation function introduces nonlinearity while avoiding the problem of neuron death caused by the ReLU activation function, improving the model's ability to learn complex patterns. The sigmoid function compresses the output to between 0 and 1, consistent with the range of DNA methylation levels. The prediction head can convert the DNA features obtained by the multi-scale CNN block into the final tissue-specific DNA methylation level. Through the above design, especially the application of activation functions, the model can capture the tissue specificity of DNA methylation, and further predict the DNA methylation level of multiple tissues in the human body.

[0097] Step S200: iteratively optimize the pre-set deep learning model using a data sample set to obtain an initial prediction model;

[0098] It should be noted that the data sample set includes a training set and a validation set, and in some embodiments, step S200 can include the following steps: one-hot encoding the DNA sequences of the training set; inputting the one-hot encoded DNA sequences as input data into the deep learning model to make the deep learning model output a first predicted methylation level; constructing a mean square error as a loss function according to the first predicted methylation level and the DNA methylation level, and optimizing and adjusting the model parameters of the deep learning model based on the loss function; using the validation set to evaluate the accuracy of the deep learning model and adjust the hyperparameters, and returning to the step of inputting the one-hot encoded DNA sequences as input data into the deep learning model until a preset training condition is reached to obtain an initial prediction model; wherein the deep learning model is configured with a species-specific batch normalization layer, the species-specific batch normalization layer includes a batch normalization layer corresponding to each species, and the deep learning model calls the batch normalization layer corresponding to the species according to the DNA sequence of the input data. The preset training condition includes but is not limited to reaching a preset training round or the result of the accuracy evaluation meeting a preset accuracy threshold.

[0099] Exemplarily, in some specific embodiments, the model training scheme is as follows:

[0100] The multi-species training set data is read in, and the DNA sequences on both sides of the CpG (using one-hot encoding, the input DNA length is 18.4 kb) are input into MethylAI, and the output is the DNA methylation level of the CpG site.

[0101] The MethylAI model is trained using the gradient descent method, the mean square error (MSE) is used as the loss function, the AdamW optimizer is used to optimize the model parameters, the learning rate is initialized to 0.001, the batch size is set to 320, and the cosine annealing with warm restarts program is used to adjust the learning rate.

[0102] After each training round (epoch), the accuracy of the MethylAI model is evaluated using the validation set segmented by the foregoing steps, and whether the hyperparameter settings are reasonable is determined. If not, the model will be continuously optimized, and the model will be retrained by returning to the previous step until the preset training condition is reached.

[0103] It should be further noted that the species-specific batch normalization layer calls the species-specific batch normalization layer according to the species of the DNA methylation data input into the model, solves the problem of data heterogeneity in multi-species training, and the loss value can be stably decreased when training the model using multiple species data.

[0104] Step S300, parameter fine-tuning is performed on the initial prediction model using the data sample set of the target species to obtain a target prediction model.

[0105] It should be noted that the data sample set includes a training set, a validation set, and a test set. In some embodiments, step S300 can include the following steps: one-hot encoding the DNA sequences of the training set of the target species; inputting the one-hot encoded DNA sequences into the initial prediction model as input data to enable the initial prediction model to output a second predicted methylation level; constructing a mean square error as a loss function according to the second predicted methylation level and the DNA methylation level, and fine-tuning the model parameters of the initial prediction model based on the loss function; performing accuracy evaluation and hyperparameter adjustment on the initial prediction model using the validation set, returning to the step of inputting the one-hot encoded DNA sequences into the initial prediction model as input data, until a preset adjustment condition is reached, verifying the prediction accuracy of the initial prediction model using the test set; when the prediction accuracy meets the accuracy requirement, the initial prediction model is used as the target prediction model, and when the prediction accuracy does not meet the accuracy requirement, returning to the step of inputting the one-hot encoded DNA sequences into the initial prediction model as input data until the prediction accuracy meets the accuracy requirement; wherein the deep learning model is configured with a species-specific batch normalization layer, and the species-specific batch normalization layer includes a batch normalization layer corresponding to each species; in the iterative optimization phase, the deep learning model calls the batch normalization layer corresponding to the species according to the DNA sequence of the input data; in the parameter fine-tuning phase, the initial prediction model only applies the batch normalization layer corresponding to the target species. The preset adjustment condition includes but is not limited to reaching a preset training round or the result of the accuracy evaluation meeting a preset accuracy threshold.

[0106] For example, in some embodiments, after the multi-species data training is completed, the MethylAI model with the best accuracy is selected, and the MethylAI model is fine-tuned using the data sample set of the target species (such as a human data set). When fine-tuning, the learning rate is reduced to 0.0001, and the remaining settings remain unchanged to perform the model training process (the specific process steps are the same as the principle of the process steps of the foregoing specific embodiments for training the initial prediction model, and will not be described again here).

[0107] After each training round (epoch) is completed, the accuracy of the MethylAI model and whether the hyperparameter setting is reasonable are evaluated using the pre-segmented validation set corresponding to the target species. If not, the model will continue to be optimized, and the model will be retrained in the previous step. Finally, the prediction accuracy of the MethylAI model is evaluated using the pre-segmented independent test set corresponding to the target species.

[0108] The prediction accuracy of the model is measured by calculating the following parameters: mean squared error (MES) of the predicted value and the true value, Pearson's correlation coefficient, and Spearman's correlation coefficient.

[0109] Step S400, DNA methylation prediction of the target species is performed by using the target prediction model.

[0110] Exemplarily, in some specific embodiments, the explainability analysis process of MethylAI can be developed in combination with the DeepSHAP algorithm, to reveal the important regulatory sequences of DNA methylation; the mQTL effect prediction analysis process based on MethylAI is developed, the prediction accuracy of MethylAI for mQTL effect is verified using the public mQTL dataset, and is applied to the mQTL effect prediction analysis of multiple disease risk genetic loci. The specific steps are as follows:

[0111] The DeepSHAP algorithm is applied to the MethylAI model to calculate the attribution score of each base of the input DNA sequence, the regulatory sequences on the DNA are obtained from the JASPAR database, and the important regulatory sequences are screened according to the attribution score.

[0112] There are a large number of DNA variants on the human genome, and a part of the DNA variants, which are called mQTL, can affect DNA methylation. The mQTL dataset is collected from the public database, and the accuracy of MethylAI in mQTL effect prediction is verified using the dataset as a standard.

[0113] There are a large number of DNA variants related to disease occurrence on the human genome, which are called disease risk genetic loci, but whether most of the disease risk genetic loci have mQTL effect is still unclear. MethylAI is used to systematically predict the mQTL effect of disease risk genetic loci, which can provide help for the research of disease mechanism, disease biomarker and drug development in the future.

[0114] To explain the principle of the technical scheme of the present application in detail, the overall process of the present application will be described in combination with some specific embodiments. It is easy to understand that the following is an explanation of the technical principle of the present application and cannot be regarded as a limitation of the present application.

[0115] First of all, it needs to be pointed out that DNA methylation is one of the most important epigenetic modifications, which plays an important role in transcriptional regulation, organ development and disease development. In the human genome, DNA methylation is commonly found in cytosine bases of CpG dinucleotides, which is catalyzed by DNA methyltransferase and generates 5-methylcytosine (5mC). Abnormal DNA methylation is closely related to many diseases, including chronic obstructive pulmonary disease, obesity and diabetes, inflammatory bowel disease, kidney disease, and hypertension.

[0116] Genome-wide association studies (GWAS) have found that genomic DNA variations are complexly related to human diseases. Studying how genomic DNA variations regulate the DNA methylation level of CpG sites is an important means to explore the mechanism of disease development. DNA variation sites associated with DNA methylation levels are called DNA methylation quantitative trait loci (mQTL). Studies have found that there are extensive mQTL sites in the human genome. Combined with methods such as genomic co-localization analysis and Mendelian randomization, causal disease variants can be screened from mQTL sites. However, DNA methylation has significant tissue specificity, and due to the difficulty of obtaining a large number of individual solid organ tissue samples, current mQTL studies mainly focus on blood tissue, and research on solid organ tissue is less.

[0117] In addition to mQTL positioning, traditional machine learning methods and deep learning methods have also been applied to study the relationship between DNA sequences and DNA methylation. Traditional machine learning methods usually include three steps of data processing, feature extraction and classification. Common features include base frequency, GC content, sequence length, repetitive sequence, transcription factor binding site, etc. Different classification algorithms are used based on these features to predict DNA methylation levels. However, the above method needs to manually design and extract features from DNA, making this way less accurate in prediction.

[0118] To overcome the above problems, researchers have begun to explore deep learning-based methods, such as the INTERACT model using convolutional neural networks (CNN) and transformer structures, which can directly predict the DNA methylation level of CpG sites in human brain tissue based on DNA sequences. However, the above model has technical problems such as short reading DNA sequence length and few types of tissues that can be predicted. Therefore, the present application develops a new deep learning model for human genomic DNA methylation, named MethylAI, which is based on the DNA sequence on both sides of the CpG site to predict the DNA methylation level of different tissue cells in the human body. The reading DNA sequence length, the number of predicted tissue types, and the prediction accuracy all surpass the INTERACT model.

[0119] The MethylAI model can predict the DNA methylation state of the site in various tissues of the human body based on the DNA sequence on both sides of the CpG site, and can be applied to predict the mQTL effect of DNA variation sites in the whole genome of human solid tissues, providing important information for disease mechanism research, disease risk prediction, and precision treatment.

[0120] In some embodiments, the present application proposes a deep learning model MethylAI based on convolutional neural networks (CNN), which can predict the DNA methylation level of CpG sites in different human tissues based on the DNA sequence on both sides of the CpG in the human genome, and can be applied to the study of disease risk genetic sites. The innovation of the MethylAI model lies in: successfully developing a multi-species joint model architecture, realizing the unified training and reasoning of a single model for multi-species DNA methylation; designing a Multi-scale Convolutional Neural Network (Multi-scale CNN) sequence feature extraction module, which supports input of DNA sequences centered on CpG sites (the optimal length is currently determined to be 18.4kb); designing a multi-tissue prediction module, which can simultaneously predict the DNA methylation level of the target CpG site in up to 1574 human samples, supporting the study of DNA methylation regulation in multiple human tissue cells and disease risk genetic sites by a single MethylAI model.

[0121] In order to obtain data for training MethylAI, the present application collected a total of 1574 DNA methylation data of human samples based on bisulfite sequencing (WGBS-seq) from public databases, which covered 51 tissues and 171 cell types of human body; in addition, the present application collected DNA methylation data of 11 mammalian species (431 samples in total), including mice, non-human primates (Macaca fascicularis / Macaca mulatta), anthropoids (Gorilla gorilla / Chimpanzee) and domestic animals (cattle / sheep / dog), which were combined with the model architecture of MethylAI multi-species joint to realize the training and reasoning of MethylAI in the multi-species DNA methylation data set, so as to further improve the prediction accuracy of MethylAI.

[0122] After successfully training MethylAI, the present application developed an explainability analysis process of MethylAI combined with DeepSHAP algorithm, which successfully revealed the important regulatory sequences of DNA methylation; then the present application constructed an mQTL effect prediction analysis process based on MethylAI, and after verification, the mQTL effect prediction accuracy of MethylAI on the important regulatory sequences of DNA methylation in multiple tissues reached 94%, which proved that MethylAI could accurately analyze the mQTL effect of disease risk genetic variation sites, and provided help for the research of disease mechanism, disease biomarker and drug development.

[0123] As shown in Figure 4 The present application provides a deep learning-based DNA methylation prediction method (MethylAI model) and its application in disease research, and the scheme can be implemented as follows:

[0124] S1: Data collection and processing. The present application collected a large amount of DNA methylation data of human samples from public databases, which covered 51 tissues and 171 cell types of human body; in addition, the present application collected DNA methylation data of 11 mammalian species. The steps are as follows:

[0125] Download DNA methylation data based on WGBS-seq technology from public databases.

[0126] Data cleaning: remove low-quality sequences, remove ambiguous bases such as N bases, and remove CpG sites with poor data quality.

[0127] Smooth the DNA methylation data using BSmooth algorithm.

[0128] Data splitting: The dataset is divided into training set, validation set and test set in the ratio of 80%:10%:10%, the training set is used for the training of MethylAI model, the validation set is used to evaluate whether the hyperparameter design of the model is reasonable, and the test set is used to evaluate the generalization ability of the model.

[0129] One-hot encoding of DNA sequences on both sides of CpG: Each base is represented by a four-dimensional vector (A: [1, 0, 0, 0], T: [0, 1, 0, 0], G: [0, 0, 1, 0], C: [0, 0, 0, 1]), and the one-hot encoded DNA sequence is directly input into the model.

[0130] S2: Development and training of the model. The present application designs and develops a deep learning model MethylAI based on convolutional neural network (CNN), which can input the DNA sequence on both sides of CpG in human genome and output the DNA methylation level of the CpG site in different tissues of human. As shown in Figure 3 The specific structure of MethylAI model includes:

[0131] Multi-scale Conv: Multi-scale convolution kernel, which gives MethylAI the ability to recognize DNA feature sequences of different lengths.

[0132] Species specific BatchNorm: Species specific batch normalization layer, which calls species specific batch normalization layer according to the species of DNA methylation data input into the model, solves the problem of data heterogeneity in multi-species training, and the loss value can be stably decreased when using multiple species data to train the model.

[0133] GELU, ELU, Sigmoid: All belong to activation function, which can introduce non-linear relationship into the model, so that it can learn the complex relationship between DNA sequence and DNA methylation level.

[0134] Conv: Convolution layer, MethylAI uses convolution layer with stride of 4 at certain positions, which optimizes the operation efficiency of convolution layer, makes MethylAI can efficiently compress DNA sequence information, saves GPU memory resources, and can read longer DNA sequence under the same hardware.

[0135] Multi-scale CNN block: Multi-scale CNN block, composed of Conv, Species specific BatchNorm, GELU, and residual connection (indicated by the "plus" symbol with an outer ring) is introduced to add the input of the module to the output, thereby providing better gradient flow during network training, effectively preventing gradient vanishing during model training and preventing model degradation. Repeat stacking 12 times to form a DNA sequence feature extraction module.

[0136] Multi-scale Feature Storage: Multi-scale Feature Storage, used to store the feature map obtained by the intermediate operation of Mult-scale CNN block, so that MethylAI can extract and store DNA features of different scales to improve prediction accuracy.

[0137] Adaptive Average Pooling: Global average pooling layer, compresses DNA sequence feature dimension information.

[0138] Multi-tissue Prediction Head: Multi-tissue prediction head, which can simultaneously predict the DNA methylation level of the input DNA sequence in multiple tissues. The multi-tissue prediction head uses 2 linear layers (Linear Layer), ELU activation function, and Sigmoid activation function. The ELU (Exponential Linear Unit) activation function introduces nonlinearity while avoiding the problem of neuron death caused by the ReLU activation function, improving the model's ability to learn complex patterns; the sigmoid function compresses the output to between 0 and 1, consistent with the range of DNA methylation levels. The prediction head can convert the DNA features obtained by the multi-scale CNN block into the final tissue-specific DNA methylation level. Through the above design, especially the application of activation function, the model can capture the tissue specificity of DNA methylation, and then predict the DNA methylation level of multiple tissues in the human body.

[0139] The training scheme of MethylAI model is as follows:

[0140] Read in the multi-species training set data, input the DNA sequence on both sides of the CpG (use one-hot encoding, input DNA length is 18.4kb) into MethylAI, and the output is the DNA methylation level of the CpG site.

[0141] The MethylAI model is trained using the gradient descent method, the mean square error (MSE) is used as the loss function, the AdamW optimizer is used to optimize the model parameters, the learning rate is initialized to 0.001, the batch size is set to 320, and the cosine annealing with warm restarts program is used to adjust the learning rate.

[0142] After each training round (epoch), the accuracy of the MethylAI model and the reasonableness of the hyperparameter settings are evaluated using the validation set segmented in step S1. If it is not reasonable, the model will be further optimized, and the model will be retrained in the previous step.

[0143] After the multi-species data training is completed, the best MethylAI model with the highest accuracy is selected, and the MethylAI model is fine-tuned using the human data set. The learning rate is reduced to 0.0001 during fine-tuning, and the remaining settings remain unchanged.

[0144] After each training round (epoch), the accuracy of the MethylAI model and the reasonableness of the hyperparameter settings are evaluated using the validation set segmented in step S1. If it is not reasonable, the model will be further optimized, and the model will be retrained in the previous step. Finally, the prediction accuracy of the MethylAI model is evaluated using the independent test set segmented in step S1.

[0145] The prediction accuracy of the model is measured by calculating the following parameters: mean square error (MSE) of the predicted value and the true value, Pearson's correlation coefficient, and Spearman's correlation coefficient.

[0146] S3: Application of MethylAI in disease research. Combined with DeepSHAP algorithm, develop the explainability analysis process of MethylAI, reveal the important regulatory sequences of DNA methylation; develop the mQTL effect prediction analysis process based on MethylAI, use the public mQTL data set to verify the prediction accuracy of MethylAI for mQTL effect, and apply it to the mQTL effect prediction analysis of multiple disease risk genetic loci. The specific steps are as follows:

[0147] Apply DeepSHAP algorithm to MethylAI model, calculate the attribution score of each base in input DNA sequence, combine JASPAR database to obtain regulatory sequences on DNA, and filter important regulatory sequences by attribution score.

[0148] There are a large number of DNA variants on the human genome, and a part of the DNA variants, which are called mQTL, can affect DNA methylation. The present application collects mQTL data sets from public databases, and uses the data sets as a standard to verify the accuracy of MethylAI in mQTL effect prediction.

[0149] A large number of DNA variants on the human genome are related to the occurrence of diseases, which are called disease risk genetic loci, but whether most of the disease risk genetic loci have mQTL effects is still unclear. The present application uses MethylAI to systematically predict the mQTL effects of disease risk genetic loci, which can provide help for the research of disease mechanism, disease biomarkers and drug development, etc.

[0150] The scene application effect of the technical solution of the embodiment of the present application will be described below in combination with specific scene examples (only for example description, and cannot be regarded as a limitation of the present application):

[0151] Embodiment one:

[0152] The present embodiment provides detailed description of the specific implementation of the DNA methylation prediction method (MethylAI model) based on deep learning as shown in S1-S2 of Figure 4

[0153] S1: Data collection and processing. The present application collects DNA methylation data of 1574 human samples from public databases, covering 51 human tissues and 171 cell types. In addition, the present application collects DNA methylation data of 11 mammalian species (431 samples in total), including mice, non-human primates (Macaca fasicularis / Macaca mulatta), anthropoid apes (Gorilla gorilla / Chimpanzee), and domestic animals (cattle / sheep / dog). The detailed steps are as follows:

[0154] 1. Download DNA methylation data based on WGBS-seq technology from public databases, including: ENCODE (https: / / www.encodeproject.org / ), Methbank (https: / / ngdc.cncb.ac.cn / methbank / ), IHEC (https: / / ihec-epigenomes.org / ), NCBI GEO (https: / / www.ncbi.nlm.nih.gov / geo / ).

[0155] 2. Data cleaning: remove low-quality sequences, remove ambiguous bases such as N bases, and remove CpG sites with poor data quality.​

[0156] 3. Use the BSsmooth algorithm to smooth the DNA methylation data, and set the parameters ns=35 and h=500 in the BSsmooth algorithm. This can effectively reduce the noise of the original DNA methylation data, increase the signal-to-noise ratio of the data, and be more conducive to the training of the model.

[0157] 4. Data segmentation: The data set is divided into training set, validation set and test set in the ratio of 80%:10%:10%. The training set is used for the training of the MethylAI model, the validation set is used to evaluate whether the hyperparameter design of the model is reasonable, and the test set is used to evaluate the generalization ability of the model.

[0158] 5. One-hot encoding of DNA sequences on both sides of CpG: Each base is represented by a four-dimensional vector (A: [1, 0, 0, 0], T: [0, 1, 0, 0], G: [0, 0, 1, 0], C: [0, 0, 0, 1]). The one-hot encoded DNA sequence will be directly input into the model.

[0159] S2: Development and training of the model. The present application designs and develops a deep learning model MethylAI based on convolutional neural network (CNN), which can input the DNA sequence on both sides of CpG in human genome and output the DNA methylation level of the CpG site in different tissues of human. As shown in Figure 3 , the specific structure and parameters of the MethylAI model are as follows:

[0160] 1. Multi-scale Conv: One-dimensional convolution kernels with lengths of 3, 9 and 21 are used respectively.

[0161] 2. Species specific BatchNorm: Cooperate with the collection of human and 11 other mammalian species, set the number of species to 12.

[0162] 3. Conv(Stride=4): Use a one-dimensional convolution kernel with a length of 9 and a stride (Stride) of 4.

[0163] 4. Conv(Stride=1): Use a one-dimensional convolution kernel with a length of 9 and a stride (Stride) of 1.

[0164] 5. Multi-scale Feature Storage: Store the lengths of 9x2 2 , 9x2 4 , 9x2 6 , 9x2 8 , 9x2 10 , 9x2 11DNA sequence features.

[0165] 6. Adaptive Average Pooling: the output dimension is set to 1.

[0166] 7. Linear Layer: two linear layers are 2700x8780 and 8780x8780, respectively.

[0167] The process of model training is as follows:

[0168] 1. Pre-training MethylAI using a multi-species dataset (human dataset + 11 mammal datasets), during the training process, using mean square error (MSE) as the loss function, using AdamW optimizer to optimize the model parameters, the learning rate (learning rate) is initialized to 0.001, the weight decay is set to 0.01, the batch size is set to 320, and the learning rate is adjusted using the cosine annealing with warm restarts procedure. Train MethylAI for 3 epochs in the multi-species dataset, and after each epoch, use the validation set segmented in step S1 to verify the prediction accuracy of the model.

[0169] 2. Fine-tune all parameters of MethylAI using the human dataset, set the species-specific batch normalization layer of MethylAI to only use the layer adapted to human data during fine-tuning. During the fine-tuning process, use mean square error (MSE) as the loss function, use AdamW optimizer to optimize the model parameters, the learning rate (learning rate) is initialized to 0.0001, the batch size, the learning rate scheduler, the weight decay, etc. Settings remain unchanged. Fine-tune MethylAI for 3 epochs in the human dataset, and after each epoch, use the validation set segmented in step S1 to verify the prediction accuracy of the model. Finally, use the test set segmented in step S1 to verify the final prediction accuracy of the model.

[0170] After training, compare the prediction accuracy of MethylAI with INTERACT model, the detailed steps are as follows:

[0171] 1. Since the INTERACT model does not support multi-species training, this step is skipped. According to the method provided by the INTERACT model in its paper, train it using the human dataset collected in this application, after the training is completed, use the test set segmented in step S1 to verify the final prediction accuracy of the model.

[0172] 2. Calculate the Pearson Correlation Coefficient (PCC) and Spearman Correlation Coefficient (SCC) of the predicted values and true values of INTERACT and MethylAI on the test set segmented in step S1, respectively. The results show that the prediction accuracy of MethylAI (PCC=0.89, SCC=0.78) is significantly higher than that of INTERACT (PCC=0.81, SCC=0.67).

[0173] Embodiment Two:

[0174] The present embodiment provides a specific implementation of training MethylAI model under newly collected or newly generated human DNA methylation data. Based on the situation that new DNA methylation data is continuously generated by scientific research, the present embodiment solves the demand for analyzing newly generated DNA methylation data using MethylAI. The detailed process is as follows:

[0175] 1. Data segmentation: Divide the new data set into training set, validation set and test set in the ratio of 80:10:10. The training set is used for fine-tuning of the MethylAI model, the validation set is used to evaluate whether the hyperparameter design of the model is reasonable, and the test set is used to evaluate the generalization ability of the model.

[0176] 2. Initialize the MethylAI model according to the following parameter settings: Multi-scale Conv: use one-dimensional convolution kernels with lengths of 3, 9, and 21, respectively; Species specific BatchNorm: set the number of species to 1; Conv(Stride=4): use a one-dimensional convolution kernel with a length of 9 and a stride (Stride) of 4; Conv(Stride=1): use a one-dimensional convolution kernel with a length of 9 and a stride of 1; Multi-scale Feature Storage: store DNA sequence features with lengths of 9x2 2 , 9x2 4 , 9x2 6 , 9x2 8 , 9x2 10 , 9x2 11 respectively; Adaptive Average Pooling: set the output dimension to 1; LinearLayer: two linear layers are 2700x8780 and 8780xN respectively, N is the number of new data.

[0177] 3. Load the weights of the Multi-scale CNN block of the multi-species pre-trained MethylAI, which are obtained by Embodiment One and can be reused after being obtained.

[0178] 4. Fine-tune the model using new data, during the fine-tuning process, use mean squared error (MSE) as the loss function, use AdamW optimizer to optimize the model parameters, the learning rate (learning rate) is set as follows, the learning rate of Multi-scale CNN block is set to 0.0001, the learning rate of Multi-tissue Prediction Head is set to 0.001; the weight decay is set to 0.01, the batch size is set to 320, and the learning rate is adjusted using the cosine annealing with warm restarts procedure. Fine-tune MethylAI for 3 epochs in the new data set, and verify the model prediction accuracy using the validation set of step 1 of the process after each epoch. Finally, verify the final prediction accuracy of the model using the test set of step 1 of the process.

[0179] Embodiment Three:

[0180] The present embodiment provides detailed steps for the MethylAI model to predict the mQTL data set, and provides prediction accuracy:

[0181] 1. According to embodiment one or embodiment two, obtain the trained MethylAI model.

[0182] 2. Apply the DeepSHAP algorithm to the MethylAI model to calculate the attribution score of each base of the input DNA sequence, obtain the regulatory sequence on the DNA from the JASPAR database (https: / / jaspar.elixir.no / ), and filter the important regulatory sequences by the size of the attribution score.

[0183] 3. Obtain the published mQTL data set from the GTEx database (https: / / www.gtexportal.org / home / ), first retain significant mQTL data according to the P value (P < 10 -10 ), then combine the important regulatory sequences obtained in step 1 to filter the mQTL variants located in the important regulatory sequences.

[0184] 4. Substitute the filtered mQTL variants into the human reference genome to generate the reference DNA sequence (ref) and the variant DNA sequence (alt), after one-hot encoding, input the MethylAI model to calculate the effect of the mQTL variant on DNA methylation (mQTL effect), if the effect is consistent with the slope direction in the mQTL data set, it means that the mQTL effect prediction is accurate, if the direction is not consistent, it means that the prediction is wrong.

[0185] 5. After the above steps, the accuracy of mQTL effect is 94.1% using MethylAI.

[0186] Embodiment Four:

[0187] The present embodiment provides the application of MethylAI model in disease risk variant analysis, and the specific implementation steps are as follows:

[0188] 1. According to the first embodiment or the second embodiment, a trained MethylAI model is obtained.

[0189] 2. The DeepSHAP algorithm is applied to the MethylAI model to calculate the attribution score of each base of the input DNA sequence, and the JASPAR database (https: / / jaspar.elixir.no / ) is used to obtain the regulatory sequence on the DNA, and the important regulatory sequence is screened according to the attribution score.

[0190] 3. The information of disease risk variants is obtained from public databases, and the commonly used databases include ClinVar (https: / / www.ncbi.nlm.nih.gov / clinvar / ), UKBB (https: / / www.ukbiobank.ac.uk / ), GWAS Catalog (https: / / www.ebi.ac.uk / gwas / ), and CAUSALdb (http: / / mulinlab.tmu.edu.cn / causaldb / index.html).

[0191] The disease risk variants related to the regulatory sequence found in step 2 are retained, and these disease risk variants are substituted into the human reference genome to generate a reference DNA sequence (ref) and a variant DNA sequence (alt). After the ref sequence and the alt sequence are input into the MethylAI model after one-hot encoding, the mQTL effect of the disease risk variant is calculated, and it is determined whether the disease risk variant can participate in the regulation of DNA methylation.

[0192] In summary, compared with the prior art, the present application at least includes the following beneficial effects:

[0193] 1. The MethylAI model can read DNA sequence with a length of 18.4 kb: In view of the problem that the previous model can read DNA sequence with a short length, the multi-scale CNN block designed in MethylAI can efficiently compress the DNA sequence information (can be compressed by 4096 times), so that the model can easily process DNA sequence with a length of 18.4 kb.

[0194] 2. The MethylAI model can identify DNA sequence features of different lengths: In view of the feature sequence of different lengths existing on the DNA sequence, the present application uses multi-scale convolution (Multi-scale Conv) in the first layer of the MethylAI model, and designs a multi-scale feature storage in the middle of the model, so that the MethylAI model has the ability to identify DNA sequence features of different lengths.

[0195] 3. The MethylAI model can accurately predict tissue-specific DNA methylation levels: The DNA methylation pattern in different tissue cells has specificity, and the previously developed model can only predict the DNA methylation level of a single tissue. The present application collects and processes data of different tissue cells of human for model training; at the same time, the prediction head of MethylAI uses the structure of linear layer-ELU activation function-linear layer-Sigmoid activation function, which makes the MethylAI model capable of identifying the tissue specificity of DNA methylation; during the model training process, the present application uses resampling technology for tissue-specific methylation sites, so that the model can fully learn the DNA sequence features of the tissue-specific methylation sites.

[0196] 4. The MethylAI model can accurately predict the mQTL effect of DNA methylation regulatory sequence and be applied to the study of disease risk sites: There are a large number of DNA variations in the human genome, only a small part of which has been determined for mQTL effect, and most of the DNA variations related to the risk of disease occurrence have unknown mQTL effect, and the role of DNA methylation in these diseases is unclear. The published mQTL dataset is used to evaluate MethylAI, which can accurately predict the mQTL effect of DNA methylation regulatory sequence (accuracy up to 94%), and can reveal the role of DNA methylation in these diseases through MethylAI.

[0197] As shown in Figure 5 The present application further provides a DNA methylation prediction device 900, which can implement the above method, and the device comprises:

[0198] A data processing module 901 is configured to obtain DNA methylation data of multiple species, and preprocess to obtain a data sample set;

[0199] The sample data of the data sample set comprises a DNA sequence of a target site and a DNA methylation level;

[0200] An iterative optimization module 902 is configured to perform iterative optimization on a preset deep learning model using the data sample set to obtain an initial prediction model;

[0201] The parameter fine-tuning module 903 is configured to fine-tune the initial prediction model by using the data sample set of the target species to obtain a target prediction model.

[0202] The model application module 904 is configured to perform DNA methylation prediction on the target species by using the target prediction model.

[0203] In some embodiments, the apparatus further comprises a model configuration module configured to perform the following operations:

[0204] The deep learning model is configured according to the sequentially connected multi-scale convolution kernel, the species-specific batch normalization layer, the GELU activation function, the DNA sequence feature extraction module, the multi-scale feature storage, and the multi-tissue prediction head.

[0205] The species-specific batch normalization layer comprises a batch normalization layer corresponding to each species, the DNA sequence feature extraction module comprises a multi-scale CNN block stacked for a preset number of times, the multi-scale feature storage is configured to store feature maps output by each multi-scale CNN block, and the multi-tissue prediction head is configured to predict and output a methylation level according to the feature maps.

[0206] In some embodiments, the apparatus further comprises a configuration module of the DNA sequence feature extraction module configured to perform the following operations:

[0207] The multi-scale CNN block is configured according to the sequentially connected first convolutional layer, the species-specific batch normalization layer, the GELU activation function, the second convolutional layer, the species-specific batch normalization layer, and the GELU activation function.

[0208] The input data of the first convolutional layer and the output data of the second species-specific batch normalization layer in the multi-scale CNN block are connected by a residual connection as input data of the second GELU activation function in the multi-scale CNN block.

[0209] The multi-scale CNN block is stacked for a preset number of times to configure the DNA sequence feature extraction module.

[0210] In some embodiments, the apparatus further comprises a configuration module of the multi-tissue prediction head configured to perform the following operations:

[0211] The multi-tissue prediction head is configured according to the sequentially connected global average pooling layer, the first linear layer, the ELU activation function, the second linear layer, and the Sigmoid activation function.

[0212] It can be understood that the contents in the above method embodiments are applicable to the apparatus embodiments, the apparatus embodiments specifically implement the functions of the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0213] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0214] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0215] like Figure 6 As shown, Figure 6 The hardware structure of an electronic device 1000 according to another embodiment is illustrated. The electronic device 1000 includes:

[0216] The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (aSIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0217] The memory 1002 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RaM). The memory 1002 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001.

[0218] Input / output interface 1003 is used to implement information input and output;

[0219] The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0220] Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004);

[0221] The processor 1001, the memory 1002, the input / output interface 1003 and the communication interface 1004 are communicatively connected with each other through the bus 1005.

[0222] The electronic device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0223] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the method described above.

[0224] It can be understood that the contents in the above method embodiments are all applicable to the present storage medium embodiment, the function realized by the present storage medium embodiment is the same as that of the above method embodiments, and the beneficial effects achieved by the present storage medium embodiment are also the same as those achieved by the above method embodiments.

[0225] The embodiment of the present application further provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to realize the method described above.

[0226] It can be understood that the contents in the above method embodiments are all applicable to the present program product embodiment, the function realized by the present program product embodiment is the same as that of the above method embodiments, and the beneficial effects achieved by the present program product embodiment are also the same as those achieved by the above method embodiments.

[0227] The memory, as a non-transitory computer readable storage medium, can be used to store non-transitory software programs and non-transitory computer executable programs. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory disposed remotely with respect to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0228] The DNA methylation prediction method, device, electronic equipment, storage medium and program product provided by the embodiments of the present application obtain DNA methylation data of multiple species, and obtain a data sample set by preprocessing; wherein, the sample data of the data sample set comprises a DNA sequence of a target site and a DNA methylation level; the data sample set is used to iteratively optimize a preset deep learning model to obtain an initial prediction model; the data sample set of the target species is used to fine-tune the parameters of the initial prediction model to obtain a target prediction model; and the target prediction model is used to predict DNA methylation of the target species. The present application integrates multiple species of public methylation data through a multi-species pre-training mechanism, constructs an initial model with generalization ability, solves the core pain point of insufficient single-species entity organ sample, makes the target species (such as a specific tissue of a human) only need a small amount of sample to fine-tune a high-precision model, and significantly reduces the dependence on large-scale entity tissue data. Moreover, the present application uses a two-stage optimization strategy (pre-training + fine-tuning): the initial model provides universal representation ability, and the target species fine-tuning focuses on learning tissue-specific patterns, thereby providing a reliable basis for disease-related mQTL causal inference.

[0229] The embodiments described in the embodiments of the present application are used to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that, with the evolution of technology and the appearance of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0230] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and can include more or fewer steps than those shown in the figures, or combine certain steps or different steps.

[0231] The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separate, that is, can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0232] Those skilled in the art can understand that all or some of the steps in the above disclosed method, the functions of the modules / units in the system and the device can be implemented as software, firmware, hardware and their appropriate combinations.

[0233] The terms "first", "second", "third", "fourth", and the like in the description and in the claims of this application, if any, are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so termed is interchangeable under appropriate circumstances such that the embodiments of the application described herein are, for example, capable of orderly or chronological mundane operation, reverse order operation, based on circuitry availability, based on stated preference or the like, and that "default" or other orderings are thus permissible. Further, the terms "comprise", "comprising", "include", "including", and the like, are specifically intended to be open-ended. That is, references to individual steps and the like do not suhstantially exclude the presence of two or more of a recited step or its integral sub-steps or additional steps whether or not readily ascertainable from the description or the like. Further, the words "a" or "an", as used herein in the disclosure and elsewhere, are used indiscriminately and are to be interpreted in the same way, i.e. as meaning "one or more".

[0234] It should be understood that, in this application, "at least one" means one or more, "multiple" means two or more. "And / or", used to describe the relationship between associated objects, means that there can be three relationships, for example, "A and / or B" can mean: only A, only B, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects. "At least one of the following" or the like means any combination of these items, including single or multiple combinations. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0235] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the above units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed objects can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0236] The units described above as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. they can be located in one place or distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0237] In addition, each function unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software function unit.

[0238] If the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in part, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes multiple instructions used to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various other media that can store programs.

[0239] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, and are not limited to the scope of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. A DNA methylation prediction method, characterized by, The method comprises the following steps: Obtaining DNA methylation data of multiple species, and preprocessing to obtain a data sample set; The sample data of the data sample set comprises a DNA sequence of a target site and a DNA methylation level; Using the data sample set to iteratively optimize a preset deep learning model to obtain an initial prediction model; Using the data sample set of the target species to fine-tune the parameters of the initial prediction model to obtain a target prediction model; Using the target prediction model to predict DNA methylation of the target species.

2. The method of claim 1, wherein, The preprocessing to obtain the data sample set comprises the following steps: Data preprocessing is performed on the DNA methylation data; The data preprocessing comprises data cleaning and smoothing processing; According to the results of the data preprocessing, the data sample set is arranged; The data sample set comprises a training set, a validation set and a test set divided based on a preset ratio.

3. The method of claim 1, wherein, The method further comprises the following steps: The deep learning model is configured according to a plurality of scale convolution kernels connected in sequence, a species-specific batch normalization layer, a GELU activation function, a DNA sequence feature extraction module, a multi-scale feature storage and a multi-tissue prediction head; The species-specific batch normalization layer comprises a batch normalization layer corresponding to each species, the DNA sequence feature extraction module comprises a plurality of scale CNN blocks stacked repeatedly for a preset number of times, the multi-scale feature storage is used to store feature maps output by each of the plurality of scale CNN blocks, and the multi-tissue prediction head is used to predict and output a predicted methylation level according to the feature maps.

4. The method of claim 3, wherein, The method further comprises the following steps: The plurality of scale CNN blocks are configured according to a first convolutional layer, the species-specific batch normalization layer, the GELU activation function, a second convolutional layer, the species-specific batch normalization layer and the GELU activation function connected in sequence; The input data of the first convolutional layer and the output data of the second species-specific batch normalization layer in the second plurality of scale CNN blocks are connected in a residual manner as the input data of the second GELU activation function in the second plurality of scale CNN blocks; The plurality of scale CNN blocks are stacked repeatedly for the preset number of times to configure the DNA sequence feature extraction module.

5. The method of claim 3, wherein, The method further comprises the following steps: The multi-tissue prediction head is configured according to a global average pooling layer, a first linear layer, an ELU activation function, a second linear layer and a Sigmoid activation function connected in sequence.

6. The method according to any one of claims 1 to 5, characterized in that, The data sample set comprises a training set and a validation set, and the iteratively optimizing a preset deep learning model using the data sample set to obtain an initial prediction model comprises the following steps: The DNA sequence of the training set is one-hot encoded; The one-hot encoded DNA sequence is input as input data into the deep learning model, so that the deep learning model outputs a first predicted methylation level; constructing mean square error as a loss function according to the first predicted methylation level and the DNA methylation level, and adjusting model parameters of the deep learning model based on the loss function by using a preset optimizer; performing accuracy evaluation and hyperparameter adjustment on the deep learning model by using the validation set, returning to the step of inputting the DNA sequence after one-hot encoding as input data into the deep learning model until a preset training condition is reached, and obtaining the initial prediction model; In the deep learning model, a species-specific batch normalization layer is configured, and the species-specific batch normalization layer includes a batch normalization layer corresponding to each species. The deep learning model calls the batch normalization layer corresponding to the species according to the DNA sequence of the input data.

7. The method according to any one of claims 1 to 5, characterized in that, The data sample set includes a training set, a validation set, and a test set. The initial prediction model is fine-tuned by using the data sample set of the target species to obtain a target prediction model, including the following steps: one-hot encoding the DNA sequence of the training set of the target species; inputting the DNA sequence after one-hot encoding as input data into the initial prediction model to make the initial prediction model output a second predicted methylation level; constructing mean square error as a loss function according to the second predicted methylation level and the DNA methylation level, and fine-tuning model parameters of the initial prediction model based on the loss function by using a preset optimizer; performing accuracy evaluation and hyperparameter adjustment on the initial prediction model by using the validation set, returning to the step of inputting the DNA sequence after one-hot encoding as input data into the initial prediction model until a preset adjustment condition is reached, and verifying the prediction accuracy of the initial prediction model by using the test set; when the prediction accuracy meets the accuracy requirement, the initial prediction model is used as the target prediction model; when the prediction accuracy does not meet the accuracy requirement, the step of inputting the DNA sequence after one-hot encoding as input data into the initial prediction model is returned until the prediction accuracy meets the accuracy requirement; In the deep learning model, a species-specific batch normalization layer is configured, and the species-specific batch normalization layer includes a batch normalization layer corresponding to each species. In the iterative optimization stage, the deep learning model calls the batch normalization layer corresponding to the species according to the DNA sequence of the input data. In the parameter fine-tuning stage, the initial prediction model only applies the batch normalization layer corresponding to the target species.

8. A DNA methylation prediction device, characterized by comprising: The device comprises: a data processing module configured to obtain DNA methylation data of multiple species, and to preprocess the data to obtain a data sample set; wherein the sample data of the data sample set includes a DNA sequence of a target site and a DNA methylation level; an iterative optimization module configured to perform iterative optimization on a preset deep learning model by using the data sample set to obtain an initial prediction model; and a fine-tuning module configured to fine-tune the initial prediction model by using a data sample set of a target species to obtain a target prediction model. a parameter fine-tuning module, configured to perform parameter fine-tuning on the initial prediction model by using the data sample set of the target species, to obtain a target prediction model; a model application module, configured to perform DNA methylation prediction on the target species by using the target prediction model.

9. An electronic device, comprising: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method in any one of claims 1 to 7 when executing the computer program.

10. A computer storage medium, which stores a processor-executable program, and the processor-executable program, when executed by a processor, is configured to implement the method in any one of claims 1 to 7.

11. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the method in any one of claims 1 to 7.

Images