Method for predicting special functional non-coding variation of cell line

Abstract

The invention provides a method for predicting specific functional non-coding variation of a cell line, and belongs to the technical field of bioinformatics. Comprising the following steps: constructing an MPRA data set based on an REVA database, the MPRA data set having seven cell lines, each cell line in the MPRA data set having positive and negative samples with a ratio of not less than 1: N1, and each sample having an expression regulation variation site verified by an experiment; annotating and screening common features required by the generation model based on different specific cell lines to obtain a common feature data set; constructing a multi-task residual convolutional neural network model, and training the multi-task residual convolutional neural network model by using the common feature data set; and predicting variation functionality and variation cell lines of unknown variation entries by using a multi-task residual convolutional neural network model. The method is used for solving the problem that the multi-cell line specificity cannot be considered in functional non-coding variation prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for predicting cell line-specific functional non-coding variation. Background Art

[0002] Whole-genome sequencing (WGS) is increasingly used to diagnose patients who carry pathogenic genetic variants. However, the interpretation of these variants is mainly limited to protein-coding regions that account for approximately 2% of the genome, because mutations affecting codons and splice sites are relatively well understood. The vast majority of the human genome consists of non-coding DNA sequences that play a key role in gene regulation and cellular function. Although these sequences do not encode proteins, they are essential for controlling when, where, and how genes are expressed. Recent studies have shown that more than 90% of disease-associated variants identified by genome-wide association studies (GWAS) and other genetic analyses are located in non-coding regions of the genome. With the rapid development of sequencing technology, large-scale experimental data (such as Hi-C, ATAC-seq, ChIP-seq, etc.) and functional validation data (such as MPRA, CRISPR screening) continue to accumulate, providing a rich information resource for the functional prediction of non-coding variants.

[0003] Functional analysis of non-coding variants is one of the core issues in modern genetics and genomics research. The vast majority of the human genome is non-coding, and its variation may affect phenotypes by regulating gene expression and is associated with a variety of complex diseases. However, functional analysis of non-coding variants faces huge challenges, especially how to accurately predict the functional effects of these variants in the context of cell type-specific regulation, which is of great significance for understanding the mechanisms of disease occurrence and development.

[0004] In the existing technology, tools such as Sei-framework, CADD, and Expecto have played an important role in the functional prediction of non-coding variants: Sei-framework is a deep learning-based genome regulatory activity prediction tool that aims to predict and explain the effects of genomic variants on regulatory activity. The Sei-framework model can predict 21,907 chromatin spectra, including transcription factor binding, histone marks, and chromatin accessibility. Sei-framework clusters the 4kb sequence of the whole genome into 40 sequence classes, each representing a specific regulatory activity pattern, such as promoters, enhancers, and tissue-specific regulatory elements. Sei-framework can also quantify the effect of variants on regulatory activity and distinguish between gain-of-function (GoF) and loss-of-function (LoF) variants. The framework plays a certain role in tissue-specific expression prediction, evolutionary constraint analysis, GWAS genetic structure analysis, and disease variation mechanism prediction.

[0005] CADD (Combined Annotation-Dependent Depletion) is a variant harmfulness prediction tool that provides a single score to assess the harmfulness of variants by integrating more than 60 genomic features (such as sequence conservation, gene model annotation, epigenetic measurements, and functional predictions). CADD uses simulated de novo variants and variants fixed in the human population as training data. The former are considered to be potentially harmful variants, and the latter are considered to be neutral or weakly harmful variants. The model is trained using L2 regularized logistic regression, and the original score is converted to a PHRED-like score. CADD can be used to predict the function of non-coding region variants and has achieved certain results.

[0006] ExPecto is a deep learning-based framework designed to predict cell-type-specific gene expression and the effects of variants on expression and disease risk from DNA sequences. The framework converts DNA sequences into regulatory features and predicts gene expression by integrating large-scale epigenetic data (such as chromatin accessibility, histone modifications, and transcription factor binding) using deep convolutional neural networks (CNNs) and linear models. ExPecto predicts the effects of variants in more than 140 million promoter-proximal regions through an in silico mutagenesis approach, helping to prioritize variants that may affect gene expression and disease risk. ExPecto has demonstrated its effectiveness in a number of application scenarios, including prioritization and experimental validation of variants associated with diseases or traits in GWAS.

[0007] MPRA (Massively Parallel Reporter Assays) and CRISPR screening technologies developed in recent years provide strong experimental support for functional studies of non-coding variants. These data can accurately verify the functional role of variants in specific cell lines, but high-throughput functional evaluation directly relying on experimental methods is costly and difficult to achieve large-scale data integration and prediction.

[0008] These methods combine sequence features, evolutionary conservation, and functional annotation information to effectively evaluate the overall functional effects of variants. However, the raw data used by these tools are usually association data, and large-scale experimental data are not fully integrated. Therefore, using association data to train models will lead to the defect of insufficient reliability. In addition, these tools mainly focus on non-specific predictions within the whole genome, lack the ability to predict the function of specific cell types, and have difficulty capturing regulatory differences in different cell types. The lack of cell type-specific prediction ability makes it difficult to accurately identify key variants in complex disease research. Summary of the invention

[0009] In view of this, the present invention provides a method for predicting cell line-specific functional non-coding variations, constructs a prediction model for predicting cell type specificity and functionality, and uses cell lines with experimentally verified expression regulatory variation sites from the REVA database as raw data for model training, which can effectively identify key variations and has the advantage of high prediction reliability.

[0010] The technical solution adopted by the embodiment of the present invention to solve the technical problem is: A method for predicting cell line-specific functional non-coding variants, comprising: Step S1, constructing an MPRA data set based on the REVA database; the MPRA data set includes seven cell lines: GM12878, GM18507, HaCaT, HEK293FT, HEK293T, HepG2, and K562; each cell line in the MPRA data set has positive and negative samples with a ratio of no less than 1:N1, and each cell line has expression regulation variation sites that have been experimentally verified; Step S2, based on different specific cell lines, annotate and screen the common features required for generating the model to obtain a common feature dataset; Step S3, constructing a multi-task residual convolutional neural network model, using the common feature data set to train the multi-task residual convolutional neural network model, the multi-task residual convolutional neural network model is used to predict the functionality of the variation and identify the cell line where the variation is located; Step S4, using the multi-task residual convolutional neural network model to predict the variant functionality of the unknown variant entry and the cell line to which the variant belongs.

[0011] Preferably, the specific implementation of step S1 includes: Step S11, obtaining experimentally verified variation data from the REVA database, and retaining variation entries of cell lines with p-values ​​less than N2 according to the significance level value, and excluding entries with p-values ​​greater than N2; Step S12, excluding mouse cell lines, and obtaining variant entries of seven human genome cell lines as research objects through variant entry screening and extraction; in the process of variant entry screening and extraction, judging whether each cell line meets the two conditions that the amount of data still contains more than N3 entries after filtering and the ratio of positive and negative samples is not less than 1:N1, firstly retaining the data corresponding to the cell lines that directly meet the two conditions, and secondly reducing the number of negative samples to match the positive samples by downsampling adjustment for the data of the cell lines that do not meet the ratio condition, if the data after downsampling exceeds N3 entries, it is also retained, and finally excluding the data of the cell lines that do not meet the two conditions; after screening and extraction, seven human genome cell lines are obtained; Step S13, assigning labels 1 to 7 to each variant entry in the cell line classification task to form the MPRA dataset.

[0012] Preferably, the value of N1 is 3.5, the value of N2 is 0.01, and the value of N3 is 600.

[0013] Preferably, the specific implementation of step S2 includes: Step S21, using Sei-framework to annotate each variant entry in the positive and negative sample datasets of each cell line in the MPRA dataset, that is, generating 21907 initial features for each variant entry; Step S22, screening out common features of seven cell lines and the MPRA dataset to form the common feature dataset, wherein the number of feature dimensions of the common feature dataset is 2192.

[0014] Preferably, the specific implementation of step S3 includes: Step S31, constructing the multi-task residual convolutional neural network model: the multi-task residual convolutional neural network model consists of an input layer, a convolution and pooling module, a residual connection module, a flattening layer, two fully connected layers, two Dropout layers and two output layers connected in sequence; The input layer is used to receive input data of dimension (2192, 1) and input it into the convolution and pooling module; The convolution and pooling module is composed of four feature extraction units in sequence. The feature extraction unit consists of a convolution layer and a maximum pooling layer. The convolution operation is used to extract features, and the pooling operation is used to reduce the data dimension. The residual connection module: after performing convolution transformation and pooling processing on the input data, add the input data to the output of the maximum pooling layer in the fourth feature extraction unit to output multi-dimensional data, wherein the convolution transformation and pooling processing are used to unify the dimension of the input data with the output of the maximum pooling layer in the fourth feature extraction unit; The flattening layer converts the multi-dimensional data into one-dimensional data and inputs them into two fully connected layers respectively; The first fully connected layer: extracts features and applies L2 regularization to prevent overfitting; the extracted feature results are input to the first Dropout layer; The first Dropout layer randomly discards neurons to prevent overfitting, and inputs the processing results into the first output layer; The first output layer performs binary classification according to the processing result of the first Dropout layer to obtain a probability value for determining whether the input variant data is functional, wherein a probability value higher than a threshold indicates functionality; The second fully connected layer: extracts features and applies L2 regularization to prevent overfitting; the extracted feature results are input to the second Dropout layer; The second Dropout layer randomly discards neurons to prevent overfitting, and inputs the processing results into the second output layer; The second output layer performs multi-classification according to the processing result of the second Dropout layer, and outputs a 7-dimensional vector for determining the cell line to which the input data belongs, wherein each element of the 7-dimensional vector is the probability that the variant belongs to each of the 7 cell lines, and the cell line corresponding to the maximum probability is the cell line to which the variant belongs; Step S32, data processing and model input: processing each piece of data in the common feature data set into input data with a dimension of (2192, 1); Step S33, model training and prediction: use the common feature data set data to train the multi-task residual convolutional neural network model, adopt the specified optimizer, and define the loss function and loss weight to optimize the model parameters; during the model training process, grid search is used to determine the optimal hyperparameters.

[0015] Preferably, in the output layer of the multi-task residual convolutional neural network model: The activation function for the binary classification is sigmoid, and the classification threshold is defined as 0.5, where an output value greater than 0.5 is considered to be functional, and an output value less than 0.5 is considered to be non-functional; The activation function for multi-classification is softmax, where the cell line with the highest probability value is the cell line that the input data most likely belongs to.

[0016] Preferably: in the convolution and pooling module of the multi-task residual convolutional neural network model, each one-dimensional convolutional layer sets different numbers of neurons and convolution kernel sizes according to its location, and uses ReLU as an activation function to increase the nonlinearity of the feature; The first Dropout layer uses a `Dense(1, activation='sigmoid')` output layer to output a predicted value between 0 and 1; The second Dropout layer uses a `Dense(7, activation='softmax')` output layer to output a probability distribution vector of length 7; The four feature extraction units of the convolution and pooling module, the pooling size pool_size of each maximum pooling layer is set to 2 at different positions, and the pooling size pool_size of the residual connection module is set to 16, which is used to gradually reduce the data dimension, reduce the amount of calculation, and match the data dimension; The dropout probability of the Dropout layer is controlled by the `dropout_rate` parameter, which is used to randomly drop some neurons to prevent the model from overfitting; The optimal hyperparameter combination is: `dropout_rate` is 0.1, the number of neurons in the first convolutional layer is 64, and each layer doubles in the first four convolutional layers, the convolution kernel size is 3, and the optimizer is Adam.

[0017] It can be seen from the above technical scheme that the method for predicting cell line-specific functional non-coding variations provided in an embodiment of the present invention is used to predict cell type specificity and functionality, and the implementation process is: first, an MPRA data set is constructed based on the REVA database; the MPRA data set contains seven cell lines, namely GM12878, GM18507, HaCaT, HEK293FT, HEK293T, HepG2 and K562, and each cell line in the MPRA data set has positive and negative samples with a ratio of not less than 1:N1, and each cell line has experimentally verified expression regulation variation sites; based on different specific cell lines, the common features required for the generation model are annotated and screened to obtain a common feature data set; a multi-task residual convolutional neural network model is constructed, and the multi-task residual convolutional neural network model is trained using the common feature data set, and the multi-task residual convolutional neural network model is used to predict the functionality of the variation and identify the cell line where the variation is located; the multi-task residual convolutional neural network model is used to predict the variation functionality of the unknown variation entry and the cell line to which the variation belongs. The present invention constructs a prediction model for predicting cell type specificity and functionality, and uses cell lines with experimentally verified expression regulatory variation sites from the REVA database as original data for model training, which can effectively identify key variations. The original data is effective, reliable and experimentally verified. The model trained using the data has the advantage of high prediction reliability, and can play an important promoting role in the study of complex diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A prediction flow chart of a method for predicting cell line-specific functional non-coding variations provided in an embodiment of the present invention.

[0019] Figure 2 A schematic diagram of one-hot encoding of a multi-task learning model provided in an embodiment of the present invention.

[0020] Figure 3a The AUC result of MIRACN in the variant functional prediction task provided by the embodiment of the present invention.

[0021] Figure 3b AUPRC results of MIRACN on the variant functional prediction task provided by the embodiment of the present invention.

[0022] Figure 3c The AUC result of MIRACN on the cell line prediction task provided by the embodiment of the present invention.

[0023] Figure 3d The Top-1 Error and Top-5 Error results of MIRACN in the cell line prediction task provided by the embodiment of the present invention.

[0024] Figure 4a The AUC result of MIRACN on the GTEx dataset provided by the embodiment of the present invention.

[0025] Figure 4b The TPR result of MIRACN on the GTEx dataset provided by the embodiment of the present invention.

[0026] Figure 5a The AUC results of MIRACN on rare variant datasets provided by the embodiments of the present invention.

[0027] Figure 5b The AUPRC results of MIRACN on rare variant datasets provided by the embodiments of the present invention.

[0028] Figure 6a Comparison of normalized scores of the cell line prediction task of MIRACN on a rare variant dataset provided by an embodiment of the present invention.

[0029] Figure 6b Functional scoring of the cell line prediction task of MIRACN on a rare variant dataset provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The technical scheme and technical effects of the present invention are further elaborated in detail below in conjunction with the accompanying drawings of the present invention.

[0031] The present invention provides a method for predicting cell line-specific functional non-coding variations, wherein a multi-task residual convolutional neural network model, named MIRACN, is established to accurately predict the functionality of non-coding variations and their specific regulatory effects in cell lines. The present invention solves the problem that the existing technology cannot simultaneously take into account the specificity of multiple cell lines in the prediction of functional non-coding variations by integrating MPRA (Massively Parallel Reporter Assays) data and multi-task learning strategies. Cell type-specific prediction capabilities are crucial for complex disease research and can help researchers identify key variations more accurately, thereby providing important clues for disease diagnosis and treatment.

[0032] like Figure 1 The following is a flow chart of a method for predicting cell line-specific functional non-coding variants. The specific implementation steps include: Step S1, constructing an MPRA data set based on the REVA database; the MPRA data set includes seven cell lines: GM12878, GM18507, HaCaT, HEK293FT, HEK293T, HepG2, and K562; each cell line in the MPRA data set has positive and negative samples with a ratio of no less than 1:N1, and each cell line has expression regulation variation sites that have been experimentally verified; Step S2, based on different specific cell lines, annotate and screen the common features required for generating the model to obtain a common feature dataset; Step S3, constructing a multi-task residual convolutional neural network model, using the common feature data set to train the multi-task residual convolutional neural network model, the multi-task residual convolutional neural network model is used to predict the functionality of the variation and identify the cell line where the variation is located; it is the probability of functional variation and the probability of belonging to the seven cell lines respectively; Step S4, using the multi-task residual convolutional neural network model to predict the variant functionality of the unknown variant entry and the cell line to which the variant belongs.

[0033] As an implementation method, the specific implementation of step S1 of constructing the MPRA data set based on the REVA database includes: Step S11, obtaining experimentally verified variant data from the REVA database, and retaining variant entries of cell lines with p-values ​​less than N2 according to the significance level value, and excluding entries with p-values ​​greater than N2 (for example, N2 is 0.01); Step S12, excluding mouse cell lines, and obtaining variant entries of seven human genome cell lines as research objects through variant entry screening and extraction; in the process of variant entry screening and extraction, it is determined whether each cell line meets the two conditions that the data volume still contains more than N3 entries (for example, N3 is 600) after filtering and the positive-negative sample ratio is not less than 1:N1 (for example, 1:3.5). First, the data corresponding to the cell lines that directly meet the two conditions are retained, and secondly, the data of the cell lines that do not meet the ratio condition are adjusted by downsampling to reduce the number of negative samples to match the positive samples (that is, the number of positive and negative samples is the same, if the sample size is greater than N3 at this time, it is retained) to retain the variant entries corresponding to the cell lines that meet the two conditions, and finally the data of the cell lines that do not meet the two conditions are excluded; after screening and extraction, seven human genome cell lines are obtained; Step S13, assigning labels 1 to 7 to each variant entry in the cell line classification task to form an MPRA dataset.

[0034] Among them, the massively parallel reporter gene experiment (MPRA) data comes from the REVA database (https: / / reva.gao-lab.org / download.php), which contains more than 11.8 million experimentally verified expression regulatory variation sites in human genomic cell lines. Specifically, we downloaded and used the GRCh38 uniqueness dataset version 1.1.1 from the REVA database for analysis. First, we screened the variation entries of all cell lines based on a strict significance level (p value < 0.01) to ensure the high quality of the data. Subsequently, based on the number and distribution of the screened data, we selected seven available human cell lines as the research objects, and intentionally excluded mouse cell lines to ensure the uniformity of the data and the relevance of the prediction of human-specific variations. During the variation extraction process, if the data volume of a cell line still contains more than 600 entries after filtering and the data distribution is relatively balanced (the ratio of positive and negative samples is not less than 1:3.5), the cell line will be retained directly; if downsampling is required to achieve sample balance (such as the GM12878 cell line), the cell line will also be retained. Otherwise, the cell line will be excluded. For example, for the GM12878 cell line, due to the imbalanced ratio of positive and negative samples, we reduced the number of negative samples to match 1,956 positive samples by downsampling; for the HEK293FT cell line, we reduced the number of negative samples to match 641 positive samples, thus achieving a balanced data set. For cell lines such as HEK293T where the data volume is balanced, the filtered data is directly used without further processing. Finally, we assigned labels 1 to 7 to these seven cell lines in the cell line classification task, and constructed a main data set containing 141,983 variant entries, including 43,487 positive samples and 98,496 negative samples. The REVA database has the advantages of large data volume, many types, and high accuracy. The MPRA data set established by the method of step S1 has the characteristics of large enough data volume, clear negative-positive ratio and compliance with the requirements, which can effectively avoid the defects of small data volume and underfitting.

[0035] As an embodiment, the specific implementation of step S2 of annotating and screening the features used in the present invention based on different specific cell lines and constructing a public feature data set includes: Step S21, using Sei-framework to annotate each variant entry in the positive and negative sample datasets of each cell line in the MPRA dataset, that is, generating 21,907 initial features for each variant entry (equivalent to the MPRA dataset being only a matrix of original data, and after annotation by Sei-framework, 21,907 feature columns will be added to the original for each variant in the MPRA dataset); Step S22, screening out common features of the seven cell lines and the MPRA dataset to form a common feature dataset, the feature dimension number of the common feature dataset is 2192.

[0036] The present invention annotates each variant in the positive and negative sample data sets of each cell line to generate the input features required for the model. The annotation process generates 21,907 features for each variant entry. In order to ensure the consistency and correlation of features among different cell lines, the present invention screened the common features in seven target cell lines (GM12878, GM18507, HaCaT, HEK293FT, HEK293T, HepG2 and K562). By focusing on these shared features, the present invention reduces the feature dimension from 21,907 to 2,192, thereby eliminating the influence of irrelevant features in other cell lines outside the research object. Subsequently, the present invention extracted these 2,192 common features from the annotation data of each cell line, providing a unified feature set for model training and evaluation.

[0037] As an implementation method, refer to Figure 1 As shown, the multi-task residual convolutional neural network model constructed in step S3 of the present invention consists of an input layer, a convolution and pooling module, a residual connection module, a flattening layer, two fully connected layers, two Dropout layers and two output layers connected in sequence: The input layer is used to receive input data of dimension (2192,1) and input it into the convolution and pooling modules; Convolution and pooling module: It is composed of four feature extraction units in sequence. The feature extraction unit consists of a convolution layer and a maximum pooling layer. The convolution operation is used to extract features, and the pooling operation is used to reduce the data dimension. Each one-dimensional convolution layer sets different numbers of neurons and convolution kernel sizes according to its location, and uses ReLU as the activation function to increase the nonlinearity of the feature. The pooling size pool_size of each maximum pooling layer is set to 2 at different locations, and the pooling size pool_size of the residual connection module is set to 16, which is used to gradually reduce the data dimension, reduce the amount of calculation, and match the data dimension. Residual connection module: after convolution transformation and pooling processing are performed on the input data, the input data is added to the output of the maximum pooling layer in the fourth feature extraction unit to output multi-dimensional data, wherein the convolution transformation and pooling processing are used to unify the dimension of the input data with the output of the maximum pooling layer in the fourth feature extraction unit; the residual connection module has a convolution kernel with a kernel size of 1 and a maximum pooling layer with a kernel size of 16; Flattening layer: converts multidimensional data into one-dimensional data and inputs them into two fully connected layers respectively; The first fully connected layer: extracts features and applies L2 regularization to prevent overfitting; the extracted feature results are input to the first Dropout layer; The first Dropout layer randomly discards neurons to prevent overfitting and improve the generalization ability of the model; the processing results are input to the first output layer; The first output layer performs binary classification based on the processing results of the first Dropout layer to obtain a probability value for determining whether the input variant data is functional, where a probability value higher than a threshold indicates functionality; the activation function for binary classification is sigmoid, and the classification threshold is defined as 0.5, where an output value greater than 0.5 indicates that the input data is functional, and an output value less than 0.5 indicates that the input data is not functional; The second fully connected layer: extracts features, applies L2 regularization to prevent overfitting, and improves the generalization ability of the model; the extracted feature results are input into the second Dropout layer; The second Dropout layer randomly discards neurons to prevent overfitting, and the processing results are input to the second output layer; The second output layer performs multi-classification based on the processing results of the second Dropout layer, and outputs a 7-dimensional vector used to determine the cell line to which the input data belongs. Each element of the 7-dimensional vector is the probability that the variant belongs to each of the seven cell lines. The activation function of the multi-classification is softmax, and the cell line with the highest probability value is the cell line to which the input data is most likely to belong. A convolution and pooling module for feature extraction is constructed by alternately stacking one-dimensional convolution layers with a specific convolution kernel size and a maximum pooling layer with a kernel size of 2. The input data is subjected to feature extraction through a one-dimensional convolution layer, and the data dimension is reduced using a maximum pooling layer with a kernel size of 2. At the same time, a residual connection is introduced to add the result of the input data after a specific convolution and pooling operation to the output of the main path to enhance the model's ability to learn features. Finally, the multidimensional data is flattened through a flattening layer and input into two `Dense` layers for full connection operations, and a `Dropout` layer is used to prevent overfitting. Finally, two prediction results are output, corresponding to functional prediction and cell line type prediction, respectively. As an implementation mode, the specific implementation of step S3 using the common feature data set to train the multi-task residual convolutional neural network model includes: Data processing and model input: Process each data in the public feature data set into input data with a dimension of (2192,1); Model training and prediction: The multi-task residual convolutional neural network model was trained using the public feature dataset data, the specified optimizer was used, and the loss function and loss weight were defined to optimize the model parameters; during the model training process, grid search was used to determine the optimal hyperparameters, and finally the dropout_rate was selected as 0.1, the number of neurons in the first convolutional layer was 64, and each layer doubled in the first four convolutional layers, the convolution kernel size was 3, and the optimizer was the parameter combination of Adam.

[0038] During the training process, the model performance is evaluated based on the loss rate on the validation set. The model with the smallest loss rate on the validation set is defined as the best model and saved.

[0039] Furthermore, the embodiment of the present invention provides an MPRA test set evaluation process for constructing a residual convolutional neural network model based on the common feature data sets of seven cell lines and analyzes its effect. In the functional prediction task, the present invention uses binary classification labels; in the cell line identification task, a one-hot encoding method is adopted. For example, the variation belonging to the GM12878 cell line is encoded as [1,0,0,0,0,0,0], and the variation belonging to the K562 cell line is encoded as [0,0,0,0,0,0,1]. Figure 2 shown.

[0040] During the data set preparation process, the present invention divides the MPRA data set into two parts: 10% of the data is randomly selected as an independent test set, which is only used for the final model evaluation and does not participate in any training process; the remaining 90% of the data is used for model training, and the generalization ability of the model is enhanced through five-fold cross-validation (CV). In each cross-validation, a subset of the training data is randomly selected as the validation set. All data are processed in batches with a batch size of 128.

[0041] In order to determine the best hyperparameter combination, the present invention uses the ParameterGrid method to perform a grid search on the hyperparameters and trains a model for each combination. During the model training process, the present invention uses an early stopping callback to prevent overfitting and records the best model under each configuration. It is specifically set as follows: if the validation set loss does not improve for 20 consecutive training cycles, the training is stopped. After the training is completed, the present invention uses AUC (for functional regulatory variation prediction) and accuracy (for cell line classification) as evaluation indicators, and saves the best performing model with this indicator. The present invention uses the GRCh38 / hg38 version of the human genome assembly data in all analyses.

[0042] We evaluate the present invention on the MPRA test set. Figure 3a and Figure 3b As shown in the figure, the experimental results show that the present invention has achieved the highest AUC (Area Under The Receiver Operating Characteristic) and AUPRC (Area Under The Precision-Recall Curve) in the variant functional prediction task compared with the mainstream methods, which are 0.832 and 0.727 respectively. For the cell type prediction task, according to Figure 3c The present invention also saw a higher AUC than Sei-framework in most cell lines. Figure 3d ,The present invention also achieves better results than the random forest model and Sei-framework in Top-1 Error and Top-5 Error.

[0043] First, based on the GTEx dataset, the performance of MIRACN is evaluated: The model was evaluated using the GTEx (Genotype-Tissue Expression) dataset, which contains functional scores calculated by 13 different tools, totaling 1,164 variant entries, each of which only includes chromosome, position, and alternative allele information. In order to extract complete variant details, the dbSNP_GRCh37 dataset was used to match variants based on chromosome and position. Due to the loose matching criteria, the alternative alleles of some variants are inconsistent with the original dataset; the present invention only retains variants whose alternative alleles are consistent with the original data. For entries where no matching information was found, the present invention replaced its alternative allele with the value provided in the GTEx dataset to ensure data consistency. During data processing, one variant was not found in either the dbSNP_GRCh37 dataset or the 1000 Genomes Project dataset, so it was removed. In addition, during the lifterover conversion process, two variants were lost because they were deleted in the new version of the genome assembly. When replacing alternative alleles, if the reference and alternative bases of some variants were the same, resulting in a zero difference, these variants were also removed. In addition, since there are 26 overlapping entries in the GTEx dataset and the MIRACN training set, in order to avoid these entries affecting the test results, the present invention removes all 26 entries. After these adjustments, the number of variant entries in the three standardized datasets is updated to 278 and 526 unique SNPs, respectively. These processing steps ensure the independence of the test set and eliminate the potential impact of the training set on the test results.

[0044] During the tool evaluation process, the present invention used three data sets, namely the largest effect size, the most likely causal relationship (Most Likely Causal, based on p-value), and the best p-value (best p-value) data sets to evaluate the ability of the tool to detect functional regulatory variations. The largest effect size data set is based on the significance of SNPs on gene expression changes (|β|), and preliminarily screens out the SNPs with the strongest association with the target gene. After removing duplicate entries, it contains 429 unique SNPs, representing the most functionally active candidate entries in the GTEx catalog. In addition to the most likely causal relationship data set, the authors of this data set also constructed two data sets to highlight different aspects of SNP functionality. Among them, one data set consists of the 1000 SNPs with the lowest p-values, representing the most statistically significant variations in gene expression associations, and contains 304 unique SNPs after removing duplicate entries; the other data set, which focuses on a wider selection of effect values, consists of the 1000 SNPs with the largest effect values, and contains 552 unique SNPs after removing duplicate entries. Each dataset reveals the functionality of SNPs from different perspectives, providing a multidimensional perspective for the comprehensive evaluation of tool performance based on indicators such as TPR and AUC.

[0045] During the evaluation process, the present invention compared six tools including MIRACN (such as DeepSEA, LINSIGHT, Eigen, etc.). According to the characteristics of different tools, the present invention set corresponding thresholds for each tool. In the evaluation of True Positive Rate (TPR), the present invention significantly outperformed other tools under all criteria (including maximum effect size, most likely causal relationship, and best p value).

[0046] Specifically, the TPR of the present invention exceeds 0.358 under all evaluation criteria and reaches a peak of 0.556 under the best p-value criterion, while the TPR of the best performing mainstream tool GWAVA under the same criterion is only 0.257. In addition, in the evaluation of AUC (area under the receiver operating characteristic curve), the present invention not only shows a higher average level, but also achieves the highest AUC score of 0.599 among all tools and evaluation criteria (see Figure 4a and 4b ). It is noteworthy that the best p-value criterion is particularly suitable for the performance of MIRACN, which achieves the highest scores in both TPR and AUC. This shows that the present invention has excellent ability in identifying regulatory variants, especially when evaluated based on the best p-value criterion, which fully demonstrates its robustness and reliability in eSNV diagnosis.

[0047] Secondly, based on the rare variant dataset, the performance of the MIRACN model is further evaluated: To further validate the performance of the present invention and systematically evaluate the ability of MIRACN in predicting the regulatory potential and cell line specificity of rare variants, the present invention uses an MPRA test dataset containing rare non-coding variants. This dataset specifically evaluates rare non-coding variants identified in a group of undiagnosed rare disease patients.

[0048] After comparing MIRACN with mainstream tools on this test dataset, it was found that MIRACN outperformed other models in many indicators. Specifically, MIRACN's AUROC (area under the receiver operating characteristic curve) and AUPRC (area under the precision-recall curve) scores were 1% higher than the best performing tool, and the AUROC and AUPRC scores were 5% and 3% higher than the worst performing tool, respectively (see Figure 5a and 5b ). In addition, the present invention also calculated the recall rate of MIRACN and mainstream models. In the case of a total of 138 positive variants in the data set, MIRACN identified 69 true positive examples, which is higher than Sei-framework (68) and DVAR (66); while CADD (Combined Annotation Dependent Depletion, variant annotation tool) only identified about 10 positive examples at thresholds of 6 and 10, and Expecto failed to identify any positive examples. These results fully demonstrate the robustness and effectiveness of MIRACN in predicting cell line-specific functional regulatory variants.

[0049] To further illustrate the capabilities of MIRACN, we focus on analyzing six cases that were correctly predicted only by MIRACN and calculate the normalized scores generated by each tool, showing that MIRACN is able to identify and correctly predict variants that were misclassified by other existing tools ( Figure 6a ). At the same time, the present invention also shows the functional scores of these variants in the corresponding cell lines ( Figure 6b). For example, the variant chr6:52175302 was predicted as a positive variant only by MIRACN, with a normalized score of 0.259 (threshold of 0.252), and was classified into the K562 cell line, corresponding to a functional score of 0.824. According to the Human Phenotype Ontology (HPO) terminology, this variant is associated with diseases such as focal segmental glomerulosclerosis and increased circulating serum creatinine concentrations. Studies have shown that renal damage associated with chronic lymphocytic leukemia (CLL) often leads to elevated serum creatinine levels. This renal damage may be caused by direct glomerular damage, tubular obstruction, or other secondary mechanisms, and is more obvious during leukemia progression or treatment. These results show that MIRACN exhibits excellent functionality in rare variant prediction, fully demonstrating its robustness and reliability in dealing with rare mutations, and can effectively identify regulatory variants even in challenging scenarios.

[0050] The advantages and positive effects of the present invention include: the present invention extracts functional regulatory variations of multiple cell lines from the MPRA dataset, and constructs high-quality training sets, validation sets, and test sets through strict screening and annotation. The MIRACN model based on the residual convolutional neural network was developed using a multi-task learning framework, which can not only simultaneously predict the functionality of the variation, but also solve the problem that existing tools cannot accurately identify cell line-specific variation. Experimental results show that the MIRACN model of the present invention can accurately identify functional regulatory variations, especially in the context of multiple cell lines, showing excellent predictive ability, significantly better than mainstream methods, and achieving ideal application effects.

[0051] The present invention constructs a multi-task residual convolutional neural network method (MIRACN) that can predict the function of non-coding variants under cell line-specific conditions. MIRACN improves the prediction accuracy and reliability of the model through strict feature selection and data preprocessing. In addition, MIRACN provides detailed variant functional annotations, including the function of variants in specific cell types, which helps researchers better understand and verify the prediction results and reduces the complexity of subsequent research.

[0052] Based on the screening process introduced in step S1 in the embodiment of the present invention, the present invention selects data of 7 cell lines for model training. These 7 cell lines are also 7 types with many common features. Therefore, using these 7 types as the classification combination, the model output result also has very high accuracy and practicality.

[0053] The present invention realizes high-precision functional analysis of non-coding variants in different cell lines, solves the problem that the prior art cannot take into account the specificity of multiple cell lines at the same time in the prediction of functional non-coding variants, and fills the shortcomings of the prior art in the study of regulatory differences between cell lines. The present invention aims to combine sequence features, functional annotations and large-scale experimental data to design a model method that integrates multi-task learning, while improving the accuracy of functional prediction of non-coding variants, while ensuring the interpretability and biological relevance of the prediction results. In addition, the present invention optimizes the training and testing process of the model by effectively utilizing large-scale experimental data (such as MPRA data), and has significant advantages in large-scale data processing, feature extraction, model design, and practical applications, providing support for the study of the genetic mechanisms of complex diseases and further promoting the application and development of precision medicine.

[0054] According to the embodiments disclosed in the present invention, the present invention also provides an electronic device, a readable storage medium and a computer program product. The electronic device is intended to represent various forms of digital computers, and the electronic device can also represent various forms of mobile devices, such as personal digital processing computing devices. The components shown herein, their connections and relationships, and their functions are only examples and are not intended to limit the implementation of the present invention disclosed herein and / or required.

[0055] The electronic device includes a computing unit, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for the operation of the device can also be stored. The computing unit, ROM, and RAM are connected to each other through a bus. The input / output (I / O) interface is also connected to the bus. Multiple components in the device are connected to the I / O interface, including: input units, such as keyboards, mice, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as disks, optical disks, etc.; and communication units, such as network cards, modems, wireless communication transceivers, etc. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0056] The computing unit can be various general and / or special processing components with processing and computing power. Some examples of computing units include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit performs the various methods and processes described above, such as methods for predicting cell line-specific functional non-coding variations. For example, in some embodiments, the method for predicting cell line-specific functional non-coding variations can be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the computing unit, one or more steps of the method for predicting cell line-specific functional non-coding variations described above can be performed. Alternatively, in other embodiments, the computing unit can be configured to perform methods for predicting cell line-specific functional non-coding variations by any other appropriate means (e.g., by means of firmware).

[0057] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0058] The program code for implementing the method disclosed in the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0059] In the context disclosed by the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0060] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., an LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0061] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0062] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services.

[0063] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in the present disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document is not limited here.

[0064] What is disclosed above is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A method for predicting cell line-specific functional non-coding variation, characterized in that: include: Step S1, constructing an MPRA data set based on the REVA database; the MPRA data set includes seven cell lines: GM12878, GM18507, HaCaT, HEK293FT, HEK293T, HepG2, and K562; each cell line in the MPRA data set has positive and negative samples with a ratio of no less than 1:N1, and each cell line has expression regulation variation sites that have been experimentally verified; Step S2, based on different specific cell lines, annotate and screen the common features required for generating the model to obtain a common feature dataset; Step S3, constructing a multi-task residual convolutional neural network model, using the common feature data set to train the multi-task residual convolutional neural network model, the multi-task residual convolutional neural network model is used to predict the functionality of the variation and identify the cell line where the variation is located; Step S4, using the multi-task residual convolutional neural network model to predict the variant functionality of the unknown variant entry and the cell line to which the variant belongs.

2. The method for predicting cell line-specific functional non-coding variation according to claim 1, characterized in that: The specific implementation of step S1 includes: Step S11, obtaining experimentally verified variation data from the REVA database, and retaining variation entries of cell lines with p-values ​​less than N2 according to the significance level value, and excluding entries with p-values ​​greater than N2; Step S12, excluding mouse cell lines, and obtaining variant entries of seven human genome cell lines as research objects through variant entry screening and extraction; in the process of variant entry screening and extraction, judging whether each cell line meets the two conditions that the amount of data still contains more than N3 entries after filtering and the ratio of positive and negative samples is not less than 1:N1, firstly retaining the data corresponding to the cell lines that directly meet the two conditions, and secondly reducing the number of negative samples to match the positive samples by downsampling adjustment for the data of the cell lines that do not meet the ratio condition, if the data after downsampling exceeds N3 entries, it is also retained, and finally excluding the data of the cell lines that do not meet the two conditions; after screening and extraction, seven human genome cell lines are obtained; Step S13, assigning labels 1 to 7 to each variant entry in the cell line classification task to form the MPRA dataset.

3. The method for predicting cell line-specific functional non-coding variation according to claim 2, characterized in that: The value of N1 is 3.5, the value of N2 is 0.01, and the value of N3 is 600.

4. The method for predicting cell line-specific functional non-coding variation according to claim 2, wherein: The specific implementation of step S2 includes: Step S21, using Sei-framework to annotate each variant entry in the positive and negative sample datasets of each cell line in the MPRA dataset, that is, generating 21907 initial features for each variant entry; Step S22, screening out common features of seven cell lines and the MPRA dataset to form the common feature dataset, wherein the number of feature dimensions of the common feature dataset is 2192.

5. The method for predicting cell line-specific functional non-coding variation according to claim 4, characterized in that: The specific implementation of step S3 includes: Step S31, constructing the multi-task residual convolutional neural network model: the multi-task residual convolutional neural network model consists of an input layer, a convolution and pooling module, a residual connection module, a flattening layer, two fully connected layers, two Dropout layers and two output layers connected in sequence; The input layer is used to receive input data of dimension (2192, 1) and input it into the convolution and pooling module; The convolution and pooling module is composed of four feature extraction units in sequence. The feature extraction unit consists of a convolution layer and a maximum pooling layer. The convolution operation is used to extract features, and the pooling operation is used to reduce the data dimension. The residual connection module: after performing convolution transformation and pooling processing on the input data, add the input data to the output of the maximum pooling layer in the fourth feature extraction unit to output multi-dimensional data, wherein the convolution transformation and pooling processing are used to unify the dimension of the input data with the output of the maximum pooling layer in the fourth feature extraction unit; The flattening layer converts the multi-dimensional data into one-dimensional data and inputs them into two fully connected layers respectively; The first fully connected layer: extracts features and applies L2 regularization to prevent overfitting; the extracted feature results are input to the first Dropout layer; The first Dropout layer randomly discards neurons to prevent overfitting, and inputs the processing results into the first output layer; The first output layer performs binary classification according to the processing result of the first Dropout layer to obtain a probability value for determining whether the input variant data is functional, wherein a probability value higher than a threshold indicates functionality; The second fully connected layer: extracts features and applies L2 regularization to prevent overfitting; the extracted feature results are input to the second Dropout layer; The second Dropout layer randomly discards neurons to prevent overfitting, and inputs the processing results into the second output layer; The second output layer performs multi-classification according to the processing result of the second Dropout layer, and outputs a 7-dimensional vector for determining the cell line to which the input data belongs, wherein each element of the 7-dimensional vector is the probability that the variant belongs to each of the 7 cell lines, and the cell line corresponding to the maximum probability is the cell line to which the variant belongs; Step S32, data processing and model input: processing each piece of data in the common feature data set into input data with a dimension of (2192, 1); Step S33, model training and prediction: use the common feature data set data to train the multi-task residual convolutional neural network model, adopt the specified optimizer, and define the loss function and loss weight to optimize the model parameters; during the model training process, grid search is used to determine the optimal hyperparameters.

6. The method for predicting cell line-specific functional non-coding variation according to claim 5, characterized in that: In the output layer of the multi-task residual convolutional neural network model: The activation function for binary classification is sigmoid, and the classification threshold is defined as 0.5, where an output value greater than 0.5 is considered to be functional, and an output value less than 0.5 is considered to be non-functional; The activation function for multi-classification is softmax, where the cell line with the highest probability value is the cell line that the input data most likely belongs to.

7. The method for predicting cell line-specific functional non-coding variation according to claim 6, characterized in that: In the convolution and pooling module of the multi-task residual convolutional neural network model, each one-dimensional convolutional layer sets different numbers of neurons and convolution kernel sizes according to its location, and uses ReLU as the activation function to increase the nonlinearity of the feature; The first Dropout layer uses a `Dense(1, activation='sigmoid')` output layer to output a predicted value between 0 and 1; The second Dropout layer uses a `Dense(7, activation='softmax')` output layer to output a probability distribution vector of length 7; The four feature extraction units of the convolution and pooling module, the pooling size pool_size of each maximum pooling layer is set to 2 at different positions, and the pooling size pool_size of the residual connection module is set to 16, which is used to gradually reduce the data dimension, reduce the amount of calculation, and make the data dimension match; The dropout probability of the Dropout layer is controlled by the `dropout_rate` parameter, which is used to randomly drop some neurons to prevent the model from overfitting; The optimal hyperparameter combination is: `dropout_rate` is 0.1, the number of neurons in the first convolutional layer is 64, and each layer doubles in the first four convolutional layers, the convolution kernel size is 3, and the optimizer is Adam.

8. An electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method according to any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images