Method and system for predicting up- and down-regulation of risk gene expression induced by non-coding mutations

Abstract

The application provides a non-coding mutation-induced risk gene expression up-regulation and down-regulation prediction method and system, comprising the following steps: step 1, obtaining a data set and performing pretreatment; step 2, information input standardization between non-coding mutations and gene transcription start points TSS; step 3, non-coding mutation information input standardization; step 4, sequence length alignment and representation learning framework; and step 5, embedded feature merging and prediction output. The application is matched with chromatin accessibility sequencing data to migrate to any tissue or cell type, make accurate and reliable prediction, and utilize an attention mechanism to endow the model with interpretability.

Description

Technical Field

[0001] This invention relates to the fields of computer and DNA sequence technology, and more specifically, to a method and system for predicting the up- and down-regulation of expression of risk genes induced by non-coding mutations. Background Technology

[0002] Understanding the effects of gene expression regulation and genomic variation is crucial for human genetics and precision medicine. Published analyses have primarily focused on protein-altering variations to predict the harmfulness of genetic variations within an individual's genome. However, non-coding variations constitute the vast majority of human genetic variation, and the enormous non-coding mutation space presents significant challenges, particularly in understanding how non-coding DNA affects gene expression in different cell types.

[0003] Understanding how mutations in non-coding DNA affect the regulation of gene expression remains a formidable challenge with profound implications for advancing research in human genetics and disease. Accurately predicting the direction of quantitative trait loci (eQTLs) in gene expression could offer a potential solution to accelerate the identification of phenotype-associated non-coding variants. However, despite the existence of methods for predicting the impact of non-coding mutations on gene expression changes, the current best-performing tool, Enformer, still cannot accurately predict the direction of eQTL effects. Furthermore, tissue-specific limitations necessitate the use of different training models for each specific tissue type in existing methods. This hinders the expansion of predictive capabilities to the single-cell resolution level.

[0004] Predicting the upregulation or downregulation of gene expression by single-point non-coding mutations from DNA sequences using computational methods is extremely challenging. The regulatory range of cis-eQTLs can reach 1,000,000 bp, posing a significant challenge to deep learning-based models. Predicting the direction of gene expression impact from single-point non-coding mutations can be expressed as follows: given a single-point mutation located in a non-coding region of a reference genome and the transcription start site (TSS) of the gene whose expression is affected by the mutation, using chromatin accessibility sequencing data (ATAC-seq) of the predicted tissue or cell type, computationally predict whether the mutation will upregulate or downregulate gene expression. This is a binary classification problem.

[0005] Several computational methods have been developed to predict the impact of non-coding mutations on gene expression, such as DeepSea, Basenji2, Expecto, and Enformer. The most advanced method currently is Enformer, which takes a DNA sequence as input and allows predictions up to 100,000 bp upstream and downstream of the non-coding mutation site. Enformer uses a network architecture built with stacked multi-layer convolutional neural networks and Transformer modules; for tissue-specific predictions, the model needs to be trained specifically for each tissue.

[0006] Existing computational methods can only predict gene expression in regulatory regions within 100,000 bp upstream or downstream of non-coding mutations, while the actual range of cis-regulation can extend to 1,000,000 bp. Furthermore, these methods are trained on specific tissues individually, indicating a lack of generalization ability, as they can only predict for cell types and tissues present in the training data and cannot be transferred to new cell types or tissues.

[0007] Predicting the orientation of quantitative trait loci (eQTLs) can accelerate the identification of associations between non-coding variants and phenotypic traits. However, even the state-of-the-art Enformer model has been reported to struggle with accurately predicting eQTL orientation, particularly when the regulatory range exceeds 1,000 bp. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for predicting the up- and down-regulation of risk gene expression induced by non-coding mutations.

[0009] The method for predicting the up- and down-regulation of risk gene expression induced by non-coding mutations according to the present invention includes:

[0010] Step 1: Obtain the dataset and preprocess it;

[0011] Step 2: Normalization of information input between non-coding mutations and gene transcription start points (TSS);

[0012] Step 3: Normalize the input of non-coding mutation information;

[0013] Step 4: Sequence length alignment and representation learning framework;

[0014] Step 5: Embedded feature merging and prediction output.

[0015] Preferably, step 1 includes:

[0016] Cis-eQTL data for 48 tissue types were organized in the GTEx v8 database. Only eQTL datasets with a causal probability greater than 0.9 after statistical fine mapping were retained. This eQTL dataset was then cross-referenced with 19 overlapping tissue-specific ATAC-seq datasets available in EpiMap, and eQTLs with multiple site substitutions were discarded.

[0017] Sequences between non-coding mutations and TSS were extracted from the hg38 human genome and classified into upregulation and downregulation based on the direction of the effect of eQTL on the quantitative expression of a given gene.

[0018] Downsampling was performed on these two categories. For each eQTL entry, the DNA sequence between non-coding mutations and TSS, along with their corresponding ATAC-seq, was used as input features for deep learning. In cases where multiple ATAC-seq results corresponded to the same tissue, the average was calculated on a base-by-base basis. Based on the different lengths of the extracted DNA sequences, the datasets were classified. The training, validation, and test datasets were randomly allocated in an 8:1:1 ratio.

[0019] Preferably, step 2 includes: obtaining information between non-coding mutations and gene transcription start points as input; extracting DNA sequences from non-coding mutation sites to TSS and ATAC-seq data of the corresponding eQTL tissue; for DNA sequences, using One-Hot encoding of four different bases and splicing them with ATAC-seq in the sequence dimension as one of the input branches of the model.

[0020] Preferably, step 3 includes: obtaining information related to non-coding mutation information as input; extracting the reference genome and its mutated DNA sequence with a length of 51bp centered on the non-coding site; splicing it with the corresponding tissue ATAC-seq as one of the input branches of the model; and then feeding the input branch into a neural network consisting of two convolutional layers for preliminary feature learning.

[0021] Preferably, step 4 includes: padding the sequence data obtained in step 2 with zeros according to the original length of the DNA sequence to input features: sequences in the range of 1bp to 1,000bp are padded to 1,000bp, sequences in the range of 1,000bp to 10,000bp are padded to 10,000bp, sequences in the range of 10,000bp to 100,000bp are padded to 100,000bp, and sequences longer than 100,000bp are padded to 1,000,000bp. After BiGRUs, average pooling is also performed on sequences longer than 1,000bp. For the aligned 1,000-dimensional output, rotational position embedding is used to indicate positional information. Then, a sparse self-attention layer is implemented based on BigBird, which includes global, sliding, and random multi-head self-attention for representation learning. The number of attention heads in each self-attention layer is set to 8.

[0022] Step 5 includes: merging the representation learning embeddings obtained in steps 3 and 4, and feeding them into a network module consisting of three fully connected layers. The final output is then processed by the SoftMax activation function for prediction and classification output.

[0023] The non-coding mutation-induced up- and down-regulation prediction system for risk gene expression provided by the present invention includes:

[0024] Module M1: Acquires the dataset and performs preprocessing;

[0025] Module M2: Normalization of input information between non-coding mutations and gene transcription start points (TSS);

[0026] Module M3: Input normalization of non-coding mutation information;

[0027] Module M4: Sequence length alignment and representation learning framework;

[0028] Module M5: Embedded feature merging and prediction output.

[0029] Preferably, the module M1 includes:

[0030] Cis-eQTL data for 48 tissue types were organized in the GTEx v8 database. Only eQTL datasets with a causal probability greater than 0.9 after statistical fine mapping were retained. This eQTL dataset was then cross-referenced with 19 overlapping tissue-specific ATAC-seq datasets available in EpiMap, and eQTLs with multiple site substitutions were discarded.

[0031] Sequences between non-coding mutations and TSS were extracted from the hg38 human genome and classified into upregulation and downregulation based on the direction of the effect of eQTL on the quantitative expression of a given gene.

[0032] Downsampling was performed on these two categories. For each eQTL entry, the DNA sequence between non-coding mutations and TSS, along with their corresponding ATAC-seq, was used as input features for deep learning. In cases where multiple ATAC-seq results corresponded to the same tissue, the average was calculated on a base-by-base basis. Based on the different lengths of the extracted DNA sequences, the datasets were classified. The training, validation, and test datasets were randomly allocated in an 8:1:1 ratio.

[0033] Preferably, module M2 includes: acquiring information between non-coding mutations and gene transcription start points as input; extracting DNA sequences from non-coding mutation sites to TSS and ATAC-seq data of the corresponding eQTL tissue; for DNA sequences, using One-Hot encoding of four different bases and splicing them with ATAC-seq in the sequence dimension as one of the input branches of the model.

[0034] Preferably, module M3 includes: acquiring information related to non-coding mutation information as input; extracting the reference genome and its mutated DNA sequence with a length of 51bp centered on the non-coding site; splicing it with the corresponding tissue ATAC-seq as one of the input branches of the model; and then feeding the input branch into a neural network consisting of two convolutional layers for preliminary feature learning.

[0035] Preferably, module M4 includes: zero-padding the sequence data obtained by module M2 with input features according to the original length of the DNA sequence: sequences in the range of 1bp to 1,000bp are padded to 1,000bp, sequences in the range of 1,000bp to 10,000bp are padded to 10,000bp, sequences in the range of 10,000bp to 100,000bp are padded to 100,000bp, and sequences longer than 100,000bp are padded to 1,000,000bp. After BiGRUs, average pooling is also performed on sequences longer than 1,000bp. For the aligned 1,000-dimensional output, rotational position embedding is used to indicate positional information. Then, a sparse self-attention layer is implemented based on BigBird, which includes global, sliding, and random multi-head self-attention for representation learning. The number of attention heads in each self-attention layer is set to 8.

[0036] The module M5 includes: the representation learning embedding obtained by merging the modules M3 and M4, and feeding it into a network module consisting of three fully connected layers. The final output is then processed by the SoftMax activation function for prediction and classification output.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] (1) The purpose of this invention is to improve the prediction accuracy of the effect of single-point non-coding mutation on the up and down regulation of gene expression, and to extend the predictable regulation range to 1,000,000 bp upstream and downstream of the non-coding site, so that the prediction is not limited by the tissue or cell type contained in the training set. It can be transferred to any tissue or cell type by matching with chromatin accessibility sequencing data (ATAC-seq) to make accurate and reliable predictions, and to give the model interpretability by using the attention mechanism.

[0039] (2) The embeddings obtained by the present invention were visualized by representation learning and ablation studies were conducted to explore the independent contribution of each model component. In order to demonstrate the fine-tuning and generalization ability of the present invention on new tissues, eQTLs of two new brain tissues that did not exist in the training set were collected for external validation. It was observed that the pre-trained embeddings of the present invention improved the reasonableness of the prediction results compared with end-to-end prediction. In addition, the present invention was evaluated on single-cell eQTLs of six different cell types by transfer learning, and satisfactory results were achieved in all cell types (AUC>0.860). The present invention also achieved accurate prediction of two eQTLs related to mental illness (rs1902660-TSPAN14 and rs4698412-CD38, respectively), indicating its potential application in interpreting the SNP mechanism of disease risk. Attached Figure Description

[0040] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0041] Figure 1 This is a diagram illustrating the internal framework of the present invention, from input data to predicted output. Detailed Implementation

[0042] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0043] Example 1

[0044] This invention provides a method for predicting the up- and down-regulation of risk gene expression induced by non-coding mutations, comprising the following steps:

[0045] Step 1: Obtain and preprocess the dataset.

[0046] Cis-eQTL data from 48 tissue types were compiled in GTEx v8, retaining only those eQTL datasets with a causal probability exceeding 0.9 after statistical fine mapping. This eQTL dataset was then cross-referenced with 19 overlapping tissue-specific ATAC-seq datasets available in EpiMap, discarding eQTLs with multiple substitutions for subsequent prediction. Sequences between non-coding mutations and TSSs were extracted from the hg38 human genome. All eQTLs were categorized into two classes, "upregulated" and "downregulated," based on their direction of influence on the quantitative expression of a given gene. Downsampling was performed on these two tissues due to potential model bias caused by excessive entries in the left ventricle of the heart and the esophageal mucosa. For each eQTL entry, the DNA sequence between the non-coding mutation and the TSS, along with its corresponding ATAC-seq, was used as input features for deep learning. Specifically, in cases where multiple ATAC-seq results corresponded to the same tissue, averaging was performed on a base-by-base basis. Because cis-eQTL has a wide range of influence, extending from within 1,000 bp to far 1,000,000 bp, blindly aligning data integrity may lead to unnecessary computational waste. Therefore, based on different ranges of extracted DNA sequence length, the dataset is divided into four distinct parts: a "small" dataset (1-1,000 bp), a "medium" dataset (1,001-10,000 bp), a "large" dataset (10,001-100,000 bp), and a "huge" dataset (100,001-1,000,000 bp). The training, validation, and test datasets are randomly assigned in an 8:1:1 ratio. We have thus retained 25,609 eQTL entries and encoded the labels as numbers (up: 1, down: 0), which serve as the training and test datasets for this invention.

[0047] Step 2: Normalize the input information between non-coding mutations and TSS.

[0048] This invention requires information from non-coding mutations to gene transcription start sites as input. It extracts DNA sequences from non-coding mutation sites to TSSs and ATAC-seq data of the corresponding eQTL tissues. For the DNA sequences, four different bases are encoded using a one-hot encoding method, and these sequences are concatenated with ATAC-seq data in the sequence dimension as one of the input branches of the model, as shown in the appendix. Figure 1 As shown in (i).

[0049] Step 3: Normalize the input of non-coding mutation information.

[0050] This invention requires information related to non-coding mutations as input. Centered on a single non-coding site and within a 51bp range, the reference genome and its mutated DNA sequence are extracted and concatenated with the corresponding tissue ATAC-seq data. This concatenation serves as one of the input branches of the model. Subsequently, this input branch is fed into a neural network consisting of two convolutional layers for preliminary feature learning, as shown in the attached diagram. Figure 1 As shown in (iv).

[0051] Step 4: Sequence length alignment and representation learning framework.

[0052] Since cis-eQTLs are SNPs within the 1,000,000 bp range of gene transcription start points, aligning all DNA sequences to 1,000,000 bp would result in computational waste for samples with medium to short DNA sequences. Therefore, the sequence data obtained in step 2 was zero-padded with input features according to the original length of the DNA sequences: sequences in the range of 1 bp to 1,000 bp were padded to 1,000 bp, sequences in the range of 1,000 bp to 10,000 bp were padded to 10,000 bp, sequences in the range of 10,000 bp to 100,000 bp were padded to 100,000 bp, and sequences longer than 100,000 bp were padded to 1,000,000 bp. After BiGRUs, average pooling was also performed on sequences longer than 1,000 bp. For the aligned 1,000-dimensional output, a rotational position embedding is used to indicate positional information. Then, a sparse self-attention layer is implemented based on BigBird, incorporating global, sliding, and random multi-head self-attention for representation learning. The number of attention heads in each self-attention layer is set to 8, as shown in the appendix. Figure 1 As shown in (ii) and (iii).

[0053] Step 5: Embedded Feature Merging and Predicted Output

[0054] The representation learning embeddings obtained from steps 3 and 4 are combined and fed into a network module consisting of three fully connected layers. The final output is then processed by the SoftMax activation function for prediction and classification, as shown in the attached diagram. Figure 1 As shown in (v).

[0055] The entire model framework's code implementation, training, and testing all use the Tensorflow2 deep learning framework. The model's main hyperparameters (including the number of network layers, batch size, learning rate, etc.) have been optimized to achieve the best values ​​for the validation set results. The Adam optimizer was used to train the neural network parameters.

[0056] Example 2

[0057] The present invention also provides a system for predicting the up- and down-regulation of non-coding mutation-induced risk gene expression. The system can be implemented by executing the steps of the method for predicting the up- and down-regulation of non-coding mutation-induced risk gene expression. That is, those skilled in the art can understand the method for predicting the up- and down-regulation of non-coding mutation-induced risk gene expression as a preferred embodiment of the system.

[0058] The non-coding mutation-induced risk gene expression up-down regulation prediction system provided by the present invention includes: module M1: acquiring and preprocessing a dataset; module M2: normalizing the input of information between non-coding mutations and gene transcription start points (TSS); module M3: normalizing the input of non-coding mutation information; module M4: sequence length alignment and representation learning framework; and module M5: embedding feature merging and prediction output.

[0059] Module M1 includes: organizing cis-eQTL data for 48 tissue types in the GTEx v8 database, retaining only eQTL datasets with a causal probability greater than 0.9 after statistical fine mapping, then cross-referencing this eQTL dataset with 19 overlapping tissue-specific ATAC-seq datasets available in EpiMap, and discarding eQTLs with multiple site substitutions; extracting sequences between non-coding mutations and TSS from the hg38 human genome, classifying them into upregulation and downregulation based on the direction of the eQTL's influence on the quantitative expression of a given gene; performing downsampling on these two categories, and for each eQTL entry, using the DNA sequence between the non-coding mutation and TSS and its corresponding ATAC-seq as input features for deep learning, averaging the results in base units when multiple ATAC-seq results correspond to the same tissue, classifying the dataset based on the different lengths of the extracted DNA sequences, and randomly allocating the training, validation, and test datasets in an 8:1:1 ratio.

[0060] The module M2 includes: acquiring information from non-coding mutations to gene transcription start points as input, extracting DNA sequences from non-coding mutation sites to TSS and ATAC-seq data of the corresponding eQTL tissues, and for DNA sequences, using one-hot encoding of four different bases and splicing them with ATAC-seq in the sequence dimension as one of the input branches of the model.

[0061] The module M3 includes: acquiring information related to non-coding mutations as input; extracting the reference genome and its mutated DNA sequence centered on the non-coding site with a length of 51 bp; splicing it with the corresponding tissue ATAC-seq as one of the input branches of the model; and then feeding the input branch into a neural network consisting of two convolutional layers for preliminary feature learning.

[0062] The module M4 includes: zero-padding the input features of the sequence data obtained by module M2 according to the original length of the DNA sequence: sequences in the range of 1bp to 1,000bp are padded to 1,000bp, sequences in the range of 1,000bp to 10,000bp are padded to 10,000bp, sequences in the range of 10,000bp to 100,000bp are padded to 100,000bp, and sequences longer than 100,000bp are padded to 1,000,000bp. After BiGRUs, average pooling is also performed on sequences longer than 1,000bp. For the aligned 1,000-dimensional output, rotational position embedding is used to indicate positional information. Then, a sparse self-attention layer is implemented based on BigBird, which includes global, sliding, and random multi-head self-attention for representation learning. The number of attention heads in each self-attention layer is set to 8.

[0063] The module M5 includes: the representation learning embedding obtained by merging the modules M3 and M4, and feeding it into a network module consisting of three fully connected layers. The final output is then processed by the SoftMax activation function for prediction and classification output.

[0064] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0065] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for predicting the up- and down-regulation of risk gene expression induced by non-coding mutations, characterized in that, include: Step 1: Obtain the dataset and preprocess it; Step 2: Normalization of information input between non-coding mutations and gene transcription start points (TSS); Step 3: Normalize the input of non-coding mutation information; Step 4: Sequence length alignment and representation learning framework; Step 5: Embedded feature merging and prediction output; Step 2 includes: obtaining information between non-coding mutations and gene transcription start points as input, extracting DNA sequences from non-coding mutation sites to TSS and ATAC-seq data of the corresponding eQTL tissues, and using one-hot encoding of four different bases for DNA sequences and splicing them with ATAC-seq in the sequence dimension as one of the input branches of the model. Step 3 includes: obtaining information related to non-coding mutation information as input; extracting the reference genome and its mutated DNA sequence with a length of 51bp as the center of the non-coding site; splicing it with the corresponding tissue ATAC-seq as one of the input branches of the model; and then feeding the input branch into a neural network consisting of two convolutional layers for preliminary feature learning. Step 4 includes: padding the sequence data obtained in step 2 with zeros according to the original length of the DNA sequence into input features: sequences in the range of 1bp to 1,000bp are padded to 1,000bp, sequences in the range of 1,000bp to 10,000bp are padded to 10,000bp, sequences in the range of 10,000bp to 100,000bp are padded to 100,000bp, and sequences longer than 100,000bp are padded to 1,000,000bp. After BiGRUs, average pooling is also performed on sequences longer than 1,000bp. For the aligned 1,000-dimensional output, rotational position embedding is used to indicate positional information. Then, a sparse self-attention layer is implemented based on BigBird, which includes global, sliding, and random multi-head self-attention for representation learning. The number of attention heads in each self-attention layer is set to 8. Step 5 includes: merging the representation learning embeddings obtained in steps 3 and 4, and feeding them into a network module consisting of three fully connected layers. The final output is then processed by the SoftMax activation function for prediction and classification output.

2. The method for predicting the up- and down-regulation of risk gene expression induced by non-coding mutations according to claim 1, characterized in that, Step 1 includes: Cis-eQTL data for 48 tissue types were organized in the GTEx v8 database. Only eQTL datasets with a causal probability greater than 0.9 after statistical fine mapping were retained. This eQTL dataset was then cross-referenced with 19 overlapping tissue-specific ATAC-seq datasets available in EpiMap, and eQTLs with multiple site substitutions were discarded. Sequences between non-coding mutations and TSS were extracted from the hg38 human genome and classified into upregulation and downregulation based on the direction of the effect of eQTL on the quantitative expression of a given gene. Downsampling was performed on these two categories. For each eQTL entry, the DNA sequence between non-coding mutations and TSS, along with their corresponding ATAC-seq, was used as input features for deep learning. In cases where multiple ATAC-seq results corresponded to the same tissue, the average was calculated on a base-by-base basis. Based on the different lengths of the extracted DNA sequences, the datasets were classified. The training, validation, and test datasets were randomly allocated in an 8:1:1 ratio.

3. A system for predicting the up- and down-regulation of risk gene expression induced by non-coding mutations, characterized in that, include: Module M1: Acquires the dataset and performs preprocessing; Module M2: Normalization of input information between non-coding mutations and gene transcription start points (TSS); Module M3: Input normalization of non-coding mutation information; Module M4: Sequence length alignment and representation learning framework; Module M5: Embedded feature merging and prediction output; The module M2 includes: acquiring information between non-coding mutations and gene transcription start points as input, extracting DNA sequences from non-coding mutation sites to TSS and ATAC-seq data of the corresponding eQTL tissues, and using one-hot encoding of four different bases for DNA sequences and splicing them with ATAC-seq in the sequence dimension as one of the input branches of the model. The module M3 includes: acquiring information related to non-coding mutations as input, extracting the reference genome and its mutated DNA sequence with a length of 51bp as the center of the non-coding site, and splicing it with the corresponding tissue ATAC-seq as one of the input branches of the model, and then feeding the input branch into a neural network consisting of two convolutional layers for preliminary feature learning. The module M4 includes: zero-padding the input features of the sequence data obtained by module M2 according to the original length of the DNA sequence: sequences in the range of 1bp to 1,000bp are padded to 1,000bp, sequences in the range of 1,000bp to 10,000bp are padded to 10,000bp, sequences in the range of 10,000bp to 100,000bp are padded to 100,000bp, and sequences longer than 100,000bp are padded to 1,000,000bp. After BiGRUs, average pooling is also performed on sequences longer than 1,000bp. For the aligned 1,000-dimensional output, rotational position embedding is used to indicate positional information. Then, a sparse self-attention layer is implemented based on BigBird, which includes global, sliding, and random multi-head self-attention for representation learning. The number of attention heads in each self-attention layer is set to 8. The module M5 includes: the representation learning embedding obtained by merging the modules M3 and M4, and feeding it into a network module consisting of three fully connected layers. The final output is then processed by the SoftMax activation function for prediction and classification output.

4. The non-coding mutation-induced up- and down-regulation prediction system for risk gene expression according to claim 3, characterized in that, The module M1 includes: Cis-eQTL data for 48 tissue types were organized in the GTEx v8 database. Only eQTL datasets with a causal probability greater than 0.9 after statistical fine mapping were retained. This eQTL dataset was then cross-referenced with 19 overlapping tissue-specific ATAC-seq datasets available in EpiMap, and eQTLs with multiple site substitutions were discarded. Sequences between non-coding mutations and TSS were extracted from the hg38 human genome and classified into upregulation and downregulation based on the direction of the effect of eQTL on the quantitative expression of a given gene. Downsampling was performed on these two categories. For each eQTL entry, the DNA sequence between non-coding mutations and TSS, along with their corresponding ATAC-seq, was used as input features for deep learning. In cases where multiple ATAC-seq results corresponded to the same tissue, the average was calculated on a base-by-base basis. Based on the different lengths of the extracted DNA sequences, the datasets were classified. The training, validation, and test datasets were randomly allocated in an 8:1:1 ratio.

Images