A chromatin accessibility and transcription factor interaction deep learning method

Abstract

The present application belongs to the field of bioinformatics, and particularly relates to a chromatin accessibility and transcription factor interaction deep learning method. First, a gene expression prediction framework based on deep learning is proposed, and a three-dimensional interaction tensor of cell x transcription factor x chromatin region is constructed to simulate cis-regulation; second, a neural network containing a learnable interaction weight matrix is designed, an attention mechanism is used to dynamically model the specific binding of transcription factors and regulatory regions, and a joint loss function is used to simultaneously optimize the prediction accuracy and correlation; finally, gene-specific training and data enhancement strategies are introduced to realize personalized modeling and robust prediction of different gene regulation modes. The present application establishes an interpretable deep learning system that can infer potential transcription factor and chromatin interactions from multi-omics data, providing a new computational tool for analyzing gene regulation mechanisms and screening key regulatory elements.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically involving a deep learning method for chromatin accessibility and transcription factor interaction. This method constructs a learnable weight matrix for chromatin accessibility and transcription factor interaction, integrates three-dimensional tensor features from multi-omics data, and employs gene-specific training and attention mechanisms to achieve high-precision and interpretable gene expression prediction. Background Technology

[0002] Cellular function and state transitions are influenced by multi-level regulatory mechanisms, encompassing complex processes ranging from epigenetic modification, chromatin accessibility, transcription factor binding, to gene expression regulation. In recent years, single-cell multi-omics technologies have enabled the simultaneous measurement of gene expression and chromatin accessibility information within the same cell, providing unprecedented opportunities to elucidate the regulatory networks of cellular state transitions. However, accurately inferring the functional interactions between transcription factors and chromatin accessibility regions from these high-dimensional, sparse multi-omics data remains a significant challenge in computational biology.

[0003] Existing methods for constructing regulatory networks that integrate chromatin accessibility with transcription factor interactions integrate multi-omics data and utilize learnable weight matrices W to model the interactions between transcription factors and chromatin accessibility regions, thereby predicting the expression levels of target genes. The core of these methods lies in using a three-dimensional interaction tensor of cell × transcription factor × chromatin region as input, learning specific binding patterns between transcription factors and chromatin regions through neural networks, and employing gene-specific training strategies to independently construct predictive models for each gene. However, this method has several shortcomings: the model architecture is too simple, containing only a single-layer fully connected network, lacking deep feature extraction capabilities; the training process only optimizes the mean squared error loss function, without considering directly improving the Pearson correlation coefficient, which is of greater interest to biologists; it lacks effective regularization mechanisms and data augmentation strategies, easily leading to overfitting; the evaluation system is incomplete, with the early stopping strategy based only on the validation set mean squared error rather than the correlation index, and it does not systematically monitor performance changes throughout the training process; interpretability is insufficient, although the weight matrix can reflect chromatin accessibility and the strength of transcription factor interactions, it does not integrate advanced interpretability techniques such as attention mechanisms; furthermore, training efficiency is low, as it does not employ modern optimizers and learning rate scheduling strategies, affecting model convergence speed and final performance. Therefore, developing a method capable of deep feature extraction, implementing effective regularization mechanisms and data augmentation strategies, and integrating advanced interpretability techniques such as attention mechanisms is a pressing technical problem to be solved in this field. Summary of the Invention

[0004] This invention proposes a deep learning method for chromatin accessibility and transcription factor interaction, aiming to address the problems of overly simplistic model architectures, lack of effective regularization mechanisms, insufficient data augmentation strategies, and inadequate interpretability in existing technologies, thereby achieving efficient and accurate gene expression prediction. The specific technical solution includes seven steps:

[0005] Step 1, Innovative Model Architecture: Design an attention-enhanced neural network that specifically combines learnable chromatin accessibility with the interaction weight matrix of transcription factors through parameterized modeling; Employ tensor feature fusion, integrating batch normalization, Dropout, and residual connections, and construct a joint loss function of mean squared error and Pearson correlation coefficient;

[0006] Step 2, Multimodal Data Fusion: Transcription factor expression and chromatin accessibility are fused into a three-dimensional matrix of cell × transcription factor × chromatin region. Regulatory regions within a 250kb range upstream and downstream are dynamically screened based on gene location. Multi-omics data are coordinated through adaptive standardization and transfer learning is initialized based on the position frequency matrix motif.

[0007] Step 3, Dynamic Training Strategy: Implement gene-specific training process, use noise injection and feature masking for data augmentation; design adaptive learning rate scheduling, combined with gradient pruning and early stopping mechanism based on Pearson coefficients of the validation set;

[0008] Step 4, Interpretability Modeling: Integrating soft attention mechanisms to focus on key regulatory regions, and visualizing chromatin accessibility and transcription factor regulatory intensity through weight matrices;

[0009] Step 5, Integration Learning Strategy: Develop a cross-gene parameter sharing strategy, implement dual training integration, and improve prediction reliability through dynamic integration decision-making and uncertainty quantification;

[0010] Step 6, Multi-objective optimization: Design a multi-objective function that weights mean squared error and Pearson loss, adopt an adaptive balancing strategy and hierarchical optimization, and introduce a feature importance-aware regularization method;

[0011] Step 7: System Evaluation Framework: Establish a multi-index evaluation system, conduct cross-cell generalization tests and perturbation simulation analysis, and construct an interpretable output system to guide experimental verification.

[0012] A deep learning method for chromatin accessibility and transcription factor interaction is proposed. Step 1 is implemented as follows: An innovative deep learning architecture is constructed, and an attention-enhanced neural network specifically designed for multi-omics data integration is developed. This method uses a three-dimensional tensor structure as its core, deeply cross-fusing transcription factor expression profiles and chromatin accessibility signals to form a high-dimensional feature representation of cell × transcription factor × chromatin region. The core innovation lies in the introduction of a learnable chromatin accessibility and transcription factor interaction weight matrix. This matrix directly models the specific binding strength between each transcription factor and each chromatin region through parameterization, simulating the cis-regulatory mechanism in biology. The network structure adopts a multi-level feature interaction design, fully preserving the complex correlation information between multi-omics data through tensor-level operations rather than simple feature concatenation. To ensure the stability of deep network training, the model integrates batch normalization layers to accelerate convergence, dropout layers to prevent overfitting, and residual connections to promote gradient flow. The specially designed joint loss function organically combines the traditional mean squared error loss with the biologically more relevant Pearson correlation coefficient loss. Through adjustable weight parameters, it achieves the dual optimization goals of prediction accuracy and relevance, making the model training process more in line with the needs of practical applications.

[0013] A deep learning method for chromatin accessibility and transcription factor interaction is presented, with step 2 implemented as follows: Multimodal data fusion and feature engineering innovation are achieved by using deep feature cross-referencing technology to integrate transcription factor expression and chromatin accessibility data in three dimensions. The cellular-level transcription factor expression matrix and chromatin accessibility matrix are multiplied using tensor multiplication to generate an interaction feature matrix of cell × transcription factor × chromatin region. This design overcomes the limitations of traditional two-dimensional feature splicing and better preserves the spatial correlation information between regulatory elements. A genome location-aware mechanism is innovatively introduced, dynamically screening regulatory regions within a 250kb range upstream and downstream of the target gene based on its chromosomal coordinates, start and termination sites, and chain direction information, simulating the spatial distribution characteristics of cis-regulatory elements in biology. An adaptive data standardization strategy is used to coordinate and normalize data from different omics, eliminating batch effects while maximizing the preservation of biological variation signals. Furthermore, a transfer learning approach is employed, using pre-trained transcription factors combined with motif information as prior knowledge to guide the initialization of model parameters, effectively improving the starting point and convergence speed of model training, reflecting the advanced concept of combining knowledge-driven and data-driven approaches in computational biology.

[0014] A deep learning method for chromatin accessibility and transcription factor interaction is proposed, with step 3 implemented as follows: A complete training optimization process is designed specifically for the characteristics of single-cell multi-omics data, employing innovative dynamic training strategies and regularization techniques. This strategy implements a gene-specific dynamic training mechanism, independently training customized models for each target gene, fully considering the heterogeneity of different gene expression regulation patterns. To enhance the model's robustness to technical noise and biological heterogeneity in single-cell sequencing data, multiple data augmentation techniques are employed, including Gaussian noise injection and random feature masking, simulating potential signal loss and measurement errors during experiments. A progressive learning rate scheduling strategy is designed during training, combined with cosine annealing and an adaptive adjustment mechanism to achieve fine control over the optimization process. Gradient pruning is introduced to prevent gradient explosion, and an intelligent early stopping monitoring mechanism based on the Pearson correlation coefficient of the validation set is established. Training is automatically terminated when model performance does not significantly improve for several consecutive training cycles, preventing overfitting and ensuring the model achieves optimal generalization performance, thus improving training efficiency and model reliability.

[0015] A deep learning method for chromatin accessibility and transcription factor interaction is presented, with step 4 implemented as follows: Integrating attention mechanisms and interpretability modeling transforms the "black box" of deep learning into an effective tool for biological insights. This method integrates a soft attention mechanism in the chromatin accessibility and transcription factor interaction modeling. Through learnable attention weights, the model adaptively focuses on the regulatory regions most relevant to the target gene, simulating the specific recognition process of transcription factors on target genes in biological systems. Visualizing the trained weight matrix directly demonstrates the relative regulatory strength of each transcription factor in different chromatin regions. Furthermore, an interaction importance scoring system is designed to quantify the contribution of each chromatin accessibility and transcription factor to gene expression based on attention weights, generating a prioritized list of ranked regulatory elements, providing a clear basis for target selection in downstream experimental validation.

[0016] A deep learning method for chromatin accessibility and transcription factor interaction is proposed, with step 5 implemented as follows: A cross-gene knowledge transfer and ensemble learning strategy is developed to improve overall prediction performance while maintaining the advantages of gene-specific modeling. A cross-gene parameter sharing strategy is developed by designing a shareable general feature extraction module in the model architecture, enabling models of different genes to learn common regulatory patterns while maintaining personalized modeling capabilities for specific genes. A dual-training ensemble scheme is implemented, with each target gene undergoing two independent training iterations. Diversity in model predictions is generated through different random initialization and data augmentation methods. A dynamic ensemble decision mechanism is then used to automatically select the model weights for the optimal training rounds, effectively balancing the risks of underfitting and overfitting. Furthermore, an uncertainty quantification framework is established. The reliability and stability of model predictions are evaluated through the analysis of the coefficient of variation and confidence intervals of multiple training results, providing important quality control indicators for the interpretation of experimental results and making the prediction results of the entire method more scientifically rigorous and practically valuable.

[0017] A deep learning method for chromatin accessibility and transcription factor interaction is proposed, with step 6 implemented as follows: End-to-end optimization and multi-objective loss design are performed to construct a complete deep learning training optimization system. An innovative multi-objective loss function is designed, weighting and fusing the traditional regression loss (mean squared error) with the correlation optimization objective (Pearson correlation coefficient loss). This allows model training to directly target the correlation coefficient index, which is of greatest interest to biologists, breaking the limitation of traditional regression tasks that only optimize error. An adaptive loss balancing strategy is adopted, dynamically adjusting the relative weights of different loss components according to the training stage to ensure smooth convergence and reach global optimum. A hierarchical gradient backpropagation mechanism is implemented, designing differentiated learning rate strategies for different network layers to improve the efficiency and accuracy of parameter optimization. Simultaneously, a feature importance-aware regularization method is integrated, applying different levels of regularization constraints based on the importance of chromatin accessibility and transcription factor interaction. This prevents overfitting while better preserving key biological signals, ensuring the final model has both good generalization ability and accurately captures core regulatory relationship features.

[0018] A deep learning method for chromatin accessibility and transcription factor interaction is proposed, with step 7 implemented as follows: A systematic evaluation and biological validation framework is established, forming a complete research loop from computational prediction to experimental validation. A multi-level, multi-dimensional evaluation system is established, simultaneously monitoring dual performance metrics (mean squared error and Pearson correlation coefficient) on the training, validation, and test sets during training to comprehensively evaluate the model's learning ability and generalization performance. Finally, a complete interpretable output system is constructed to achieve deep integration and positive interaction between the computational model and experimental research. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the overall architecture of a deep learning method for chromatin accessibility and transcription factor interaction.

[0020] Figure 2 End-to-end deep neural network architecture diagram. Detailed Implementation

[0021] The following, in conjunction with the accompanying drawings and embodiments, provides a detailed description of the deep learning method for chromatin accessibility and transcription factor interaction according to the present invention. In specific implementation, this method can be based on the Python programming language and relies on the PyTorch and its extended library, the PyTorch Geometric deep learning framework. The overall process is as follows: Figure 1 It mainly includes the following 7 detailed steps:

[0022] Step 1: Innovative Model Architecture: Construct an end-to-end deep neural network to integrate the transcription factor expression matrix. ( N Each cell, M (transcription factors) and chromatin accessibility matrix ( P (a total of candidate cis-regulatory elements) are fused into a three-dimensional regulation tensor. Each element Indicated in cells n The Middle m The transcription factor and the first p The joint activity of several candidate cis-regulatory elements. The core of the model is a learnable interaction weight matrix. This matrix encodes the specific binding strength between transcription factors and candidate cis-regulatory elements in a parameterized manner. The forward propagation process involves first calculating the modulated features through an attention-enhanced interaction layer. ,in This represents element-wise multiplication, where σ is the sigmoid function that constrains the weights to the [0,1] interval. The modulated features are then flattened into a two-dimensional matrix and subjected to a nonlinear transformation using a multilayer perceptron with two hidden layers, ultimately outputting the predicted gene expression values. The entire network structure is as follows: Figure 2As shown, the system comprises four parts: a feature interaction layer, a flattening layer, a batch normalization layer, and a fully connected network module. The feature interaction layer weights the input features based on their importance, enhancing the model's sensitivity to key features. The flattening layer flattens the 3D tensor [batch_size, num_tf, num_peak] into a 2D tensor [batch_size, num_tf × num_peak], facilitating processing by subsequent fully connected layers. The first batch normalization layer normalizes the flattened features; each subsequent fully connected layer is followed by a batch normalization layer to accelerate training and improve stability. The fully connected network module employs a three-layer decreasing linear transformation structure. The first layer has an input dimension of num_tf × num_peak and an output dimension of 256, using the ReLU activation function and Dropout with a default parameter of 0.2 to prevent overfitting. The second layer has an input dimension of 256 and an output dimension of 128, also using the ReLU activation function and Dropout with a default parameter of 0.2. The output layer has an input dimension of 128 and an output dimension of 1. A multi-objective joint optimization strategy is adopted, and the loss function is defined as follows: ,in For mean square error loss, For Pearson correlation coefficient loss, For mixed regularization terms, α , β , γ To balance hyperparameters, batch normalization layers (to accelerate convergence) and Dropout layers (to prevent overfitting) are embedded in the network structure, and residual connections are introduced between hidden layers to alleviate the gradient vanishing problem.

[0023] Step 2, Multimodal Data Fusion: For each target gene g Based on transcription start site TSS g A dynamic selection window is defined on the reference genome coordinates, centered on [the reference genome]. TSS g -250kb, TSS g [+250kb]. Within this window, all chromatin-accessible regions detected in experiments using high-throughput sequencing to detect transposase-accessible chromatin were extracted, forming a region set. P g Simultaneously, filtering is performed from the expression matrix. M Given 10 known transcription factors, construct a cell × transcription factor expression matrix. The data fusion process is achieved through outer product operations: for each cell... n Calculate the vector outside ,generate M × P g The regulatory activity matrix, in whichA g [ n , :] for cells n In the region set P g Accessibility vector on the surface. Stack the regulatory activity matrices of all cells to form a three-dimensional tensor. .

[0024] Step 3, Dynamic Training Strategy: A gene-specific independent training paradigm is adopted, with each gene corresponding to an independent model instance. The training process employs a two-stage strategy: the first stage involves 100 rounds of warm-up training, optimizing only the weight matrix. W The parameters of the fully connected layer are fixed, and the learning rate is 0.01. In the second stage, all parameters are jointly optimized, and the learning rate is decayed to 0.001. The optimizer is AdamW. The adaptive strategy includes the following four items: (1) Data augmentation, which is performed during training with probability. p = 0.3 Apply two enhancements to the input tensor: Gaussian noise injection and random feature masking. (2) Learning rate scheduling, using cosine annealing strategy. (3) Early stopping mechanism, monitoring the Pearson correlation coefficient of the validation set, if it does not improve beyond the threshold of 0.001 for 50 consecutive rounds, then terminate training early. (4) Gradient clipping, calculating the global gradient norm, if it is greater than 1, then scaling is performed.

[0025] Step 4, interpretability modeling: A multi-head self-attention mechanism is introduced after the first hidden layer of the fully connected layer. The attention weight matrix reveals the synergistic or antagonistic relationships between different transcription factors, while the value matrix preserves the characteristic representation of chromatin accessibility and transcription factor interaction. By visualizing the attention weights, the core transcription factor regulatory modules can be identified. After training, the weight matrix W provides the most direct biological interpretation: (1) Transcription factor regulatory strength, calculate the global regulatory strength of each transcription factor, and identify key transcription factors. (2) Importance of candidate cis-regulatory elements, calculate the regulatory importance of each candidate cis-regulatory element region, and identify functional regulatory elements. (3) Specific binding pattern, for each transcription factor, extract the 10 candidate cis-regulatory elements with the highest weights, analyze the sequence characteristics and epigenetic markers of these regions, and verify the biological rationality of the model findings. Based on the weight matrix W, a gene regulatory network G = (V, E) is constructed, where the node set V contains all transcription factors and candidate cis-regulatory elements, and each edge in the edge set E contains all transcription factors and candidate cis-regulatory elements. e m,p The weight is | W m,p |

[0026] Step 5, Ensemble Learning Strategy: Design a shared-specific dual-path architecture. The base layer contains a feature extractor shared by all transcription factors, and the gene-specific layer contains an adapter module for each gene. The loss function adds parameter sharing constraints. Two models are trained independently for each gene, with differences arising from: (1) different random seed initializations; (2) different training data subsampling (80% of the samples); and (3) different data augmentation combinations. Soft voting ensemble is used in the prediction phase, and the prediction variance is calculated as an uncertainty estimate. The total uncertainty is decomposed into cognitive uncertainty (model parameter uncertainty) and random uncertainty (inherent noise in the data). Cognitive uncertainty is estimated through multiple sampling using Dropout; random uncertainty is estimated by adding trainable noise parameters. ε study.

[0027] Step 6: Multi-objective optimization: Design a dynamic weight adjustment mechanism based on the loss change rate. In the... t In each training round, the rate of change of each loss component relative to the previous round is calculated. The weights are updated and then normalized to a sum of 1. This strategy automatically allocates more weights to loss terms that improve slowly, achieving balanced optimization. The parameters are divided into three groups: Θ1 = W interaction weights, Θ2 = parameters of the previous MLP layer, and Θ3 = parameters of the next MLP layer. An alternating optimization scheme is used: Θ1 and Θ2 are updated in odd-numbered rounds, while Θ3 is fixed; Θ3 is updated in even-numbered rounds, while Θ1 and Θ2 are fixed. The learning rate is adjusted by group in each round. The regularization strength is adjusted based on the feature importance score. The importance score is calculated using an ensemble gradient method. The regularization term is modified so that important features are penalized less, retaining more information; while less important features are subject to strong sparsity constraints, improving model simplicity.

[0028] Step 7, System Evaluation Framework: During training, simultaneously monitor dual performance metrics (mean squared error and Pearson correlation coefficient) on the training, validation, and test sets to comprehensively evaluate the model's learning ability and generalization performance. Leave-one-cell-type cross-validation is employed, where all samples of one cell type are used as the test set each time, and samples of other cell types are used as the training set to evaluate the model's predictive ability on unseen cell types.

[0029] After systematic experimental verification on the T2D dataset, the deep learning method for chromatin accessibility and transcription factor interaction proposed in this invention shows significant performance advantages, with a Pearson correlation coefficient of 0.779, which is significantly improved compared to 0.717 of the existing representative method regX.

[0030] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A deep learning method for chromatin accessibility and transcription factor interaction, characterized in that, Includes the following steps: Step 1: Innovative Model Architecture: Construct an end-to-end deep neural network to integrate the transcription factor expression matrix. and chromatin accessibility matrix Fusion into a three-dimensional control tensor Each element This represents the joint activity of the m-th transcription factor and the p-th candidate cis-regulatory element in cell n; the core of the model is a learnable interaction weight matrix. This matrix encodes the specific binding strength between transcription factors and candidate cis-regulatory elements in a parameterized manner. The forward propagation process is as follows: First, through an attention-enhanced interaction layer, the modulated feature X' = X ⊗ σ(W) is calculated, where ⊗ represents element-wise multiplication, and σ is a sigmoid function that constrains the weights to the [0,1] interval. Then, the modulated feature is flattened into a two-dimensional matrix and subjected to a nonlinear transformation through a multilayer perceptron containing two hidden layers, finally outputting the gene expression prediction value ŷ∈ R N ; The overall network structure is shown in Figure 2. The system comprises four parts: a feature interaction layer, a flattening layer, a batch normalization layer, and a fully connected network module. The feature interaction layer weights the input features based on their importance, enhancing the model's sensitivity to key features. The flattening layer flattens the 3D tensor [batch_size, num_tf, num_peak] into a 2D tensor [batch_size, num_tf × num_peak], facilitating processing by subsequent fully connected layers. The first batch normalization layer normalizes the flattened features; each subsequent fully connected layer is followed by a batch normalization layer to accelerate training and improve stability. The fully connected network module employs a three-layer decreasing linear transformation structure. The first layer has an input dimension of num_tf × num_peak and an output dimension of 256, using the ReLU activation function and Dropout with a default parameter of 0.2 to prevent overfitting. The second layer has an input dimension of 256 and an output dimension of 128, also using the ReLU activation function and Dropout with a default parameter of 0.

2. The output layer has an input dimension of 128 and an output dimension of 1. A multi-objective joint optimization strategy is employed, with the loss function defined as... ,in For mean square error loss, For Pearson correlation coefficient loss, The network structure includes a mixed regularization term, with α, β, and γ being balanced hyperparameters. Batch normalization layers (to accelerate convergence) and Dropout layers (to prevent overfitting) are embedded in the network structure, and residual connections are introduced between hidden layers to alleviate the gradient vanishing problem. Step 2, Multimodal Data Fusion: Transcription factor expression and chromatin accessibility are fused into a three-dimensional matrix of cell × transcription factor × chromatin region, and regulatory regions within a 250kb range upstream and downstream are dynamically screened based on gene location; Multi-omics data are coordinated through adaptive standardization, and transfer learning is initialized based on the location frequency matrix basis order. Step 3, Dynamic Training Strategy: Implement gene-specific training process, and use noise injection and feature masking for data augmentation; Design an adaptive learning rate scheduler that combines gradient clipping with an early stopping mechanism based on the Pearson coefficient of the validation set; Step 4, Interpretability Modeling: Integrating soft attention mechanisms to focus on key regulatory regions, and visualizing chromatin accessibility and transcription factor regulatory intensity through weight matrices; Step 5, Integration Learning Strategy: Develop a cross-gene parameter sharing strategy, implement dual training integration, and improve prediction reliability through dynamic integration decision-making and uncertainty quantification; Step 6, Multi-objective optimization: Design a multi-objective function that weights mean squared error and Pearson loss, adopt an adaptive balancing strategy and hierarchical optimization, and introduce a feature importance-aware regularization method; Step 7: System Evaluation Framework: Establish a multi-index evaluation system, conduct cross-cell generalization tests and perturbation simulation analysis, and construct an interpretable output system to guide experimental verification.

2. The deep learning method for chromatin accessibility and transcription factor interaction according to claim 1, characterized in that, The implementation process of step 2 is as follows: Multimodal data fusion: for each target gene g Based on transcription start site TSS g A dynamic selection window is defined on the reference genome coordinates, centered on [the reference genome]. TSS g -250kb, TSS g [+250kb]; Within this window, extract all chromatin-accessible regions detected in experiments using high-throughput sequencing technology to detect transposase-accessible chromatin, forming a region set. P g Simultaneously, filtering from the expression matrix M Given 10 known transcription factors, construct a cell × transcription factor expression matrix. ; The data fusion process is achieved through outer product operations: for each cell n Calculate the outer product of vectors E g [ n , :]⊗ A g [ n , :], generate M × P g The regulatory activity matrix, in which A g [ n , :] for cells n In the region set P g Accessibility vectors on the surface; stacking the regulatory activity matrices of all cells to form a three-dimensional tensor. .

3. The deep learning method for chromatin accessibility and transcription factor interaction according to claim 1, characterized in that, The implementation process of step 3 is as follows: Dynamic training strategy: Employs a gene-specific independent training paradigm, with each gene corresponding to an independent model instance; the training process uses a two-stage strategy: the first stage involves 100 rounds of warm-up training, optimizing only the weight matrix. W In the first stage, the parameters of the fully connected layer are fixed with a learning rate of 0.

01. In the second stage, all parameters are jointly optimized with the learning rate decaying to 0.

001. The optimizer used is AdamW. The adaptive strategy includes the following four items: (1) Data augmentation, which uses probability during training. p = 0.3 applies two enhancements to the input tensor: Gaussian noise injection and random feature masking; (2) Learning rate scheduling, using cosine annealing strategy; (3) Early stopping mechanism, monitoring the Pearson correlation coefficient of the validation set, if it does not improve beyond the threshold of 0.001 for 50 consecutive rounds, then terminate training early; (4) Gradient clipping, calculating the global gradient norm, if it is greater than 1, then scaling is performed.

4. The deep learning method for chromatin accessibility and transcription factor interaction according to claim 1, characterized in that, The implementation process of step 4 is as follows: Interpretable modeling: A multi-head self-attention mechanism is introduced after the first hidden layer of the fully connected layer; the attention weight matrix reveals the synergistic or antagonistic relationships between different transcription factors, while the value matrix preserves the characteristic representation of chromatin accessibility and transcription factor interaction; by visualizing the attention weights, the core transcription factor regulatory modules can be identified. After training, the weight matrix W provides the most direct biological interpretation: (1) the intensity of transcription factor regulation, calculating the global regulatory intensity of each transcription factor and identifying key transcription factors; (2) Importance of candidate cis-regulatory elements: Calculate the regulatory importance of each candidate cis-regulatory element region and identify functional regulatory elements; (3) Specific binding pattern: For each transcription factor, the 10 candidate cis-regulatory elements with the highest weights are extracted, and the sequence characteristics and epigenetic markers of these regions are analyzed to verify the biological rationality of the model findings. A gene regulatory network G = (V, E) is constructed based on a weight matrix W. The node set V contains all transcription factors and candidate cis-regulatory elements, and each edge in the edge set E... e m,p The weight is | W m,p | 5. The deep learning method for chromatin accessibility and transcription factor interaction according to claim 1, characterized in that, The implementation process of step 5 is as follows: Ensemble learning strategy: Design a shared-specific dual-path architecture, with the base layer containing a feature extractor shared by all transcription factors, and the gene-specific layer containing an adapter module for each gene; increase parameter sharing constraints in the loss function; train two models independently for each gene, with differences arising from: (1) different random seed initialization; (2) different training data subsampling (80% of samples); (3) different data augmentation combinations; use soft voting ensemble in the prediction phase, and calculate the prediction variance as an uncertainty estimate; decompose the total uncertainty into cognitive uncertainty (model parameter uncertainty) and random uncertainty (inherent noise in the data); estimate cognitive uncertainty through multiple sampling using Dropout; and estimate random uncertainty by adding trainable noise parameters. ε study.

6. The deep learning method for chromatin accessibility and transcription factor interaction according to claim 1, characterized in that, The implementation process of step 6 is as follows: Multi-objective optimization: Design a dynamic weight adjustment mechanism based on the rate of change of loss; in the first... t In each round of training, calculate the rate of change of each loss component relative to the previous round; Update the weights, then normalize them so that the sum of the weights is 1; This strategy automatically allocates more weights to the loss term that improves slowly, achieving balanced optimization. The parameters are divided into three groups: Θ1 = W (interaction weights), Θ2 = parameters of the MLP's pre-layer, and Θ3 = parameters of the MLP's post-layer. An alternating optimization scheme is adopted: Θ1 and Θ2 are updated in odd-numbered rounds, while Θ3 is fixed; Θ3 is updated in even-numbered rounds, while Θ1 and Θ2 are fixed. The learning rate is adjusted by group in each round. The regularization strength is adjusted based on the feature importance score. The importance score is calculated using an ensemble gradient method. The regularization term is modified so that important features are less penalized, retaining more information, while less important features are subject to strong sparsity constraints, improving model simplicity.

7. The deep learning method for chromatin accessibility and transcription factor interaction according to claim 1, characterized in that, The implementation process of step 7 is as follows: System evaluation framework: During training, dual performance metrics (mean squared error and Pearson correlation coefficient) are monitored simultaneously on the training set, validation set, and test set to comprehensively evaluate the model's learning ability and generalization performance; leave-one-cell-type cross-validation is adopted, in which all samples of one cell type are used as the test set each time, and samples of other types are used as the training set to evaluate the model's predictive ability on unseen cell types.

Images