A single-cell Hi-C profile prediction method based on single-cell RNA expression data
Through deep learning methods, single-cell RNA and single-cell Hi-C data are integrated into a unified low-dimensional space, solving the problems of high cost of single-cell Hi-C technology and difficulty in data integration, and achieving accurate prediction of Hi- from RNA data C map improves the accuracy and efficiency of the research and provides new tools for genomics and transcriptomics research.
Patent Information
- Application Number
- CN202411098137.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-12
AI Technical Summary
Existing single-cell Hi-C technologies are costly, complex and data noise when capturing chromatin heterogeneity at the single-cell level, and face the difficulty of increasing sparsity and noise when integrating data, especially single-cell There are challenges in integrating RNA and single-cell Hi-C data into a unified low-dimensional space.
Using a deep learning-based approach, the data are mapped to the same low-dimensional space by constructing the pseudo-distance between single-cell RNA expression data and single-cell Hi-C map features, using multimodal alignment. The network and the variational autoencoder perform data integration, and use RNA expression data to predict the Hi-C map.
Reduces the cost and technical difficulty of data acquisition, improves the accuracy and stability of prediction models, enhances data integration capabilities, provides new tools and ideas to study gene expression regulation, cell differentiation and disease mechanisms, suitable for different types and Single-cell data from source.
Smart Images

Figure CN118645154B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of bioinformatics and computational biology, and specifically relates to a single-cell Hi-C map prediction method based on single-cell RNA expression data. Background Art
[0002] Chromatin conformation plays a key role in gene expression regulation, cell differentiation and disease occurrence. Traditional chromatin conformation capture technologies (such as Hi-C) provide three-dimensional chromatin structural information at the whole genome level, but usually require a large number of cells and it is difficult to capture heterogeneity at the single cell level. The development of single-cell technology enables researchers to explore the three-dimensional organization of chromatin at the single cell level. Chromatin is a complex structure composed of DNA, histones and other chromatin proteins, and is the storage and expression unit of genetic information in the cell nucleus. Chromatin not only has a linear structure, but more importantly, it forms a complex three-dimensional conformation in the cell nucleus. This three-dimensional structure plays a vital role in biological processes such as gene expression regulation, cell cycle regulation, and DNA repair. Scientists have successively developed a variety of improved technologies based on 3C, including 4C, 5C and Hi-C. 4C technology can realize the interaction analysis of chromatin fragments with other fragments in the whole genome, 5C technology can detect the interaction between multiple chromatin fragments, and Hi-C technology can detect the interaction between all chromatin fragments in the whole genome. The advantage of Hi-C technology is that it provides high-resolution chromatin interaction data on a genome-wide scale, but it requires a large number of cell samples and cannot capture chromatin heterogeneity at the single-cell level.
[0003] Breakthroughs in single-cell technology have enabled scientists to study a variety of omics data, including genomes, transcriptomes, and proteomes, at the single-cell level. Single-cell RNA sequencing (scRNA-seq) can detect the gene expression of each cell at the single-cell level, revealing the functional state of the cell and the gene regulation mechanism. Through scRNA-seq technology, researchers can identify different cell types and subtypes and understand the heterogeneity and dynamic changes of cells. Single-cell Hi-C technology (scHi-C) can capture the three-dimensional structure of chromatin at the single-cell level. Compared with traditional Hi-C, scHi-C technology can reveal chromatin heterogeneity at the single-cell level, but the experimental cost is high, the technology is complex, the data noise is large, and the coverage is low, which limits its application in large-scale studies.
[0004] Data integration methods play an important role in genomics and transcriptomics research. They can extract comprehensive information from multi-dimensional data and reveal the laws of complex biological systems. Existing multi-omics data integration methods include integration based on paired data and unpaired data. Paired data-based integration methods require that different omics data come from the same batch of cells, such as Seurat v3 and LIGER, which use canonical correlation analysis (CCA) or non-negative matrix factorization (NMF) to achieve joint data embedding. Unpaired data-based integration methods process data from different experiments and map different omics data into the same space by constructing a common low-dimensional space. For example, Cobolt and MultiVI use variational autoencoders (VAEs) to learn multi-omics data and reduce the different modalities of all cells to a unified low-dimensional representation. GLUE (graph-linked unifiedembedding) learns feature embedding from prior knowledge through graph variational autoencoders, and then reconstructs and links omics data based on cell embedding.
[0005] Although these data integration methods have made significant progress in processing multi-omics data, they still face challenges when applied to the integration of single-cell RNA and single-cell Hi-C data. The sparsity and noise of single-cell Hi-C data increase the difficulty of data integration. The feature spaces of single-cell RNA data and single-cell Hi-C data are quite different. How to accurately represent the features of these two types of data in a unified low-dimensional space is a key problem that data integration methods need to solve.
[0006] To address these challenges, this paper proposes a method for integrating single-cell RNA and single-cell Hi-C data based on deep learning. By constructing a unified low-dimensional space, single-cell RNA expression data is mapped to the chromatin conformation space, achieving the goal of predicting Hi-C maps from RNA data. This method can make full use of the association between RNA expression data and chromatin conformation data, improve the accuracy and stability of the prediction model, reduce the cost and technical difficulty of data acquisition, and provide new tools and ideas for studying gene expression regulation, cell differentiation and disease mechanisms. Summary of the invention
[0007] The present invention first provides a single-cell Hi-C map prediction method based on single-cell RNA expression data, which comprises the step of constructing a pseudo-distance between scRNA expression data features and scHi-C map features.
[0008] In certain embodiments, the step of constructing pseudo-distances between scRNA expression data features and scHi-C profile features comprises mapping scRNA data and scHi-C data into the same low-dimensional space using multimodal alignment as an iterative optimization process guided by feature embeddings encoded by the graph.
[0009] In certain embodiments, the mapping includes (1) data preprocessing; and (2) constructing a cell embedding space.
[0010] In some embodiments, the (1) data preprocessing includes:
[0011] (1.1) Data acquisition: Obtain single-cell RNA sequencing data (scRNA-seq) and single-cell Hi-C data (scHi-C) from public databases or experimental data;
[0012] (1.2) Data filtering: Perform quality control on the raw data and remove low-quality cells and genes. For RNA data, filter out low-expressed genes; for Hi-C data, remove fragments with low interaction frequency;
[0013] (1.3) Data normalization: Normalize RNA data and Hi-C data to eliminate technical noise and batch effects. Commonly used methods include log normalization and z-score normalization;
[0014] (1.4) Dimensionality reduction: Use methods such as principal component analysis (PCA) and t-SNE to reduce the dimensionality of the data to facilitate subsequent model training and prediction.
[0015] In certain embodiments, the (2) constructing a cell embedding space comprises:
[0016] (2.1) Graph structure representation: RNA data and Hi-C data are constructed into a graph structure. The nodes of RNA data represent genes, and the edges represent the co-expression relationship between genes; the nodes of Hi-C data represent genomic regions, and the edges represent the interactions between regions.
[0017] (2.2) Association graph construction: By calculating the similarity between the features of RNA data and Hi-C data, an association graph of RNA and Hi-C data is constructed. The similarity calculation can use cosine similarity or Euclidean distance.
[0018] (2.3) Multimodal data integration: Using the association graph and dimensionality reduction results, unsupervised integration algorithms such as GLUE are used to integrate scRNA and scHi-C data into a common low-dimensional space.
[0019] The present invention also provides a method for integrating single-cell RNA expression data with single-cell Hi-C maps, which comprises the following steps:
[0020] (1) Normalization and dimensionality reduction of RNA expression data and Hi-C maps.
[0021] (2) Construct a pseudo-distance represented by a graph structure to characterize the relationship between the two features.
[0022] (3) Map the two into the same low-dimensional space in an unsupervised manner.
[0023] The present invention also provides a local single-cell Hi-C map prediction algorithm, which comprises the following steps:
[0024] (1) Training using a one-to-one dataset based on RNA data and Hi-C data.
[0025] (2) The transformer model is used as the encoder and decoder to learn the mapping relationship between expression data and graph.
[0026] (3) The input data is converted into a two-dimensional matrix and divided into fixed-size blocks to enhance the performance of the attention mechanism.
[0027] The present invention also provides a pseudo time series analysis, model attribution and label migration method, which comprises the following steps:
[0028] (1) Observe the dynamic changes of chromatin conformation at the single-cell level;
[0029] (2) Predict chromatin regulatory factors;
[0030] (3) Observe differences in chromatin conformation under disease conditions;
[0031] (4) Apply the integrated gradient algorithm to analyze the model and identify regulatory factors that affect chromatin conformation in specified genomic regions.
[0032] The present invention also provides an application of the above prediction method, wherein the application is a) achieving the goal of predicting Hi-C maps from RNA data; or b) studying gene expression regulation; or c) studying cell differentiation; or d) studying disease mechanisms, and the application is for non-diagnostic or non-therapeutic purposes.
[0033] Optionally, the application also includes pseudo-time analysis, model attribution and label migration in dynamic changes of chromatin conformation, and prediction of regulatory factors;
[0034] The present invention also provides a model, which has a functional module for implementing the above method.
[0035] In some embodiments, the model includes an encoder and a decoder.
[0036] In some embodiments, the encoder is a transformer; and the decoder is an autoencoder.
[0037] Finally, the present invention provides applications of the above-mentioned model, wherein the applications are: a) achieving the goal of predicting Hi-C maps from RNA data; or b) studying gene expression regulation; or c) studying cell differentiation; or d) studying disease mechanisms, and the applications are non-diagnostic or non-therapeutic purposes.
[0038] Optionally, the application also includes pseudo-time analysis, model attribution and label migration in dynamic changes of chromatin conformation, and prediction of regulatory factors;
[0039] Compared with the prior art, the present invention has at least the following beneficial effects:
[0040] The SEE framework uses a large amount of single-cell RNA (scRNA) expression data and a small amount of single-cell Hi-C (scHi-C) maps as input for training.
[0041] The present invention proposes a method for predicting single-cell Hi-C maps based on single-cell RNA expression data, which has many beneficial effects. The model can predict the Hi-C map of a specific genomic region based on RNA expression data at the single-cell level. For downstream applications, the trained model can infer the dynamic changes in chromatin conformation, and identify chromatin regulatory factors that affect the chromatin conformation of a specific genomic region, and can explore the impact of dynamic changes in chromatin conformation on the occurrence and development of diseases. By combining deep learning models and multi-omics data integration technology, the present invention not only technically achieves the goal of inferring the three-dimensional structure of chromatin from RNA data, but also significantly reduces experimental costs in application, improves data analysis efficiency, and enhances data integration capabilities, providing new tools and ideas for genomics and transcriptomics research. The following are the specific beneficial effects brought about by the present invention:
[0042] 1. Reduce data acquisition costs
[0043] The high experimental cost of single-cell Hi-C technology limits its application in large-scale research. Each single-cell Hi-C experiment requires not only complex experimental operations, but also high reagent and equipment costs. In contrast, single-cell RNA sequencing technology has developed relatively maturely and can generate rich transcriptome data in large-scale cell populations at a lower cost. The present invention predicts single-cell Hi-C maps by utilizing existing single-cell RNA data, avoiding actual Hi-C experiments, thereby significantly reducing the cost of data acquisition. Researchers can obtain important information about the three-dimensional structure of chromatin through the method of the present invention without adding additional experimental costs.
[0044] 2. Improve data analysis efficiency
[0045] The traditional chromatin conformation capture experimental process is complicated and time-consuming, and data processing and analysis require a lot of computing resources. Through the method of the present invention, researchers can use existing single-cell RNA data to quickly predict the three-dimensional structure of chromatin, saving a lot of time and resources. After training, the deep learning model can predict new data at a high speed, greatly improving the efficiency of data analysis. This advantage makes the present invention have significant application value in large-scale data processing and high-throughput analysis.
[0046] 3. Enhance data integration capabilities
[0047] In genomics and transcriptomics research, the integrated analysis of multi-omics data can provide more comprehensive biological information. The present invention integrates single-cell RNA and Hi-C data by constructing a unified low-dimensional space, thereby realizing the joint analysis of different omics data. Using the methods of graph neural networks and variational autoencoders, the characteristic relationships between RNA data and Hi-C data can be fully mined and represented. This multi-omics integration method not only improves the accuracy and stability of data analysis, but also provides a new perspective for studying the complex relationship between gene expression and chromatin structure.
[0048] 4. Improve research accuracy and depth
[0049] The present invention uses a deep learning model to learn the complex relationship between gene expression and chromatin three-dimensional structure from single-cell RNA data, and can predict the interaction frequency of chromatin with high accuracy. The model generates high-quality Hi-C maps by capturing long-range dependencies and complex feature patterns in RNA data, thereby improving the accuracy of the research. In addition, by integrating different types of single-cell data, the present invention can deeply reveal the gene expression regulation mechanism, cell differentiation process, and chromatin structure changes in disease states, providing new tools and methods for life science research.
[0050] 5. Wide applicability and strong generalization ability
[0051] The method of the present invention has strong generalization ability and can be applied to single-cell data of different types and sources. Through experimental verification on multiple public data sets, the model of the present invention shows good stability and accuracy under different data sets and experimental conditions. Whether in human, animal or plant cells, the method of the present invention can effectively predict the three-dimensional structure of chromatin and has wide applicability. This feature makes the present invention have important application prospects in basic research, clinical application and biotechnology development.
[0052] 6. Promote the development of personalized medicine and precision medicine
[0053] The three-dimensional structure of chromatin plays an important role in gene expression regulation and the occurrence and development of diseases. Through the method of the present invention, the spatial conformational changes of disease-related genes can be revealed at the single-cell level, thereby understanding the molecular mechanism and regulatory network of the disease. This ability is of great significance for the development of personalized medicine and precision medicine. Researchers and clinicians can use the method of the present invention to analyze the single-cell RNA data of patients and predict their three-dimensional chromatin structure, so as to formulate personalized treatment plans and improve the accuracy of disease diagnosis and treatment.
[0054] 7. Promote the comprehensive use of multi-omics data
[0055] The integrated analysis of multi-omics data is an important trend in modern biological research. Through the method of the present invention, researchers can jointly analyze single-cell RNA data with other omics data (such as proteomics, metabolomics, etc.) to fully understand the functional status and regulatory mechanism of cells. The method of the present invention provides an effective tool for the comprehensive utilization of multi-omics data and promotes the development of multi-omics research.
[0056] 8. Provide new research methods for basic research
[0057] The present invention not only has significant advantages in application, but also provides a new research method for basic research. By inferring the three-dimensional structure of chromatin from single-cell RNA data, researchers can deeply explore the dynamic changes of chromatin during gene expression regulation, cell differentiation and development. In particular, the method of the present invention has important scientific research value in studying the relationship between chromatin structure and function and revealing the mechanism of chromatin regulation.
[0058] In summary, the technical solution of the present invention has significant beneficial effects in reducing data acquisition costs, improving data analysis efficiency, enhancing data integration capabilities, and improving research accuracy and depth. In addition, the method of the present invention is widely applicable and has strong generalization capabilities, which promotes the development of personalized medicine and precision medicine, promotes the comprehensive utilization of multi-omics data, and provides new research methods for basic research. Through the method of the present invention, researchers can better understand the three-dimensional structure of the genome and its role in gene expression regulation and cell function, and further promote the development of life science research. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 : SEE model workflow diagram;
[0060] Figure 2 : Integration method architecture diagram;
[0061] Figure 3:Data flow diagram for mapping scRNA and scHi-C data;
[0062] Figure 4 : Model architecture diagram;
[0063] Figure 5 : The loss function change curve during the model training process shows the optimization of the loss function during the training process;
[0064] Figure 6 :Comparison chart of the predicted results and the actual Hi-C map, showing the accuracy and stability of the model prediction results;
[0065] Figure 7 :Combined analysis of scHi-C maps and scRNA expression data of the MBP gene locus on a pseudo-time scale;
[0066] Figure 8 : Differences in the changes of characteristics of QKI, DOCK10, and PIP4K2A gene loci on a pseudo-time scale;
[0067] Fig. 9 : Differences in the changes of characteristics of QKI, DOCK10, and PIP4K2A gene loci on a pseudo-time scale;
[0068] Fig.10 : Chromatin conditioning factors that influence the dynamics of interactions;
[0069] Fig.11 :Correlation analysis between the oscillation pattern of gene expression counts and the oscillation pattern of chromatin conformation volume;
[0070] Fig.12 : Robust interactions found around the gene BIN1 in astrocyte subtypes. DETAILED DESCRIPTION
[0071] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.
[0072] List of abbreviations, English and definitions of key terms:
[0073] 1. SEE (a Single-cell method for predicting Chromatin Conformationbased on gene Expression);
[0074] 2. scRNA: single-cell RNA;
[0075] 3. scHi-C: single cell Hi-C;
[0076] Example 1
[0077] A deep learning-based prediction model that can infer single-cell Hi-C profiles from single-cell RNA expression data. The overall process and architecture of the model are as follows: Figure 1 The specific technical solution includes the following steps:
[0078] like Figure 2 As shown, the present invention proposes a method for integrating cell embedding of single-cell RNA (scRNA) expression data and single-cell Hi-C (scHi-C) maps. In order to model the cell state, the present invention normalizes and scales the scRNA expression data and the scHi-C map to obtain their respective low-dimensional space information. Using previous biological knowledge, the present invention constructs pseudo-distances between scRNA expression data features and scHi-C map features in a graph representation. Specifically, the vertices are genes and chromatin regions (i.e., 10-kilobase (kb) bins), and positive edges connect a gene and its genomic location. Then, under the guidance of the feature embedding encoded by the graph, multimodal alignment is used as an iterative optimization process to map the scRNA data and scHi-C data into the same low-dimensional space. The SEE method is built on a scRNA expression dataset and a scHi-C map dataset. The mapping process of scRNA and scHi-C data is as follows Figure 3 exhibit.
[0079] 1. Data preprocessing:
[0080] (1.1) Data acquisition: Obtain single-cell RNA sequencing data (scRNA-seq) and single-cell Hi-C data (scHi-C) from public databases or experimental data.
[0081] (1.2) Data filtering: Perform quality control on the raw data and remove low-quality cells and genes. For RNA data, filter out low-expressed genes; for Hi-C data, remove fragments with low interaction frequency.
[0082] (1.3) Data normalization: Normalize RNA data and Hi-C data to eliminate technical noise and batch effects. Commonly used methods include log normalization and z-score normalization.
[0083] (1.4) Dimensionality reduction: Use methods such as principal component analysis (PCA) and t-SNE to reduce the dimensionality of the data to facilitate subsequent model training and prediction.
[0084] 2. Constructing cell embedding space:
[0085] (2.1) Graph structure representation: RNA data and Hi-C data are constructed into a graph structure. The nodes of RNA data represent genes, and the edges represent the co-expression relationship between genes; the nodes of Hi-C data represent genomic regions, and the edges represent the interactions between regions.
[0086] (2.2) Association graph construction: By calculating the similarity between the features of RNA data and Hi-C data, an association graph of RNA and Hi-C data is constructed. The similarity calculation can use cosine similarity or Euclidean distance.
[0087] (2.3) Multimodal data integration: Using the association graph and dimensionality reduction results, unsupervised integration algorithms such as GLUE are used to integrate scRNA and scHi-C data into a common low-dimensional space.
[0088] Example 2
[0089] Model training: A specific model using transformer as encoder and decoder to learn dependencies between features (gene expression) in low information density input data (scRNA expression data), such as Figure 4 As shown. In order to enhance the performance of the attention mechanism in utilizing a large number of features, the input data is converted into a two-dimensional matrix and then divided into multiple small blocks of fixed size, each of which is encoded separately at the input layer of the model. Since the focus of model learning is on initial information extraction to obtain factors that affect chromatin conformation, a more complex encoder network is required to extract information from the input data. See Figure 4 .
[0090] (3.1) Model architecture design: Design a deep learning model based on transformer and autoencoder. The model consists of two parts: encoder and decoder. The encoder is responsible for extracting the features of RNA data, and the decoder is responsible for mapping the features into Hi-C maps.
[0091] (3.2) Encoder: We use transformer as encoder to capture long-range dependencies and complex feature patterns in RNA data. Transformer consists of multiple layers of self-attention mechanism and can effectively handle high-dimensional and sparse data.
[0092] (3.3) Decoder: An autoencoder is used as a decoder to map the encoded low-dimensional representation back to a high-dimensional space to generate a Hi-C map. The decoder consists of a multi-layer fully connected neural network that can reconstruct a high-dimensional interaction frequency matrix.
[0093] (3.4) Loss function: The focal loss function is used for model training to deal with the problem of data imbalance. The focal loss function can better handle rare categories and improve the performance of the model on sparse data.
[0094] (3.5) Training process: The Adam optimizer is used to train the model, and the generalization ability of the model is improved through cross-validation and parameter tuning. The training data set is divided into training set, validation set and test set, which are used for model training, validation and evaluation respectively.
[0095] Example 3
[0096] Prediction and verification:
[0097] (4.1) Model prediction: Apply the trained model to new single-cell RNA data to predict the corresponding Hi-C map. The input data is the preprocessed RNA expression data, and the output is the predicted chromatin interaction frequency matrix.
[0098] (4.2) Result verification: The prediction accuracy and stability of the model were verified by comparing with the actual Hi-C data. Commonly used evaluation indicators include Pearson correlation coefficient, mean square error (MSE) and adjusted Rand index (ARI).
[0099] (4.3) Model performance evaluation: Experiments are conducted on multiple public datasets to evaluate the generalization ability of the model under different datasets and experimental conditions. Through cross-validation and multiple experiments, the model is ensured to have good stability and accuracy. Figure 5-Figure 6 . Figure 5 By comparing the focal loss and mean-square error (MSE) experiments, it can be observed that the focal loss retains the overall accuracy of the MSE loss and produces better training results in cell subtypes (Astro and OPC) with relatively small amounts of data. Figure 6 It indicates that at the Slc1a2 and Slc1a3 gene loci, the predicted Hi-C maps are similar to the original results.
[0100] Downstream analysis
[0101] (4.4) Pseudo-temporal dynamics analysis of gene expression and chromatin conformation: First, pseudo-temporal trajectory analysis of single-cell RNA (scRNA) expression data is performed to obtain pseudo-time labels of samples. Then, the scRNA expression data and the predicted single-cell Hi-C (scHi-C) maps are reordered according to the pseudo-time labels.
[0102] (4.5) Calculate chromatin volume: Use the predicted Hi-C map to calculate the three-dimensional volume of chromatin in each cell. The specific method can be based on the three-dimensional reconstruction of the chromatin interaction frequency matrix to calculate the volume occupied by chromatin in three-dimensional space.
[0103] (4.6) Quantify the dynamic score of chromatin interactions: Analyze the reordered Hi-C maps and calculate the dynamic score of chromatin interactions. The dynamic score reflects the changes in chromatin interactions at different pseudo-time points.
[0104] (4.7) Visualize TAD rearrangement: Analyze and display TAD rearrangement on pseudo-time-sequential trajectories to study the dynamic changes of TAD boundaries and internal interactions. Figure 7 Take MBP as an example. Figure 7 The UMAP plot in Figure a shows that in the ODC cell subtype, the cells corresponding to the scRNA data are displayed as an obvious arc, and this arc is significantly correlated with the cell pseudo time calculated using Scanpy. Subsequently, this experiment reordered the cells according to the pseudo time and visualized the gene expression values of the cells in the new order, as shown in Figure 5. Figure 7 As shown in panel b, a clear upward trend in gene expression can be observed over pseudo time, indicating that MBP expression changes during ODC development.
[0105] (4.8) Quantify the oscillation of chromatin interactions at the locus level: Analyze the predicted Hi-C maps and quantify the oscillation of chromatin interactions at specific loci on the pseudo-time trajectory. Calculate the interaction frequency of each locus at different pseudo-time points and analyze its oscillation pattern. See Figure 8-Figure 9 .like Figure 8 As shown in , in order to study the intrinsic dynamics of chromatin conformation, this paper measures the distribution of changes in chromatin interactions. Since the threshold for determining whether an interaction oscillates is sample-specific during the calculation of the switch score, it is more suitable for evaluating the dynamic changes between different interactions in the same sample. However, due to this processing method, the switch score is not conducive to evaluating the dynamic changes between different samples. Fig. 9 Differentially expressed genes (DEGs) were selected from (Astro: SLC1A2, SLC1A3 and GPM6A; ODC: MBP, QKI and DOCK10); lowly expressed genes (not DEGs) were selected from (Astro: MBP, QKI and DOCK10; ODC: SLC1A2, SLC1A3 and GPM6A)). As shown, for the same gene, higher vibration scores can be observed in the characterized cell subtypes, indicating that the interaction is more dynamic compared with other cell subtypes.
[0106] (4.9) Obtaining important chromatin regulatory factors: Use the integral gradient algorithm to calculate the importance of each feature (chromatin regulatory factor) on the input RNA expression data corresponding to a single interaction on the local Hi-C map; then, traverse all interactions on the local Hi-C map to perform similar calculations, and average the results as the contribution of all features on the RNA expression data to the local Hi-C map. See Fig.10 .like Fig.10 As shown, Fig.10 Middle a picture and Fig.10 Figure c shows the three-dimensional structural changes of the PDGFRA (a) / MBP (c) site during differentiation from OPCs to ODCs. Fig.10 Middle B picture and Fig.10 Figure d in the middle shows that for the same genomic region (PDGFRA (b) and MBP (d)), different cell subtypes have different regulatory patterns. Each heat map consists of 20 significantly different chromatin regulators at three locations of the target gene (promoter, gene body, and terminator), and the color on the heat map represents the importance score of the chromatin regulator calculated by the IG algorithm.
[0107] (4.10) Quantify the dynamic changes of chromatin conformation coupled with gene expression: Quantify the dynamic changes of chromatin conformation coupled with gene expression by integrating gene expression data and chromatin conformation data on pseudo-time-sequential trajectories. Analyze the coordinated change patterns of the interaction between gene expression and chromatin. See Fig.11 .like Fig.11 As shown, by analyzing up to 15 differentially expressed genes (DEGs) in ODC cell subtypes, this paper found a positive correlation between the oscillation amplitude of chromatin volume and the oscillation amplitude of expression level.
[0108] (4.11) Disease-associated single nucleotide polymorphisms can be explained by the dynamic features of chromatin conformation: Disease-associated single nucleotide polymorphisms (SNPs) can be explained by the dynamic features of chromatin conformation. The predicted Hi-C profiles are combined with disease-associated gene or variant data to study the role of chromatin conformation in disease. See Fig.12 The results are as follows Fig.12 As shown, rs10207628 has robust interactions with expression regulatory regions in the Astro cell subtype. Therefore, we infer that these interactions are not easily altered in disease states, thereby ensuring stable changes in BIN1 expression when TFEB expression levels are upregulated.
[0109] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A single-cell Hi-C profile prediction method based on single-cell RNA sequencing data, characterized in that: The method comprises the following steps: (1) Data preprocessing: (1.1) Data acquisition: Obtain single-cell RNA sequencing data (scRNA-seq) and single-cell Hi-C data (scHi-C) from public databases or experimental data; (1.2) Data filtering: Perform quality control on raw data and remove low-quality cells and genes; for RNA sequencing data, filter out low-expressed genes; for Hi-C data, remove fragments with low interaction frequency; (1.3) Data normalization: Normalize RNA sequencing data and Hi-C data to eliminate technical noise and batch effects; (1.4) Dimensionality reduction: Use principal component analysis (PCA) and t-SNE methods to reduce the dimension of the data to facilitate subsequent model training and prediction; (2) Constructing cell embedding space: (2.1) Graph structure representation: RNA sequencing data and Hi-C data are constructed into a graph structure. The nodes of RNA sequencing data represent genes, and the edges represent the co-expression relationship between genes; the nodes of Hi-C data represent genomic regions, and the edges represent the interactions between regions; (2.2) Association graph construction: By calculating the similarity between the features of RNA sequencing data and Hi-C data, an association graph between RNA sequencing data and Hi-C data is constructed; (2.3) Multimodal data integration: Using the association graph and dimensionality reduction results, the GLUE unsupervised integration algorithm was used to integrate scRNA-seq and scHi-C data into a common low-dimensional space; (3) Model training: (3.1) Model architecture design: Design a deep learning model based on transformer and autoencoder. The model consists of two parts: encoder and decoder. The encoder is responsible for extracting the features of RNA sequencing data, and the decoder is responsible for mapping the features into Hi-C maps. (3.2) Encoder: Use transformer as encoder to capture long-range dependencies and complex feature patterns in RNA sequencing data; The transformer consists of a multi-layer self-attention mechanism that can effectively handle high-dimensional and sparse data; (3.3) Decoder: Use the autoencoder as the decoder to map the encoded low-dimensional representation back to the high-dimensional space to generate the Hi-C map; the decoder consists of a multi-layer fully connected neural network that can reconstruct the high-dimensional interaction frequency matrix; (3.4) Loss function: The focal loss function is used for model training to deal with the problem of data imbalance; the focal loss function can better handle rare categories and improve the performance of the model on sparse data; (3.5) Training process: The Adam optimizer is used to train the model, and the generalization ability of the model is improved through cross-validation and parameter tuning. The training data set is divided into training set, validation set and test set, which are used for model training, validation and evaluation respectively.
2. An application of the prediction method according to claim 1, characterized in that: The application is a) achieving the goal of predicting Hi-C maps from RNA sequencing data; or b) studying gene expression regulation; or c) studying cell differentiation, or d) studying disease mechanisms, and the application is non-diagnostic and non-therapeutic.
3. A model, characterized in that: The model has a functional module for implementing the prediction method according to claim 1.
4. The model according to claim 3, characterized in that The model includes an encoder and a decoder.
5. The model according to claim 4, characterized in that The encoder is a transformer; the decoder is an autoencoder.
6. An application of a model as claimed in any one of claims 3 to 5, characterized in that: The application is a) achieving the goal of predicting Hi-C maps from RNA sequencing data; or b) studying gene expression regulation; or c) studying cell differentiation, or d) studying disease mechanisms, and the application is non-diagnostic and non-therapeutic.
Citation Information
Patent Citations
Accurate single-cell multi-omics matching data generation method
CN115862746A
ScRNA-seq data dimension reduction method based on graph neural network
CN116386729A
Deep learning and alignment of spatially-resolved whole transcriptomes of single cells
US20220068438A1
Transomic systems and methods of their use
WO2024059658A2
