A hierarchical single-cell chromatin accessibility data analysis and annotation method
By using a hierarchical neural network model and specific peak screening, the problem of insufficient cell-specific signals in the parsing of single-cell chromatin accessibility data was solved, achieving efficient cell type annotation and improved biological credibility.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2026-04-02
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for parsing single-cell chromatin accessibility data fail to adequately characterize cell-specific signals, resulting in unstable annotation results. Furthermore, extracting true biological features from sparse signals is challenging, impacting interpretability and reliability.
A hierarchical neural network model was used to process single-cell chromatin accessibility data. Cell type-specific features were extracted by specific peak screening and structured sequence construction, combined with TF-IDF transformation and Welch t test. Feature integration was performed at the local and global levels, and cell type annotation was performed using a classification model.
It achieves comprehensive coverage of major and low-abundance cell types, improves the stability and interpretability of annotation, enables accurate annotation of multiple cell types at the cell atlas level, and reveals the intrinsic correlation patterns of chromatin accessibility signals.
Smart Images
Figure CN122290718A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of single-cell chromatin accessibility data analysis technology, and in particular to a hierarchical method for parsing and annotating single-cell chromatin accessibility data. Background Technology
[0002] Epigenomics, as a crucial field for elucidating gene expression regulation, is widely applied in disease research. With the development of sequencing technology, single-cell chromatin accessibility sequencing has become a core tool for characterizing cellular epigenetic heterogeneity. Among related technologies, an analytical system has been constructed through the synergistic application of multiple techniques. Specifically, this process encompasses open region detection and type identification, including peak selection, dimensionality reduction representation, and cluster annotation.
[0003] However, existing scATAC-seq data parsing methods directly use raw peak information, which does not adequately characterize cell-specific signals. This may lead to the loss of key information, thus affecting the stability and reliability of annotation results. Furthermore, limitations in sequencing depth and noise increase the difficulty of extracting true biological features from sparse signals, and the insufficient interpretability of existing methods restricts their reliability in practical applications. Summary of the Invention
[0004] The main objective of this invention is to provide a hierarchical method for parsing and annotating single-cell chromatin accessibility data.
[0005] Another objective of this invention is to provide a hierarchical device for parsing and annotating single-cell chromatin accessibility data.
[0006] The third objective of this invention is to provide an electronic device.
[0007] To achieve the above objectives, a first aspect of the present invention proposes a hierarchical method for parsing and annotating single-cell chromatin accessibility data, comprising:
[0008] Single-cell chromatin accessibility data were obtained, and genomic region features were screened based on cell type specificity to determine multiple feature sets; The genomic region features of each cell are converted into structured sequences, which contain sub-units corresponding to the plurality of feature sets and identifiers for identifying type information; The structured sequence is processed by a hierarchical neural network model. Local features are extracted within each subunit to obtain a subunit representation, and global features are integrated between subunits to obtain a cell-level representation vector. Based on the cell-level representation vector, the cell type of each cell is predicted by a classification model, and cell type annotation results are generated.
[0009] Optionally, the step of obtaining single-cell chromatin accessibility data and screening genomic region features based on cell type specificity to determine multiple feature sets includes: TF-IDF transformation is applied to the single-cell chromatin accessibility matrix in the single-cell chromatin accessibility data to reduce the weight of non-specific peaks that frequently appear in most cells and amplify the ability of rare peaks to identify specific cell types. For each cell type, the sample cells are divided into two groups: those belonging to that type and those not belonging to that type. A two-sided Welch t test is performed on each peak after TF-IDF conversion to calculate the degree of difference between that cell type and other cells. The peaks are sorted according to the obtained t values. Peaks with high t values indicate that they make a significant contribution to distinguishing this type of cell. These peaks are selected as a specific peak set for constructing subsequent structured sequences.
[0010] Optionally, the process of sorting the peaks according to the obtained t-values, wherein peaks with high t-values indicate that they significantly contribute to distinguishing this type of cell, and selecting them as a specific peak set for constructing subsequent structured sequences, includes: Selecting peaks from the rarest to the most common, in order of frequency, from the peaks that have never been selected. Each peak forms a specific set of peaks for each cell type, balancing the specific information among different cell types; Obtain non-overlapping values for all cell types. The set of peaks remains fixed throughout the model training and inference process to ensure that the constructed structured sequences can effectively reflect the specific chromatin accessibility patterns of each cell type.
[0011] Optionally, the step of converting the genomic region features of each cell into a structured sequence, the structured sequence comprising sub-units corresponding to the plurality of feature sets and identifiers for identifying type information, including: The accessibility data of each cell is converted into a serialized representation called a cell sequence. Each cell sequence consists of multiple sub-units, and each sub-unit corresponds to a set of peaks for a specific cell type to preserve its cell type specificity. For cells The Each subunit is constructed including The token is used to mark the cell type information corresponding to the sub-unit, and serves as a sequence of identifiers for attention calculation in the subsequent transformer module; All sub-units are connected in a fixed order to form a complete cell sequence, so as to transform the sparse accessibility vector of each cell into a serializable, structured input containing type context, providing a standardized input format for subsequent hierarchical feature parsing and cell-level representation learning.
[0012] Optionally, the process of processing the structured sequence using a hierarchical neural network model, extracting local features within sub-units to obtain sub-unit representations, and integrating global features between sub-units to obtain cell-level representation vectors, includes: Each sub-unit is first transformed into a continuous vector representation through an embedding layer. Each token contains not only accessibility information but also its identity information in the peak set, which facilitates the subsequent transformer to capture local dependencies. A peak-level transformer is applied to each clause to calculate attention weights within the same set of peaks to model the relationships between peaks and adaptively assign weights to different peaks to highlight the contribution of key peaks to cell type discrimination. Extract the corresponding output from each clause. The token vector serves as the embedding representation of the cell-type-specific peak set, preparing for subsequent peak-set level integration.
[0013] Optionally, the application of a peak-level transformer on each clause to calculate attention weights within the same peak set, modeling the relationships between peaks to adaptively assign weights to different peaks and highlight the contribution of key peaks to cell type discrimination, includes: Through formula Calculate the key matrix and value matrix of the query matrix; Through formula Calculate attention weights and use the formula We obtain a weighted output to reduce the impact of nonspecific peaks.
[0014] Optionally, the corresponding output is extracted from each clause. The token vector serves as the embedding representation of this cell-type-specific peak set, preparing for subsequent peak-set level integration, including: The overall representation of all sub-units is compared with the global representation. Token embeddings are concatenated to form a peak-set input matrix, which integrates all sub-units at the peak-set level to generate a global representation for each cell; The peak-set input matrix is input into the peak-set-level transformer to perform global attention modeling of the entire peak set to model the dependencies between different cell-type-specific peak sets in order to achieve cell-level feature integration. extract The token's output is used as a cell-level representation for downstream annotation tasks to generate a global representation for each cell and capture interaction patterns across the peak set globally.
[0015] Optionally, the step of predicting the cell type of each cell based on the cell-level representation vector using a classification model and generating cell type annotation results includes: The obtained cell-level representation is input into a single hidden-layer MLP for cell type annotation. The hidden layer dimension of the MLP is set to 64, and the output layer dimension is the number of cell types. ; Through formula Computational cells The predicted probability distribution for each cell type is used to achieve the final cell type annotation and output the multi-class predicted probability for each cell.
[0016] To achieve the above objectives, a second aspect of the present invention provides a hierarchical single-cell chromatin accessibility data parsing and annotation apparatus, comprising: The first module is used to acquire single-cell chromatin accessibility data, screen genomic region features based on cell type specificity, and determine multiple feature sets; The second module is used to convert the genomic region features of each cell into a structured sequence, the structured sequence containing sub-units corresponding to the plurality of feature sets and identifiers for identifying type information; The third module is used to process the structured sequence through a hierarchical neural network model, extract local features within the sub-units to obtain sub-unit representations, and integrate global features between the sub-units to obtain cell-level representation vectors. The fourth module is used to predict the cell type of each cell based on the cell-level representation vector using a classification model, and generate cell type annotation results.
[0017] To achieve the above objectives, a third aspect of this application provides an electronic device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory to implement the method described in the first aspect.
[0018] The embodiments of the present invention have the following beneficial effects: (1) Sequence construction covering all cell types: Through specific peak screening and multi-subunit structured sequence construction, the embodiments of this application achieve comprehensive coverage of major and low-abundance cell types, providing a foundation for subsequent hierarchical analysis and accurate annotation.
[0019] (2) Hierarchical analysis of chromatin accessibility patterns: The embodiments of this application use a hierarchical transformer to capture local accessibility patterns at the peak level and integrate global features at the peak set level, while retaining cell type-specific information.
[0020] (3) Cell atlas-level multi-class cell annotation capability: After training on cell atlas-level scATAC-seq data, the embodiments of this application can directly perform multi-class annotation on new data. Experimental verification shows that this method can cover the annotation of 31 common human cell types and maintain stability in the recognition of various cell types.
[0021] (4) Highly interpretable scATAC-seq data analysis: The embodiments of this application provide attention interpretability information during the parsing and annotation process, revealing the intrinsic relationship between peaks and the interaction patterns between different peak sets, thereby enhancing annotation transparency and biological credibility. Attached Figure Description
[0022] The above-described and additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 A flowchart illustrating a hierarchical method for parsing and annotating single-cell chromatin accessibility data, provided in an embodiment of the present invention; Figure 2 This is a framework diagram of a hierarchical single-cell chromatin accessibility data parsing and annotation method provided in an embodiment of the present invention. Detailed Implementation
[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0025] Epigenomics research aims to elucidate the fine-grained regulation of gene expression through mechanisms such as chromatin structure, accessibility changes, and regulatory element activity, providing a crucial information foundation for studying the molecular mechanisms of cell differentiation, tissue development, and disease progression. Among these, chromatin openness is considered a key epigenetic characteristic reflecting transcriptional regulatory potential. Transposase-mediated chromatin accessibility sequencing (ATAC-seq) technology can detect open chromatin regions across the entire genome, providing an effective means to analyze regulatory elements such as promoters and enhancers and their dynamic changes. With continuous optimization of sequencing technology and experimental procedures, ATAC-seq has evolved from population-level analysis to single-cell resolution, forming single-cell chromatin accessibility sequencing (scATAC-seq) technology. scATAC-seq technology can characterize chromatin openness at the single-cell level, enabling the analysis of epigenetic regulatory heterogeneity among different cells in complex tissues. For computational analysis of scATAC-seq data, existing technologies have proposed a variety of methods to support cell population structure resolution, feature representation learning, and related biological interpretation, providing a computational foundation for mining cellular heterogeneity information from chromatin accessibility signals.
[0026] However, with the increasing scale and complexity of data, existing methods for parsing and annotating scATAC-seq data face several technical challenges. First, the number of cells increases significantly in atlas-level data, while scATAC-seq data itself has a high-dimensional, extremely sparse, and near-binary feature structure, imposing stringent requirements on the scalability and robustness of computational methods. Second, high-dimensional sparse data is prone to loss of key information during feature compression or extraction, especially cell-specific or cell-type-specific accessibility signals, which are difficult to fully characterize. Simultaneously, the highly heterogeneous cell type composition in complex tissues, and the significant abundance imbalance among different cell types, often leads existing methods to favor the dominant cell populations in the sample, limiting the identification and differentiation capabilities of low-abundance cell types. Furthermore, the difficulty of extracting true biological features from sparse signals is further increased due to limitations in sequencing depth and experimental noise, thus affecting the stability and reliability of annotation results. Finally, the insufficient interpretability of existing methods during the parsing process also limits their credibility and generalizability in biological research and practical applications.
[0027] There is an urgent need to develop a scATAC-seq data parsing and annotation method that can operate stably on cell atlas-level data while maintaining the ability to identify low-abundance populations and interpretability of the inference process. To address these issues, this invention proposes a hierarchical single-cell chromatin accessibility data parsing and annotation method. This method parses scATAC-seq data hierarchically, extracts chromatin accessibility features related to cell type, and achieves the parsing of cell type-specific information. Even in the presence of cell type abundance imbalances, it achieves stable annotation for both major and low-abundance cell types. After training on cell atlas-level scATAC-seq data, it can annotate multiple cell types in unseen data without additional training. Furthermore, the method reveals the intrinsic structural patterns of chromatin accessibility signals during parsing and annotation, improving the interpretability of the annotation process.
[0028] The following description, with reference to the accompanying drawings, describes a hierarchical method and apparatus for parsing and annotating single-cell chromatin accessibility data according to an embodiment of the present invention.
[0029] Example 1 This embodiment provides a hierarchical method for parsing and annotating single-cell chromatin accessibility data. For example... Figure 1 and Figure 2 As shown, the method includes the following steps: S1: Obtain single-cell chromatin accessibility data, screen genomic region features based on cell type specificity, and determine multiple feature sets.
[0030] In this embodiment of the invention, step S1 is achieved through the following steps: applying TF-IDF transformation to the single-cell chromatin accessibility matrix in the single-cell chromatin accessibility data to reduce the weight of non-specific peaks that frequently appear in most cells and amplify the ability of rare peaks to identify specific cell types; for each cell type, dividing the sample cells into two groups, belonging to that type and not belonging to that type, and performing a two-sided Welch t test on each peak after TF-IDF transformation to calculate its degree of difference from other cells in that cell type; sorting the peaks according to the obtained t values and using the peaks with high t values, indicating that they have made significant contributions to distinguishing that type of cell, as a specific peak set for constructing subsequent structured sequences.
[0031] Specifically, in the scATAC-seq data, the single-cell chromatin accessibility matrix It typically exhibits high-dimensional, extremely sparse, and near-binary structural features. Indicates the number of cells. (This represents the number of peaks). This data characteristic makes it easy to introduce noise when directly using raw peak information for downstream analysis, and it is difficult to fully capture cell type-specific signals. Therefore, in this embodiment, peaks are first filtered in the training data for each cell type to generate a specific peak set for constructing subsequent structured sequences. Specifically, the matrix is first... TF-IDF (Term Frequency-Inverse Document Frequency) transformation is applied to reduce the weight of non-specific peaks that frequently occur in most cells and amplify the ability of rare peaks to identify specific cell types.
[0032] Subsequently, for each cell type, the sample cells were divided into two groups: those belonging to that type and those not belonging to that type. A two-sided Welch t-test was performed on each peak after TF-IDF conversion to calculate its degree of difference from other cells within that cell type. Peaks were ranked according to the obtained t-values; peaks with high t-values indicated that they significantly contributed to distinguishing that cell type. To further balance the specific information between different cell types, this embodiment of the application uses a sequence from rare to common, sequentially selecting the most common peaks from the unselected peaks. Each peak forms a specific set of peaks for each cell type.
[0033] Ultimately, all cell types obtained non-overlapping [cell types]. A set of peaks is maintained throughout the model training and inference process, ensuring that the constructed structured sequences effectively reflect the specific chromatin accessibility patterns of each cell type. This step guarantees that only the most cell type-specific features are retained in the sparse, high-dimensional scATAC-seq data, providing a foundation for subsequent construction of structured sequences and, to some extent, suppressing the impact of noise on model performance.
[0034] S2, convert the genomic region features of each cell into a structured sequence, the structured sequence containing sub-units corresponding to the plurality of feature sets and identifiers for identifying type information.
[0035] In this embodiment of the invention, step S2 is achieved through the following steps: converting the accessibility data of each cell into a serialized representation called a cell sequence, wherein each cell sequence consists of multiple sub-units and each sub-unit corresponds to a peak set of a specific cell type to preserve its cell type specificity; for cells The Each subunit is constructed including The token is a sequence of identifiers used to label the cell type information corresponding to the sub-unit, serving as the identifier for attention calculation in the subsequent transformer module. All sub-units are connected in a fixed order to form a complete cell sequence, thereby transforming the sparse accessibility vector of each cell into a serializable, structured input containing type context, providing a standardized input format for subsequent hierarchical feature parsing and cell-level representation learning.
[0036] Specifically, after completing peak screening, this embodiment of the application will use the accessibility data of each cell. This is converted into a serialized representation called a cell sequence. Each cell sequence consists of multiple sub-units, each corresponding to a set of peaks for a specific cell type. This structured representation effectively preserves the accessibility information of each cell across different cell type peaks and introduces explicit cell type context during the serialization process. Specifically, for cells... The Each subunit is used to construct the following sequence:
[0037] in Indicates the first A peak in the cell Accessible in the middle Indicates something is unattainable; The token is used to identify the cell type information corresponding to the sub-unit, serving as an identifier for attention calculation in the subsequent transformer module. All sub-units are connected in a fixed order to form a complete cell sequence.
[0038] Through this step, the embodiments of this application transform the sparse accessibility vector of each cell into a serializable, structured input that includes type context, providing a standardized input format for subsequent hierarchical feature parsing and cell-level representation learning, while maintaining the traceability of information specific to different cell types.
[0039] S3, the structured sequence is processed by a hierarchical neural network model, local feature extraction is performed within the sub-unit to obtain the sub-unit representation, and global feature integration is performed between the sub-units to obtain the cell-level representation vector.
[0040] In this embodiment of the invention, step S3 is implemented through the following steps: Each sub-unit is first transformed into a continuous vector representation through an embedding layer, and each token contains not only accessibility information but also its identity information within the peak set, facilitating subsequent transformer capture of local dependencies; a peak-level transformer is applied to each clause to calculate attention weights within the same peak set, modeling the relationships between peaks to adaptively assign weights to different peaks, highlighting the contribution of key peaks to cell type discrimination; the corresponding [value / value] is extracted from the output of each clause. The token vector serves as the embedding representation of the cell-type-specific peak set, preparing for subsequent peak-set level integration.
[0041] Specifically, after constructing the structured cell sequence, this embodiment of the application analyzes each sub-unit at the peak level to capture local chromatin accessibility patterns and correlations between peaks. Specifically, each sub-unit... First, it is transformed into a continuous vector representation through an embedding layer:
[0042] in, Mapping binary accessibility signals to vectors Encode the peak index information into a vector. This is the embedding dimension. In this way, each token not only contains accessibility information, but also its identity information in the peak set, which facilitates the subsequent capture of local dependencies by the transformer.
[0043] Subsequently, in this embodiment of the application, a peak-level transformer is applied to each clause to calculate attention weights within the same set of peaks and model the relationships between peaks:
[0044]
[0045] In this way, the model can adaptively assign weights to different peaks, highlighting the contribution of key peaks to cell type discrimination while reducing the influence of non-specific peaks. Finally, the corresponding values are extracted from the output of each clause. The vector of tokens This serves as an embedded representation of the cell-type-specific peak set, preparing for subsequent peak-set level integration.
[0046] After completing the local representation of each sub-unit, this embodiment integrates all sub-units at the peak-set level to generate a global representation of each cell. First, the overall representation of all sub-units is compared with the global representation. token embedding representation Concatenate to form the peak-set input matrix:
[0047] Then, the input is fed into a peak-set-level transformer, which performs global attention on the entire peak set, modeling the dependencies between different cell-type-specific peak sets:
[0048] Through this layer, the embodiments of this application can capture interaction patterns across peak sets globally, achieving cell-level feature integration. Finally, extraction... output of token As a cell-level representation, it is used for downstream annotation tasks.
[0049] S4. Based on the cell-level representation vector, the cell type of each cell is predicted by the classification model, and cell type annotation results are generated.
[0050] In this embodiment of the invention, step S4 is achieved through the following steps: the obtained cell-level representation is input into a single hidden-layer MLP, cell type annotation is performed, and the hidden layer dimension of the MLP is set to 64 and the output layer dimension is the number of cell types. ; through formula Computational cells The predicted probability distribution for each cell type is used to achieve the final cell type annotation and output the multi-class predicted probability for each cell.
[0051] Specifically, the cell-level representation obtained in the embodiments of this application Cell type annotation is performed using a single-hidden-layer MLP (Multi-Layer Perceptron). The hidden layer dimension of the MLP is set to 64, and the output layer dimension is the number of cell types. :
[0052] in Represents cells Predicted probability distribution for each cell type.
[0053] Example 2 This invention also provides a device for parsing and annotating single-cell chromatin accessibility data, the device comprising: The first module is used to acquire single-cell chromatin accessibility data, screen genomic region features based on cell type specificity, and determine multiple feature sets; The second module is used to convert the genomic region features of each cell into a structured sequence, the structured sequence containing sub-units corresponding to the plurality of feature sets and identifiers for identifying type information; The third module is used to process the structured sequence through a hierarchical neural network model, extract local features within the sub-units to obtain sub-unit representations, and integrate global features between the sub-units to obtain cell-level representation vectors. The fourth module is used to predict the cell type of each cell based on the cell-level representation vector using a classification model, and generate cell type annotation results.
[0054] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0055] Example 3 To implement the methods of the above embodiments, the present invention also provides an electronic device, which includes a memory and a processor; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the various steps of the methods described above.
[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
[0057] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0058] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. A hierarchical method for parsing and annotating single-cell chromatin accessibility data, characterized in that, Includes the following steps: Single-cell chromatin accessibility data were obtained, and genomic region features were screened based on cell type specificity to determine multiple feature sets; The genomic region features of each cell are converted into structured sequences, which contain sub-units corresponding to the plurality of feature sets and identifiers for identifying type information; The structured sequence is processed by a hierarchical neural network model. Local features are extracted within each subunit to obtain a subunit representation, and global features are integrated between subunits to obtain a cell-level representation vector. Based on the cell-level representation vector, the cell type of each cell is predicted by a classification model, and cell type annotation results are generated.
2. The method as described in claim 1, characterized in that, The process of obtaining single-cell chromatin accessibility data involves screening genomic region features based on cell type specificity to determine multiple feature sets, including: TF-IDF transformation is applied to the single-cell chromatin accessibility matrix in the single-cell chromatin accessibility data to reduce the weight of non-specific peaks that frequently appear in most cells and amplify the ability of rare peaks to identify specific cell types. For each cell type, the sample cells are divided into two groups: those belonging to that type and those not belonging to that type. A two-sided Welch t test is performed on each peak after TF-IDF conversion to calculate the degree of difference between that cell type and other cells. The peaks were sorted according to the obtained t-values. Peaks with high t-values indicated that they made a significant contribution to distinguishing this type of cell and were selected as a specific peak set for constructing subsequent structured sequences.
3. The method as described in claim 2, characterized in that, The peaks are sorted according to the obtained t-values, where peaks with high t-values indicate that they significantly contribute to distinguishing this type of cell. These peaks are selected as a specific peak set for constructing subsequent structured sequences, including: Selecting peaks from the rarest to the most common, in order of frequency, from the peaks that have never been selected. Each peak forms a specific set of peaks for each cell type, balancing the specific information among different cell types; All cell types get disjoint sets of peaks The sets of peaks remain fixed throughout the model training and inference process to ensure that the structured sequences built effectively reflect the specific chromatin accessibility patterns of each cell type.
4. The method of claim 1, wherein, The process of converting the genomic region features of each cell into a structured sequence, wherein the structured sequence contains sub-units corresponding to the plurality of feature sets and identifiers for identifying type information, including: The accessibility data of each cell is converted into a serialized representation called a cell sequence, where each cell sequence consists of multiple sub-units, and each sub-unit corresponds to a set of peaks for a specific cell type to preserve its cell type specificity; For cells The Each subunit is constructed including The token is used to mark the cell type information corresponding to the sub-unit, and serves as a sequence of identifiers for attention calculation in the subsequent transformer module; All sub-units are connected in a fixed order to form a complete cell sequence, so as to transform the sparse accessibility vector of each cell into a serializable, structured input containing type context, providing a standardized input format for subsequent hierarchical feature parsing and cell-level representation learning.
5. The method as described in claim 1, characterized in that, The process of processing the structured sequence using a hierarchical neural network model involves extracting local features within each sub-unit to obtain a sub-unit representation, and integrating global features between sub-units to obtain a cell-level representation vector, including: Each sub-unit is first transformed into a continuous vector representation through an embedding layer. Each token contains not only accessibility information but also its identity information in the peak set, which facilitates the subsequent transformer to capture local dependencies. A peak-level transformer is applied to each clause to calculate attention weights within the same set of peaks to model the relationships between peaks and adaptively assign weights to different peaks to highlight the contribution of key peaks to cell type discrimination. From the output of each clause, a vector of corresponding tokens is extracted as an embedded representation of the cell-type-specific peak set in preparation for subsequent peak-set level integration.
6. The method of claim 5, wherein, The process involves applying a peak-level transformer to each clause, calculating attention weights within the same peak set to model the relationships between peaks, and adaptively assigning weights to different peaks to highlight the contribution of key peaks to cell type discrimination. This includes: The query matrix is calculated by the formula and the value matrix is calculated by the formula Through formula Calculate attention weights and use the formula We obtain a weighted output to reduce the impact of nonspecific peaks.
7. The method of claim 5, wherein, The corresponding output is extracted from each clause. The token vector serves as the embedding representation of this cell-type-specific peak set, preparing for subsequent peak-set level integration, including: The overall representation of all sub-units is compared with the global representation. Token embeddings are concatenated to form a peak-set input matrix, which integrates all sub-units at the peak-set level to generate a global representation for each cell; The peak-set input matrix is input into the peak-set-level transformer to perform global attention modeling of the entire peak set to model the dependencies between different cell-type-specific peak sets in order to achieve cell-level feature integration. extract The output of the token is used as a cell-level representation for downstream annotation tasks to generate a global representation for each cell and capture interaction patterns across the peak set globally.
8. The method as described in claim 1, characterized in that, The process of predicting the cell type of each cell using a classification model based on the cell-level representation vector, and generating cell type annotation results, includes: The obtained cell-level representation is input into a single hidden-layer MLP for cell type annotation, where the hidden layer dimension of the MLP is set to 64 and the output layer dimension is the number of cell types. ; Through formula Computational cells The predicted probability distribution for each cell type is used to achieve the final cell type annotation and output the multi-class predicted probability for each cell.
9. A hierarchical single-cell chromatin accessibility data analysis and annotation apparatus, characterized by, include: The first module is used to acquire single-cell chromatin accessibility data, screen genomic region features based on cell type specificity, and determine multiple feature sets; The second module is used to convert the genomic region features of each cell into a structured sequence, the structured sequence containing sub-units corresponding to the plurality of feature sets and identifiers for identifying type information; The third module is used to process the structured sequence through a hierarchical neural network model, extract local features within the sub-units to obtain sub-unit representations, and integrate global features between the sub-units to obtain cell-level representation vectors. The fourth module is used to predict the cell type of each cell based on the cell-level representation vector using a classification model, and generate cell type annotation results.
10. An electronic device, comprising: Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the method as described in any one of claims 1-8.