A method for integrating single-cell multi-omics mismatched data based on prior constraints
By employing a priori constraint-based single-cell multi-omics mismatched data integration method, and utilizing RNA and ATAC heterogeneous encoders to extract features and construct a Gene-Peak relationship graph, this approach solves the semantic mismatch and cell type alignment bias problems in cross-modal integration in existing technologies, achieving high-precision multi-omics data alignment and preservation of cell heterogeneity.
Patent Information
- Application Number
- CN202611083745.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to fully model cross-modal deep regulatory semantics when integrating multi-omics data, leading to semantic mismatches and cell type alignment biases. Furthermore, they face modeling difficulties and noise interference in high-dimensional, extremely sparse ATAC scenarios, resulting in insufficient cell type differentiation capabilities.
We employ a prior-constrained single-cell multi-omics non-matching data integration method. We extract features using RNA heterogeneous encoders and ATAC heterogeneous encoders respectively, construct a Gene-Peak relationship graph, and use multi-head cross-modal attention features and a prior-constrained Transformer model for cross-modal integration. We combine reconstruction loss and adversarial loss for training to ensure modality alignment and preservation of cell heterogeneity.
It achieves high-precision multi-omics data alignment in complex scenarios, maintains cellular heterogeneity, improves cell type differentiation and tag transfer accuracy, and effectively identifies cellular heterogeneity under different subtypes and conditions.
Smart Images

Figure CN123120053A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computational biology, specifically to a method for integrating single-cell multi-omics non-matching data based on prior constraints. Background Technology
[0002] In related technologies, existing techniques, when integrating multi-omics data, mostly rely on gene activity alignment based on ATAC (Assay for Transposase-Accessible Chromatin) data. This makes it difficult to fully model deep regulatory semantics across modalities, and is prone to semantic mismatches and cell type alignment biases. In particular, compressing ATAC peak signals into gene-level features can also result in the loss of distant regulatory information, affecting the ability to distinguish fine-grained cell states.
[0003] While some existing methods can integrate peak-level information, they still face modeling difficulties and noise interference in high-dimensional, extremely sparse ATAC scenarios, and are insufficient in characterizing complex topological relationships. Furthermore, they are prone to clustering distortion or cell aliasing in imbalanced cell populations and cross-conditional tasks. It should be noted that the information disclosed in the above background section is only for enhancing the understanding of the background of this invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0004] This invention provides a method for integrating single-cell multi-omics non-matching data based on prior constraints. It can achieve high-precision alignment while preserving cellular heterogeneity and maintaining stable performance in complex scenarios such as imbalance and cross-conditions, thus achieving accurate and efficient multi-omics data integration and overcoming the shortcomings of existing technologies to a certain extent.
[0005] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.
[0006] According to a first aspect of the present invention, a method for integrating single-cell multi-omics mismatched data based on prior constraints is provided, the method comprising: Acquire the unmatched single-cell multi-omics data to be processed, and preprocess the data according to preset rules to obtain the dataset after screening out abnormal data; the unmatched single-cell multi-omics data to be processed includes: unmatched transcriptome data scRNA-seq and chromatin accessibility data scATAC-seq. Genes and open chromatin peaks in the dataset are extracted based on the regulatory relationship between genes and chromatin accessibility. Connection weights are configured according to the regulatory strength between Genes and Peaks to construct a Gene-Peak relationship graph. RNA features were extracted from transcriptome RNA data in the dataset using an RNA heterogeneous encoder; and ATAC features were extracted from ATAC data in the dataset using an ATAC heterogeneous encoder. RNA features were used to represent continuous semantic features of gene expression; ATAC features were used to represent high-dimensional sparse, near-binary peak signals and accessibility features. RNA features and ATAC features are mapped to a shared latent space to form an initial cross-modal low-dimensional embedding. The RNA low-dimensional embedding, ATAC low-dimensional embedding, and Gene-Peak relationship graph are then input into a data integration model based on prior constraints for cross-modal integration to obtain integrated data.
[0007] In some exemplary embodiments, preprocessing is performed according to preset rules to obtain a dataset after filtering out abnormal data, including: To complete the ATAC data filtering, remove low-quality cells with a sequencing depth of less than 1800 and / or more than 100,000 abnormal cells from the single-cell chromatin accessibility scATAC-seq data. The filtered ATAC data is then subjected to linear dimensionality reduction using the Latent Semantic Indexing (LSI) method; and... To filter RNA data, remove low-quality cells with fewer than 1,000 unique molecular identifiers and / or more than 25,000 abnormal cells from single-cell transcriptome scRNA-seq data; remove cells with abnormal chromatin structure and nucleosome signal exceeding 2; and remove cells with poor chromatin openness and transcription start site enrichment score not exceeding 1. The filtered RNA data were then normalized and logarithmized sequentially, and linear dimensionality reduction was performed using principal component analysis (PCA).
[0008] In some exemplary embodiments, RNA low-dimensional embeddings, ATAC low-dimensional embeddings, and Gene-Peak relationship graphs are input into a prior-constrained data integration model for cross-modal integration, including: Using RNA low-dimensional embedding as the query vector and ATAC low-dimensional embedding as the key and value vectors, a multi-head cross-modal attention feature is established. The Gene-Peak relationship graph is converted into an attention mask, and the attention weight matrix after prior constraints is determined by combining multi-head cross-modal attention features, including:
[0009] in, This represents the genes defined in the attention mask. and open chromatin peaks The potential for regulation between them; Represents cells The low-dimensional embedding obtained by the RNA characterization after passing through a heterogeneous encoder; Represents cells The ATAC representation is used to obtain a low-dimensional embedding after passing through a heterogeneous encoder. Using the attention weight matrix with prior constraints, feature fusion is performed on the value vector and query vector to obtain the initial fused representation. ; The initial fused representation is used to maintain the semantic stability of the original modality through residual connections and layer normalization to obtain the final fused representation. ,include:
[0010] Wherein, FFN represents a feedforward network; Representation layer normalization.
[0011] In some exemplary embodiments, the method further includes: training a data integration model based on prior constraints, including: Single-cell multi-omics mismatched data were obtained from the sequencing dataset and preprocessed according to preset rules to obtain a dataset after filtering out abnormal data. The single-cell multi-omics mismatched data included mismatched transcriptome data scRNA-seq and chromatin accessibility data scATAC-seq. Genes and open chromatin peaks in the dataset are extracted based on the regulatory relationship between genes and chromatin accessibility. Connection weights are configured according to the regulatory strength between Genes and Peaks to construct a Gene-Peak relationship graph. An RNA heterogeneous encoder was constructed, and features were extracted from the transcriptomic RNA data in the dataset to obtain RNA features; and, An ATAC heterogeneous encoder is constructed, and features are extracted from the ATAC data in the dataset to obtain ATAC features. Among them, RNA features are used to represent continuous semantic features of gene expression, and ATAC features are used to represent high-dimensional sparse, near-binary peak signals and accessibility features. RNA features and ATAC features are mapped to a shared latent space to form an initial cross-modal low-dimensional embedding. A reconstruction loss is constructed to decode and reconstruct the RNA and ATAC modalities separately, in order to constrain the potential representations to retain their respective key biological information; The initial model is input with RNA low-dimensional embedding, ATAC low-dimensional embedding and Gene-Peak relationship graph. The RNA low-dimensional embedding is used as the query vector and the ATAC low-dimensional embedding is used as the key vector and value vector to establish multi-head cross-modal attention features. The Gene-Peak relationship graph is converted into an attention mask, and the attention weight matrix with prior constraints is determined by combining multi-head cross-modal attention features. Using the attention weight matrix with prior constraints, feature fusion is performed on the value vector and query vector to obtain the initial fused representation. ; The initial fused representation is used to maintain the semantic stability of the original modality through residual connections and layer normalization to obtain the final fused representation. ; Construct an adversarial loss function to bridge the gap between the two modalities; backpropagate the initial model based on the adversarial loss function until the model converges to obtain a trained data integration model based on prior constraints. The distance between RNA data and ATAC data was calculated using the k-nearest neighbor method, the data were classified using majority voting, and the labels were transferred based on the classification results.
[0012] In some exemplary implementations, the reconstruction loss includes:
[0013] in, Represents cells Characterization on RNA; Represents cells Characterization in the characterization view and chromatin accessibility view on ATAC; Represents cells The low-dimensional embedding obtained by the RNA characterization after passing through a heterogeneous encoder; Represents cells The ATAC representation is used to obtain a low-dimensional embedding after passing through a heterogeneous encoder.
[0014] In some exemplary implementations, an adversarial loss function is constructed to bridge the gap between two modalities, including: Introducing a domain discriminant in a shared latent representation space RNA modality samples are labeled as the source domain and ATAC modality samples as the target domain. A binary classification cross-entropy domain adversarial loss is used to measure the separability of the two modality representations, expressed by the formula:
[0015] in, This represents the final fusion representation. It expresses expectation.
[0016] According to a second aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the above-described method for integrating single-cell multi-omics non-matching data based on prior constraints is implemented.
[0017] According to a third aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the above-described method for integrating single-cell multi-omics non-matching data based on prior constraints by executing the executable instructions.
[0018] According to a fourth aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for integrating single-cell multi-omics non-matching data based on prior constraints.
[0019] The single-cell multi-omics mismatched data integration method based on prior constraints provided in the embodiments of the present invention directly uses ATAC Peak data during the data integration process, and uses heterogeneous encoders to extract features from transcriptome RNA data and ATAC data respectively. This enables the RNA heterogeneous encoder to extract continuous semantic representations of gene expression, while the ATAC encoder extracts high-dimensional sparse, near-binary peak signals and accessibility representations. Simultaneously, a Gene-Peak relationship graph is constructed to guide potential intercellular connections, thus providing important guidance for modality alignment and distinguishing cellular heterogeneity. The data integration model aligns modalities by designing prior constraints, effectively preserving cell specificity while achieving modality integration. This invention achieves higher alignment accuracy in integrating single-cell multi-omics mismatched data and can clearly distinguish cells of different subtypes and under different conditions, effectively identifying cellular heterogeneity.
[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description
[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.
[0022] Figure 1 The illustration shows a schematic diagram of an exemplary embodiment of the present invention, namely, a method for integrating single-cell multi-omics mismatched data based on prior constraints; Figure 2 The diagram illustrates a model training method flow according to an exemplary embodiment of the present invention. Figure 3The diagram illustrates a model architecture of an exemplary embodiment of the present invention. Figure 4 This diagram illustrates the visualization results of integrating transcriptome and chromatin accessibility data UMAP using the present invention and existing GLUE methods. Figure 5 The diagram illustrates the composition of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation
[0023] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.
[0024] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.
[0025] In related technologies, due to significant differences in feature space, signal sparsity, and complex cross-modal regulatory relationships among unpaired data from different omics, existing methods struggle to fully identify potential differences and associations between cells, making it difficult to achieve high-quality alignment during the integration of multi-omics data. Furthermore, due to the complex representation of cellular features, existing methods fail to preserve the characteristic differences between different cells while aligning data, resulting in poor heterogeneity discrimination capabilities.
[0026] For example, some technical solutions have implemented SCALEX (Single-Cell Atlas Exploration), a method focused on aligning single-cell transcriptome data across batches, platforms, and species. Its core employs a variational autoencoder (VAE) framework, decoupling biological variation and batch effects to map data from different sources into a unified low-dimensional space. SCALEX introduces an adaptive weighting mechanism to dynamically balance the intensity of batch correction and biological signal preservation, thus avoiding the loss of true biological differences due to overcorrection. However, this approach is highly dependent on the quality of preprocessing steps for handling highly sparse data and has limited scalability in multimodal integration. Other solutions implement the uniPort unified embedding model. This approach aims to map different modalities into a shared latent space through a hybrid architecture of generative adversarial networks and variational autoencoders. This model can achieve cross-omics cell type alignment, missing modality completion, and joint multimodal dimensionality reduction, significantly improving the robustness of cross-technology data integration. However, it still relies on optimized preprocessing steps when handling highly sparse or noisy data. In addition, existing cross-modal integration methods for scRNA-seq and scATAC-seq, such as GLUE (Graph-Linked Unified Embedding), emphasize leveraging biological prior knowledge to improve alignment reliability. This method introduces a peak-gene guidance graph, incorporating inter-modal feature associations into a unified graph learning framework, balancing structural constraints and representation learning during cross-modal mapping. Its advantages lie in strong interpretability and stable performance on routine integration tasks. However, a significant drawback is the increasing computational cost in scenarios with ultra-high-dimensional peak features. Furthermore, the existing Cobolt method aims to achieve consistent representations of different omics from a unified latent space. This method is based on a multimodal variational autoencoder framework, achieving cross-modal embedding and clustering analysis through joint probabilistic modeling. It exhibits good versatility in multi-data source fusion and can support downstream cell state analysis. However, when faced with extremely sparse and near-binary ATAC features, the quality of the integrated low-dimensional representation may be limited, and model training and feature selection strategies have a significant impact on the final results. In addition, the existing LIGER (LinkedInference of Genomic Experimental Relationships) method is based on integrative non-negative matrix factorization (iNMF), which achieves cross-data alignment by decomposing shared factors and data-specific factors.Its method is intuitive and its engineering implementation is mature, making it widely used in gene-level integration tasks. However, LIGER typically relies on converting ATAC into gene activity for alignment, which may result in the loss of peak-level information such as distant regulation, and it suffers from insufficient expressive power in complex cross-modal semantic mapping.
[0027] To address the shortcomings and deficiencies of existing technologies, this example implementation provides a method for integrating mismatched single-cell multi-omics data based on prior constraints. (Reference) Figure 1 As shown, the method may specifically include: Step S11: Obtain the unmatched single-cell multi-omics data to be processed, and preprocess the data to be processed according to preset rules to obtain the dataset after screening out abnormal data; wherein, the unmatched single-cell multi-omics data to be processed includes: unmatched transcriptome data scRNA-seq and chromatin accessibility data scATAC-seq. Step S12: Extract the Gene and Peak from the dataset based on the regulatory relationship between Gene and chromatin accessibility, and configure the connection weights according to the regulatory strength between Gene and Peak to construct the Gene-Peak relationship graph. Step S13: Use an RNA heterogeneous encoder to extract features from the transcriptome RNA data in the dataset to obtain RNA features; and use an ATAC heterogeneous encoder to extract features from the ATAC data in the dataset to obtain ATAC features; wherein, RNA features are used to represent continuous semantic features of gene expression; ATAC features are used to represent high-dimensional sparse, near-binary peak signals and accessibility features. Step S14: Map RNA features and ATAC features to a shared latent space to form an initial cross-modal low-dimensional embedding; and input the RNA low-dimensional embedding, ATAC low-dimensional embedding, and Gene-Peak relationship graph into a data integration model based on prior constraints for cross-modal integration to obtain integrated data.
[0028] The following will describe in more detail each step of the single-cell multi-omics non-matching data integration method based on prior constraints in this exemplary embodiment, with reference to the accompanying drawings and embodiments.
[0029] In this example implementation, the above method may include: pre-training a data integration model based on prior constraints, including: Step S101: Obtain single-cell multi-omics non-matching data from the sequencing dataset and preprocess it according to preset rules to obtain a dataset after screening out abnormal data; wherein, single-cell multi-omics non-matching data includes: non-matching transcriptome data scRNA-seq and chromatin accessibility data scATAC-seq. Step S102: Extract the Gene and Peak from the dataset based on the regulatory relationship between Gene and chromatin accessibility, and configure the connection weights according to the regulatory strength between Gene and Peak to construct the Gene-Peak relationship graph. Step S103: Construct an RNA heterogeneous encoder and use it to extract features from the transcriptome RNA data in the dataset to obtain RNA features; and, Step S104: Construct an ATAC heterogeneous encoder and use it to extract features from the ATAC data in the dataset to obtain ATAC features. Among them, RNA features are used to represent continuous semantic features of gene expression; ATAC features are used to represent high-dimensional sparse, near-binary peak signals and accessibility features. Step S105: Map RNA features and ATAC features to the shared latent space to form an initial cross-modal low-dimensional embedding; Step S106: Construct reconstruction loss, decode and reconstruct the RNA and ATAC modalities respectively, in order to constrain the potential representation to retain their respective key biological information; Step S107: Input the RNA low-dimensional embedding, ATAC low-dimensional embedding and Gene-Peak relationship graph into the initial model, use the RNA low-dimensional embedding as the query vector, and use the ATAC low-dimensional embedding as the key vector and value vector to establish multi-head cross-modal attention features. Step S108: Convert the Gene-Peak relationship graph into an attention mask, and combine it with multi-head cross-modal attention features to determine the attention weight matrix after prior constraints. Step S109: Using the attention weight matrix with prior constraints, feature fusion is performed on the value vector and query vector to obtain the initial fused representation. ; Step S110: For the initial fused representation, the final fused representation is obtained by maintaining the semantic stability of the original modality through residual connections and layer normalization. ; Step S111: Construct an adversarial loss function to narrow the gap between the two modalities; backpropagate the initial model based on the adversarial loss function until the model converges to obtain a trained data integration model based on prior constraints. Step S112: Calculate the distance between RNA data and ATAC data using the k-nearest neighbor method, classify the data using majority voting, and perform label transfer based on the classification results.
[0030] Specifically, users can pre-train single-cell multi-omics mismatched data fusion models based on prior constraints on user-side smart terminals or server-side devices. (Reference) Figure 2As shown, non-matching transcriptome data (scRNA-seq) and chromatin accessibility data (scATAC-seq) can first be obtained from the sequencing dataset as raw data. Then, the raw data can be preprocessed to remove abnormal cells, thus filtering out low-quality data.
[0031] Specifically, data preprocessing may include: Step S21: Delete low-quality cells with a sequencing depth of less than 1800 and / or abnormal cells with a sequencing depth of more than 100,000 in the single-cell chromatin accessibility scATAC-seq data to complete ATAC data filtering. Step S22, and perform linear dimensionality reduction on the filtered ATAC data using the Latent Semantic Indexing (LSI) method; and, Step S23: Delete low-quality cells with fewer than 1,000 unique molecular identifiers and / or more than 25,000 abnormal cells in the single-cell transcriptome scRNA-seq data; delete cells with abnormal chromatin structure and nucleosome signal greater than 2; delete cells with poor chromatin openness and transcription start site enrichment score not exceeding 1, in order to complete the filtering of RNA data. Step S24 involves performing normalization and logarithmic transformation on the filtered RNA data, followed by linear dimensionality reduction using principal component analysis (PCA).
[0032] Of course, in other exemplary embodiments of this disclosure, the sequencing depth and unique molecular identifiers described above can also be set to other specific values according to actual business needs, and this invention does not impose any special limitations on them.
[0033] For example, based on the regulatory relationship between genes and chromatin accessibility, all Genes and Peaks in the data features are collected, and weights are calculated according to the regulatory strength to construct a Gene-Peak relationship graph.
[0034] Specifically, it can be based on the potential regulatory relationship between genes and chromatin accessibility, by statistically analyzing all Genes and Peaks as feature nodes, and configuring edge weights according to the predicted regulatory strength, thereby constructing a Gene-Peak relationship graph.
[0035] For example, the regulatory strength between Gene and Peak can be configured using genomic distance decay weights, co-accessibility scores (association strength), and / or Hi-C chromatin interaction contact frequency.
[0036] For example, an RNA heterogeneous encoder is constructed to extract features from the screened transcriptome RNA data, which is used to extract continuous semantic representations of gene expression.
[0037] Meanwhile, an ATAC heterogeneous encoder is constructed to extract features from the filtered ATAC data, which is used to extract high-dimensional sparse, near-binary peak signals and accessibility representations.
[0038] Then, the outputs of the two heterogeneous encoders can be mapped to a shared latent space to form an initial cross-modal low-dimensional embedding.
[0039] A reconstruction loss is established, and the RNA and ATAC modalities are decoded and reconstructed separately, constraining the latent representations to retain their respective key biological information. The reconstruction loss formula is expressed as:
[0040] in, Represents cells Characterization on RNA; Represents cells Characterization in the characterization view and chromatin accessibility view on ATAC; Represents cells The low-dimensional embedding obtained by the RNA characterization after passing through a heterogeneous encoder; Represents cells The ATAC representation is used to obtain a low-dimensional embedding after passing through a heterogeneous encoder.
[0041] Specifically, RNA heterogeneous encoders and ATAC heterogeneous encoders can employ multilayer perceptron (MLP) encoders to map high-dimensional modality-specific features to a low-dimensional latent embedding space. The corresponding formula is expressed as:
[0042]
[0043] in, and These are the weight matrix and bias parameters for each modal encoder; This represents the nonlinear activation function (ReLU).
[0044] For example, a cross-modal Transformer ensemble model based on prior constraints can be constructed. Specifically, a multi-head cross-modal attention mechanism can be established using RNA low-dimensional embeddings as query vectors and ATAC low-dimensional embeddings as key and value vectors.
[0045] The gene-peak graph is used as prior data and transformed into an attention mask to suppress attention connections without biological basis. For example, if there is an edge between gene g and peak p, then they have a candidate regulatory association. Specifically, if there is a connection between gene g and peak p, it is set to 1; otherwise, it is set to 0; and an attention mask matrix is constructed accordingly.
[0046] Construct a prior-constrained attention loss, and constrain the attention distribution to the prior. Figure 1 To ensure that cross-modal associations satisfy the gene-peak regulation relationship, the attention weight matrix with prior constraints is obtained. The formula is expressed as:
[0047] in, This represents the genes defined in the attention mask. and open chromatin peaks The potential for regulation between them; Represents cells The low-dimensional embedding obtained by the RNA characterization after passing through a heterogeneous encoder; Represents cells The ATAC representation is used to obtain a low-dimensional embedding after passing through a heterogeneous encoder. refer to Figure 3 As shown, the initial fusion representation is obtained by weighting and summing the value vectors of the Transformer model using the acquired attention weights. The final fused representation is obtained by maintaining the semantic stability of the original modality through residual connections and layer normalization. The formula is:
[0048] Wherein, FFN represents a feedforward network; Representation layer normalization.
[0049] Construct an adversarial loss and introduce a domain discriminator in a shared latent representation space. RNA modality samples are labeled as the source domain and ATAC modality samples as the target domain. A binary classification cross-entropy domain adversarial loss is used to measure the separability of the two modality representations; its expression is:
[0050] in, This represents the final fusion representation. It expresses expectation.
[0051] For example, the low-dimensional mapping of filtered and processed transcriptome RNA data and chromatin accessibility ATAC data, the Gene-Peak relationship graph, is input into the integration model. Adversarial loss is used to narrow the distance between the two modalities. During training, the parameters of the integration model are adjusted until the value of the loss function decreases to a minimum, that is, the cell achieves the best alignment effect, and the integrated potential representation is obtained.
[0052] The KNN classifier is trained using the low-dimensional RNA embedding vector obtained from the model output and used to predict ATAC cell type labels. The RNA is then transferred to the ATAC with the highest similarity to its low-dimensional vector. The label transfer accuracy is calculated using the original ATAC cell type labels and the predicted cell type labels.
[0053] In this exemplary embodiment, the method described above may further include: In step S11, the unmatched single-cell multi-omics data to be processed is obtained, and the data to be processed is preprocessed according to preset rules to obtain a dataset after screening out abnormal data; wherein, the unmatched single-cell multi-omics data to be processed includes: unmatched transcriptome data scRNA-seq and chromatin accessibility data scATAC-seq.
[0054] For example, preprocessing is performed according to preset rules to obtain a dataset after filtering out outlier data, including: To complete the ATAC data filtering, remove low-quality cells with a sequencing depth of less than 1800 and / or more than 100,000 abnormal cells from the single-cell chromatin accessibility scATAC-seq data. The filtered ATAC data is then subjected to linear dimensionality reduction using the Latent Semantic Indexing (LSI) method; and... To filter RNA data, remove low-quality cells with fewer than 1,000 unique molecular identifiers and / or more than 25,000 abnormal cells from single-cell transcriptome scRNA-seq data; remove cells with abnormal chromatin structure and nucleosome signal exceeding 2; and remove cells with poor chromatin openness and transcription start site enrichment score not exceeding 1. The filtered RNA data were then normalized and logarithmized sequentially, and linear dimensionality reduction was performed using principal component analysis (PCA).
[0055] For example, the trained model can be deployed on a server or a user-side terminal device. The user can select the single-cell multi-omics non-matching data to be processed on the terminal. This non-matching single-cell multi-omics data includes: non-matching transcriptome data (scRNA-seq) and chromatin accessibility data (scATAC-seq).
[0056] Taking the model deployed locally on the terminal device as an example, the above preprocessing rules can be used first to filter out low-quality data and delete abnormal cell data from the single-cell multi-omics non-matching data to be processed.
[0057] In step S12, the Gene and Peak of open chromatin peaks in the dataset are extracted based on the regulatory relationship between the Gene and Peak, and the connection weights are configured according to the regulatory strength between the Gene and Peak to construct the Gene-Peak relationship graph.
[0058] For example, a Gene-Peak relationship graph can be constructed by statistically analyzing all Genes and Peaks in the data based on the regulatory relationship of genes on chromatin accessibility and calculating weights according to the regulatory strength.
[0059] In step S13, RNA features are extracted from the transcriptome RNA data in the dataset using an RNA heterogeneous encoder to obtain RNA features; and ATAC features are extracted from the ATAC data in the dataset using an ATAC heterogeneous encoder to obtain ATAC features; wherein, RNA features are used to represent continuous semantic features of gene expression; and ATAC features are used to represent high-dimensional sparse, near-binary peak signals and accessibility features.
[0060] For example, a pre-constructed RNA heterogeneous encoder is used to extract features from preprocessed transcriptome RNA data to extract continuous semantic representations of gene expression. Simultaneously, a pre-constructed ATAC heterogeneous encoder is used to extract features from preprocessed ATAC data to extract high-dimensional sparse, near-binary peak signals and accessibility representations.
[0061] In step S14, RNA features and ATAC features are mapped to a shared latent space to form an initial cross-modal low-dimensional embedding; and the RNA low-dimensional embedding, ATAC low-dimensional embedding, and Gene-Peak relationship graph are input into a data integration model based on prior constraints for cross-modal integration to obtain integrated data.
[0062] For example, step S14 includes: Step S141: Using RNA low-dimensional embedding as the query vector and ATAC low-dimensional embedding as the key vector and value vector, establish multi-head cross-modal attention features. Step S142: Convert the Gene-Peak relationship graph into an attention mask, and combine it with multi-head cross-modal attention features to determine the attention weight matrix after prior constraints; Step S143: Using the attention weight matrix with prior constraints, perform feature fusion on the value vector and query vector to obtain the initial fused representation. ; Step S144: For the initial fused representation, the final fused representation is obtained by maintaining the semantic stability of the original modality through residual connections and layer normalization. .
[0063] Specifically, refer to Figure 3 As shown, the encoded features output by two heterogeneous encoders can be mapped to a shared latent space to form corresponding initial cross-modal low-dimensional embeddings, which are then used as input parameters for the model. Simultaneously, the Gene-Peak graph is used as an input parameter for the model and transformed into an attention mask (prior mask).
[0064] A multi-head cross-modal attention mechanism is established using RNA low-dimensional embedding as the query vector Q and ATAC low-dimensional embedding as the key vector K and value vector V. The multi-head cross-modal attention features are obtained by performing a dot product operation on the query vector and the key vector.
[0065] Multi-head cross-modal attention features are multiplied element-wise with an attention mask to obtain a priori-constrained attention weight matrix; thus, the attention mask is used to suppress biologically unfounded attention connections. The formula includes:
[0066] in, This represents the genes defined in the attention mask. and open chromatin peaks The potential for regulation between them; Represents cells The low-dimensional embedding obtained by the RNA characterization after passing through a heterogeneous encoder; Represents cells The ATAC representation is used to obtain a low-dimensional embedding after passing through a heterogeneous encoder.
[0067] Then, using the attention weight matrix, the value vector and the query vector are weighted and summed to obtain the initial fused representation. The formula includes:
[0068] Then, the initial fused representation is stabilized by residual connections and layer normalization to obtain the final fused representation. The formula includes:
[0069] Wherein, FFN represents a feedforward network; Representation layer normalization.
[0070] In this exemplary embodiment, taking the dataset obtained from 10x Genomics, which jointly measures gene expression RNA and chromatin accessibility ATAC in human peripheral blood mononuclear cells (PBMCs), as an example, the scRNA-seq data of this dataset measures 4464 cells, each containing 17441 genes, and the scATAC-seq data of this dataset measures 3157 cells, each containing 137442 peak data points. The aforementioned method for integrating single-cell multi-omics mismatched data based on prior constraints may specifically include: Step 1: Perform low-quality filtering on multi-omics non-matching data and remove abnormal cells.
[0071] 1.1) Delete low-quality cells with a sequencing depth of less than 1800 or abnormal cells with a sequencing depth of more than 100,000 from the single-cell chromatin accessibility scATAC-seq data; 1.2) Delete low-quality cells with fewer than 1,000 unique molecular identifiers or abnormal cells with more than 25,000 unique molecular identifiers in single-cell transcriptome scRNA-seq data; delete cells with abnormal chromatin structure and nucleosome signal greater than 2; and delete cells with poor chromatin openness and transcription start site enrichment fraction not exceeding 1. 1.3) Normalize and logarithmize the filtered RNA data, and use PCA to perform linear dimensionality reduction. Use LSI to perform dimensionality reduction on the filtered ATAC data.
[0072] Step 2: Construct the Gene-Peak relationship graph.
[0073] Based on the regulatory relationship between gene and chromatin accessibility, candidate Gene-Peak pairs are first generated according to genomic distance. Then, the regulatory strength is calculated by combining cross-cell accessibility-expression correlation and distance decay. After significance screening and normalization, a weighted Gene-Peak relationship graph is constructed. Step 3: Construct a heterogeneous encoder.
[0074] 3.1) Construct an RNA encoder to extract continuous semantic representations of gene expression; 3.2) Construct an ATAC encoder to extract high-dimensional sparse, near-binary peak signals and extract accessibility representations; 3.3) The outputs of the two encoders are mapped to a shared latent space to form an initial cross-modal low-dimensional embedding; 3.4) Establish reconstruction loss The RNA and ATAC modalities were decoded and reconstructed separately, and the latent representations were constrained to retain their respective key biological information.
[0075] Step 4: Construct a cross-modal Transformer integrated model with prior constraints.
[0076] 4.1) Using RNA low-dimensional embedding as the query vector and ATAC low-dimensional embedding as the key-value vector, a multi-head cross-modal attention mechanism is established; 4.2) The Gene-Peak prior map is converted into an attention mask to suppress attention connections without biological basis; 4.3) Constructing a priori constrained attention loss Constraints on attention distribution and priors Figure 1 This ensures that cross-modal associations satisfy the gene-peak regulation relationship, resulting in a priori constrained attention weight matrix.
[0077] 4.4) Generate the fusion representation: Use the attention weights mentioned above to weight and sum the value vectors of the Transformer model to obtain the initial fusion representation. The final fused representation is obtained by maintaining the semantic stability of the original modality through residual connections and layer normalization. .
[0078] 4.5) Constructing counter-loss Introducing a domain discriminator in a shared latent representation space RNA modality samples are labeled as the source domain and ATAC modality samples are labeled as the target domain. A binary classification cross-entropy domain is used to counteract the loss and measure the separability of the two modality representations. Step 5: Train and integrate the data.
[0079] 5.1) The filtered and processed transcriptome RNA data and chromatin accessibility ATAC data from step (1) are used to construct the Gene-Peak relationship map and input into the integration model through a heterogeneous encoder. The adversarial loss is used to narrow the distance between the two modalities. During training, the parameters of the integration model are adjusted until the value of the loss function decreases to the minimum, that is, the cell achieves the best alignment effect, and the integrated potential representation is obtained. 5.2) Using the knn method, the distance between RNA data and ATAC data is calculated, and the data is classified by majority voting to achieve the purpose of label transfer.
[0080] Specifically, 1) This invention enhances the ability to identify cell features of different modalities by using heterogeneous encoders to model the distribution characteristics of scRNA-seq and scATAC-seq respectively, and achieves more accurate cross-modal alignment in a unified latent space.
[0081] 2) This invention introduces a priori-constrained cross-modal Transformer, which explicitly integrates the Gene-Peak regulatory relationship into the attention calculation process, thereby achieving highly consistent integration of different sequencing data in a biological sense.
[0082] 3) This invention incorporates an adversarial domain alignment mechanism, which removes batch effects and platform differences while preserving cell type heterogeneity, thereby improving generalization and robustness in complex scenarios (such as unbalanced cell populations and cross-conditional data).
[0083] 4) This invention, by setting reconstruction constraints and multi-objective joint optimization, improves alignment accuracy while preserving peak-level chromatin accessibility and key transcriptome information as much as possible, effectively preserving cellular characteristics.
[0084] For example, the present invention will be further described in conjunction with specific verification data.
[0085] I. Simulation Conditions: The computer hardware for the simulation experiment consisted of an AMD EPYC 7K62 CPU (48 cores), 500 GB DDR4 RAM, and an NVIDIA RTX 5090 GPU (32 GB). Computer software: Python 3.10 and Rstudio integrated development environment on Linux system.
[0086] II. Simulation Content: Simulation 1: This invention is compared with five existing methods: uniPort, SCALEX, GLUE, Cobolt, and LIGER. Batch entropy is used as an indicator to evaluate the integration performance of RNA and ATAC, and Mean Accuracy is used as an indicator to evaluate the accuracy of tag transfer from RNA to ATAC. The dataset uses human peripheral blood mononuclear cells PBMC3k and mouse spleen cells.
[0087] Table 1
[0088] Table 1 presents a performance comparison of the present invention and five existing methods on human and mouse datasets. As shown in Table 1, the present invention achieved the highest tag transfer accuracy and best integration performance in human peripheral blood mononuclear cells. In mouse spleen cells, the present invention also maintained the best tag transfer accuracy and integration performance. GLUE performance remained stable but was second best compared to the present invention, failing to maintain stability. The integration performance of Cobolt and uniPort needs further improvement. SCALEX and LIGER showed poor integration performance on human peripheral blood mononuclear cells, indicating insufficient feature capture and unstable performance.
[0089] Simulation 2: To compare the ability of our method to balance modality alignment and preserve cellular heterogeneity with five existing methods (uniPort, Seurat, Harmony, SCALEX, and scCorrector), we used Batch entropy as an indicator to evaluate the integration performance of RNA and ATAC, and Silhouette Coefficient as an indicator to evaluate the distinctness of different cell types. The dataset used was PBMC10k. We performed cell type imbalance processing, removing specific cell types from RNA (RNA-unbalanced) and characteristic cell types from ATAC (ATAC-unbalanced).
[0090] Table 2
[0091] Table 2 shows the comparison results of the ability of the present invention and five existing methods to distinguish cellular heterogeneity in human datasets. As can be seen from Table 2, in human peripheral blood mononuclear cells under ATAC-unbalanced conditions, the present invention achieved the highest modality alignment performance and cellular heterogeneity discrimination ability. In RNA-unbalanced conditions, the present invention's modality alignment performance was slightly inferior to GLUE. This is because GLUE is more likely to over-align some independent cell types to similar subtypes, which, while improving batch entropy, may introduce biological semantic shifts. uniPort performed similarly to GLUE, achieving suboptimal alignment while preserving cellular heterogeneity, but uniPort was slightly worse in modality alignment. Cobolt, while having poor modality alignment ability, had a strong ability to distinguish cellular heterogeneity. SCALEX and LIGER performed poorly, failing to achieve effective alignment.
[0092] Simulation 3: Using the CITE-seq and ASAP-seq PBMC datasets containing stimulated T cells; where T cells in the ATAC modality served as the control group and T cells in the RNA modality served as the stimulation group, they were integrated using the methods of this invention and GLUE (Graph-Linked Unified Embedding, a single-cell multi-omics integration algorithm). The integrated low-dimensional embedding data was clustered and visualized using the UMAP algorithm, and each cell was stained using chromatin accessibility, transcriptomics, and different cell types. The results are as follows: Figure 4 As shown. Among them, Figure 4 The images, from left to right, represent the UMAP visualization of this invention, the UMAP visualization of GLUE, and a cell type coloring legend at the bottom.
[0093] from Figure 4 As can be seen, while the existing GLUE method achieves good alignment, it completely fails to distinguish T cells under different conditions. The two modalities only achieve uniform alignment, and the distinction between different cell types is rather vague. This invention, while achieving modal alignment, distinguishes between conditional differences. T cells from both modalities belong to the same cluster, without completely unconditional alignment, and are finely distinguished according to subtypes. This demonstrates that this invention can distinguish cellular heterogeneity while achieving good alignment results.
[0094] The method provided in this invention eliminates the inconsistency in feature dimensions between RNA and ATAC through independent nonlinear linear transformation and dimensionality reduction operations of heterogeneous encoders, thereby obtaining potential cell representations at the same feature level. This provides a unified input for subsequent prior-guided Transformers and preserves high-quality biological heterogeneity for the adversarial alignment module. The constructed Gene-Peak relationship graph guides the potential connections between cells, and masking removes redundant information from the attention mechanism, explicitly guiding subsequent modality alignment. By combining prior constraints and adversarial loss to align modalities, modality integration is achieved while effectively preserving cell specificity. This invention achieves higher alignment accuracy in integrating single-cell multi-omics mismatched data and can clearly distinguish cells of different subtypes and under different conditions, effectively identifying cellular heterogeneity.
[0095] It should be noted that the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may, for example, be executed synchronously or asynchronously in multiple modules.
[0096] It should be noted that although several modules or units of the device for performing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.
[0097] Figure 5 A schematic diagram of an electronic device suitable for implementing embodiments of the present invention is shown.
[0098] It should be noted that, Figure 5 The electronic device 1000 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.
[0099] like Figure 5 As shown, the electronic device 1000 includes a Central Processing Unit (CPU) 1001, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 1002 or programs loaded from storage section 1008 into Random Access Memory (RAM) 1003. The RAM 1003 also stores various programs and data required for system operation. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An Input / Output (I / O) interface 1005 is also connected to the bus 1004. Furthermore, the electronic device 1000 also includes an FPGA device and a System-on-a-Chip (SoC) device.
[0100] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. Removable media 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1010 as needed so that computer programs read from them can be installed into storage section 1008 as needed.
[0101] In particular, according to embodiments of the present invention, the processes described below with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by central processing unit (CPU) 1001, it performs various functions defined in the system of this application.
[0102] It should be noted that the storage medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, wherein computer-readable program code is carried. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.
[0103] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0104] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.
[0105] It should be noted that, as another aspect, this application also provides a storage medium, which may be included in an electronic device or may exist independently without being assembled into the electronic device. The aforementioned storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the methods described in the following embodiments. For example, the electronic device may perform... Figure 1 The steps of the method shown.
[0106] In one embodiment, this application provides a computer program product including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0107] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.
[0108] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.
[0109] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.
Claims
1. A method for integrating single-cell multi-omics non-matched data based on prior constraints, characterized in that, The method includes: Acquire the unmatched single-cell multi-omics data to be processed, and preprocess the data according to preset rules to obtain the dataset after screening out abnormal data; the unmatched single-cell multi-omics data to be processed includes: unmatched transcriptome data scRNA-seq and chromatin accessibility data scATAC-seq. Genes and open chromatin peaks in the dataset are extracted based on the regulatory relationship between genes and chromatin accessibility. Connection weights are configured according to the regulatory strength between Genes and Peaks to construct a Gene-Peak relationship graph. RNA features were extracted from transcriptome RNA data in the dataset using an RNA heterogeneous encoder; and ATAC features were extracted from ATAC data in the dataset using an ATAC heterogeneous encoder. RNA features were used to represent continuous semantic features of gene expression; ATAC features were used to represent high-dimensional sparse, near-binary peak signals and accessibility features. RNA features and ATAC features are mapped to a shared latent space to form an initial cross-modal low-dimensional embedding. The RNA low-dimensional embedding, ATAC low-dimensional embedding, and Gene-Peak relationship graph are then input into a data integration model based on prior constraints for cross-modal integration to obtain integrated data.
2. The method of claim 1, wherein, Preprocessing is performed according to preset rules to obtain a dataset after filtering out outlier data, including: To complete the ATAC data filtering, remove low-quality cells with a sequencing depth of less than 1800 and / or more than 100,000 abnormal cells from the single-cell chromatin accessibility scATAC-seq data. The filtered ATAC data is then subjected to linear dimensionality reduction using the Latent Semantic Indexing (LSI) method; and... To filter RNA data, remove low-quality cells with fewer than 1,000 unique molecular identifiers and / or more than 25,000 abnormal cells from single-cell transcriptome scRNA-seq data; remove cells with abnormal chromatin structure and nucleosome signal exceeding 2; and remove cells with poor chromatin openness and transcription start site enrichment score not exceeding 1. The filtered RNA data were then normalized and logarithmized sequentially, and linear dimensionality reduction was performed using principal component analysis (PCA).
3. The method of claim 1, wherein, RNA low-dimensional embeddings, ATAC low-dimensional embeddings, and Gene-Peak relationship graphs are input into a prior-constrained data integration model for cross-modal integration, including: Using RNA low-dimensional embedding as the query vector and ATAC low-dimensional embedding as the key and value vectors, a multi-head cross-modal attention feature is established. The Gene-Peak relationship graph is converted into an attention mask, and the attention weight matrix after prior constraints is determined by combining multi-head cross-modal attention features, including: wherein, represents the regulatory potential between genes defined in the attention mask and open chromatin peaks; represents the low-dimensional embedding of the RNA representation of the cells after passing through the isometric encoder; represents the low-dimensional embedding of the ATAC representation of the cells after passing through the isometric encoder; The value vector and the query vector are fused by using the attention weight matrix after the prior constraint to obtain an initial fusion representation ; The final fused representation is obtained by keeping the original modality semantic stable through residual connection and layer normalization on the initial fused representation comprising: where FFN denotes a feed forward network; denotes layer normalization.
4. The method of claim 1, wherein, The method further includes: training a data integration model based on prior constraints, including: Single-cell multi-omics mismatched data were obtained from the sequencing dataset and preprocessed according to preset rules to obtain a dataset after filtering out abnormal data. The single-cell multi-omics mismatched data included mismatched transcriptome data scRNA-seq and chromatin accessibility data scATAC-seq. Genes and open chromatin peaks in the dataset are extracted based on the regulatory relationship between genes and chromatin accessibility. Connection weights are configured according to the regulatory strength between Genes and Peaks to construct a Gene-Peak relationship graph. An RNA heterogeneous encoder was constructed, and features were extracted from the transcriptomic RNA data in the dataset to obtain RNA features; and, An ATAC heterogeneous encoder is constructed, and features are extracted from the ATAC data in the dataset to obtain ATAC features. Among them, RNA features are used to represent continuous semantic features of gene expression, and ATAC features are used to represent high-dimensional sparse, near-binary peak signals and accessibility features. RNA features and ATAC features are mapped to a shared latent space to form an initial cross-modal low-dimensional embedding. A reconstruction loss is constructed to decode and reconstruct the RNA and ATAC modalities separately, in order to constrain the potential representations to retain their respective key biological information; The initial model is input with RNA low-dimensional embedding, ATAC low-dimensional embedding and Gene-Peak relationship graph. The RNA low-dimensional embedding is used as the query vector and the ATAC low-dimensional embedding is used as the key vector and value vector to establish multi-head cross-modal attention features. The Gene-Peak relationship graph is converted into an attention mask, and the attention weight matrix with prior constraints is determined by combining multi-head cross-modal attention features. The value vector and the query vector are fused by using the attention weight matrix after the prior constraint to obtain an initial fusion representation ; The final fused representation is obtained by keeping the original modality semantic stable through residual connection and layer normalization on the initial fused representation ; Construct an adversarial loss function to bridge the gap between the two modalities; backpropagate the initial model based on the adversarial loss function until the model converges to obtain a trained data integration model based on prior constraints. The distance between RNA data and ATAC data was calculated using the k-nearest neighbor method, the data were classified using majority voting, and the labels were transferred based on the classification results.
5. The method of claim 4, wherein, Reconstruction losses include: wherein, representing cells characterization on RNA; representing cells characterization in characterization view and chromatin accessibility view on ATAC; representing cells low-dimensional embedding of RNA characterization of cells after passing through an isometric encoder; representing cells low-dimensional embedding of ATAC characterization of cells after passing through an isometric encoder.
6. The method of claim 4, wherein, Construct an adversarial loss function to bridge the gap between two modalities, including: Introducing domain discriminators in shared latent representation space The RNA modality sample is marked as a source domain, the ATAC modality sample is marked as a target domain, and a binary classification cross-entropy is used to define a domain adversarial loss for measuring the separability of the two modality representations, and the formula is represented as: wherein, represents the final fused representation, represents the expectation.
7. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the single-cell multi-omics non-matching data integration method based on prior constraints as described in any one of claims 1 to 6.
