Single-cell multi-omics triplets prediction method and system based on graph causal learning

Abstract

The application relates to the technical field of bioinformatics and artificial intelligence, and provides a single-cell multi-omics triple prediction method and system based on graph causal learning, which comprises the following steps: obtaining multi-omics original data and performing pretreatment, constructing a heterogeneous graph, and generating single-cell multi-omics triple nodes according to the heterogeneous graph; performing multi-layer propagation and feature aggregation on the triple nodes through a graph neural network, and splicing to obtain a splicing vector; obtaining overall semantic features through an overall semantic module; obtaining directional representation through a directional causal module; fusing the overall semantic features and the directional representation to obtain a triple prediction vector; and inputting the triple prediction vector into a classifier to obtain a classification probability vector. The application realizes accurate prediction of various cell gene chromatin open region triple structures, can fuse multi-omics information, distinguish various triple modes, has causal reasoning and generalization capabilities, and improves the biological interpretability and generalization capability of the model.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and artificial intelligence, and in particular to a single-cell multi-omics triplet prediction method and system based on graph causal learning. Background Technology

[0002] Single-cell transcriptome sequencing can resolve gene expression information at the cellular level, while single-cell chromatin accessibility sequencing characterizes transcriptional regulatory potential by identifying open chromatin regions. However, existing studies often analyze these two types of omics data separately or only perform modality alignment at the low-dimensional representation level, lacking systematic modeling of the tripartite relationship between cells, genes, and open chromatin regions. For example, single-cell transcriptome sequencing can reveal the transcriptional state of cells, but research on upstream regulatory areas like chromatin is limited. Chromatin openness is a prerequisite for transcription factor binding and activation with gene promoters; therefore, research on open chromatin regions can provide a potential regulatory mechanism level, revealing upstream regulatory modes of gene expression. Joint analysis of chromatin and transcriptome within the same cell helps understand cellular transcriptional regulation, especially in areas such as tumors, cell development, and cell atlases, providing new perspectives for understanding disease mechanisms and developing diagnostic and therapeutic strategies. However, single-cell sequencing data is characterized by large scale, high dimensionality, high sparsity, and high heterogeneity. How to effectively integrate expression and regulatory data at the single-cell level and develop new algorithms to effectively process and utilize this data has become an important research direction in bioinformatics today.

[0003] In single-cell chromatin accessibility sequencing analysis, open regions of chromatin form peaks on the genome. These peaks typically correspond to promoters and transcription factor binding sites of active genes, providing information about intracellular gene expression and regulation. Traditional regulatory relationship construction relies heavily on whole-cell omics data, failing to capture cell-type or state-specific gene regulatory networks. However, exploring ternary relationships at the single-cell level—cell-gene-peak—can more accurately characterize active regulatory pathways in different cells. However, existing graph learning methods, in ternary pattern classification prediction, mostly perform binary classification only for single ternary structures (such as triangular closed loops), failing to distinguish multiple biological patterns; high-frequency unilateral structures are excessive, while rare but important regulatory patterns are easily overlooked. Most methods rely on statistical co-occurrence relationships, making them susceptible to confounding factors and lacking explicit modeling of causal relationships within structures. Summary of the Invention

[0004] This invention aims to at least solve one of the technical problems existing in related technologies. To this end, this invention provides a single-cell multi-omics triple prediction method and system based on graph causal learning. By constructing heterogeneous graphs, fusing overall semantic features and directional causal features, combining a dual contrastive learning mechanism, and introducing a causal counterfactual enhancement strategy, it achieves accurate prediction of various cell-gene-chromatin open region triple structures. It can integrate multi-omics information, distinguish various triple patterns, and has causal reasoning and generalization capabilities, thereby improving the biological interpretability and generalization ability of the model.

[0005] This invention provides a single-cell multi-omics triple prediction method based on graph causal learning, comprising:

[0006] S1: Acquire multi-omics raw data including single-cell transcriptome sequencing and single-cell chromatin accessibility sequencing and preprocess them to obtain preprocessed cell nodes, peak nodes and gene nodes.

[0007] S2: Construct a heterogeneous graph based on the preprocessed cell nodes, peak nodes, and gene nodes, and generate single-cell multi-omics triplet nodes based on the heterogeneous graph;

[0008] S3: Multi-layer propagation and feature aggregation are performed on single-cell multi-omics triple nodes on heterogeneous graphs through graph neural networks to obtain node embedding features, and the node embedding features are concatenated to obtain concatenated vectors.

[0009] S4: Construct a causal prediction model and train it through dual contrastive learning and causal counterfactual methods. The causal prediction model includes an overall semantic module and a directional causal module.

[0010] S5: The global semantic features of the spliced ​​vector are fused through the overall semantic module to obtain the overall semantic features; the directional features of the single-cell multi-omics triples are extracted through the directional causal module to obtain the directional representation; the overall semantic features and the directional representation are fused to obtain the triple prediction vector.

[0011] S6: Input the triplet prediction vector into the classifier to obtain the single-cell multi-omics triplet classification probability vector.

[0012] Furthermore, the preprocessing includes:

[0013] Peak screening was performed on single-cell chromatin accessibility sequencing, and the retained peak features were subjected to feature dimensionality reduction to obtain a set of peak feature nodes;

[0014] Gene screening was performed on single-cell transcriptome sequencing, and gene-related peaks were selected to obtain a set of gene nodes.

[0015] Furthermore, step S2 includes:

[0016] S21: Construct an edge relationship matrix based on the preprocessed cell nodes, peak nodes, and gene nodes, according to the biological regulatory relationships;

[0017] S22: An embedding learning method based on PyTorch-BigGraph is used to learn the vectorized representation of cell nodes, peak nodes, and gene nodes to obtain node embeddings;

[0018] S23: Construct a heterogeneous graph based on node embeddings and edge relationship matrices;

[0019] S24: Construct seven types of three-node structures based on the heterogeneous graph.

[0020] Furthermore, the seven types of three-node structures include triangular structures, cell-centered bilateral structures, peak-centered bilateral structures, gene-centered bilateral structures, gene-cell unilateral structures, gene-peak unilateral structures, and cell-peak unilateral structures.

[0021] Furthermore, step S3 includes:

[0022] S31: Extract the initial cellular features, initial peak features, and initial gene features of the triplet node;

[0023] S32: Cell node embedding features are obtained by multi-layer propagation of initial cell features through graph convolutional neural networks; peak node embedding features are obtained by multi-layer propagation of initial peak features through graph convolutional neural networks; gene node embedding features are obtained by multi-layer propagation of initial gene features through graph convolutional neural networks.

[0024] S33: Concatenate the cell node embedding features, peak node embedding features, and gene node embedding features into vectors to obtain a concatenated vector.

[0025] Furthermore, the dual contrastive learning includes node consistency constraints and structural pattern-level semantic aggregation;

[0026] The node consistency constraints include:

[0027] S411: Randomly mask some edges of a heterogeneous graph to obtain a perturbation graph;

[0028] S412: Extract node embeddings of heterogeneous graphs through a graph convolutional neural network with shared parameters to obtain heterogeneous graph node embeddings; extract node embeddings of perturbed graphs through a graph convolutional neural network with shared parameters to obtain perturbed graph node embeddings.

[0029] S413: Calculate the cosine similarity between the node embeddings of the heterogeneous graph and the node embeddings of the perturbed graph to obtain the node consistency contrast loss;

[0030] The structural pattern-level semantic aggregation includes:

[0031] S421: Calculate the mean node embedding of triple samples in the heterogeneous graph to obtain the prototype vector of the heterogeneous graph; calculate the mean node embedding of triple samples in the perturbation graph to obtain the prototype vector of the perturbation graph.

[0032] S422: Calculate the prototype consistency loss based on the heterogeneous graph prototype vector and the perturbation graph prototype vector;

[0033] S423: Obtain multi-class probability distributions by calculating the similarity between triples and heterogeneous graph prototype vectors, as well as the similarity between triples and perturbation graph prototype vectors.

[0034] S424: Calculate cross-entropy loss based on multi-class probability distributions.

[0035] Furthermore, counterfactual causality includes:

[0036] S431: Replace the target nodes of some edges in the edge set to generate counterfactual triples;

[0037] S432: Use KL divergence to measure the difference in embedding distributions between counterfactual triples and generate counterfactual loss.

[0038] Furthermore, the causal prediction model is jointly optimized by node consistency comparison loss, prototype consistency loss, cross-entropy loss, and counterfactual loss.

[0039] Furthermore, obtaining directional representations includes:

[0040] S521: Construct the cell peak direction difference vector, the peak gene direction difference vector, and the gene cell direction difference vector;

[0041] S522: Directional characterization is obtained by nonlinearly mapping and splicing the cell peak direction difference vector, peak gene direction difference vector, and gene cell direction difference vector.

[0042] This invention also provides a single-cell multi-omics triple prediction system based on graph causal learning, for executing the above-mentioned single-cell multi-omics triple prediction method based on graph causal learning, comprising:

[0043] The acquisition module acquires multi-omics raw data including single-cell transcriptome sequencing and single-cell chromatin accessibility sequencing and preprocesses it to obtain preprocessed cell nodes, peak nodes and gene nodes.

[0044] A node generation module, which constructs a heterogeneous graph based on preprocessed cell nodes, peak nodes and gene nodes, and generates single-cell multi-omics triplet nodes based on the heterogeneous graph.

[0045] The feature splicing module performs multi-layer propagation and feature aggregation on single-cell multi-omics triple nodes on the heterogeneous graph through a graph neural network to obtain node embedded features, and splices the node embedded features to obtain a splicing vector.

[0046] The model building module constructs a causal prediction model and trains it through dual contrastive learning and causal counterfactual methods. The causal prediction model includes an overall semantic module and a directional causal module.

[0047] The prediction module fuses the global semantic features of the spliced ​​vector through the overall semantic module to obtain the overall semantic features; it extracts the directional features of the single-cell multi-omics triples through the directional causal module to obtain the directional representation; and it fuses the overall semantic features and the directional representation to obtain the triple prediction vector.

[0048] The classification module inputs the triplet prediction vector into the classifier to obtain the single-cell multi-omics triplet classification probability vector.

[0049] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0050] This invention achieves accurate prediction of multiple cell-gene-chromatin open region ternary structures by constructing heterogeneous graphs, fusing global semantics and directional causal features, combining a dual contrastive learning mechanism, and introducing a causal counterfactual enhancement strategy. It can simultaneously distinguish seven cell-peak-gene ternary structure patterns, enabling more refined biological structure prediction. Causal reasoning and counterfactual enhancement significantly alleviate prediction bias caused by class imbalance and structural redundancy, improve the model's biological interpretability and generalization ability, and the training embedding results can be used for the analysis of cell regulatory mechanisms and disease research.

[0051] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0053] Figure 1 This is a flowchart illustrating a single-cell multi-omics triplet prediction method based on graph causal learning provided by the present invention.

[0054] Figure 2 This is a technical roadmap for a single-cell multi-omics triplet prediction model based on graph causal learning provided by this invention.

[0055] Figure 3 This is a visualization diagram of the t-distributed random neighborhood embedding of ternary relation embeddings for seven categories according to an embodiment of the present invention.

[0056] Figure 4 This is a schematic diagram illustrating the visualization of cell embeddings obtained from model training using dimensionality reduction techniques in an embodiment of the present invention.

[0057] Figure 5 This is a schematic diagram illustrating the combined representation of cell embedding and gene embedding using dimensionality reduction techniques in an embodiment of the present invention.

[0058] Figure 6 This is a visualization diagram illustrating how UMAP is used in an embodiment of the present invention to map the embedded data of cell, gene, and chromatin accessibility regions to a unified low-dimensional space.

[0059] Figure 7 This is a schematic diagram of the structure of a single-cell multi-omics triplet prediction system based on graph causal learning provided by the present invention.

[0060] Figure label:

[0061] 101. Acquisition Module; 102. Node Generation Module; 103. Feature Concatenation Module; 104. Model Building Module; 105. Prediction Module; 106. Classification Module. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but cannot be used to limit the scope of this invention.

[0063] In the description of the embodiments of the present invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is 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 a suitable manner in any 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.

[0064] The following is combined with Figures 1 to 7 This invention describes a single-cell multi-omics triplet prediction method and system based on graph causal learning.

[0065] like Figure 1 As shown, a single-cell multi-omics triple prediction method based on graph causal learning includes:

[0066] S1: Acquire multi-omics raw data including single-cell transcriptome sequencing (scRNA-seq) and single-cell chromatin accessibility sequencing (scATAC-seq) and preprocess them to obtain preprocessed cell nodes, peak nodes and gene nodes.

[0067] Peak selection was performed on single-cell chromatin accessibility sequencing, and the retained peak features were subjected to feature dimensionality reduction to obtain a set of peak feature nodes.

[0068] The frequency of each peak in all cells was counted, and low-frequency peaks that appeared in fewer than 3 cells were filtered out to remove low-quality signals.

[0069] Principal component analysis (PCA) is used to reduce the dimensionality of the retained peak features, thereby reducing the feature dimension and lowering the complexity of the model input while preserving as much information as possible.

[0070] Specifically, the chromatin accessibility data matrix of single cells is centered, then the covariance matrix is ​​calculated, and the covariance matrix is ​​decomposed into eigenvalues. The eigenvectors corresponding to the first few eigenvalues ​​are selected as principal components.

[0071] Gene screening was performed on single-cell transcriptome sequencing, and peaks associated with genes were identified to obtain a set of hypervariable gene nodes.

[0072] By using DeepMAPS (a deep learning-based multi-omics analysis platform for single-cell data), spatial relationships, chromatin accessibility, and gene expression characteristics are integrated to identify and quantify the association between peaks and genes, and only peaks that are significantly associated with genes are retained.

[0073] Specifically, two key matrices are given: the gene expression matrix and the gene expression matrix. and chromatin accessibility matrix ;

[0074]

[0075]

[0076] in, For cells genes Gene expression, open chromatin regions In cells Medium chromatin accessibility For the number of genes, Cell count, This represents the number of open regions of chromatin.

[0077] For calculating chromatin open regions On genes Regulation potential weight The distance between the two in the genome plays a decisive role. If Center and gene Distance of transcription start site More than 150kb (kilobase, kilobase pairs), or Located within other nearby genes, the potential regulatory weight is defined as:

[0078]

[0079] Long-distance interactions are usually negligible, and peaks located near other genes are more likely to regulate neighboring genes.

[0080] when When located in the exon region of a gene:

[0081]

[0082] in, is the exon length.

[0083] Based on observations in batch chromatin accessibility sequencing data, highly expressed genes have open chromatin regions in their exons. Normalization by exon length helps to account for the probability of background readouts.

[0084] In other cases:

[0085]

[0086] in, The half-life is the distance value, reflecting the principle that the regulatory effect weakens exponentially with increasing distance.

[0087] In obtaining Afterwards, through Computational cells middle On genes Regulation potential score Finally, the regulatory potential scores of peaks regulating the same gene are summed, and the chromatin accessibility matrix is ​​then used to... The results are converted into a potential gene regulation matrix, yielding a certain number of representative peaks.

[0088] A screening strategy based on highly variable genes (HVGs) was employed to identify gene-related peaks, resulting in a set of highly variable gene nodes. This strategy is based on the core assumption that genes with higher expression variability are more discriminative in distinguishing different cell types or states. The variability score was calculated by statistically analyzing the relationship between the mean expression of each gene across all cells and its variability. Based on this, all genes were sorted according to their variability scores, and the genes with the highest variability were selected as the target for analysis.

[0089] S2: Construct a heterogeneous graph based on the preprocessed cell nodes, peak nodes, and gene nodes, and generate single-cell multi-omics triplet nodes based on the heterogeneous graph;

[0090] S21: Based on the preprocessed cell nodes, peak nodes, and gene nodes, construct an edge relationship matrix according to biological regulatory relationships; including three types of relationships: cell–peak, cell–gene, and peak–gene.

[0091] S22: An embedding learning method based on PyTorch-BigGraph is used to learn the vectorized representation of cell nodes, peak nodes, and gene nodes to obtain node embeddings;

[0092] The constructed graph data is converted into a triplet text format conforming to PBG. Before training begins, the system automatically analyzes the basic statistical information of the graph structure and empirically estimates the weight decay parameters based on this information to avoid embedding overfitting. Specifically, a relatively large weight decay factor is used for graphs with fewer edges, while a smaller regularization coefficient is used for graphs with dense edges to retain more structural details. Subsequently, the model iteratively optimizes the embedding representation of each type of node using triplet information and a predefined multi-class cross-entropy loss. In each training step, PBG adjusts the embedding of the corresponding node according to the triplet relationships, so that the edge connections are approximately preserved in the vector space. After training, vector representations of three types of nodes—cell, peak, and gene—are obtained, namely, cell node embeddings, peak node embeddings, and gene node embeddings.

[0093] S23: Construct a heterogeneous graph based on node embeddings and edge relationship matrices;

[0094] Define heterogeneous graph , , Let be the set of edges. For a set of nodes, It contains three types of nodes: cell, peak, and gene; It contains three types of edges: cell–peak, cell–gene, and peak–gene;

[0095] S24: Construct seven types of three-node structures based on heterogeneous graphs;

[0096] like Figure 2 As shown in Figure (a), the seven types of three-node structures include the Triangle structure, the Cell-Center structure (a two-sided structure centered on the cell), the Peak-Center structure (a two-sided structure centered on the peak), the Gene-Center structure (a two-sided structure centered on the gene), the Gene-Cell structure (a one-sided structure centered on the gene and cell), the Gene-Peak structure (a one-sided structure centered on the gene and peak), and the Cell-Peak structure (a one-sided structure centered on the cell and peak).

[0097] Triangle Structure: Triples that simultaneously meet the following conditions are selected: the peak is associated with a gene; the peak is connected to a cell; and the cell and gene have a regulatory relationship. Cell-gene-peak triplets that meet these conditions are retained as triangle structure samples, representing the fully connected biological mechanisms between the three types of nodes.

[0098] Cell-Center: For each cell node, its known associated peak nodes and gene nodes are extracted; and all peak-gene pairs are examined. If no known association exists between peak-gene pairs, the cell-gene-peak triplet is retained. The Cell-Center reflects situations where cells may indirectly influence certain genes through mechanisms that have not yet been identified.

[0099] Peak-Center: Obtain the cell and gene sets associated with each peak. For any cell-gene pair, if there is no direct connection, a ternary structure is formed with the peak as the center. The Peak-Center reflects a possible missing link in the peak regulatory pathway, which helps guide the model to identify unknown cell-gene regulatory pathways.

[0100] Gene-Center: Identify the set of peaks and cell nodes associated with each gene. For each peak-cell pair, if there is no known association between them, construct a ternary relationship centered on the gene.

[0101] The remaining edges in the heterogeneous graph are single-sided structures, and there are three types: gene-cell, gene-peak, and cell-peak.

[0102] S3: Multi-layer propagation and feature aggregation are performed on single-cell multi-omics triple nodes on heterogeneous graphs through graph neural networks to obtain node embedding features, and the node embedding features are concatenated to obtain concatenated vectors.

[0103] S31: Extract the initial cellular features, initial peak features, and initial gene features of the triplet node;

[0104] S32: Cell node embedding features are obtained by multi-layer propagation of initial cell features through graph convolutional neural networks; peak node embedding features are obtained by multi-layer propagation of initial peak features through graph convolutional neural networks; gene node embedding features are obtained by multi-layer propagation of initial gene features through graph convolutional neural networks.

[0105] The calculation expression is:

[0106]

[0107] in, For the first Layer node features It is a non-linear activation function. To normalize the Laplace curve, For the first Layer node features For weights.

[0108] Cell node features are encoded using a multi-layer graph convolutional neural network to obtain cell node embedding features. Peak node embedding features are obtained by encoding peak node features through a multi-layer graph convolutional neural network. Gene node embedding features are obtained by encoding gene node features through a multi-layer graph convolutional neural network. .

[0109] Each node's embedding features contain contextual information about itself and its multi-level neighbors.

[0110] S33: Concatenate the cell node embedding features, peak node embedding features, and gene node embedding features into vectors to obtain the concatenated vector. .

[0111] S4: Construct a causal prediction model and train it using dual contrastive learning and causal counterfactual methods. The causal prediction model includes a holistic semantic module and a directional causal module. The causal prediction model is as follows: Figure 2 As shown in Figure (b), Figure 2 In the text, CAT stands for splicing operation.

[0112] like Figure 2 As shown in Figure (c), dual contrastive learning includes node consistency constraints and structural pattern-level semantic aggregation;

[0113] Node consistency constraints include:

[0114] S411: Randomly mask some edges of a heterogeneous graph to obtain a perturbation graph;

[0115] S412: Extract node embeddings of the heterogeneous graph using a graph convolutional neural network with shared parameters to obtain the heterogeneous graph node embeddings. Extract node embeddings of the perturbed graph using a graph convolutional neural network with shared parameters to obtain the perturbed graph node embeddings. The calculation expression is as follows:

[0116]

[0117]

[0118] in, Embedding matrix for heterogeneous graph nodes For graph convolutional neural networks, It is a heterogeneous graph. The node feature matrix, Embedding matrix for nodes in perturbation graph This is a perturbation diagram;

[0119] Heterogeneous graphs and perturbation graphs share GCN parameters, and node-level consistency constraints are applied to the node characteristics of heterogeneous graphs and perturbation graphs.

[0120] S413: Calculate the cosine similarity between the node embeddings of the heterogeneous graph and the node embeddings of the perturbed graph to obtain the node consistency contrast loss;

[0121] The cosine similarity between heterogeneous graph node embeddings and perturbed graph node embeddings is expressed as follows:

[0122]

[0123] in, For cosine similarity, nodes of heterogeneous graphs Embedded vector, For nodes in the perturbation graph Embedded vector, Let L2 norm be the vector. It is the vector inner product.

[0124] The expression for calculating the node-level contrastive loss is:

[0125]

[0126] in, For nodes Node-level contrast loss, For temperature coefficient, For nodes The set of negative sample pairs This refers to all nodes that have been traversed.

[0127] The consistency loss of each node in the triplet is summarized under both views. The expression for calculating the node consistency comparison loss is as follows:

[0128]

[0129] in, The loss is due to node consistency comparison. It is a set of nodes.

[0130] The structural pattern-level semantic aggregation includes:

[0131] S421: Calculate the mean node embedding of triple samples in the heterogeneous graph to obtain the prototype vector of the heterogeneous graph; calculate the mean node embedding of triple samples in the perturbation graph to obtain the prototype vector of the perturbation graph.

[0132] S422: Calculate the prototype consistency loss based on the heterogeneous graph prototype vector and the perturbation graph prototype vector. The calculation expression is as follows:

[0133]

[0134] in, For prototype consistency loss, For the heterogeneous graph One prototype vector, For the perturbation diagram One prototype vector, The number of prototypes;

[0135] S423: By calculating the similarity between triples and the prototype vectors of heterogeneous graphs, as well as the similarity between triples and the prototype vectors of perturbation graphs, the multi-class probability distribution is obtained. The calculation expression is as follows:

[0136]

[0137] in, for Time of the first The probability of a prototype, For the first One prototype vector, For the first One prototype vector, for The triple vector at time t;

[0138] S424: Calculate the cross-entropy loss based on the multi-class probability distribution. The calculation expression is as follows:

[0139]

[0140] in, For cross-entropy loss, This is an indicator function for the actual label.

[0141] The classification accuracy is optimized by using cross-entropy loss, achieving intra-class aggregation and inter-class separation.

[0142] like Figure 2 As shown in diagram (d), causal counterfactuals include:

[0143] S431: Replace the target nodes of some edges in the edge set to generate counterfactual triples;

[0144] Given a set of edges , , For the first edge, This is the second edge. For the first A subset of edges is randomly selected for intervention. By modifying the target nodes of these edges, different causal relationships are simulated. The resulting set of edges after intervention is as follows:

[0145]

[0146] in, For the set of edges after intervention, The first edge after intervention, The second edge after intervention. For the first day after intervention Edge;

[0147] exist and The counterfactual node representation and counterfactual triple representation after intervention are obtained through a graph convolutional neural network with shared parameters.

[0148] S432: Use KL divergence to measure the difference in embedding distributions between counterfactual triples and generate counterfactual loss.

[0149] Assumption It is a node The representation after intervention. It is a node The original representation, defined by KL divergence, amplifies the sensitivity to the causal effects of key edges, resulting in the following counterfactual loss:

[0150]

[0151] in, For counterfactual losses, Let KL divergence function be used. For the softmax function, the softmax function is a function that maps a real number vector to a probability distribution;

[0152] Counterfactual loss is obtained by causally comparing the original features and counterfactual features.

[0153] The triplet samples are proportionally divided into training, validation, and test sets; triplet samples and their negative sample sets are batch-sampled to construct two-view and counterfactual graph batches. A learning rate is used. End-to-end training was performed with batch size of 500 and hidden dimension of 128. The causal prediction model was jointly optimized using node consistency contrast loss, prototype consistency loss, cross-entropy loss, and counterfactual loss. The calculation expression is as follows:

[0154]

[0155] in, For the total loss, As the first hyperparameter, This is the second hyperparameter. This is the counterfactual loss hyperparameter.

[0156] Node-level consistency loss suppresses structural redundancy and overfitting; structural pattern-level semantic aggregation alleviates class imbalance and maintains consistency between the two views; counterfactual enhancement stabilizes training and improves generalization and interpretability, ultimately resulting in robust triplet pattern recognition performance.

[0157] S5: The global semantic features of the spliced ​​vector are fused through the overall semantic module to obtain the overall semantic features; the directional features of the single-cell multi-omics triples are extracted through the directional causal module to obtain the directional representation; the overall semantic features and the directional representation are fused to obtain the triple prediction vector.

[0158] S51: By fusing the global semantic features of the concatenated vectors through the overall semantic module, the overall semantic features are obtained, and the calculation expression is as follows:

[0159]

[0160] in, This is a vector concatenation operation. As for the overall semantic features, This is the first multilayer perceptron.

[0161] The overall semantic module is a non-linear mapping network; by fusing the full-graph contextual semantics between nodes through the overall semantic module, it emphasizes the joint expression between cells, peaks and genes.

[0162] S52: Extract directional features of single-cell multi-omics triples through the directional causal module to obtain directional characterization;

[0163] S521: Construct the cell peak direction difference vector, peak gene direction difference vector, and cell gene direction difference vector. The calculation expression is as follows:

[0164]

[0165] in, This is the difference vector of the cell peak direction. The peak gene direction difference vector, This is the directional difference vector of the cell genes.

[0166] S522: Directional characterization is obtained by nonlinearly mapping and concatenating the cell peak direction difference vector, peak gene direction difference vector, and gene-cell direction difference vector. The calculation expression is as follows:

[0167]

[0168] in, As a directional representation, For the second multilayer perceptron, This is the feature mapping function.

[0169] S53: By fusing overall semantic features and directional representations, a triplet prediction vector is obtained, calculated as follows:

[0170]

[0171] in, This is the prediction vector for the triplet.

[0172] S6: Input the triplet prediction vector into the classifier to obtain the single-cell multi-omics triplet classification probability vector.

[0173] The triplet prediction vector is fed into a softmax classifier, which outputs a probability vector that belongs to one of the seven types of three-node structures. Multi-task classification and single-cell multimodal analysis are then performed to obtain the prediction results of the seven types of three-node structures.

[0174] like Figure 3 As shown, the t-distributed random neighborhood embedding (t-SNE) of the seven-class ternary relation embeddings obtained from training is visualized. The seven ternary patterns are labeled with different colors to show the distribution of data points with different prediction labels in two-dimensional space. Figure 3 Data points of different triplet patterns exhibit intra-class clustering and inter-class separation in low-dimensional space, demonstrating that this invention can overcome the limitations of traditional methods and achieve accurate differentiation of seven biological triplet patterns.

[0175] like Figure 4 As shown, dimensionality reduction techniques are used to visualize the cell embeddings obtained from model training and to analyze the distribution patterns of different cell types. Each cell type is labeled with a different color, and different cell types exhibit different clustering patterns in the low-dimensional space, reflecting their similarities and differences. Figure 4 The five cell types are: Hair Shaft-cuticle.cortex (hair shaft-keratin / cortical cells), IRS (inner root sheath cells), Medula (hair medulla cells), TAC-1 (Transit-Amplifying Cells, transitional proliferating cells 1), and TAC-2 (transital proliferating cells 2). Regarding the distribution of cell types, TAC-1 and TAC-2 cell types are more dispersed, reflecting the high heterogeneity of these cells in the data space and their complex or diverse biological characteristics.

[0176] like Figure 5 As shown, a joint representation of cell and gene embeddings is visualized using dimensionality reduction techniques. Figure 5The distribution of genes overlaps with that of cell types, suggesting a potential intrinsic link between genes and cell types. To analyze the correlation between specific genes and cell types, relevant marker genes are labeled in the visualization to reveal the expression patterns of key genes in different cell types. Figure 5 The distribution of Igf1 (Insulin-like growth factor 1), Tsacc (TSSK6-activating co-chaperone), Tmpo (Thymopoietin / LAP2), and Syt7 (Synaptotagmin-7) are shown in the figure.

[0177] like Figure 6 As shown, by mapping the embedding data of cells, genes, and chromatin accessibility regions to a unified low-dimensional space and using UMAP (Uniform Manifold Approximation and Projection) for visualization, the two types of omics data, "single-cell transcriptome (genes)" and "single-cell chromatin accessibility (peaks)," are unified with the features of the "cell" node into the same embedding space, achieving cross-modal fusion of the three types of nodes; the spatial distribution reveals the distribution and potential relationships of cell types, genes, and peaks in the low-dimensional space.

[0178] like Figure 7 As shown, a single-cell multi-omics triplet prediction system based on graph causal learning is used to execute a single-cell multi-omics triplet prediction method based on graph causal learning, including:

[0179] The acquisition module 101 acquires multi-omics raw data including single-cell transcriptome sequencing and single-cell chromatin accessibility sequencing and preprocesses it to obtain preprocessed cell nodes, peak nodes and gene nodes.

[0180] The node generation module 102 constructs a heterogeneous graph based on the preprocessed cell nodes, peak nodes, and gene nodes, and generates single-cell multi-omics triplet nodes based on the heterogeneous graph.

[0181] The feature splicing module 103 performs multi-layer propagation and feature aggregation on the single-cell multi-omics triplet nodes on the heterogeneous graph through a graph neural network to obtain node embedded features, and splices the node embedded features to obtain a splicing vector.

[0182] The model building module 104 constructs a causal prediction model and trains the causal prediction model through dual contrastive learning and causal counterfactual learning. The causal prediction model includes an overall semantic module and a directional causal module.

[0183] The prediction module 105 obtains the overall semantic features by fusing the global semantic features of the spliced ​​vector through the overall semantic module; it obtains the directional features of the single-cell multi-omics triplet by extracting the directional features of the single-cell multi-omics triplet through the directional causal module; and it obtains the directional representation by fusing the overall semantic features and the directional representation to obtain the triplet prediction vector.

[0184] The classification module 106 inputs the triplet prediction vector into the classifier to obtain the single-cell multi-omics triplet classification probability vector.

[0185] Through the collaborative work of the above modules, this invention constructs heterogeneous graphs, integrates global semantics and directional causal features, combines a dual contrastive learning mechanism, and introduces a causal counterfactual enhancement strategy to achieve accurate prediction of multiple cell-gene-chromatin open region ternary structures. It can simultaneously distinguish seven cell-peak-gene ternary structure patterns, achieving more refined biological structure prediction. Causal reasoning and counterfactual enhancement significantly alleviate prediction bias caused by class imbalance and structural redundancy, improve the model's biological interpretability and generalization ability, and the training embedding results can be used for cell regulatory mechanism analysis and disease research. It can be applied to single-cell multi-omics analysis scenarios such as cell type identification, disease mechanism research, and drug target screening.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A single-cell multi-omics triple prediction method based on graph causal learning, characterized in that, include: S1: Acquire multi-omics raw data including single-cell transcriptome sequencing and single-cell chromatin accessibility sequencing and preprocess them to obtain preprocessed cell nodes, peak nodes and gene nodes. S2: Construct a heterogeneous graph based on the preprocessed cell nodes, peak nodes, and gene nodes, and generate single-cell multi-omics triplet nodes based on the heterogeneous graph; S3: Multi-layer propagation and feature aggregation are performed on single-cell multi-omics triple nodes on heterogeneous graphs through graph neural networks to obtain node embedding features, and the node embedding features are concatenated to obtain concatenated vectors. S4: Construct a causal prediction model and train it through dual contrastive learning and causal counterfactual methods. The causal prediction model includes an overall semantic module and a directional causal module. The dual contrastive learning includes node consistency constraints and structural pattern-level semantic aggregation; The node consistency constraints include: S411: Randomly mask some edges of a heterogeneous graph to obtain a perturbation graph; S412: Extract node embeddings of heterogeneous graphs through a graph convolutional neural network with shared parameters to obtain heterogeneous graph node embeddings; extract node embeddings of perturbed graphs through a graph convolutional neural network with shared parameters to obtain perturbed graph node embeddings. S413: Calculate the cosine similarity between the node embeddings of the heterogeneous graph and the node embeddings of the perturbed graph to obtain the node consistency contrast loss; The structural pattern-level semantic aggregation includes: S421: Calculate the mean node embedding of triple samples in the heterogeneous graph to obtain the prototype vector of the heterogeneous graph; calculate the mean node embedding of triple samples in the perturbation graph to obtain the prototype vector of the perturbation graph. S422: Calculate the prototype consistency loss based on the heterogeneous graph prototype vector and the perturbation graph prototype vector; S423: Obtain multi-class probability distributions by calculating the similarity between triples and heterogeneous graph prototype vectors, as well as the similarity between triples and perturbation graph prototype vectors. S424: Calculate cross-entropy loss based on multiple probability distributions; Counterfactual causality includes: S431: Replace the target nodes of some edges in the edge set to generate counterfactual triples; S432: Use KL divergence to measure the difference in embedding distributions between counterfactual triples and generate counterfactual loss; The causal prediction model is jointly optimized by node consistency contrast loss, prototype consistency loss, cross-entropy loss, and counterfactual loss. S5: The global semantic features of the spliced ​​vector are fused through the overall semantic module to obtain the overall semantic features; the directional features of the single-cell multi-omics triples are extracted through the directional causal module to obtain the directional representation; the overall semantic features and the directional representation are fused to obtain the triple prediction vector. S6: Input the triplet prediction vector into the classifier to obtain the single-cell multi-omics triplet classification probability vector.

2. The single-cell multi-omics triple prediction method based on graph causal learning according to claim 1, characterized in that, The preprocessing includes: Peak screening was performed on single-cell chromatin accessibility sequencing, and the retained peak features were subjected to feature dimensionality reduction to obtain a set of peak feature nodes; Gene screening was performed on single-cell transcriptome sequencing, and gene-related peaks were selected to obtain a set of gene nodes.

3. The single-cell multi-omics triple prediction method based on graph causal learning according to claim 1, characterized in that, Step S2 includes: S21: Construct an edge relationship matrix based on the preprocessed cell nodes, peak nodes, and gene nodes, according to the biological regulatory relationships; S22: An embedding learning method based on PyTorch-BigGraph is used to learn the vectorized representation of cell nodes, peak nodes, and gene nodes to obtain node embeddings; S23: Construct a heterogeneous graph based on node embeddings and edge relationship matrices; S24: Construct seven types of three-node structures based on the heterogeneous graph.

4. The single-cell multi-omics triple prediction method based on graph causal learning according to claim 3, characterized in that, The seven types of three-node structures include triangular structures, cell-centered bilateral structures, peak-centered bilateral structures, gene-centered bilateral structures, gene-cell unilateral structures, gene-peak unilateral structures, and cell-peak unilateral structures.

5. The single-cell multi-omics triple prediction method based on graph causal learning according to claim 1, characterized in that, Step S3 includes: S31: Extract the initial cellular features, initial peak features, and initial gene features of the triplet node; S32: Cell node embedding features are obtained by multi-layer propagation of initial cell features through graph convolutional neural networks; peak node embedding features are obtained by multi-layer propagation of initial peak features through graph convolutional neural networks; gene node embedding features are obtained by multi-layer propagation of initial gene features through graph convolutional neural networks. S33: Concatenate the cell node embedding features, peak node embedding features, and gene node embedding features into vectors to obtain a concatenated vector.

6. The single-cell multi-omics triple prediction method based on graph causal learning according to claim 1, characterized in that, Obtaining directional characterization includes: S521: Construct the cell peak direction difference vector, the peak gene direction difference vector, and the gene cell direction difference vector; S522: Directional characterization is obtained by nonlinearly mapping and splicing the cell peak direction difference vector, peak gene direction difference vector, and gene cell direction difference vector.

7. A single-cell multi-omics triplet prediction system based on graph causal learning, characterized in that, The single-cell multi-omics triple prediction method based on graph causal learning, as described in any one of claims 1 to 6, is used to perform the following: The acquisition module acquires multi-omics raw data including single-cell transcriptome sequencing and single-cell chromatin accessibility sequencing and preprocesses it to obtain preprocessed cell nodes, peak nodes and gene nodes. A node generation module, which constructs a heterogeneous graph based on preprocessed cell nodes, peak nodes and gene nodes, and generates single-cell multi-omics triplet nodes based on the heterogeneous graph. The feature splicing module performs multi-layer propagation and feature aggregation on single-cell multi-omics triple nodes on the heterogeneous graph through a graph neural network to obtain node embedded features, and splices the node embedded features to obtain a splicing vector. The model building module constructs a causal prediction model and trains it through dual contrastive learning and causal counterfactual methods. The causal prediction model includes an overall semantic module and a directional causal module. The prediction module fuses the global semantic features of the spliced ​​vector through the overall semantic module to obtain the overall semantic features; it extracts the directional features of the single-cell multi-omics triples through the directional causal module to obtain the directional representation; and it fuses the overall semantic features and the directional representation to obtain the triple prediction vector. The classification module inputs the triplet prediction vector into the classifier to obtain the single-cell multi-omics triplet classification probability vector.

Images