Single-cell multi-omics gene regulatory network reasoning method based on comparative learning

Abstract

The invention discloses a single-cell multi-omics gene regulation network reasoning method based on comparative learning, and belongs to the technical field of biological information, and the method comprises the following steps: obtaining single-cell RNA sequencing data and single-cell chromatin accessibility data, and carrying out preprocessing to obtain a prior network of multi-omics data; loading the prior regulation and control network into a graph structure, and converting the graph structure into a feature space matrix; calculating a Pearson's correlation coefficient of two nodes on each side in the graph structure, when the Pearson's correlation coefficient is greater than a preset threshold value, taking the Pearson's correlation coefficient as a node pair, and storing the node pair and the corresponding feature vector and hidden vector into a data set; the data set is divided into a plurality of batches to be randomly input into the comparison learning model, and the node type of each node, the pairing degree of node pairs and node feature reconstruction vectors are obtained; and extracting the mapping of two nodes in the node feature reconstruction vector in the feature space by adopting a regulation and control network interpreter, and calculating the structural entropy of the nodes based on the mapping so as to determine whether the nodes are key nodes or not.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a single-cell multi-omics gene regulatory network inference method based on contrastive learning. Background Art

[0002] Gene expression regulation determines cell function and state. Gene regulatory networks (GRNs) constitute a complex system of gene expression control in cells, which reflects the regulatory interactions between transcription factors (TFs) and their target genes. In this system, genes regulate the expression of other genes through interactions between proteins encoded by transcription. GRNs inference refers to the process of summarizing gene regulatory interactions. It involves converting complex regulatory processes into interpretable network structures through computational methods.

[0003] Studying gene regulatory networks is an important way to reveal the mechanism of disease treatment and key genes. Especially in cancer research, certain genes, known as cancer driver genes, play a central role in the occurrence and development of tumors. Mutations in these genes can enable cancer cells to acquire the ability to proliferate uncontrolled, escape immune surveillance and other defense mechanisms, thus becoming a key driver of cancer progression. For example, mutations in genes such as TP53, PIK3CA, KMT2C, ARID1A, KMT2D, LRP1B, PTEN, RB1, FAT4, and KRAS are closely related to the formation of various types of cancer. In-depth exploration of gene regulatory networks can clarify the transcriptional regulatory mechanism of genes, thereby revealing the therapeutic targets and pathological mechanisms of diseases. By comparing samples extracted from healthy states and analyzing gene regulatory networks with gene regulatory networks analyzed from diseased state samples, key genes in diseases can be analyzed.

[0004] At present, the inference of gene regulatory networks has gone through three major development stages, namely, using multi-cell omics data, single-cell single-omics data, and single-cell multi-omics cell data for gene regulatory network inference. The comprehensive impact of multi-omics features on gene expression can significantly narrow the search space, making the integration of diverse data a key research direction in GRN inference.

[0005] There are currently two main data integration methods for single-cell multi-omics cell data. The first method involves inputting multi-omics data into a deep learning model with limited preprocessing to integrate multi-omics data into the feature space. For example, GLUE is a variational autoencoder model that can integrate multi-omics data. By embedding a priori knowledge network and adjusting the variational autoencoder model, GLUE significantly reduces the impact of data heterogeneity on the model. The second method is heterogeneous network embedding, which is a computer method that can model multiple mechanisms simultaneously. Heterogeneous networks abstract various types of data and their relationships into multiple nodes and edges, which are then processed using machine learning methods to extract useful information. GEEK is a classic heterogeneous network evaluation framework that constructs various "meta-paths" to extract more information, such as protein, gene-binding associations, and chromatin accessibility to predict gene expression levels.

[0006] However, these existing methods focus on constructing a single regulatory network using multi-omics data. The method of constructing a regulatory network is complex and requires a lot of computing power. In addition, the existing methods only study the mechanism of the network at the overall level, and it is necessary to conduct an interpretable analysis of the internal mechanism of the network. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the single-cell multi-omics gene regulatory network inference method based on contrastive learning provided by the present invention solves the problem that the prior art method for constructing regulatory networks is complex and requires a lot of computing power.

[0008] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0009] A single-cell multi-omics gene regulatory network inference method based on contrastive learning is provided, which comprises the steps of:

[0010] S1. Obtain single-cell RNA sequencing data and single-cell chromatin accessibility data, and preprocess them to obtain a priori networks of multi-omics data;

[0011] S2, loading the prior regulatory network into a graph structure including transcription factor nodes and target gene nodes, and using the node2vec method to convert the graph structure into a feature space matrix including several hidden vectors;

[0012] S3, calculate the Pearson correlation coefficient of the two nodes on each edge in the graph structure, and when it is greater than a preset threshold, store the node pair and its corresponding feature vector and hidden vector in the data set as a node pair;

[0013] S4, divide the data in the data set into several batches, randomly input the contrastive learning model, and obtain the node type of each node, the pairing degree of the node pair, and the node feature reconstruction vector;

[0014] S5. Use the control network interpreter to extract the mapping of two nodes in the node feature reconstruction vector in the feature space, and calculate the structural entropy of the node based on the mapping to determine whether the node is a key node.

[0015] Furthermore, the method of preprocessing to obtain the prior network of multi-omics data includes:

[0016] Cluster the single-cell RNA sequencing data and single-cell chromatin accessibility data, and use the MACS2 tool to analyze the peaks of the clustered single-cell chromatin accessibility data to generate peak coordinate data;

[0017] Obtain sequence-specific motifs related to transcription factors in the Cis-BP database and human gene data and their corresponding position information in the human genome database;

[0018] The bedtools tool was used to pair the peak coordinate information with sequence-specific motifs and the peak coordinate information with human genes;

[0019] Sequence-specific motifs were connected to genes based on the pairing information to obtain a multi-omics network file containing gene, motif information, and peak coordinate data;

[0020] Obtain transcription factor and motif datasets, delete genes and motifs in the multi-omics network file that are not in the transcription factor and motif datasets, and obtain the prior regulatory network of the multi-omics data.

[0021] The beneficial effects of the above technical solution are as follows: This preprocessing solution constructs an accurate multi-omics network file by clustering single-cell RNA sequencing data and single-cell chromatin accessibility data and combining transcription factor motif information with gene location information, thereby improving the accuracy of regulatory network construction, providing important information for revealing the gene regulatory mechanism, and providing strong data support for the next step of the invention.

[0022] Furthermore, the method of converting the graph structure into a feature space matrix using the node2vec method includes:

[0023] Set the number of visits to each node in each graph structure, then randomly select a node in the graph structure, start the random walk strategy, and obtain the walk paths of all nodes in the graph structure;

[0024] The walking paths of all nodes are input into the Skip-gram model to obtain the hidden vector of the walking path of each node, and all the hidden vectors are used to form a feature space matrix.

[0025] The beneficial effects of the above technical solution are as follows: the present invention adopts the node2vec method, by setting the number of node visits and starting the random walk strategy, obtains the walk paths of all nodes in the graph structure, and then inputs the walk paths into the Skip-gram model to obtain the hidden vector of each node and form a feature space matrix. This method can not only effectively capture the local and global relationships between nodes in the graph structure, but also map multiple node information to the same feature space, so that subsequent models can better understand and predict the regulatory relationship between nodes. In addition, this method is applicable to large-scale and complex graph structures, and improves computational efficiency.

[0026] Furthermore, the objective function of the Skip-gram model is:

[0027]

[0028] Where L is the objective function; P(m,k) is the probability that the next node of node k is node m; V is the node set consisting of all nodes in the graph structure; w is any node in V that is not node k; score(.) is the similarity or proximity function; exp(.) is the power function of the natural logarithm; F(α) is the Skip-gram model;

[0029] The loss function of the Skip-gram model is:

[0030]

[0031] Among them, L M is the loss function.

[0032] The beneficial effects of the above technical solution are as follows: the Skip-gram model adopted in this solution can effectively quantify and predict the pairing probability between nodes by defining the conditional probability and similarity function between nodes, combined with the negative log-likelihood loss function, thereby establishing the correlation between nodes in the graph structure. This method not only improves the prediction accuracy, but also is applicable to large-scale graph data. By learning the low-dimensional representation of nodes, it reduces data complexity and improves computational efficiency.

[0033] Furthermore, the feature vector represents a single value of the chromatin accessibility score or gene expression level of each node.

[0034] Furthermore, the contrastive learning model includes a twin encoder, a multi-head attention mechanism, and a multi-task decoder;

[0035] Methods for obtaining the node type of each node, the matching degree of node pairs, and the node feature reconstruction vector using the contrastive learning model include:

[0036] The feature vectors and hidden vectors of the two nodes u and v of each node pair in the data set are split into corresponding nodes, and input into a twin encoder respectively to calculate the hidden vector R u and R v ;

[0037] Use multi-head attention mechanism to calculate the latent vector R u The attention weights are extracted from the attention matrix and passed to another multi-head attention mechanism to calculate the latent vector R v The attention weight vector of

[0038] According to the attention weight vector, the three output heads of the multi-task decoder are used to output the node type, the pairing degree of node pairs and the node feature reconstruction vector respectively.

[0039] The beneficial effects of the above technical solution are as follows: the contrastive learning model adopted in this solution can simultaneously capture the local and global features of node pairs through the structure of twin encoders, multi-head attention mechanisms and multi-task decoders, pay attention to important information in different subspaces, extract attention matrices and transfer them, and realize multi-objective joint learning. This method not only improves the accuracy of prediction, but also retains the unique information of each node and enhances the representation ability of the model.

[0040] Furthermore, the expressions of the attention weight and the attention weight vector are respectively:

[0041]

[0042] Among them, O u is the attention weight; L i (.) is a linear neural network; Q is the query vector; K is the key vector; T is the transpose; V u is a value vector; d k is the dimension of the key vector and query vector, which is used to scale the attention score to stabilize the gradient during training; QK T is the attention matrix; is the attention weight vector;

[0043] The expression for the output head prediction node type is:

[0044]

[0045] in, is the output head of the predicted node type; Softmax(.) is a mathematical function that converts any real vector into a probability distribution; relu(.) is the activation function; c is the middle value; Dropout(.) is the Dropout layer;

[0046] The expression of the pairing degree of the output head prediction node pair is:

[0047]

[0048] in, is the pairing degree of the node pair; O p is the intermediate parameter;

[0049] The expression for reconstructing the node feature reconstruction vector by the output head is:

[0050]

[0051] Among them, O v Reconstruct the node feature vectors for nodes u and v.

[0052] The beneficial effects of the above technical solution are as follows: the contrastive learning model adopted in this solution can effectively capture the complex relationship between node pairs, focus on important information, and quantify the correlation between node pairs through the QS attention mechanism. This method not only improves the accuracy and robustness of prediction, but also enhances the generalization ability and anti-overfitting ability of the model. In addition, the reconstruction of the node feature reconstruction vector by the output head retains the unique information of each node, improves the representation ability of the model, and then infers the gene regulatory network.

[0053] Furthermore, the loss function L of the reconstruction part of the node feature reconstruction vector is N for:

[0054]

[0055] Where N is the total number of nodes of all types in the graph structure; exp(.) is the power function of natural logarithm; sim(.) is the cosine similarity function; τ is the temperature coefficient of the cosine similarity function; anchor is the midpoint between nodes u and v; x u is the feature vector and hidden vector of node u; x v are the feature vector and hidden vector of node v; mean(.) is the average function; sum(.) is the sum function.

[0056] The beneficial effects of the above technical solution are: the loss function L of this solution is N By adopting the cosine similarity function, temperature coefficient and midpoint "anchor" design, the accuracy of feature reconstruction can be improved, the sensitivity and flexibility of the model to similarity measurement can be enhanced, and the common features between node pairs can be captured. This method not only improves the ability of the reconstruction vector to characterize the common characteristics of node pairs, but also improves the efficiency and accuracy of node feature reconstruction by minimizing the reconstruction error and optimizing the parameters of the model.

[0057] Furthermore, the expression of the structural entropy of the node based on the mapping calculation is:

[0058]

[0059] Among them, x j is the feature vector and hidden vector of node j; w o is the mapping x of node j o With x j The Euclidean distance between j is the structural entropy of node j; D j is the degree of node j; j is any one of the two nodes corresponding to the node feature reconstruction vector.

[0060] The beneficial effects of the present invention are as follows: after the feature space matrix is ​​obtained by this scheme, most of the irrelevant node pairs can be removed through the Pearson correlation coefficient of any two nodes, which can reduce the magnitude of the data set, thereby overcoming the problem that the method of constructing the regulatory network is complicated and requires a lot of computing power; since the quantitative correlation in the data set is relatively large, when performing comparative learning, it can better match the unlabeled regulatory relationship, thereby ensuring the recognition accuracy of the type and the degree of node pair matching.

[0061] When conducting regulatory network inference, this scheme can associate multi-omics information through preprocessing, and then use the node2vec method to obtain information of different dimensions of cells, which can ensure the accuracy of subsequent reasoning processes; this scheme is combined with the comparative learning method to match unlabeled regulatory relationships, which can further improve the accuracy of node (gene and transcription factor) types; finally, the structural entropy obtained by the regulatory network interpreter of this scheme can quickly determine the key nodes. Since the key nodes are selected through structural entropy, it is also explainable. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Flowchart of the single-cell multi-omics gene regulatory network inference method based on contrastive learning.

[0063] Figure 2 The performance comparison chart of node2vec and metapath2vec on the GM12878 dataset, where (a) is the performance comparison chart based on the AUPR indicator; (b) is the performance comparison chart based on the F-Score indicator.

[0064] Figure 3 The t-SNE visualization comparison chart of node2vec and metapath2vec on the GM12878 dataset; (a) is the t-SNE visualization comparison chart of node2vec; (b) is the t-SNE visualization comparison chart of metapath2vec. DETAILED DESCRIPTION

[0065] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0066] refer to Figure 1 , Figure 1 A flowchart of a single-cell multi-omics gene regulatory network inference method based on contrastive learning is shown; Figure 1 As shown, the method S includes steps S1 to S5.

[0067] In step S1, single-cell RNA sequencing data and single-cell chromatin accessibility data are obtained and preprocessed to obtain a priori network of multi-omics data.

[0068] The single-cell RNA sequencing data and single-cell chromatin accessibility data collected in this protocol come from "Integrated single-cell analysis maps the continuous regulatory landscape of human hemato poietic differentiation", Chinese translation: "Integrated single-cell analysis maps the continuous regulatory landscape of human hemato poietic differentiation", author: Buenrostro et al., journal: Cell, publication date: 2018-05.

[0069] In one embodiment of the present invention, the method for preprocessing to obtain a priori network of multi-omics data includes steps S11 to S15:

[0070] S11. Cluster the single-cell RNA sequencing data and the single-cell chromatin accessibility data, and use the MACS2 tool to analyze the peaks of the clustered single-cell chromatin accessibility data to generate peak coordinate data;

[0071] During implementation, the detailed clustering process of this solution is as follows:

[0072] First, the matrices of single-cell RNA data and single-cell chromatin accessibility data are read; then the LIGER tool is used to standardize the data; next, joint non-negative matrix factorization (iNMF) is performed on the standardized data to extract common and dataset-specific genes; then the cells are clustered according to the genes obtained by iNMF, and the cluster correspondence between datasets is achieved.

[0073] The clustering of this scheme is equivalent to combining and analyzing the two types of data to obtain which cells in the two types of data are in which class, and then putting the two types of data of the cell into the corresponding clusters respectively.

[0074] S12, obtaining sequence-specific motifs related to transcription factors in the Cis-BP database and human gene data and their corresponding position information in the human genome database;

[0075] S13, using bedtools to pair the peak coordinate information with sequence-specific motifs and the peak coordinate information with human genes;

[0076] S14, linking sequence-specific motifs to genes according to the pairing information to obtain a multi-omics network file containing gene, motif information and peak coordinate data;

[0077] S15. Obtain transcription factor and motif datasets, delete genes and motifs in the multi-omics network file that are not in the transcription factor and motif datasets, and obtain the prior regulatory network of the multi-omics data.

[0078] In step S2, the prior regulatory network is loaded into a graph structure including transcription factor nodes and target gene nodes, and the node2vec method is used to convert the graph structure into a feature space matrix including several hidden vectors; this scheme preferably uses the networkx tool to load the prior regulatory network into a graph structure including transcription factor nodes and target gene nodes.

[0079] During implementation, the present solution preferably uses the node2vec method to convert the graph structure into a feature space matrix. Steps S21 and S22 include:

[0080] S21. Set the number of times each node in each graph structure is visited, then randomly select a node in the graph structure, start the random walk strategy, and obtain the walk paths of all nodes in the graph structure; this scheme preferably starts the random walk strategy according to probabilities p and q, where probability p controls the probability of visiting duplicate nodes, and probability q determines the probability of visiting visited nodes or unvisited nodes.

[0081] S22. Input the walking paths of all nodes into the Skip-gram model to obtain the hidden vector of the walking path of each node, and use all the hidden vectors to form a feature space matrix.

[0082] The objective function of the preferred Skip-gram model in this scheme is:

[0083]

[0084] Where L is the objective function; P(m,k) is the probability that the next node of node k is node m; V is the node set consisting of all nodes in the graph structure; w is any node in V that is not node k; score(.) is the similarity or proximity function; exp(.) is the power function of the natural logarithm; F(α) is the Skip-gram model;

[0085] The loss function of the Skip-gram model is:

[0086]

[0087] Among them, L M is the loss function.

[0088] This scheme optimizes the setting of the Skip-gram model optimization objective and the damage function, and takes the maximization of the edges in the coverage structure graph as the optimization objective. The more edges are covered, the more regulatory relationships there are, so this scheme may find more gene regulatory relationships that have not been proposed.

[0089] In step S3, the Pearson correlation coefficient of the two nodes on each edge in the graph structure is calculated. When it is greater than a preset threshold, the node pair and its corresponding feature vector and hidden vector are stored in the data set as a node pair.

[0090] In step S4, the data in the data set are divided into several batches and randomly input into the contrastive learning model to obtain the node type of each node, the matching degree of the node pair and the node feature reconstruction vector;

[0091] In one embodiment of the present invention, the contrastive learning model includes a twin encoder, a multi-head attention mechanism and a multi-task decoder; wherein two twin encoders and two multi-head attention mechanisms are provided.

[0092] The method of using the contrastive learning model to obtain the node type of each node, the matching degree of the node pair and the node feature reconstruction vector includes steps S41 to S43:

[0093] S41, split the feature vectors and hidden vectors of the two nodes u and v of each node pair in the data set into corresponding nodes, namely the x mentioned later u and x v ; Then input a twin encoder respectively to calculate the latent vector R u and R v .

[0094] S42. Use multi-head attention mechanism to calculate the latent vector R u The attention weights are extracted from the attention matrix and passed to another multi-head attention mechanism to calculate the latent vector R v The attention weight vector.

[0095] Among them, the expressions of attention weight and attention weight vector are:

[0096]

[0097] Among them, O u is the attention weight; L i (.) is a linear neural network; Q is the query vector; K is the key vector; T is the transpose; V u is a value vector; d k is the dimension of the key vector and query vector, which is used to scale the attention score to stabilize the gradient during training; QK T is the attention matrix; is the attention weight vector;

[0098] S43. According to the attention weight vector, the three output heads of the multi-task decoder are used to output the node type, the pairing degree of the node pair and the node feature reconstruction vector respectively.

[0099] This scheme shares the linear transformation weights of Query and Key through two self-attention mechanisms, while Value has independent linear transformation weights. This scheme optimizes the attention mechanism in this way to adapt to the task of quantifying node relationships.

[0100] When implemented, the expression of the preferred output header prediction node type of this scheme is:

[0101]

[0102] in, is the output head of the predicted node type; Softmax(.) is a mathematical function that converts any real vector into a probability distribution; relu(.) is the activation function; c is the middle value; Dropout(.) is the Dropout layer.

[0103] The expression of the pairing degree of the output head prediction node pair is:

[0104]

[0105] in, is the pairing degree of the node pair; O p is the intermediate parameter;

[0106] The expression for reconstructing the node feature reconstruction vector by the output head is:

[0107]

[0108] Among them, O vReconstruct the node feature vectors for nodes u and v.

[0109] The loss function L for the reconstruction of the node feature reconstruction vector N for:

[0110]

[0111] Where N is the total number of nodes of all types in the graph structure; exp(.) is the power function of natural logarithm; sim(.) is the cosine similarity function; τ is the temperature coefficient of the cosine similarity function; anchor is the midpoint between nodes u and v; x u is the feature vector and hidden vector of node u; x v are the feature vector and hidden vector of node v; mean(.) is the average function; sum(.) is the sum function.

[0112] In step S5, the control network interpreter is used to extract the mapping of two nodes in the node feature reconstruction vector in the feature space, and the structural entropy of the node is calculated based on the mapping to determine whether the node is a key node.

[0113] During implementation, the scheme preferably calculates the structural entropy of the node based on the mapping as follows:

[0114]

[0115] Among them, x j is the feature vector and hidden vector of node j; w o is the mapping x of node j o With x j The Euclidean distance between j is the structural entropy of node j; D j is the degree of node j; j is any one of the two nodes corresponding to the node feature reconstruction vector.

[0116] The regulatory network interpreter of this scheme has two purposes: first, it can confirm which nodes in the network are key nodes; second, it can identify important modules in the regulatory network. When identifying key nodes, it is believed that lower structural entropy indicates that the node is more stable. This scheme can set a threshold, and when the structural entropy is lower than this threshold, it can indicate that the node corresponding to the structural entropy is a key node.

[0117] The following is a detailed description of the effect of the single-cell multi-omics gene regulatory network inference method of this scheme combined with specific comparative experiments:

[0118] Experiment,parameters and dataset settings

[0119] This protocol evaluates the effectiveness of this protocol in inferring single-cell multi-omics gene regulatory networks through performance comparison. To evaluate the performance of the framework, this protocol compares the evaluation framework with seven baseline methods: scMTNI, GNAT, INDEP, SCENIC, Lasso, AMuSR-tuned, and MRTLE.

[0120] scMTNI is a multi-task learning evaluation framework that aims to infer regulatory networks for specific cell types by optimizing hyperparameters such as marginal probabilities, sparsity penalties, and the strength of prior networks. MRTLE is an algorithm based on probabilistic graphical models that simultaneously infers genome-scale regulatory networks across species using phylogenetic structures, multi-species transcriptomic data, and sequence-specific patterns. The GNAT algorithm uses tissue hierarchies to share relevant information and infer tissue-specific gene co-expression networks, each of which is modeled using a Gaussian Markov Random Field (GMRF). The AMuSR algorithm uses sparse block sparse regression to estimate the activity of transcription factors and infer gene regulatory networks from expression datasets, decomposing the model coefficient matrix through multi-task learning to capture conservation and dataset-specific interactions. INDEP is a single-task evaluation framework version of scMTNI that infers regulatory networks for each cell type independently, using a dependency network model, and learning the regulatory network of each cell type through a score-based search algorithm. The LASSO method uses linear regression with L1 regularization to predict the expression profile of the target gene based on the expression profile of the candidate regulator, and infers the regulator with a non-zero coefficient as the regulator of the gene. The SCENIC algorithm uses GENIE3 or GRNBoost2 as part of the Arboreto evaluation framework to infer transcription factor-target gene relationships, using a tree ensemble approach that can be directly applied to each cell type-specific dataset.

[0121] In performance evaluation, this scheme uses two metrics, AUPR (area under the precision-recall curve) and F-Score, to evaluate its effectiveness in inferring regulatory networks. To ensure robustness, all experimental models start with five different random numbers, and the comparative experimental results represent the average results of these five iterations. The model's prediction uses a top-k validation method, which selects the first k edges (sorted by weight) inferred by the evaluation framework of this scheme for validation and compares them with the real network containing exactly k edges.

[0122] The node2vec algorithm (multi-omics fusion algorithm, i.e., random walk) is an integral part of the evaluation framework for generating embedding vectors, and its key parameters include embedding_dim, walk_length, context_size, walks_per_node, num_negative_samples, p, and q. Here, embedding_dim represents the dimension of the embedding space, walk_length refers to the length of the random walk, context_size refers to the scope of the considered context, walks_per_node refers to the number of walks per node, num_negative_samples refers to the number of negative samples, p is the probability of returning to the previous node during the walk, and q is the balance between exploration and exploitation during the random walk.

[0123] The hyperparameters of the evaluation framework of this scheme are configured as follows: batch_size is set to 64, α is set to 128, pearson_threshold is set to 0.75, hidden_layer_size is set to 128, learning_rate is set to 0.0001, embedding_dim is set to 128, walk_length is set to 10, context_size is set to 5, walks_per_node is set to 8, num_negative_samples is set to 1, p is set to 0.25, and q is set to 4. All experiments were performed on a server equipped with an NVIDIA RTX3080 GPU. These parameter settings are the best results obtained after grid search. When using the metapath2vec method in the comparison experiment, this scheme uses the best performing templates among the Gene-Genome and bin-Gene templates reported in the study.

[0124] This protocol uses a simulated dataset to evaluate the effectiveness of the framework in inferring regulatory networks. This dataset is simulated using the BoolODE method and represents single-cell data of 2,000 cells with a sparsity of 80%. It includes three different cell types: hematopoietic stem cells (HSCs), common myeloid progenitors (CMPs), and granulomonocytic progenitors (GMPs). Each cell type is represented by a sub-dataset containing 15 transcription factors and 65 genes, each of which is accompanied by its corresponding prior regulatory network. The evaluation framework uses this dataset to infer regulatory networks and evaluates the performance of the inferred networks using the AUPR and F-Score metrics.

[0125] The evaluation framework uses human hematopoietic differentiation data to evaluate the effectiveness of various modules in the evaluation framework of this solution. This dataset integrates single-cell accessibility (scATAC-seq) and single-cell RNA sequencing (scRNA-seq) data, covering a diverse set of immune phenotype cells. The dataset is divided into eight different cell clusters, each accompanied by its existing regulatory network. These prior regulatory networks are based on the motifs of transcription factors within a 5000bp window around the transcription start site (TSS) of their genes, indicating that there are regulatory interactions between genes and transcription factors. Therefore, the regulatory network of the dataset contains two types of nodes, namely transcription factors (TFs) and genes, and TF-Gene interaction pairs. The evaluation framework of this solution uses this dataset to infer regulatory networks and evaluates the performance of the inferred networks through AUPR and F-Score metrics. The evaluation framework uses the GM12878 dataset. This dataset contains two types of nodes, genomic bins and genes, and three pairing relationships: genomic bin-genomic bin, gene-gene, and genomic bin-gene. It integrates information on chromatin accessibility (DNase-seq), protein-protein interactions, and gene expression levels. The prior regulatory network of this dataset believes that gene pairing is determined by protein interactions, protein epigenetic modifications establish pairing between genomic bins, and genomic bin-gene pairing depends on whether the gene exists in the genomic bin. This scheme uses the Top-K method to test the AUPR and F-Score indicators of the two embedding and two inference methods in this dataset.

[0126] Comparative experiment

[0127] The performance of the proposed evaluation framework in inferring regulatory networks was tested on three simulated datasets and compared with seven baseline methods. The comparison results are shown in Tables 1 and 2.

[0128] Table 1 Comparison of this scheme with 7 existing methods on simulated data sets (regarding AUPR index)

[0129]

[0130] Table 2 Comparison of this scheme with 7 existing methods on simulated data sets (about F1score index)

[0131]

[0132] As shown in Tables 1 and 2, this scheme outperforms other methods in both AUPR and F-score. Compared with the multi-task model scMTNI, this scheme achieved significant improvements of 2.7%, 2.4%, and 1.4% on the three datasets, respectively, which shows that the multi-task configuration of this scheme is more effective in inferring regulatory networks.

[0133] Similarly, the proposed model maintains an advantage of more than 1.5% in comparison with MRTLE and INDEP, which focus on regulatory networks related to cell evolution, further confirming the superior performance of the evaluation framework in feature extraction. Compared with single-omics machine learning methods such as Lasso and SCENIC, the proposed model also shows significant performance improvement, indicating that the joint inference of multi-omics data can achieve better performance in regulatory network inference.

[0134] In order to evaluate the performance of the evaluation framework of this scheme in integrating multi-omics data, the GM12878 dataset was used to compare the node2vec method (walking path strategy) and the metapath2vec method used in this scheme. The performance of the node2vec method and the metapath2vec method were explained by the AUPR and F-Score indicators. The two indicators obtained were respectively referred to Figure 2 (a) and (b) in .

[0135] like Figure 2 As shown in Figure 2, when using the same network inference method, metapath2vec performs significantly worse than node2vec on 23 chromatin datasets. This is mainly because metapath2vec is based on fixed-pattern feature vector embedding, which may miss some potential regulatory relationships and cannot fully learn the structural characteristics of the network in complex network inference tasks.

[0136] The node2vec (walking strategy) of this solution outperforms metapath2vec in both AUPR and F-score. This superior performance shows that node2vec is better at capturing complex structures and semantic relationships in gene regulatory networks because its feature vector embedding is more flexible and detailed than the fixed pattern method of metapath2vec.

[0137] This scheme also performs dimensionality reduction and visualization analysis on the eigenvector matrix integrated by node2vec and metapath2vec methods, referring to Figure 3 (a) and (b) in Figure 3 Feature 1 in t-SNE is the first coordinate axis; Feature 2 is the second coordinate axis of t-SNE; red dots represent genes, and blue dots represent genomic fragments. Figure 3 It can be seen that the feature vector matrix constructed by the node2vec method is significantly better than that of the metapath2vec method. This further confirms the effectiveness of this scheme in integrating multiple omics data to infer regulatory networks.

[0138] In general, this scheme can efficiently extract structured features from multi-omics data to facilitate the inference of regulatory networks.

Claims

1. A single-cell multi-omics gene regulatory network inference method based on contrastive learning, characterized in that: Includes steps: S1. Obtain single-cell RNA sequencing data and single-cell chromatin accessibility data, and preprocess them to obtain a priori networks of multi-omics data; S2, loading the prior regulatory network into a graph structure including transcription factor nodes and target gene nodes, and using the node2vec method to convert the graph structure into a feature space matrix including several hidden vectors; S3, calculate the Pearson correlation coefficient of the two nodes on each edge in the graph structure, and when it is greater than a preset threshold, store the node pair and its corresponding feature vector and hidden vector in the data set as a node pair; S4, divide the data in the data set into several batches, randomly input the contrastive learning model, and obtain the node type of each node, the pairing degree of the node pair, and the node feature reconstruction vector; S5. Use the control network interpreter to extract the mapping of two nodes in the node feature reconstruction vector in the feature space, and calculate the structural entropy of the node based on the mapping to determine whether the node is a key node.

2. The single-cell multi-omics gene regulatory network inference method according to claim 1, characterized in that: Methods for preprocessing multi-omics data to obtain prior networks include: Cluster the single-cell RNA sequencing data and single-cell chromatin accessibility data, and use the MACS2 tool to analyze the peaks of the clustered single-cell chromatin accessibility data to generate peak coordinate data; Obtain sequence-specific motifs related to transcription factors in the Cis-BP database and human gene data and their corresponding position information in the human genome database; The bedtools tool was used to pair the peak coordinate information with sequence-specific motifs and the peak coordinate information with human genes; Sequence-specific motifs were connected to genes based on the pairing information to obtain a multi-omics network file containing gene, motif information, and peak coordinate data; Obtain transcription factor and motif datasets, delete genes and motifs in the multi-omics network file that are not in the transcription factor and motif datasets, and obtain the prior regulatory network of the multi-omics data.

3. The single-cell multi-omics gene regulatory network inference method according to claim 1, characterized in that: Methods for converting graph structures into feature space matrices using the node2vec method include: Set the number of visits to each node in each graph structure, then randomly select a node in the graph structure, start the random walk strategy, and obtain the walk paths of all nodes in the graph structure; The walking paths of all nodes are input into the Skip-gram model to obtain the hidden vector of the walking path of each node, and all the hidden vectors are used to form a feature space matrix.

4. The single-cell multi-omics gene regulatory network inference method according to claim 3, characterized in that: The objective function of the Skip-gram model is: Where L is the objective function; P(m,k) is the probability that the next node of node k is node m; V is the node set consisting of all nodes in the graph structure; w is any node in V that is not node k; score(.) is the similarity or proximity function; exp(.) is the power function of the natural logarithm; F(α) is the Skip-gram model; The loss function of the Skip-gram model is: Among them, L M is the loss function.

5. The single-cell multi-omics gene regulatory network inference method according to claim 1, characterized in that: The feature vector represents a single value for each node's chromatin accessibility score or gene expression level.

6. The single-cell multi-omics gene regulatory network inference method according to claim 1 or 2, characterized in that: The contrastive learning model includes a twin encoder, a multi-head attention mechanism, and a multi-task decoder; Methods for obtaining the node type of each node, the matching degree of node pairs, and the node feature reconstruction vector using the contrastive learning model include: The feature vectors and hidden vectors of the two nodes u and v of each node pair in the data set are split into corresponding nodes, and input into a twin encoder respectively to calculate the hidden vector R u and R v ; Use multi-head attention mechanism to calculate the latent vector R u The attention weights are extracted from the attention matrix and passed to another multi-head attention mechanism to calculate the latent vector R v The attention weight vector of According to the attention weight vector, the three output heads of the multi-task decoder are used to output the node type, the pairing degree of node pairs and the node feature reconstruction vector respectively.

7. The single-cell multi-omics gene regulatory network inference method according to claim 6, characterized in that: The expressions of the attention weight and the attention weight vector are respectively: Among them, O u is the attention weight; L i (.) is a linear neural network; Q is the query vector; K is the key vector; T is the transpose; V u is a value vector; d k is the dimension of the key vector and query vector; QK T is the attention matrix; is the attention weight vector; The expression for the output head prediction node type is: in, is the output head of the predicted node type; Softmax(.) is a mathematical function that converts any real vector into a probability distribution; relu(.) is an activation function; c is the middle value; Dropout(.) is the Dropout layer; The expression of the pairing degree of the output head prediction node pair is: in, is the pairing degree of the node pair; O p is the intermediate parameter; The expression for reconstructing the node feature reconstruction vector by the output head is: Among them, O v Reconstruct the node feature vectors for nodes u and v.

8. The single-cell multi-omics gene regulatory network inference method according to claim 7, characterized in that: The loss function L for reconstructing the node feature reconstruction vector N for: Where N is the total number of nodes of all types in the graph structure; exp(.) is the power function of natural logarithm; sim(.) is the cosine similarity function; τ is the temperature coefficient of the cosine similarity function; anchor is the midpoint between nodes u and v; x u is the feature vector and hidden vector of node u; x v are the feature vector and hidden vector of node v; mean(.) is the average function; sum(.) is the sum function.

9. The single-cell multi-omics gene regulatory network inference method according to claim 7, characterized in that: The expression of the structural entropy of the node based on the mapping calculation is: Among them, x j is the feature vector and hidden vector of node j; w o is the mapping x of node j o With x j The Euclidean distance between j is the structural entropy of node j; D j is the degree of node j; j is any one of the two nodes corresponding to the node feature reconstruction vector.

Images