Single-cell transcriptional gene regulatory network inference method, model, system and storage medium

Abstract

The application relates to the technical field of biological information, in particular to a single-cell transcription gene gene regulation network inference method, model, system and storage medium. A correlation vector of each gene pair in to-be-measured single-cell transcription gene data is determined, and each gene pair is represented in the form of a histogram; time characteristics in the correlation vector are captured through a time sequence convolution network, and space characteristics in the histogram are captured through a convolution network; the time characteristics and the space characteristics are weighted and fused to obtain fused characteristics; and a gene regulation network inference result of the to-be-measured single-cell transcription gene data is predicted based on the fused characteristics. The application aims to solve the problem of how to predict single-cell transcriptome data in a gene regulation network.

Description

Technical Field

[0001] This application relates to the field of bioinformatics, and in particular to a method, model, system, and storage medium for inferring gene regulatory networks of single-cell transcription genes. Background Technology

[0002] Gene regulatory networks are network structures composed of complex interactions between genes. By studying gene regulatory networks, researchers can reveal the information transmission pathways between genes, which is of great significance for understanding the various stages of the cell life cycle, damage repair mechanisms, and the basic mechanisms of life activities such as apoptosis.

[0003] With the development of computer vision technology, some studies (Yuan Y, Bar-Joseph Z. Deep learning for inferring gene relationships from single-cell expression data[J]. Proceedings of the National Academy of Sciences, 2019, 116(52): 27151-27158.) have proposed to express the expression data of paired population genes in the form of visual histograms and further analyze them through convolutional neural networks to capture the potential regulatory relationships between genes.

[0004] However, single-cell RNA sequencing differs from traditional population sequencing. The inherent high dimensionality, strong noise, and sparsity of single-cell RNA data make it difficult for traditional convolutional neural networks to accurately capture the regulatory relationships between genes, making it a technical challenge to accurately infer gene regulatory networks from them.

[0005] Therefore, this application aims to propose a novel gene regulatory network inference method based on single-cell RNA to achieve accurate prediction of gene regulatory networks from single-cell transcriptome data. Summary of the Invention

[0006] The main objective of this application is to provide a method for inferring gene regulatory networks from single-cell transcription genes, aiming to solve the problem of how to predict gene regulatory networks from single-cell transcriptome data.

[0007] To achieve the above objectives, this application provides a method for inferring the gene regulatory network of cellular transcription genes, the method comprising:

[0008] Determine the correlation vector for each gene pair in the single-cell transcriptional gene data to be tested, and represent each gene pair in the form of a histogram;

[0009] The construction of the correlation vector includes: first calculating the ordered correlation vector between the gene pairs based on the statistical standard method, and then using a sliding window mechanism to save the pseudo-temporal information in the gene pairs;

[0010] The histogram includes the main image corresponding to the gene pair, and 2r+2 neighboring images of the gene pair with strong positive covariance.

[0011] Temporal features in the relevant vectors are captured using a temporal convolutional network, and spatial features in the histogram are captured using a convolutional network.

[0012] The temporal and spatial features are weighted and fused to obtain the fused features;

[0013] Based on the fusion features, the gene regulatory network inference results of the single-cell transcriptional gene data to be tested are predicted.

[0014] Optionally, the expression for the relevant vector is:

[0015]

[0016] in,

[0017]

[0018]

[0019]

[0020]

[0021] In the formula, Represents the set of related vectors. These represent the correlation vectors for each gene pair. , , The number of related vectors, The length of each related vector, Gene expression vectors, arranged in pseudo-time. The data fragment of transcription factor u extracted from the sliding window m. The sliding window interval corresponds to the transcription factor, and s is the sliding window size. This refers to the data fragment of the target gene v extracted from the sliding window n. This represents the sliding window interval corresponding to the target gene.

[0022] Optionally, the step of capturing the temporal features in the relevant vectors through a temporal convolutional network specifically includes:

[0023] Capturing dependencies in time series data through convolution operations; and...

[0024] The receptive field is expanded using dilated convolution, and the calculation expression for dilated convolution includes:

[0025]

[0026] In the formula, Indicates at time step The output of , where w is the convolution kernel. The convolution kernel is represented by the first... Each weight, Indicates the input sequence at time step The value, The expansion rate is denoted as .

[0027] Optionally, the step of weightedly fusing the temporal features and the spatial features to obtain the fused features specifically includes:

[0028] The attention weights corresponding to the temporal and spatial features are calculated using the dot product method.

[0029] The attention weights are multiplied one by one by each element of the temporal feature and each element of the spatial feature to obtain the temporal weighted output and the spatial weighted output, respectively.

[0030] The time-weighted output and the spatial-weighted output are concatenated to obtain the fused feature.

[0031] Furthermore, to achieve the above objectives, this application also provides a single-cell gene regulatory network inference model, which includes:

[0032] The data preprocessing module determines the correlation vector of each gene pair in the single-cell transcriptional gene data to be tested, and represents each gene pair in the form of a histogram.

[0033] The construction of the correlation vector includes: first calculating the ordered correlation vector between the gene pairs based on the statistical standard method, and then using a sliding window mechanism to save the pseudo-temporal information in the gene pairs;

[0034] The histogram includes the main image corresponding to the gene pair, and 2r+2 neighboring images of the gene pair with strong positive covariance.

[0035] The spatiotemporal feature learning module includes a temporal convolutional network and a convolutional network. The temporal convolutional network captures the temporal features in the relevant vectors, and the convolutional network captures the spatial features in the histogram.

[0036] The feature fusion module weights and fuses the temporal features and the spatial features to obtain fused features;

[0037] The prediction module predicts the gene regulatory network inference results of the single-cell transcriptional gene data to be tested based on the fusion features.

[0038] Optionally, the convolutional network includes two convolutional layers with a filter size of 3*3, an output feature map with 32 channels, a max pooling layer with a pooling size of 2*2, and a fully connected layer with 128 nodes.

[0039] In addition, to achieve the above objectives, this application also provides a computer system comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the gene regulatory network inference method for single-cell transcription genes as described in any of the preceding claims.

[0040] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the gene regulatory network inference method for single-cell transcription genes as described in any of the preceding claims.

[0041] This application has at least the following beneficial effects:

[0042] 1. The single-cell transcriptome data is decomposed into relevant vectors and input into a temporal convolutional network for temporal feature extraction, and into histograms and input into a convolutional network for spatial feature extraction. This enables the model to comprehensively analyze the complex patterns and potential relationships in gene expression data from both temporal and spatial dimensions.

[0043] 2. When transforming the histogram, the influence of gene pairs on potential neighboring gene pairs is considered simultaneously. The 2r+2 neighborhood images of genes with strong positive covariance with gene pairs are also input into the convolutional network as histogram data to extract spatial features of the data more comprehensively.

[0044] 3. Use a dual-channel model that includes TCN and CNN. TCN captures temporal features, CNN learns spatial features, and attention mechanism is used to effectively fuse the two, thereby enhancing the model's inference performance. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the first embodiment of the gene regulatory network inference method for single-cell transcription genes involved in this application.

[0046] Figure 2This is a schematic diagram of the framework of the convolutional network involved in the embodiments of this application;

[0047] Figure 3 This is a schematic diagram showing the comparison results of TCGRN and unsupervised methods on six datasets of mouse hematopoietic stem cells involved in the embodiments of this application;

[0048] Figure 4 This is a schematic diagram showing the comparison results of TCGRN and supervised methods on mouse hematopoietic stem cell datasets of different lineage sizes involved in the embodiments of this application.

[0049] Figure 5 This is a schematic diagram showing the comparative experimental results of different models involved in the embodiments of this application on human embryonic stem cell datasets of different sizes;

[0050] Figure 6 This is a comparative diagram of the TCGRN-C model without an attention mechanism and the TCGRN model with an attention mechanism involved in the embodiments of this application.

[0051] Figure 7 This diagram illustrates a comparison between the TCGRN-IM model, which changes the input of a CNN from a histogram to a correlation vector, and the TCGRN model, which retains the histogram as the input of the CNN, as described in this application.

[0052] Figure 8 This is a schematic diagram of the hardware operating environment of the computer system involved in the embodiments of this application.

[0053] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0054] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0055] First Embodiment

[0056] As one implementation scheme, refer to Figure 1 This embodiment provides a method for inferring the gene regulatory network of single-cell transcription genes, the method comprising the following steps:

[0057] Step S10: Determine the correlation vector of each gene pair in the single-cell transcription gene data to be tested, and represent each gene pair in the form of a histogram;

[0058] In this step, the single-cell transcription gene data to be tested is preprocessed. On the one hand, the original data is converted into relevant vectors to capture the temporal information in the data. On the other hand, each gene pair is converted into a histogram image and combined with the images of its related neighboring gene pairs for spatial feature extraction in subsequent steps.

[0059] In this embodiment, the construction of the correlation vector includes: first calculating the ordered correlation vector between the gene pairs based on the statistical standard method, and then using a sliding window mechanism to save the pseudo-temporal information in the gene pairs;

[0060] It should be noted that preprocessing single-cell RNA sequencing data into correlation vectors before inputting them into the TCN can more directly capture the synergistic changes in gene expression patterns, thereby more effectively characterizing the regulatory relationships between genes. By calculating correlations within a sliding window, noise in single-cell data can be smoothed to some extent. The resulting correlation sequences also provide a more informative and suitable input format for the TCN to learn how gene interactions are established and evolve, ultimately helping to improve the accuracy of gene regulatory network inference.

[0061] Furthermore, since the single-cell transcriptional gene data to be tested is closely related to the cell differentiation process, and each cell differentiation stage reflects potential temporal information, in order to describe the relative order of cells in the differentiation process, pseudotime analysis is used to determine the pseudotime information of each gene pair in the single-cell transcriptional gene data to be tested.

[0062] In some alternative implementations, pseudo-time information can be calculated using the SlingShot method, which infers pseudo-time by mapping cells to a low-dimensional space and constructing trajectories in that space, thereby clarifying the position of each cell during differentiation.

[0063] Furthermore, during data preprocessing, statistical standards are used to measure the correlation between gene pairs, extracting a series of ordered correlation vectors reflecting the relationship between gene pairs, and a sliding window mechanism is used to store pseudo-temporal information. In some optional implementations, the expression for the correlation vector is:

[0064]

[0065] in,

[0066]

[0067]

[0068]

[0069]

[0070] In the formula, Represents the set of related vectors. These represent the correlation vectors for each gene pair. , , The number of related vectors, The length of each related vector, Gene expression vectors, arranged in pseudo-time. The data fragment of transcription factor u extracted from the sliding window m. The sliding window interval corresponds to the transcription factor, and s is the sliding window size. This refers to the data fragment of the target gene v extracted from the sliding window n. This represents the sliding window interval corresponding to the target gene.

[0071] Further and optionally, for each extracted transcription factor (TF) and target gene data segment, their Pearson correlation coefficients are calculated to characterize the correlation between gene pairs.

[0072] In this embodiment, the histogram includes the main image corresponding to the gene pair, and 2r+2 neighboring images of the gene pair with strong positive covariance.

[0073] For the histogram of gene pairs, the histogram transformed in this embodiment not only transforms the gene pairs themselves, but also needs to consider the potential neighbor pairs of the gene pairs.

[0074] Specifically, let's say we have a gene pair, gene u and gene v, and generate the following for gene u and gene v. Pixel image and genes and The neighborhood image includes:

[0075] (1) ,in and These are the first r genes that have strong positive covariance with gene u and gene v, respectively (r is 10 by default).

[0076] (2) The self-images of genes u and v: and .

[0077] That is, 2r+2 neighborhood images of genes with strong positive covariance with gene pairs are generated, and these images are input into the CNN along with the main image to comprehensively extract the spatial features of the data.

[0078] Step S20: Capture the temporal features in the relevant vectors through a temporal convolutional network, and capture the spatial features in the histogram through a convolutional network;

[0079] Step S30: The temporal features and spatial features are weighted and fused to obtain fused features;

[0080] In this embodiment, a hybrid network model (TCGRN) containing a temporal convolutional network (TCN) and a convolutional network (CNN) is used for feature extraction: TCN extracts temporal features from relevant vectors, while CNN captures spatial features in the histogram, and the two types of features are weighted and fused.

[0081] TCGRN treats the inference problem as a binary classification task, namely, determining whether there is a regulatory relationship between gene pairs.

[0082] In some alternative implementations, in addition to capturing dependencies in time series through convolution operations, TCNs also utilize dilated convolutions to expand the receptive field, thereby reducing the number of gradient propagations. Specifically, this includes:

[0083] Capturing dependencies in time series data through convolution operations; and...

[0084] The receptive field is expanded using dilated convolution, and the calculation expression for dilated convolution includes:

[0085]

[0086] in, Indicates at time step The output of , where w is the convolution kernel. The convolution kernel is represented by the first... Each weight, Indicates the input sequence at time step The value, The expansion rate is denoted as .

[0087] In the formula, the size of the convolution kernel The dilation rate determines the number of input elements involved in each convolution operation. The number of elements the convolution kernel skips in the input sequence is determined by this. Dilated convolution expands the receptive field by introducing holes in the convolution kernel, thereby capturing longer temporal dependencies.

[0088] In some alternative implementations, three different expansion rates are used to capture time dependence at different scales.

[0089] In this embodiment, TCGRN converts the raw data into a histogram to learn the spatial features of the data because image data has unique spatial properties and structure, which CNNs can utilize to achieve efficient and accurate feature extraction. Furthermore, the convolutional kernels in CNNs share the same weights across all locations in the image, which not only reduces the number of model parameters, thus improving training efficiency, but also enhances the model's translation invariance. Through the stacking of convolutional and pooling layers, CNNs can extract low-level, mid-level, and high-level features of the image layer by layer.

[0090] In some alternative implementations, an attention mechanism is used to integrate the temporal and spatial features of the data, specifically including the following steps:

[0091] Step S31: Calculate the attention weights corresponding to the temporal features and the spatial features using the dot product method;

[0092] For example, let a represent the attention weight, e represent the temporal feature of the TCN output, and c represent the spatial feature of the CNN output. The expression for calculating the attention weight is:

[0093]

[0094] Step S32: Multiply the attention weights by each element of the temporal feature and each element of the spatial feature to obtain the temporal weighted output and the spatial weighted output, respectively.

[0095] For example, the obtained attention weight 'a' is used to weight 'e' and 'c' to generate a weighted feature representation, as shown in the following equation, where, This represents the time-weighted output of TCN. This represents the spatially weighted output of the CNN. This indicates element-wise multiplication.

[0096]

[0097]

[0098] Step S33: The time-weighted output and the spatial-weighted output are concatenated to obtain the fused feature.

[0099] Step S40: Based on the fusion features, predict the gene regulatory network inference results of the single-cell transcriptional gene data to be tested.

[0100] In this step, the aforementioned fusion features are used by the model to predict gene regulatory networks, thereby predicting the gene regulatory network inference results of the single-cell transcriptional gene data to be tested.

[0101] Furthermore, in some optional implementations, before performing step S40, the sigmoid activation function is used, and the binary cross-entropy loss is applied to train the model, as shown in the following formula:

[0102]

[0103] In the formula, N represents the number of samples. The label representing sample i, This represents the probability that sample i is predicted to be a positive label by the sigmoid function.

[0104] In the technical solution provided in this embodiment, single-cell transcriptome data is decomposed into relevant vectors and input into a temporal convolutional network for temporal feature extraction, and also into histograms and input into a convolutional network for spatial feature extraction. This allows the model to comprehensively analyze the complex patterns and potential relationships in gene expression data from both temporal and spatial dimensions. In addition, when transforming the histogram, the influence of gene pairs on potential neighboring gene pairs is considered simultaneously. The 2r+2 neighborhood images of genes with strong positive covariance with gene pairs are also input into the convolutional network as histogram data to extract spatial features of the data more comprehensively.

[0105] Second Embodiment

[0106] As one implementation, this embodiment provides a single-cell gene regulatory network inference model obtained based on the method described in the first embodiment, wherein the single-cell gene regulatory network inference model includes:

[0107] The data preprocessing module 100 determines the correlation vector of each gene pair in the single-cell transcriptional gene data to be tested, and represents each gene pair in the form of a histogram.

[0108] The construction of the correlation vector includes: first calculating the ordered correlation vector between the gene pairs based on the statistical standard method, and then using a sliding window mechanism to save the pseudo-temporal information in the gene pairs;

[0109] The histogram includes the main image corresponding to the gene pair, and 2r+2 neighboring images of the gene pair with strong positive covariance.

[0110] The spatiotemporal feature learning module 200 includes a temporal convolutional network and a convolutional network. The temporal convolutional network captures the temporal features in the relevant vectors, and the convolutional network captures the spatial features in the histogram.

[0111] The feature fusion module 300 weightedly fuses the temporal features and the spatial features to obtain fused features;

[0112] The prediction module 400 predicts the gene regulatory network inference result of the single-cell transcription gene data to be tested based on the fusion features.

[0113] In some alternative implementations, refer to Figure 2 The diagram shows the framework of the convolutional network, which includes two convolutional layers with a filter size of 3*3, 32 channels in the output feature map, a max pooling layer with a pooling size of 2*2, and a fully connected layer with 128 nodes.

[0114] Third Embodiment

[0115] In this embodiment, the gene regulatory network inference method based on single-cell transcription genes in the first embodiment, or the single-cell gene regulatory network inference model TCGRN in the second embodiment, will be compared with other advanced methods.

[0116] (1) Dataset and Experimental Design

[0117] The dataset used in this embodiment is scRNA-seq data from BEELINE

[74] . The experiment was conducted based on four single-cell RNA sequencing datasets, including three lineages of mouse hematopoietic stem cells (mHSCs): erythrocytes (E), lymphoid cells (L), and granulocyte-macrophages (GM), as well as a human embryonic stem cell (hESC) dataset, to evaluate the generalization ability of TCGRN.

[0118] In the experiment, genes expressed in less than 10% of cells were first removed, and the variance and p-value of the remaining genes were calculated, focusing only on genes with p-values ​​below 0.01. Subsequently, the top 500 and top 1000 highly variable genes were selected from each dataset to construct datasets of different sizes. For each independent dataset, a non-specific ChIP-seq network was used as the reference network. The naming convention for each dataset was "original dataset name - network size". Detailed information about the data is shown in Table 1 below, which lists the number of genes and cells for each dataset.

[0119] Table 1. Detailed information on single-cell transcriptome datasets

[0120] mHSC-GM-500 108 889 17 154 0.0839 mHSC-L-500 114 847 22 147 0.0586 mHSC-E-500 169 1071 21 279 0.0786 hESC-500 188 758 22 236 0.0571 mHSC-GM-1000 321 889 48 566 0.0367 mHSC-L-1000 403 847 63 845 0.0333 mHSC-E-1000 355 1071 45 677 0.0424 hESC-1000 541 758 55 808 0.0272

[0121] When labeling samples, gene pairs of transcription factors and target genes recorded in the standard network are marked as positive samples (1), and gene pairs without regulatory relationship and not labeled in the standard network are marked as negative samples (0).

[0122] As can be seen from the network density in the data table, GRN is a sparse network, therefore the number of positive labels is far less than the number of negative labels. To comprehensively evaluate the model's performance, each dataset is divided into training, validation, and test sets in a 3:1:1 ratio, ensuring that the ratio of positive to negative samples in each subset remains consistent with the original dataset. Due to the imbalance of positive and negative samples in the data, the experiment not only uses AUROC as the evaluation metric but also introduces AUPRC to evaluate the model's performance on imbalanced data.

[0123] (2) Comparative Experiment and Analysis

[0124] To evaluate the performance of TCGRN in gene regulatory network inference, this embodiment selected three unsupervised methods and three supervised learning methods for comparative analysis.

[0125] Among unsupervised methods, PIDC, GENIE3, and GRNBoost2 were selected as comparison models:

[0126] PIDC is an algorithm based on multivariate information theory and partial information decomposition (PID). It identifies regulatory relationships between genes by exploring the statistical dependencies between gene triples in a single-cell gene expression dataset.

[0127] GENIE3 is a machine learning method based on random forests, which constructs GRNs based on regression weight coefficients.

[0128] GRNBoost2 is a regression-based inference method that predicts the expression profile of each gene by training a tree-based regression model, thereby generating partial GRNs. All generated partial GRNs are arranged and integrated according to their importance to ultimately determine the complete GRN.

[0129] In the supervised methods, CNNC, DeepDRIM, and DGRNS were selected as comparison models:

[0130] CNNC represents each pair of genes as a histogram image and uses a convolutional neural network to learn data features, predicting the regulatory relationships between genes through a classifier.

[0131] DeepDRIM adds neighborhood images to each gene pair in addition to histograms representing gene pairs, in order to reduce false positives caused by intergenetic transmission.

[0132] DGRNS is a hybrid learning framework that encodes raw data and combines recurrent neural networks with convolutional neural networks to distinguish whether gene pairs are related.

[0133] See Figure 3The diagram shows the comparison results of TCGRN and unsupervised methods on six datasets of mouse hematopoietic stem cells.

[0134] As shown in the figure, TCGRN, proposed in this chapter, achieves the best results on all six datasets in terms of both AUROC and AUPRC evaluation metrics. On the mESC-E-500 dataset, TCGRN's AUROC value is approximately 24% higher than the second-highest unsupervised method. Due to the imbalance of positive and negative samples in the datasets, the AUPRC values ​​of various methods are generally low. However, on the mESC-E-500 dataset, TCGRN's AUPRC value is still about 18% higher than the second-highest unsupervised method. Experimental results demonstrate that TCGRN can effectively learn data features, thereby improving the accuracy of GRN inference.

[0135] In addition, on the same six datasets, the comparison results between TCGRN and supervised methods on mouse hematopoietic stem cell lineage datasets of different sizes are as follows: Figure 4 As shown in the diagram. Experimental results demonstrate that TCGRN outperforms other models on all datasets, both in terms of AUROC and AUPRC values. On the mHSC-GM-1000 dataset, TCGRN's AUROC score is approximately 13% higher than the second-ranked method, DGRNS. On the mHSC-E-500 dataset, TCGRN's AUPRC score is approximately 11.5% higher than DGRNS. These experimental results demonstrate the effectiveness of TCGRN in GRN inference.

[0136] On the other hand, considering that many deep learning models may perform well on specific types of data but poorly on others, this embodiment further evaluates the models on a mouse dataset and tests them on human datasets of different sizes:

[0137] Experimental results are as follows Figure 5 As shown in the figure, the four colors represent four different methods, and each set of data corresponds to a dataset, visually demonstrating the experimental results of each method on two datasets. The figure shows that TCGRN outperforms the comparison models in both AUROC and AUPRC metrics. On the hESC-500 dataset, TCGRN's AUROC value is approximately 9% higher than the second-ranked DGRNS model, and its AUPRC value is approximately 10% higher. The experimental results demonstrate that TCGRN maintains high performance across datasets of different species and sizes, exhibiting strong generalization ability.

[0138] Fourth embodiment

[0139] To verify the impact of the attention mechanism on model performance in terms of the temporal and spatial features of the fused data, a set of ablation experiments were designed in this embodiment to analyze the feature fusion module involved in the first or second embodiment.

[0140] Specifically, the attention mechanism in the feature fusion part of TCGRN was replaced with simple concatenation, and the method was compared with TCGRN on mouse hematopoietic stem cell datasets of different lineage sizes, with AUROC and AUPRC as evaluation metrics.

[0141] Reference Figure 6 The diagram illustrates a comparison between the TCGRN-C model without and with an attention mechanism. As can be seen, on all experimental datasets, TCGRN's AUROC and AUPRC values ​​are higher than TCGRN-C. For example, on the mHSC-L-500 dataset, TCGRN's AUROC value is about 5% higher than TCGRN-C, and its AUPRC value is improved by about 6%. The experimental results demonstrate that the attention mechanism can better capture the complex relationships between temporal and spatial features, thereby enhancing the model's inference performance.

[0142] Fifth Embodiment

[0143] To verify the impact of inputting image-based data into the model's learning space features on the overall performance of the model, this embodiment designs a set of ablation experiments to analyze the role of histogram image data in feature extraction.

[0144] The specific design involves keeping the other operations unchanged, changing the input of the CNN in TCGRN from image data to a relevant vector, and comparing this method with TCGRN on mouse hematopoietic stem cell datasets of different lineage sizes.

[0145] Reference Figure 7 The diagram shows a comparison between the TCGRN-IM model, which changes the input of the CNN from a histogram to a correlation vector, and the TCGRN model, which retains the histogram as the input of the CNN.

[0146] As shown in the figure, across all datasets, the AUROC and AUPRC values ​​of TCGRN-IM are significantly lower than those of the TCGRN model. For example, on the mHSC-GM-1000 dataset, the AUROC value of TCGRN is about 12% higher than that of TCGRN-IM; on the mHSC-GM-500 dataset, the AUPRC value of TCGRN is about 16% higher than that of TCGRN-IM. This phenomenon indicates that converting the raw data into histogram images before inputting them into a CNN can more fully utilize spatial structure information, enabling the model to learn more complex and globally correlated features.

[0147] As one implementation scheme, Figure 8 This is a schematic diagram of the hardware operating environment of the computer system involved in the embodiments of this application.

[0148] like Figure 8 As shown, the computer system may include: a processor 1001, such as a CPU; a memory 1005; a user interface 1003; a network interface 1004; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0149] Those skilled in the art will understand that Figure 8 The computer system architecture shown does not constitute a limitation on the computer system and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0150] like Figure 8 As shown, the memory 1005, as a storage medium, may include an operating system, a network communication module, a user interface module, and computer programs. The operating system is a program that manages and controls the hardware and software resources of the computer system, as well as the operation of the computer programs and other software or programs.

[0151] exist Figure 8 In the computer system shown, the user interface 1003 is mainly used to connect to the terminal and communicate with the terminal; the network interface 1004 is mainly used to communicate with the backend server; and the processor 1001 can be used to call the computer program stored in the memory 1005.

[0152] In this embodiment, the computer system includes: a memory 1005, a processor 1001, and a computer program stored in the memory and executable on the processor, wherein:

[0153] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations:

[0154] Determine the correlation vector for each gene pair in the single-cell transcriptional gene data to be tested, and represent each gene pair in the form of a histogram;

[0155] The construction of the correlation vector includes: first calculating the ordered correlation vector between the gene pairs based on the statistical standard method, and then using a sliding window mechanism to save the pseudo-temporal information in the gene pairs;

[0156] The histogram includes the main image corresponding to the gene pair, and 2r+2 neighboring images of the gene pair with strong positive covariance.

[0157] Temporal features in the relevant vectors are captured using a temporal convolutional network, and spatial features in the histogram are captured using a convolutional network.

[0158] The temporal and spatial features are weighted and fused to obtain the fused features;

[0159] Based on the fusion features, the gene regulatory network inference results of the single-cell transcriptional gene data to be tested are predicted.

[0160] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations:

[0161] Capturing dependencies in time series data through convolution operations; and...

[0162] The receptive field is expanded using dilated convolution, and the calculation expression for dilated convolution includes:

[0163]

[0164] In the formula, Indicates at time step The output of , where w is the convolution kernel. The convolution kernel is represented by the first... Each weight, Indicates the input sequence at time step The value, The expansion rate is denoted as .

[0165] When processor 1001 calls a computer program stored in memory 1005, it performs the following operations:

[0166] The attention weights corresponding to the temporal and spatial features are calculated using the dot product method.

[0167] The attention weights are multiplied one by one by each element of the temporal feature and each element of the spatial feature to obtain the temporal weighted output and the spatial weighted output, respectively.

[0168] The time-weighted output and the spatial-weighted output are concatenated to obtain the fused feature.

[0169] Furthermore, those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in a computer system to implement the process steps of the embodiments of the above methods.

[0170] Therefore, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the various steps of the gene regulatory network inference method for single-cell transcription genes as described in the above embodiments.

[0171] The computer-readable storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0172] It should be noted that, since the storage medium provided in the embodiments of this application is the storage medium used to implement the methods of the embodiments of this application, those skilled in the art can understand the specific structure and variations of the storage medium based on the methods described in the embodiments of this application, and therefore will not be repeated here. All storage media used in the methods of the embodiments of this application fall within the scope of protection of this application.

[0173] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0175] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0176] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0177] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. This application can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

[0178] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0179] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for inferring gene regulatory networks of single-cell transcription genes, characterized in that, The method includes the following steps: Determine the correlation vector for each gene pair in the single-cell transcriptional gene data to be tested, and represent each gene pair in the form of a histogram; The construction of the correlation vector includes: first calculating the ordered correlation vector between the gene pairs based on the statistical standard method, and then using a sliding window mechanism to save the pseudo-temporal information in the gene pairs; The histogram includes the main image corresponding to the gene pair, and 2r+2 neighboring images of the gene pair with strong positive covariance. Temporal features in the relevant vectors are captured using a temporal convolutional network, and spatial features in the histogram are captured using a convolutional network. The temporal and spatial features are weighted and fused to obtain the fused features; Based on the fusion features, the gene regulatory network inference results of the single-cell transcriptional gene data to be tested are predicted.

2. The method as described in claim 1, characterized in that, The expression for the relevant vector is: ； in, ； ； ； ； In the formula, Represents the set of related vectors. These represent the correlation vectors for each gene pair. , , The number of related vectors, The length of each related vector, Gene expression vectors, arranged in pseudo-time. The data fragment of transcription factor u extracted from the sliding window m. The sliding window interval corresponds to the transcription factor, and s is the sliding window size. This refers to the data fragment of the target gene v extracted from the sliding window n. This represents the sliding window interval corresponding to the target gene.

3. The method as described in claim 1, characterized in that, The step of capturing the temporal features in the relevant vectors using a temporal convolutional network specifically includes: Capturing dependencies in time series data through convolution operations; and... The receptive field is expanded using dilated convolution, and the calculation expression for dilated convolution includes: ； In the formula, Indicates at time step The output of , where w is the convolution kernel. The convolution kernel is represented by the first... Each weight, Indicates the input sequence at time step The value, The expansion rate is denoted as .

4. The method as described in claim 1, characterized in that, The step of weightedly fusing the temporal features and the spatial features to obtain the fused features specifically includes: The attention weights corresponding to the temporal and spatial features are calculated using the dot product method. The attention weights are multiplied one by one by each element of the temporal feature and each element of the spatial feature to obtain the temporal weighted output and the spatial weighted output, respectively. The time-weighted output and the spatial-weighted output are concatenated to obtain the fused feature.

5. A single-cell gene regulatory network inference system, characterized in that, The single-cell gene regulatory network inference system includes: The data preprocessing module determines the correlation vector of each gene pair in the single-cell transcriptional gene data to be tested, and represents each gene pair in the form of a histogram. The construction of the correlation vector includes: first calculating the ordered correlation vector between the gene pairs based on the statistical standard method, and then using a sliding window mechanism to save the pseudo-temporal information in the gene pairs; The histogram includes the main image corresponding to the gene pair, and 2r+2 neighboring images of the gene pair with strong positive covariance. The spatiotemporal feature learning module includes a temporal convolutional network and a convolutional network. The temporal convolutional network captures the temporal features in the relevant vectors, and the convolutional network captures the spatial features in the histogram. The feature fusion module weights and fuses the temporal features and the spatial features to obtain fused features; The prediction module predicts the gene regulatory network inference results of the single-cell transcriptional gene data to be tested based on the fusion features.

6. The system as described in claim 5, characterized in that, The convolutional network includes two convolutional layers with a filter size of 3*3, an output feature map with 32 channels, a max pooling layer with a pooling size of 2*2, and a fully connected layer with 128 nodes.

7. A computer system, characterized in that, The computer system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the gene regulatory network inference method for single-cell transcription genes as described in any one of claims 1 to 4.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the gene regulatory network inference method for single-cell transcribed genes as described in any one of claims 1 to 4.

Images