Single-cell multi-omics data clustering method based on cross-omics feature fusion

Abstract

The invention relates to a single-cell multi-omics data clustering method based on cross-omics feature fusion, and belongs to the field of single-cell omics analysis of computational biology. According to the technical scheme, the method comprises the following steps: step 1, preprocessing unicellular original data of different omics; step 2, using the preprocessed data to train a plurality of denoising auto-encoders to serve as single view feature extractors of each omics; step 3, training a de-noising auto-encoder as a common view feature extractor by combining the preprocessed different omics data; step 4, utilizing an attention mechanism to enhance public view features; step 5, enhancing cross-omics public representation by using comparative learning; and 6, performing clustering analysis and cell type identification by directly using the fused low-dimensional feature representation obtained by the model. The method is used for solving the problem of clustering analysis of single-cell multi-omics data in the field of bioinformatics, and is particularly applied to multi-omics data fusion and cell type identification.

Description

Technical Field

[0001] The present invention relates to the field of single-cell omics analysis in computational biology, and in particular to a single-cell multi-omics data clustering method based on cross-omics feature fusion. Background Art

[0002] Single-cell omics analysis is one of the core tasks in the field of computational biology, which aims to accurately measure and analyze the gene expression, protein levels and other molecular characteristics of single cells through high-throughput technology. This analysis method can reveal the heterogeneity between cells and provide important data support for understanding biological processes, disease mechanisms and the development of precision medicine.

[0003] At present, a large number of multi-omics data and matching technologies that can capture different information in a cell have emerged, and single-cell omics analysis is no longer limited to the analysis of a single omics level. Traditional clustering methods only cluster data for a single omics. With the maturity of omics technology, whether it is the increase in the number of cells that can be sequenced or the high-dimensional and high-sparse characteristics of omics itself, these methods are not only time-consuming, but also have unsatisfactory clustering effects.

[0004] Due to its powerful ability in feature extraction, deep learning methods have become the mainstream method for single-cell data analysis. In order to further improve the integration effect of multi-omics data, many dual-modal data integration methods based on deep learning have been proposed. However, the existing multi-omics clustering methods use the idea of ​​splicing for data analysis, whether before or after entering the encoder. Although this can obtain the consensus expression of different omics, it often ignores the individual information of a single omics. There is a complementary or mutually restrictive relationship between the individual information of a single omics. If the consensus expression can be enhanced and restricted by individual information in the process of multi-omics fusion, it will be of great significance for downstream clustering analysis and cell type identification. This patent first preprocesses the data set by normalizing, logarithmic standardization and other operations to make it comparable in the same numerical space, and then uses three autoencoders with zero-inflated binomial distribution as reconstruction loss to extract features from the preprocessed data, including omics consensus representation and individual representation. The consensus representation is enhanced through the attention mechanism and the omics individual representation information is added through contrastive learning, with the goal of obtaining a better low-dimensional embedding representation. A single-cell multi-omics data clustering method based on cross-omics feature fusion is proposed. Summary of the invention

[0005] The purpose of the present invention is to provide a single-cell multi-omics data clustering method based on cross-omics feature fusion.

[0006] The technical solution adopted by the present invention is: a single-cell multi-omics data clustering method based on cross-omics feature fusion, comprising the following steps:

[0007] A single-cell multi-omics data clustering method based on cross-omics feature fusion, comprising the following steps:

[0008] Step 1, data preprocessing: Given a single-cell paired multi-omics dataset, each of which has the same number of cells, represented by D = {X r ,X a ,X p},in represents RNA gene expression data, represents chromatin accessibility (ATAC) data, Represents protein expression (ADT) data; First, the data of different omics are normalized using data preprocessing methods, and the normalization factor s for each cell is calculated. i ; The count data for each cell were then log-transformed and scaled to zero mean and unit variance;

[0009] Step 2: Single view denoising autoencoder training: add noise to the preprocessed data of different omics to generate noisy data The denoising autoencoder is used for feature extraction. The encoder maps the noisy data to the latent space, and the decoder reconstructs the original data to obtain the low-dimensional feature representation of a single view. The reconstruction error of each autoencoder is calculated using the ZINB loss function, which takes into account the zero-inflation of single-cell data and is optimized via (1) and (2):

[0010]

[0011] ZINB(X;π,λ,θ)=π·δ0(X)+(1-π)·NB(X;μ X ,θ) (2)

[0012] In order to estimate these parameters in the ZINB loss function, three independent fully connected layers, M, Θ, and Π, are added to the last hidden layer of each decoder to calculate the mean, dispersion, and probability of zero counts of the data distribution;

[0013] Step 3: Joint view autoencoder training: The raw data of the two omics are spliced ​​to form a joint data view, and the encoder is trained to extract latent features, which is expressed as Also use (1) and (2) to calculate the reconstruction error of the joint data and optimize the encoder parameters;

[0014] Step 4: Enhance common view features through attention mechanism

[0015] Step 4.1, deconstruction of consensus representation: deconstruct the latent feature representation obtained by training the joint view data through the autoencoder, denoted as Z ra =[Z r ,Z a ],Z rp =[Z r ,Z p ];in, Represents the feature matrix of RNA gene expression data, n is the number of samples, d r For the dimensions of RNA data features, chromatin accessibility data and protein expression data are processed in the same way as RNA;

[0016] Step 4.2, cross-omics data fusion: Use equation (3) to map the consensus representation Z to different feature spaces;

[0017]

[0018] in, , is the joint representation of cross-omics data in parameter R space; similarly, using Generate a joint representation of cross-omics data under parameters Q1 and Q2;

[0019] Step 4.3, calculate the global relationship between samples: use formula (4) to calculate the global relationship matrix S between samples;

[0020]

[0021] Where d is the dimension mapped to the low-dimensional feature space;

[0022] Step 4.4, generate enhanced consensus representation: according to the global relationship matrix obtained in step 4.3, use formula (5) to enhance the feature representation of joint data in parameter R space;

[0023]

[0024] in, represents the representation of the jth sample in the R matrix, S ij represents the relationship weight between the i-th sample and the j-th sample; the final enhanced consensus representation matrix of all samples is:

[0025] Step 5: Enhance common representations across omics using contrastive learning

[0026] Step 5.1, calculate the cosine similarity between the consensus representation and the individual view-specific representation: By calculating the consensus representation and omics-specific expression The cosine similarity between (v∈D, where D = {RNA, ATAC, ADT}) evaluates the similarity between them; Formula (6) measures the consistency of the representation of the same sample in different omics, and the larger the value, the higher the similarity;

[0027]

[0028] Step 5.2, optimize the similarity between the consensus representation and the specific representation: design a contrastive learning function guided by the global relationship matrix, as shown in formula (7);

[0029]

[0030] In order to avoid taking irrelevant samples as negative sample pairs, which would lead to conflicts in clustering objectives, the global relationship matrix S is used to guide the selection of negative sample pairs;

[0031] Step 6: directly use the fused low-dimensional feature representation obtained by the model to perform clustering analysis and cell type identification:

[0032] Step 6.1, use KL divergence to approximate cell pairing similarity: calculate the pairing similarity of cells i and i' in the latent space of the autoencoder. The similarity is represented by the t-distribution kernel function in formula (8), which is inversely proportional to the square of the Euclidean distance between the two points;

[0033]

[0034] Among them, ||Z i -Z l || 2 represents the square of the Euclidean distance between cell i and cell l in the latent space, and the denominator is normalized; the target distribution P is defined as q ii' The square of is used to emphasize the pairings with high similarity and weaken the influence of the pairings with low similarity. Formula (9) normalizes the target distribution to ensure that P forms a valid probability distribution for the pairing similarities of all cells. Formula (10) adds the loss function of KL divergence to ensure that the potential representation is that similar cells are close and different cells are far away.

[0035]

[0036] Step 6.2, Deep K-Means Clustering: First, transform the latent representation Divided into K initial clusters, the centers of the clusters are {V1,V2,...,V K}, the loss function of deep K-Means clustering is defined as formula (11);

[0037]

[0038] in: is the potential representation of the cell, V j is the center of cluster j, The cell i and the cluster center V are calculated j The Euclidean distance between them, τ is a hyperparameter that controls the Euclidean distance term, and w ij is the weight between cell i and cluster j, which is determined by the Gaussian kernel function of formula (12);

[0039]

[0040] Weight w ij Reflects the relationship between cell i and cluster center V j The correlation degree of V with the highest correlation degree is selected. j The category label of is used as the category label of cell i.

[0041] A further solution of the above technical solution is that the method of adding noise in step 2 includes generating random noise, or Gaussian noise, or Poisson noise, or salt and pepper noise, or multiplicative noise.

[0042] A further solution of the above technical solution is that the step of enhancing the cross-omics common representation through contrastive learning in step 5 strengthens the consistency between samples by designing a contrastive loss function including negative sampling, wherein the selection of negative sample pairs is guided by a global relationship matrix.

[0043] A further solution of the above technical solution is that the contrastive learning loss function defined in step 5.2 also includes a temperature scaling parameter for balancing the contribution of positive and negative sample pairs, and the selection of the temperature parameter is dynamically adjusted according to the training stage. A further solution of the above technical solution is that the noise adding method includes random noise generation, or Gaussian noise, or Poisson noise, or salt and pepper noise, or multiplicative noise, and the intensity and range of noise addition are adjustable.

[0044] The advantages of the present invention are as follows:

[0045] First, the present invention can simultaneously process and fuse data from different omics layers (such as RNA, protein, epigenetics, etc.), can fully utilize the complementarity between different omics data, and improve the accuracy of cell recognition.

[0046] Second, the present invention introduces global relationship-guided contrastive learning and cross-view feature aggregation, which can not only optimize the clustering of cells at the local level, but also learn the structural relationships between cells from a global perspective.

[0047] Third, the present invention automatically extracts multi-level and multi-omics features from raw data, reducing the need for manual intervention while improving the efficiency of data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the overall process of the single-cell multi-omics data clustering method based on the fusion of cross-omics features;

[0049] Figure 2 It is an overall framework for single-cell multi-omics data clustering methods based on cross-omics feature fusion;

[0050] Figure 3 It is an algorithmic framework for the attention module that enhances consensus representation;

[0051] Figure 4 It is a schematic diagram of positive and negative sample selection for contrastive learning guided by global relations. DETAILED DESCRIPTION

[0052] like Figures 1 to 4 As shown, the single-cell multi-omics data clustering method based on cross-omics feature fusion described in this embodiment includes the following steps:

[0053] Step 1, data preprocessing: Given a single-cell paired multi-omics dataset, each of which has the same number of cells, represented by D = {X r ,X a ,X p},in represents RNA gene expression data, represents chromatin accessibility (ATAC) data, Represents protein expression (ADT) data; first, the normalization factor s for each cell is calculated using a data preprocessing method i , normalize each cell to eliminate the technical bias caused by sequencing technology and the biological differences of the cells themselves; then logarithmically transform the count data of each cell and scale it to zero mean and unit variance;

[0054] Step 2: Single view denoising autoencoder training: add noise to the preprocessed data of different omics to generate noisy data The denoising autoencoder is used for feature extraction. The encoder maps the noisy data to the latent space, and the decoder reconstructs the original data to obtain the low-dimensional feature representation of a single view. Because of the high dimensionality and high sparsity of single-cell data, the ZINB loss function is used to calculate the reconstruction error of each autoencoder. This loss function takes into account the zero expansion of single-cell data and is optimized through (1) and (2):

[0055]

[0056] ZINB(X;π,λ,θ)=π·δ0(X)+(1-π)·NB(X;μ X,θ) (14)

[0057] In order to estimate these parameters in the ZINB loss function, three independent fully connected layers, M, Θ, and Π, are added to the last hidden layer of each decoder to calculate the mean, dispersion, and probability of zero counts of the data distribution;

[0058] Step 3: Joint view autoencoder training: The raw data of the two omics are spliced ​​to form a joint data view, and the encoder is trained to extract latent features, which is expressed as (1) and (2) are also used to calculate the reconstruction error of the joint data and optimize the encoder parameters; during reconstruction, the consensus matrix is ​​reconstructed for each omics data separately to make it consistent with the feature representation of a single view;

[0059] Step 4: Set the attention mechanism to enhance the joint view features

[0060] Step 4.1, deconstruction of joint view representation: deconstruct the latent feature representation obtained by training the joint view data through the autoencoder, denoted as Z ra =[Z r ,Z a ],Z rp =[Z r ,Z p ];in, Represents the feature matrix of RNA gene expression data, n is the number of samples, d r For the dimensions of RNA data features, chromatin accessibility data and protein expression data are processed in the same way as RNA;

[0061] Step 4.2, cross-omics data fusion: Using formula (3) and the attached Figure 3 As shown, the consensus representation Z is mapped to different feature spaces;

[0062]

[0063] in, , is the joint representation of cross-omics data in parameter R space; similarly, using

[0064] Generate a joint representation of cross-omics data under parameters Q1 and Q2; this allows data from different omics to be mapped into the same feature space, generating a more representative feature representation;

[0065] Step 4.3, calculate the global relationship between samples: In order to quantify the similarity between samples, use formula (4) to calculate the relative distance or similarity between samples in the latent space to generate a matrix S;

[0066]

[0067] Where d is the dimension mapped to the low-dimensional feature space, and each element represents the similarity and relationship strength between samples; the global relationship matrix can capture the correlation between different omics levels;

[0068] Step 4.4, generate enhanced consensus representation: according to the global relationship matrix obtained in step 4.3, use formula (5) to enhance the feature representation of joint data in parameter R space;

[0069]

[0070] in, represents the representation of the jth sample in the R matrix, S ij represents the relationship weight between the i-th sample and the j-th sample; the final enhanced consensus representation matrix of all samples is: The correlation information of different omics is transferred to the cross-omics feature fusion model through the global relationship matrix, which enhances the representation of consensus features and ensures that when fusing multi-omics data, similar cell samples are closer in the representation space, while different samples are effectively distinguished;

[0071] Step 5: Enhance common representations across omics using contrastive learning

[0072] Step 5.1, calculate the cosine similarity between the consensus representation and the individual view-specific representation: By calculating the consensus representation and omics-specific expression The cosine similarity between (v∈D, where D = {RNA, ATAC, ADT}) evaluates the similarity between them; Formula (6) measures the consistency of the representation of the same sample in different omics, and the larger the value, the higher the similarity;

[0073]

[0074] Step 5.2, optimize the similarity between the consensus representation and the specific representation: design a contrastive learning function guided by the global relationship matrix, as shown in formula (7);

[0075]

[0076] In order to avoid taking irrelevant samples as negative sample pairs, which would lead to conflicts in clustering objectives, we use the global relationship matrix S to guide the selection of negative sample pairs. Specifically, when S ij When it is small, it means that sample i and sample j belong to different clusters, and their influence on the loss function will increase. Even if the cosine similarity between them is small, it will have a greater impact on the loss value. The selection of negative samples is particularly important here. Figure 4As shown in the figure, the positive sample selection of contrastive learning is the same sample under different views (single view and joint view). If other samples are directly set as negative samples, similar samples will be defined as negative samples. Therefore, the global relationship matrix is ​​used here to guide the selection of negative samples, which can avoid the model from mistakenly pairing unrelated cells together.

[0077] Step 6: directly use the fused low-dimensional feature representation obtained by the model to perform clustering analysis and cell type identification:

[0078] Step 6.1, use KL divergence to approximate cell pairing similarity: calculate the pairing similarity of cells i and i' in the latent space of the autoencoder. The similarity is represented by the t-distribution kernel function in formula (8), which is inversely proportional to the square of the Euclidean distance between the two points;

[0079]

[0080] Among them, ||Z i -Z l || 2 represents the square of the Euclidean distance between cell i and cell l in the latent space, q ii' reflects the similarity between the two; the smaller the distance, q ii' The larger the value, the higher the similarity of similar cells. The denominator normalizes it to ensure that the sum of the similarities of each cell is 1. The target distribution P is defined as q ii' The square of is used to emphasize the pairings with high similarity and weaken the influence of the pairings with low similarity. Formula (9) normalizes the target distribution to ensure that P forms a valid probability distribution for the pairing similarities of all cells. Formula (10) adds the loss function of KL divergence to ensure that the potential representation is that similar cells are close and different cells are far away.

[0081]

[0082]

[0083] Step 6.2, Deep K-Means Clustering: First, transform the latent representation Divided into K initial clusters, the centers of the clusters are {V1,V2,...,V K}, the loss function of deep K-Means clustering is defined as formula (11);

[0084]

[0085] in: is the potential representation of the cell, V j is the center of cluster j, The cell i and the cluster center V are calculatedj The Euclidean distance between them, τ is a hyperparameter that controls the Euclidean distance term, and w ij is the weight between cell i and cluster j, which is determined by the Gaussian kernel function of formula (12);

[0086]

[0087] This weight reflects the relationship between cell i and cluster center V j The cells closer to the cluster center will have a higher weight; the weight is expanded by raising it to the power of α to ensure that the most relevant clusters dominate the optimization process; the model sets an early stopping mechanism here, first clustering to get the initial cluster center point, and then judging the difference between the predicted label and the previous round of predicted label and its proportion of all cells. If the specified threshold is not reached, the clustering is continued in a cycle to calculate the proportion described above; if the specified threshold is reached, it means that the model has converged to a stable state. After the clustering process is completed, the system will output the evaluation indicators of the clustering results and generate a UMAP visualization graph. This information will be used for subsequent cell type identification and downstream analysis.

[0088] Experimental results show that this method can achieve high-quality single-cell data clustering and downstream analysis tasks such as cell type identification. By combining advanced autoencoders, attention mechanisms, KL divergence and deep K-Means, the method of the present invention can effectively solve the challenges in single-cell multi-omics data clustering and provide a powerful tool for single-cell omics research.

Claims

1. A single-cell multi-omics data clustering method based on cross-omics feature fusion, characterized in that: The following steps are involved: Step 1, data preprocessing: Given a single-cell paired multi-omics dataset, each of which has the same number of cells, represented by D = {X r ,X a ,X p },in represents RNA gene expression data, represents chromatin accessibility (ATAC) data, Represents protein expression (ADT) data; First, the data of different omics are normalized using data preprocessing methods, and the normalization factor s for each cell is calculated. i ; The count data for each cell were then log-transformed and scaled to zero mean and unit variance; Step 2: Single view denoising autoencoder training: add noise to the preprocessed data of different omics to generate noisy data The denoising autoencoder is used for feature extraction. The encoder maps the noisy data to the latent space, and the decoder reconstructs the original data to obtain the low-dimensional feature representation of a single view. The reconstruction error of each autoencoder is calculated using the ZINB loss function, which takes into account the zero-inflation of single-cell data and is optimized via (1) and (2): ZINB(X;π,λ,θ)=π·δ0(X)+(1-π)·NB(X;μ X ,i) (2) In order to estimate these parameters in the ZINB loss function, three independent fully connected layers, M, Θ, and Π, are added to the last hidden layer of each decoder to calculate the mean, dispersion, and probability of zero counts of the data distribution; Step 3: Joint view autoencoder training: The raw data of the two omics are spliced ​​to form a joint data view, and the encoder is trained to extract latent features, which is expressed as Also use (1) and (2) to calculate the reconstruction error of the joint data and optimize the encoder parameters; Step 4: Enhance common view features through attention mechanism Step 4.1, deconstruction of consensus representation: deconstruct the latent feature representation obtained by training the joint view data through the autoencoder, denoted as Z ra =[Z r ,Z a ],Z rp =[Z r ,Z p ];in, Represents the feature matrix of RNA gene expression data, n is the number of samples, d r For the dimensions of RNA data features, chromatin accessibility data and protein expression data are processed in the same way as RNA; Step 4.2, cross-omics data fusion: Use equation (3) to map the consensus representation Z to different feature spaces; in, is the joint representation of cross-omics data in parameter R space; similarly, Generate a joint representation of cross-omics data under parameters Q1 and Q2; Step 4.3, calculate the global relationship between samples: use formula (4) to calculate the global relationship matrix S between samples; Where d is the dimension mapped to the low-dimensional feature space; Step 4.4, generate enhanced consensus representation: according to the global relationship matrix obtained in step 4.3, use formula (5) to enhance the feature representation of joint data in parameter R space; in, represents the representation of the jth sample in the R matrix, S ij represents the relationship weight between the i-th sample and the j-th sample; the final enhanced consensus representation matrix of all samples is: Step 5: Enhance common representations across omics using contrastive learning Step 5.1, calculate the cosine similarity between the consensus representation and the individual view-specific representation: By calculating the consensus representation and omics-specific expression The cosine similarity between (v∈D, where D = {RNA, ATAC, ADT}) evaluates the similarity between them; Formula (6) measures the consistency of the representation of the same sample in different omics, and the larger the value, the higher the similarity; Step 5.2, optimize the similarity between the consensus representation and the specific representation: design a contrastive learning function guided by the global relationship matrix, as shown in formula (7); In order to avoid taking irrelevant samples as negative sample pairs, which would lead to conflicts in clustering objectives, the global relationship matrix S is used to guide the selection of negative sample pairs; Step 6: directly use the fused low-dimensional feature representation obtained by the model to perform clustering analysis and cell type identification: Step 6.1, use KL divergence to approximate cell pairing similarity: calculate the pairing similarity of cells i and i' in the latent space of the autoencoder. The similarity is represented by the t-distribution kernel function in formula (8), which is inversely proportional to the square of the Euclidean distance between the two points; Among them, ||Z i -Z l || 2 represents the square of the Euclidean distance between cell i and cell l in the latent space, and the denominator is normalized; the target distribution P is defined as q ii' The square of is used to emphasize the pairings with high similarity and weaken the influence of the pairings with low similarity. Formula (9) normalizes the target distribution to ensure that P forms a valid probability distribution for the pairing similarities of all cells. Formula (10) adds the loss function of KL divergence to ensure that the potential representation is that similar cells are close and different cells are far away. Step 6.2, Deep K-Means Clustering: First, transform the latent representation Divided into K initial clusters, the centers of the clusters are {V1,V2,...,V K }, the loss function of deep K-Means clustering is defined as formula (11); in: is the potential representation of the cell, V j is the center of cluster j, The cell i and the cluster center V are calculated j The Euclidean distance between them, τ is a hyperparameter that controls the Euclidean distance term, and w ij is the weight between cell i and cluster j, which is determined by the Gaussian kernel function of formula (12); Weight w ij Reflects the relationship between cell i and cluster center V j The correlation degree of V with the highest correlation degree is selected. j The category label of is used as the category label of cell i.

2. The single-cell multi-omics data clustering method based on cross-omics feature fusion according to claim 1, characterized in that: The method of adding noise in step 2 includes random noise generation, or Gaussian noise, or Poisson noise, or salt and pepper noise, or multiplicative noise.

3. The single-cell multi-omics data clustering method based on cross-omics feature fusion according to claim 1, characterized in that: The step of contrastive learning in step 5 to enhance the cross-omics common representation enforces the consistency between samples by designing a contrastive loss function including negative sampling, where the selection of negative sample pairs is guided by the global relationship matrix.

4. The single-cell multi-omics data clustering method based on cross-omics feature fusion according to claim 1, characterized in that: The contrastive learning loss function defined in step 5.2 also includes a temperature scaling parameter for balancing the contribution of positive and negative sample pairs. The selection of the temperature parameter is dynamically adjusted according to the training stage.

Images