A single-cell multi-omics data alignment method and system based on prior knowledge injection

Abstract

This invention discloses a method and system for aligning single-cell multi-omics data based on prior knowledge injection, relating to the fields of bioinformatics and multimodal data processing. The method involves acquiring raw single-cell multi-omics data and inputting it into an adversarial autoencoder network, mapping it to a shared latent space to obtain an initial latent representation. A cell type prior knowledge matrix is ​​then obtained, and based on this matrix, the initial latent representation is orthogonally decoupled, dividing it into a priori-guided biological feature subspace and an orthogonal technical noise subspace. A discriminator is used to apply local adversarial alignment constraints, and gradient blocking is applied to the biological feature subspace to eliminate batch effects. Data is reconstructed using a decoder based on the decoupled subspace. This invention achieves decoupling between sequencing technical bias and real biological variation at the underlying mathematical topology, overcoming the biological topological collapse defect caused by global alignment, and preserving the weak heterogeneity signal of rare cell subpopulations while eliminating background noise.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and multimodal data processing, and in particular to a method and system for aligning single-cell multi-omics data based on prior knowledge injection. Background Technology

[0002] With the continuous development of precision medicine, single-cell multi-omics technology has become an important means to analyze tumor heterogeneity and discover rare cell subpopulations. By jointly analyzing single-cell transcriptome and epigenetic data of the same biological sample, researchers can more comprehensively reveal the dynamic evolution and regulatory network of cells, which has extremely high clinical translational value for the accurate identification of weak signals such as stem cell-like cells and drug-resistant precursor cells.

[0003] However, different omics data exhibit significant differences in technology platforms, scale distributions, and background noise levels, making it difficult to effectively fuse multimodal data. To eliminate this data heterogeneity and batch effect, the conventional approach in this field is to use deep learning models such as variational autoencoders or generative adversarial networks to map the original multimodal data into a purely data-driven shared latent space. Subsequently, the model uses contrastive learning or adversarial loss functions to force the statistical distribution distance between different omics data in the latent space to be narrowed, thereby achieving cross-modal data alignment and fusion.

[0004] In practical applications, the accuracy of existing alignment models is often insufficient to meet the requirements for identifying rare cells in a small number of samples. Those skilled in the art generally believe that the fundamental reason for the limited alignment accuracy is the excessive background noise caused by the batch effect, or the insufficient nonlinear fitting ability of the neural network. Based on this understanding, existing improvement directions are highly focused on increasing the number of layers in the neural network, introducing complex attention mechanisms, or adopting more stringent global filtering algorithms in the data preprocessing stage.

[0005] However, existing unsupervised adversarial networks blindly pursue the consistency of global statistical distribution in a shared latent space. When the model forcibly flattens the data distribution of different modalities to eliminate batch effects, it is prone to irreversible biological topological collapse. Specifically, the model may incorrectly overlap rare cells with ordinary cells that have similar noise patterns but drastically different biological properties, resulting in over-alignment. This leads to a contradiction in existing technologies: eliminating batch effects inevitably erases rare biological signals, making it impossible to accurately restore the heterogeneous characteristics of cells while reducing noise.

[0006] Therefore, a method and system for aligning single-cell multi-omics data based on prior knowledge injection is provided. Summary of the Invention

[0007] This invention overcomes the shortcomings of the prior art and provides a method and system for single-cell multi-omics data alignment based on prior knowledge injection.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a single-cell multi-omics data alignment method based on prior knowledge injection, comprising the following steps:

[0009] S1. Obtain raw single-cell multi-omics data from the same biological sample;

[0010] S2. Input the raw data of single-cell multi-omics into a preset adversarial autoencoder network, and map the raw data of single-cell multi-omics to a shared latent space through the encoder in the adversarial autoencoder network to obtain an initial latent representation;

[0011] S3. Obtain a pre-constructed cell type prior knowledge matrix. Through a preset prior integration layer, orthogonally decouple the initial potential representation based on the cell type prior knowledge matrix, and divide the initial potential representation into a prior-guided biological feature subspace and a technical noise subspace orthogonal to the biological feature subspace.

[0012] S4. Within the technical noise subspace, the discriminator in the adversarial autoencoder network is used to perform local adversarial alignment constraints, and gradient blocking is applied to the biological feature subspace to eliminate batch effects.

[0013] S5. Based on the decoupled biological feature subspace and the technical noise subspace after being constrained by local adversarial alignment, the data is reconstructed through the decoder in the adversarial autoencoder network to output fused single-cell data that eliminates technical noise and preserves the topology of rare cells.

[0014] In a preferred embodiment of the present invention, obtaining the pre-constructed cell type prior knowledge matrix includes:

[0015] Cell type marker genes were extracted through large-scale text mining;

[0016] Calculate the consensus strength of the cell type marker genes as markers of specific cell types in the literature;

[0017] An initial matrix is ​​constructed based on the consensus strength, and the initial matrix is ​​row normalized to generate the cell type prior knowledge matrix.

[0018] In a preferred embodiment of the present invention, the orthogonal decoupling of the initial latent representation based on the cell type prior knowledge matrix, dividing the initial latent representation into a prior-guided biological feature subspace and a technical noise subspace orthogonal to the biological feature subspace, includes:

[0019] Using the prior knowledge matrix of the cell type as an orthogonal basis, an orthogonal projection operator is constructed.

[0020] The initial latent representation is linearly transformed by the orthogonal projection operator and projected onto the first direction to generate the biological feature subspace.

[0021] Calculate the orthogonal complement space of the initial latent representation in the first direction as the technical noise subspace.

[0022] In a preferred embodiment of the present invention, the step of using the discriminator in the adversarial autoencoder network to perform local adversarial alignment constraints within the technical noise subspace and performing gradient blocking on the biological feature subspace includes:

[0023] The features in the technical noise subspace are input into the discriminator, and the adversarial loss between the feature distribution in the technical noise subspace and the standard normal distribution is calculated.

[0024] When updating the network through backpropagation based on the adversarial loss, the gradient is only backpropagated to the network weights corresponding to the technical noise subspace, and the backpropagation of the gradient to the network weights corresponding to the biological feature subspace is blocked by a masking operation.

[0025] In a preferred embodiment of the present invention, the single-cell multi-omics raw data includes at least single-cell transcriptome data and chromatin accessibility epigenome data;

[0026] Before inputting the raw single-cell multi-omics data into a preset adversarial autoencoder network, the method further includes:

[0027] The raw data from the single-cell multi-omics study were subjected to quality control and count normalization.

[0028] A biological noise correction reference is established based on the preset grouping effect and batch parameters, and the normalized single-cell multi-omics raw data is subjected to noise reduction preprocessing based on the biological noise correction reference.

[0029] In a preferred embodiment of the present invention, the adversarial autoencoder network further includes a classifier; the method further includes:

[0030] The features in the biological feature subspace are input into the classifier;

[0031] The classifier predicts cell types based solely on features within the biological feature subspace and outputs the cell type prediction results.

[0032] In a preferred embodiment of the present invention, the data reconstruction based on the decoupled biological feature subspace and the technical noise subspace after local adversarial alignment constraints, through the decoder in the adversarial autoencoder network, includes:

[0033] The features in the biological feature subspace and the technical noise subspace are input into the decoder to reconstruct the expression matrix, and the reconstruction loss is calculated.

[0034] The total loss function of the adversarial autoencoder network includes at least the reconstruction loss and the adversarial loss generated by the discriminator, and the adversarial autoencoder network is jointly optimized and trained based on the total loss function.

[0035] In a preferred embodiment of the present invention, it further includes:

[0036] Cluster analysis was performed on the fused single-cell data to extract data features of the target cell subpopulation;

[0037] The target cell subpopulation includes stem cell-like cells or drug-resistant precursor cells.

[0038] A single-cell multi-omics data alignment system based on prior knowledge injection includes:

[0039] The data acquisition module is used to acquire raw single-cell multi-omics data from the same biological sample.

[0040] The latent space mapping module is used to input the raw data of single-cell multi-omics into a preset adversarial autoencoder network, and to map the raw data of single-cell multi-omics to a shared latent space through the encoder in the adversarial autoencoder network to obtain an initial latent representation.

[0041] An orthogonal decoupling module is used to obtain a pre-constructed cell type prior knowledge matrix. Through a preset prior integration layer, the initial potential representation is orthogonally decoupled based on the cell type prior knowledge matrix, and the initial potential representation is divided into a prior-guided biological feature subspace and a technical noise subspace orthogonal to the biological feature subspace.

[0042] The local adversarial alignment module is used to perform local adversarial alignment constraints within the technical noise subspace using the discriminator in the adversarial autoencoder network, and to perform gradient blocking on the biological feature subspace to eliminate batch effects.

[0043] The data reconstruction module is used to reconstruct data based on the decoupled biological feature subspace and the technical noise subspace after local adversarial alignment constraints, through the decoder in the adversarial autoencoder network, and output fused single-cell data that eliminates technical noise and preserves the rare cell topology.

[0044] An electronic device, comprising:

[0045] At least one processor; and,

[0046] A memory that is communicatively connected to the at least one processor;

[0047] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the single-cell multi-omics data alignment method based on prior knowledge injection as described above.

[0048] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0049] This invention introduces a prior integration layer into the latent space and orthogonally decouples the initial latent representation based on the cell type prior knowledge matrix, dividing it into a prior-guided biological feature subspace and an orthogonal technical noise subspace. At the underlying mathematical topology, it physically isolates the actual biological variations in the data from the technical biases introduced by the sequencing platform, enabling the network to independently process noise without interfering with biological features. Compared to existing unsupervised adversarial networks that blindly pursue global statistical distribution consistency and forcibly flatten multimodal data, this invention completely avoids the biological topological collapse phenomenon caused by over-alignment. Thus, while eliminating batch effects, it perfectly preserves the weak heterogeneity signals of rare subpopulations such as stem cell-like cells and drug-resistant precursor cells, fundamentally breaking the zero-sum game between data denoising and biological feature fidelity.

[0050] This invention utilizes a discriminator for local adversarial alignment constraints within the technical noise subspace and simultaneously applies gradient blocking to the biological feature subspace. This ensures that when the adversarial network approximates a standard normal distribution to eliminate batch effects, the gradient backpropagation of its loss function is strictly limited to the noise dimension, and the network weights in the biological dimension are completely frozen and protected. This achieves precise removal of batch effects and absolute stability of biological anchors. Traditional global adversarial fusion algorithms inevitably discard weak real biological signals as noise during repeated iterations, leading to feature collapse. This invention, through this counterintuitive local update strategy, significantly improves the resolution of the integrated cell clusters, enhances the discovery efficiency of highly specific biomarkers, and strengthens the model's analytical limits with small sample sizes.

[0051] This invention extracts cell type marker genes through large-scale text mining and calculates consensus strength, then normalizes and generates a cell type prior knowledge matrix. This transforms the discrete biological consensus from massive amounts of literature into continuous mathematical weights that can be directly read by deep learning networks, providing an objective and dynamic absolute coordinate system for the latent space. This ensures that the prior guidance injected into the latent space has a solid data-driven foundation. Existing auxiliary classification methods that rely on manually set labels or static rules often struggle to adapt to the highly heterogeneous and dynamically evolving tumor microenvironment and are prone to introducing human bias. The automated matrix construction of this invention not only significantly reduces the operational threshold for cross-omics analysis but also significantly improves the robustness and reproducibility of the model when facing different sequencing platforms and unknown rare cell types.

[0052] This invention performs noise reduction preprocessing by establishing a biological noise correction reference based on preset grouping effects and batch parameters before inputting multi-omics data into the adversarial network. Before the data enters the nonlinear mapping, some known systematic technical biases are pre-exposed, providing a cleaner input benchmark for subsequent latent space mapping and effectively improving the initial signal-to-noise ratio of multimodal data. Unlike the stringent global filtering in traditional methods that do not distinguish between signal and noise, this invention's targeted correction based on grouping effects is gentler and more precise. It not only avoids prematurely deleting weak expressions of rare cells but also forms a microscopic synergistic interference with the backend orthogonal decoupling mechanism, further accelerating the convergence speed of the adversarial network and ensuring the effectiveness of accurate matching between transcriptome and epigenetics data in a unified latent space. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of the single-cell multi-omics data alignment method based on prior knowledge injection of the present invention;

[0055] Figure 2 This is a diagram of the orthogonal decoupling network architecture of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0058] Application Overview:

[0059] The field of single-cell multi-omics data fusion has long faced a zero-sum game contradiction between eliminating technical noise and preserving rare biological signals. Conventional pure data-driven global adversarial alignment techniques can eliminate batch effects, but unsupervised global statistical distribution consistency constraints can easily lead to biological topological collapse, causing rare cell subpopulations with similar noise patterns to erroneously overlap with ordinary cells. Techniques employing rigorous global filtering preprocessing can preserve some weak signals, but struggle to cope with the complex systematic biases introduced by multiple sequencing platforms. Current technologies have not yet solved the typical problem of how to decouple sequencing technical noise from real biological variations at the latent space. This invention addresses the problem of rare cell signals being erased in micro-sample multimodal data fusion scenarios by introducing a cell type prior knowledge matrix and combining it with latent space orthogonal decoupling techniques. Compared to existing technologies, the solution provided by this invention breaks the blind constraints of global adversarial alignment, achieving independent processing of batch effect stripping and biological feature anchoring, preventing noise reduction operations from interfering with weak heterogeneous signals.

[0060] The method provided by this invention relies on an electronic device with tensor computing capabilities. The electronic device includes a high-performance computing server or a cloud-based distributed computing node. The electronic device is equipped with a central processing unit (CPU) and a memory connected to the CPU via a system bus. The memory is used to persistently store the original single-cell multi-omics data matrix and the computer program instructions for implementing the method of this invention. The CPU instantiates an adversarial autoencoder network and an orthogonal projection operator in memory by calling the computer program instructions in the memory. Then, it performs tensor operations and gradient updates on the input gene expression matrix and chromatin accessibility peak matrix, and outputs the fused single-cell feature tensor to the downstream clustering analysis module or persistently stores it in an external database.

[0061] Example 1:

[0062] like Figure 1 and Figure 2 As shown, a single-cell multi-omics data alignment method based on prior knowledge injection includes the following steps:

[0063] S1. Obtain raw single-cell multi-omics data from the same biological sample;

[0064] S2. Input the raw data of single-cell multi-omics into a preset adversarial autoencoder network, and map the raw data of single-cell multi-omics to a shared latent space through the encoder in the adversarial autoencoder network to obtain an initial latent representation;

[0065] S3. Obtain a pre-constructed cell type prior knowledge matrix. Through a preset prior integration layer, orthogonally decouple the initial potential representation based on the cell type prior knowledge matrix, and divide the initial potential representation into a prior-guided biological feature subspace and a technical noise subspace orthogonal to the biological feature subspace.

[0066] S4. Within the technical noise subspace, the discriminator in the adversarial autoencoder network is used to perform local adversarial alignment constraints, and gradient blocking is applied to the biological feature subspace to eliminate batch effects.

[0067] S5. Based on the decoupled biological feature subspace and the technical noise subspace after being constrained by local adversarial alignment, the data is reconstructed through the decoder in the adversarial autoencoder network to output fused single-cell data that eliminates technical noise and preserves the topology of rare cells.

[0068] Introducing prior knowledge into deep learning models to guide the fusion of multi-omics data faces challenges in underlying architecture design. If prior labels are directly concatenated at the model input or output, the model cannot intervene in the manifold distribution of the latent space during the intermediate stage of feature extraction. If conventional nonlinear concatenation operations are used in the latent space, the nonlinear transformation will destroy the original manifold topology of the latent space and cause gradient vanishing. It is necessary to design a lightweight injection mechanism that can both apply biological guidance and maintain the underlying mathematical orthogonality of the latent space to bridge the technical gap between pure data-driven and knowledge-driven approaches.

[0069] Preferably, in step S1, the raw data of single-cell multi-omics includes the gene expression matrix generated by single-cell transcriptome sequencing and the peak matrix generated by single-cell chromatin accessibility sequencing.

[0070] Specifically, the data structure of the gene expression matrix is ​​a two-dimensional tensor, where the row dimension represents the number of cell samples and the column dimension represents the number of genes.

[0071] Specifically, the data structure of the peak matrix is ​​also a two-dimensional tensor, and its column dimension represents the number of peaks in the open regions of chromatin.

[0072] Preferably, in step S2, the adversarial autoencoder network includes modality-specific coding branches corresponding to different omics data, and each modality-specific coding branch adopts a multilayer perceptron structure;

[0073] Specifically, the gene expression matrix and the peak matrix are respectively input into their corresponding modality-specific coding branches to perform nonlinear dimensionality reduction operations, compressing and mapping multi-omics data of different dimensions to a shared latent space with a uniform dimension size;

[0074] Furthermore, the encoder outputs an initial latent representation characterizing the low-dimensional features of the cell. The data structure of the initial latent representation is a low-dimensional feature tensor, whose column dimension is much smaller than that of the original single-cell multi-omics data.

[0075] Preferably, in step S3, the cell type prior knowledge matrix is ​​a two-dimensional numerical matrix containing the weights of cell types and marker genes.

[0076] Specifically, the initial latent representation is divided into spatial dimensions through a preset prior integration layer, and the projection direction is constructed based on the cell type prior knowledge matrix;

[0077] Furthermore, the initial latent representation is linearly decomposed along the projection direction to generate a biological feature subspace guided by the cell type prior knowledge matrix;

[0078] In detail, the complement space orthogonal to the biological feature subspace is calculated as the technical noise subspace. The biological feature subspace is used to carry the true heterogeneity characteristics of cells, and the technical noise subspace is used to accommodate the batch effects and background noise generated by the sequencing platform.

[0079] Preferably, in step S4, the discriminator receives the feature tensor in the technical noise subspace and calculates its divergence difference with the standard normal distribution;

[0080] Specifically, when updating the weights during network backpropagation, gradient blocking is implemented on the biological feature subspace to cut off the backpropagation path of the error gradient generated by the discriminator to the network weights corresponding to the biological feature subspace.

[0081] Furthermore, gradient blocking is implemented so that the adversarial alignment operation updates parameters only within the technical noise subspace, preventing the denoising process from altering the distribution of biological characteristics of rare cells.

[0082] Preferably, in step S5, the decoder in the adversarial autoencoder network performs a dimension-upgrading mapping to restore the low-dimensional feature subspace tensor to the original dimension expression matrix;

[0083] Specifically, the decoder outputs fused single-cell data that eliminates technical noise and preserves the topology of rare cells, and the fused single-cell data is represented as a high-dimensional feature matrix after noise reduction.

[0084] Furthermore, the high-dimensional feature matrix is ​​used by downstream computing nodes to perform cell subpopulation clustering and biomarker extraction.

[0085] Example 2:

[0086] This embodiment, based on embodiment 1, elaborates in detail on the specific implementation methods of obtaining the pre-constructed cell type prior knowledge matrix and orthogonal decoupling and gradient blocking.

[0087] Preferably, obtaining the pre-constructed cell type prior knowledge matrix includes extracting cell type marker genes through large-scale text mining and calculating the consensus strength of the cell type marker genes as markers of specific cell types in the literature;

[0088] Specifically, using public biomedical literature databases as data sources, and employing named entity recognition algorithms in natural language processing, gene name entities and cell type entities are extracted from literature abstracts or texts.

[0089] Specifically, an initial matrix is ​​constructed based on the consensus strength, and the initial matrix is ​​subjected to row normalization to generate the cell type prior knowledge matrix;

[0090] Furthermore, the data structure of the initial matrix is ​​a two-dimensional numerical matrix. ,in This represents the total number of preset cell types. Represents the total number of marker genes extracted;

[0091] In detail, the statistics of the first The marker gene and the first The number of co-occurring documents in which two cell types appear together in the same document, or the co-occurrence weight score calculated using the term frequency-inverse document frequency algorithm, is used as an element in the initial matrix. ;

[0092] Specifically, the calculation formula for performing row normalization on the initial matrix is ​​as follows: ,in These are the normalized weight values;

[0093] Furthermore, based on the weight values The cell type prior knowledge matrix constitutes the The cell type prior knowledge matrix This transforms discrete biological literature consensus into continuous mathematical weights, providing an objective coordinate benchmark for subsequent latent space projection.

[0094] Preferably, the orthogonal decoupling of the initial latent representation based on the cell type prior knowledge matrix includes using the cell type prior knowledge matrix as an orthogonal basis to construct an orthogonal projection operator;

[0095] Specifically, the data structure of the initial latent representation input to the prior integration layer is a feature matrix. ,in Represents the number of cell samples. The characteristic dimension representing the potential space;

[0096] Furthermore, a learnable feature embedding matrix is ​​initialized in the adversarial autoencoder network. The feature embedding matrix The feature embedding matrix is ​​updated via backpropagation during network training. The cell type prior knowledge matrix Mapping to the latent space yields the latent prior matrix. ,in ;

[0097] Furthermore, the orthogonal projection operator is constructed based on the latent prior matrix, and the calculation formula is as follows: ,in Represents the orthogonal projection matrix. The transpose of the latent prior matrix. The generalized inverse matrix represents the matrix product. Using the generalized inverse matrix can avoid the collapse of the inversion operation caused by matrix singularity or ill-conditioning, and ensure the engineering stability of the projection operator.

[0098] In detail, the initial latent representation is linearly transformed by the orthogonal projection operator to project the initial latent representation onto a first direction to generate the biological feature subspace;

[0099] Specifically, the calculation formula for generating the biological feature subspace is as follows: ,in This represents a priori-guided biological feature subspace matrix that preserves true heterogeneity features strongly correlated with cell type.

[0100] Further, the orthogonal complement space of the initial latent representation in the first direction is calculated as the technical noise subspace;

[0101] In detail, the calculation formula for generating the technical noise subspace is as follows: ,in The technical noise subspace matrix is ​​strictly orthogonal to the biological feature subspace matrix in mathematical topology and is used to independently accommodate sequencing batch effects.

[0102] Preferably, the step of using the discriminator in the adversarial autocoding network to perform local adversarial alignment constraints within the technical noise subspace includes inputting features in the technical noise subspace into the discriminator and calculating the adversarial loss between the feature distribution in the technical noise subspace and the standard normal distribution.

[0103] Specifically, the formula for calculating the adversarial loss is as follows: ,in This represents the output probability of the discriminator. This represents a matrix with a mean of 0 and a covariance matrix that is the identity matrix. The standard normal distribution The technical noise feature vector representing the sampling;

[0104] Furthermore, when updating the network through backpropagation based on the adversarial loss, the gradient is only backpropagated to the network weights corresponding to the technical noise subspace, and the backpropagation of the gradient to the network weights corresponding to the biological feature subspace is blocked by a masking operation.

[0105] In detail, based on the orthogonal projection matrix Construct orthogonal complement matrix ,in It is the identity matrix;

[0106] Specifically, during backpropagation, the gradient matrix of the adversarial loss with respect to the initial latent representation is calculated. At that time, matrix multiplication is performed. ,in This represents the effective gradient matrix actually used to update the encoder weights;

[0107] Furthermore, the masking operation causes the gradient of the adversarial loss to act only on the noise dimension that eliminates batch effects, freezing the network weights corresponding to the biological feature subspace and preventing the adversarial alignment process from altering the weak biological signals of rare cells.

[0108] Example 3:

[0109] Based on Examples 1 and 2, this embodiment elaborates in detail on the preprocessing optimization of single-cell multi-omics raw data, the joint optimization training of adversarial autoencoder networks, and the extraction steps of downstream target cell subpopulations.

[0110] Preferably, before inputting the raw single-cell multi-omics data into the preset adversarial autoencoder network, the raw single-cell multi-omics data is subjected to quality control and count normalization processing.

[0111] Specifically, low-quality cells with fewer than a preset gene expression threshold are filtered out, as are dead cells with a higher proportion of mitochondrial gene expression than a preset ratio threshold.

[0112] Furthermore, the count normalization process employs a logarithmic normalization algorithm, the calculation formula of which is as follows: ,in For cells Zhonggen The original count value, The normalized expression value;

[0113] In detail, a biological noise correction reference is established based on the preset grouping effect and batch parameters, and the normalized single-cell multi-omics raw data is subjected to noise reduction preprocessing based on the biological noise correction reference.

[0114] Furthermore, the grouping effect characterizes known biological covariates under different experimental conditions, and the batch parameter characterizes the systematic bias generated by different sequencing batches;

[0115] In detail, construct a linear mixture model. ,in This represents the normalized raw data from the single-cell multi-omics dataset. The design matrix representing the grouping effect, Represents the biological fixed effects coefficient. The design matrix representing the batch parameters, Represents the batch random effects coefficient. Represents the residual term;

[0116] Specifically, the batch random effects coefficients are obtained by solving the linear mixed model using the restricted maximum likelihood estimation method. The estimated value will As a reference for the biological noise correction;

[0117] Furthermore, the biological noise correction reference is subtracted from the normalized single-cell multi-omics raw data to output the denoised expression matrix. The denoising preprocessing operation removes some systematic technical biases before the data enters the nonlinear mapping, thereby improving the initial signal-to-noise ratio of the multimodal data.

[0118] Preferably, the adversarial autoencoder network further includes a classifier, into which features from the biological feature subspace are input;

[0119] Specifically, the classifier predicts cell types based solely on features in the biological feature subspace and outputs cell type prediction results.

[0120] Furthermore, the classifier includes a fully connected layer and a softmax activation function, and the data structure of the cell type prediction result is a probability distribution vector. Its elements characterize the predicted probability of a cell belonging to various preset cell types;

[0121] In detail, obtain the cell type pseudo-labels corresponding to the raw single-cell multi-omics data. The cell type pseudo-label It is a supervision signal generated through a pre-trained auxiliary annotation model or an unsupervised pre-clustering algorithm;

[0122] Specifically, the probability distribution vector is calculated. With the cell type pseudo-tag The cross-entropy loss between them is used as the classification loss. This enables semi-supervised training of adversarial autoencoder networks;

[0123] Specifically, features from the biological feature subspace and the technical noise subspace are input into the decoder to reconstruct the expression matrix. The decoder includes modality-specific decoding branches corresponding to the gene expression matrix and the peak matrix, respectively, and the reconstruction loss is calculated.

[0124] Furthermore, the decoder linearly superimposes and recombines the biological feature subspace and the technical noise subspace, and then maps them back to the original high-dimensional space through the corresponding modality-specific decoding branches, outputting a reconstructed expression matrix containing the reconstructed gene expression matrix and the reconstructed peak matrix. ;

[0125] In detail, the reconstruction loss is calculated based on the underlying statistical distribution characteristics of different omics data; for the reconstructed gene expression matrix, the difference between it and the original gene expression matrix is ​​calculated using a zero-inflated negative binomial distribution loss function; for the reconstructed peak matrix, the difference between it and the original peak matrix is ​​calculated using a binary cross-entropy loss function; and the weighted sum of the ZINB loss and the BCE loss is used as the reconstruction loss. To accurately fit the high sparsity and technically missing features of single-cell multi-omics data;

[0126] Specifically, the total loss function of the adversarial autoencoder network includes at least the reconstruction loss and the adversarial loss generated by the discriminator, and the adversarial autoencoder network is jointly optimized and trained based on the total loss function;

[0127] Furthermore, the formula for calculating the total loss function is as follows: ,in , , These represent the preset reconstruction weights, adversarial weights, and classification weights hyperparameters, respectively.

[0128] In detail, the hyperparameters of reconstruction weights, adversarial weights, and classification weights are obtained iteratively on the validation set using a grid search algorithm combined with cross-validation;

[0129] Specifically, in order to ensure a balance between noise reduction and feature preservation in the model, the reconstructed weights The value range is set to 1.0 to 10.0, and the adversarial weight The value range is set to 0.1 to 1.0, and the classification weight... The value range is set to 1.0 to 5.0;

[0130] Furthermore, in a preferred implementation scenario, the reconstruction weights are set. The adversarial weight is 5.0. The classification weight is 0.5. The joint optimization training operation, rated at 2.0, enables the model to predict cell types while eliminating technical noise.

[0131] Preferably, cluster analysis is performed on the fused single-cell data to extract data features of the target cell subpopulation;

[0132] Specifically, the target cell subpopulation includes stem cell-like cells or drug-resistant precursor cells;

[0133] Furthermore, the Louvain community detection algorithm is used to perform unsupervised clustering on the fused single-cell data to generate multiple cell clusters;

[0134] In detail, the fold change and significance P-value of gene expression levels in each of the cell clusters are calculated, and a set of characteristic genes whose expression levels meet the preset difference threshold are screened out.

[0135] Specifically, the set of characteristic genes is compared with the reference gene set of known stem cell-like cells or drug-resistant precursor cells, and the cell cluster with the highest matching degree is extracted as the data feature of the target cell subpopulation.

[0136] Furthermore, the matching degree is calculated using the Jaccard similarity coefficient, and the calculation formula is as follows: ,in This represents the set of characteristic genes of the cell cluster. A reference gene set representing known stem cell-like cells or drug-resistant precursor cells. This represents the number of genes in the intersection of two sets. This represents the number of genes in the union of two sets.

[0137] In detail, the Jaccard similarity coefficient The cell cluster with the largest numerical value is extracted as the target cell subpopulation;

[0138] Furthermore, the extracted data features of the target cell subpopulation are represented as a gene expression profile matrix of a specific cell population, which is used to guide downstream drug target screening or drug resistance mechanism analysis calculations.

[0139] Example 4:

[0140] A single-cell multi-omics data alignment system based on prior knowledge injection includes a data acquisition module, a latent space mapping module, an orthogonal decoupling module, a local adversarial alignment module, and a data reconstruction module.

[0141] Preferably, the data acquisition module is used to acquire raw single-cell multi-omics data of the same biological sample;

[0142] Specifically, the latent space mapping module is used to input the single-cell multi-omics raw data into a preset adversarial autoencoder network, and the encoder in the adversarial autoencoder network maps the single-cell multi-omics raw data to a shared latent space to obtain an initial latent representation;

[0143] Furthermore, the orthogonal decoupling module is used to obtain a pre-constructed cell type prior knowledge matrix, and through a preset prior integration layer, orthogonally decouples the initial latent representation based on the cell type prior knowledge matrix, dividing the initial latent representation into a prior-guided biological feature subspace and a technical noise subspace orthogonal to the biological feature subspace.

[0144] In detail, the local adversarial alignment module is used to perform local adversarial alignment constraints within the technical noise subspace using the discriminator in the adversarial autoencoder network, and to perform gradient blocking on the biological feature subspace to eliminate batch effects;

[0145] Specifically, the data reconstruction module is used to reconstruct data based on the biological feature subspace and the technical noise subspace after decoupling and undergoing local adversarial alignment constraints, through the decoder in the adversarial autoencoder network, and output fused single-cell data that eliminates technical noise and retains the rare cell topology.

[0146] Furthermore, the specific working principle of each virtual functional module in the single-cell multi-omics data alignment system based on prior knowledge injection corresponds one-to-one with the execution logic of the single-cell multi-omics data alignment method based on prior knowledge injection described in Examples 1 to 3. The detailed mathematical calculations and parameter transmission processes of each virtual functional module will not be elaborated here.

[0147] Example 5:

[0148] An electronic device includes at least one processor and a memory communicatively connected to said at least one processor.

[0149] Preferably, the memory stores a computer program that can be executed by the at least one processor;

[0150] Specifically, the computer program is executed by the at least one processor to enable the at least one processor to execute the single-cell multi-omics data alignment method based on prior knowledge injection as described in any one of Embodiments 1 to 3;

[0151] Furthermore, the electronic device also includes a communication interface and a communication bus, through which the processor, the memory, and the communication interface communicate and transmit data with each other;

[0152] Specifically, the communication interface is used to receive raw single-cell multi-omics data transmitted from external sequencing equipment, and to send the fused single-cell data output by the processor to a downstream analysis terminal or an external database;

[0153] Specifically, the processor is a central processing unit, a graphics processing unit, or a dedicated tensor processor, and the memory is a high-speed random access memory or a non-volatile memory to meet the hardware resource requirements of large-scale single-cell matrix operations.

[0154] Preferably, this embodiment also provides a computer-readable storage medium storing a computer program;

[0155] Specifically, when the computer program is executed by the processor, it implements the single-cell multi-omics data alignment method based on prior knowledge injection as described in any one of Embodiments 1 to 3.

[0156] Furthermore, the computer-readable storage medium includes physical media capable of persistently storing program code, such as read-only memory, random access memory, solid-state drive, magnetic disk, or optical disk.

[0157] Experimental example:

[0158] This embodiment provides an experimental example to verify the technical effectiveness of the single-cell multi-omics data alignment method based on prior knowledge injection in rare cell identification and multimodal data fusion.

[0159] A single-cell multi-omics test dataset containing known stem cell-like cells and drug-resistant precursor cells was constructed. The single-cell multi-omics test dataset includes single-cell transcriptome sequencing data and chromatin accessibility sequencing data.

[0160] A comparative and an experimental group are set up. The comparative group uses a conventional global adversarial alignment model to process the single-cell multi-omics test dataset. The experimental group uses the single-cell multi-omics data alignment method based on prior knowledge injection described in Embodiments 1 to 3 of this invention to process the single-cell multi-omics test dataset.

[0161] The conventional global adversarial alignment model applies a uniform adversarial loss constraint to all feature dimensions in the latent space, without introducing a prior knowledge matrix or performing orthogonal decoupling operations.

[0162] Louvain clustering analysis was performed on the globally aligned data output by the comparative model and the fused single-cell data output by the experimental group, respectively, and the clustering resolution index was calculated.

[0163] The comparative example causes feature collapse when eliminating batch effects, resulting in the overlap between the stem cell-like cells and ordinary cells, and its clustering resolution baseline value is a preset value.

[0164] The experimental group freezes the biological feature subspace through the orthogonal decoupling module and the local adversarial alignment module, so that the stem cell-like cells form independent clusters;

[0165] The clustering resolution of the fused single-cell data output by the experimental group was improved by 40% compared with the comparative group. The orthogonal decoupling operation eliminated technical noise while preserving the weak heterogeneity signal of rare cells.

[0166] Based on the data output from the comparative example and the experimental group, cell type-specific regulatory networks were constructed, and the prediction accuracy of the regulatory networks was calculated.

[0167] The prediction accuracy of the control network is quantitatively evaluated using the area under the precision-recall curve.

[0168] The fused single-cell data output by the experimental group preserved the true biological manifold topology, and the prediction accuracy of the constructed regulatory network reached over 85%.

[0169] The experimental group's regulatory network prediction accuracy was better than that of the comparative group, and the local adversarial alignment constraint operation improved the accuracy of regulatory factor matching in cross-modal data.

[0170] Differential expression analysis was performed on the data output from the comparative and experimental groups to statistically analyze the number of highly specific liquid biopsy biopsy markers discovered and the validation cycle.

[0171] The experimental group used the cell type prior knowledge matrix to provide an objective coordinate benchmark, which accelerated the extraction process of characteristic genes of the target cell subpopulation.

[0172] The biomarker detection efficiency of the experimental group was three to five times higher than that of the comparative group;

[0173] The proposed single-cell multi-omics data alignment method based on prior knowledge injection breaks the zero-sum game between data denoising and biological feature fidelity preservation, and improves the clinical translation feasibility of single-cell multi-omics analysis with trace samples.

[0174] In summary, the single-cell multi-omics data alignment method, system, electronic device, and storage medium provided by this invention, by introducing a cell type prior knowledge matrix into the latent space of the adversarial autoencoder network and performing orthogonal decoupling, divides the initial latent representation into a priori-guided biological feature subspace and an orthogonal technical noise subspace. Within the technical noise subspace, a discriminator performs local adversarial alignment constraints and applies gradient blocking to the biological feature subspace, enabling the adversarial autoencoder network to update only the parameters of the noise dimension when eliminating batch effects. This overcomes the defect of conventional global adversarial alignment, which blindly pursues statistical distribution consistency and causes biological topological collapse. The orthogonal decoupling and local adversarial alignment mechanism achieves decoupling and separation of sequencing technology bias and real biological variation at the underlying mathematical topology, reducing background noise while preserving the weak heterogeneity signals of rare subpopulations such as stem cell-like cells and drug-resistant precursor cells, thus improving the clustering resolution and downstream biomarker discovery efficiency of single-cell multi-omics data fusion.

[0175] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for aligning single-cell multi-omics data based on prior knowledge injection, characterized in that, Includes the following steps: S1. Obtain raw single-cell multi-omics data from the same biological sample; S2. Input the raw data of single-cell multi-omics into a preset adversarial autoencoder network, and map the raw data of single-cell multi-omics to a shared latent space through the encoder in the adversarial autoencoder network to obtain an initial latent representation; S3. Obtain a pre-constructed cell type prior knowledge matrix. Through a preset prior integration layer, orthogonally decouple the initial potential representation based on the cell type prior knowledge matrix, and divide the initial potential representation into a prior-guided biological feature subspace and a technical noise subspace orthogonal to the biological feature subspace. S4. Within the technical noise subspace, the discriminator in the adversarial autoencoder network is used to perform local adversarial alignment constraints, and gradient blocking is applied to the biological feature subspace to eliminate batch effects. S5. Based on the decoupled biological feature subspace and the technical noise subspace after being constrained by local adversarial alignment, the data is reconstructed through the decoder in the adversarial autoencoder network to output fused single-cell data that eliminates technical noise and preserves the topology of rare cells.

2. The single-cell multi-omics data alignment method based on prior knowledge injection according to claim 1, characterized in that, The acquisition of the pre-constructed cell type prior knowledge matrix includes: Cell type marker genes were extracted through large-scale text mining; Calculate the consensus strength of the cell type marker genes as markers of specific cell types in the literature; An initial matrix is ​​constructed based on the consensus strength, and the initial matrix is ​​row normalized to generate the cell type prior knowledge matrix.

3. The method of claim 1, wherein, The orthogonal decoupling of the initial latent representation based on the prior knowledge matrix of the cell type, dividing the initial latent representation into a prior-guided biological feature subspace and a technical noise subspace orthogonal to the biological feature subspace, includes: Using the prior knowledge matrix of the cell type as an orthogonal basis, an orthogonal projection operator is constructed. The initial latent representation is linearly transformed by the orthogonal projection operator and projected onto the first direction to generate the biological feature subspace. Calculate the orthogonal complement space of the initial latent representation in the first direction as the technical noise subspace.

4. The single-cell multi-omics data alignment method based on prior knowledge injection according to claim 1, characterized in that, Within the technical noise subspace, the discriminator in the adversarial autoencoder network performs local adversarial alignment constraints, and gradient blocking is applied to the biological feature subspace, including: The features in the technical noise subspace are input into the discriminator, and the adversarial loss between the feature distribution in the technical noise subspace and the standard normal distribution is calculated. When updating the network through backpropagation based on the adversarial loss, the gradient is only backpropagated to the network weights corresponding to the technical noise subspace, and the backpropagation of the gradient to the network weights corresponding to the biological feature subspace is blocked by a masking operation.

5. The method of claim 1, wherein, The single-cell multi-omics raw data includes at least single-cell transcriptome data and chromatin accessibility epigenome data; Before inputting the raw single-cell multi-omics data into a preset adversarial autoencoder network, the method further includes: The raw data from the single-cell multi-omics study were subjected to quality control and count normalization. A biological noise correction reference is established based on the preset grouping effect and batch parameters, and the normalized single-cell multi-omics raw data is subjected to noise reduction preprocessing based on the biological noise correction reference.

6. The single-cell multi-omics data alignment method based on prior knowledge injection according to claim 1, characterized in that, The adversarial autoencoder network further includes a classifier; the method further includes: The features in the biological feature subspace are input into the classifier; The classifier predicts cell types based solely on features within the biological feature subspace and outputs the cell type prediction results.

7. The single-cell multi-omics data alignment method based on prior knowledge injection according to claim 1, characterized in that, The data reconstruction based on the decoupled biological feature subspace and the technical noise subspace after local adversarial alignment constraints, through the decoder in the adversarial autoencoder network, includes: The features in the biological feature subspace and the technical noise subspace are input into the decoder to reconstruct the expression matrix, and the reconstruction loss is calculated. The total loss function of the adversarial autoencoder network includes at least the reconstruction loss and the adversarial loss generated by the discriminator, and the adversarial autoencoder network is jointly optimized and trained based on the total loss function.

8. The single-cell multi-omics data alignment method based on prior knowledge injection according to claim 1, characterized in that, Also includes: Cluster analysis was performed on the fused single-cell data to extract data features of the target cell subpopulation; The target cell subpopulation includes stem cell-like cells or drug-resistant precursor cells.

9. A system for single-cell multi-omics data alignment based on prior knowledge injection, based on the method of any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire raw single-cell multi-omics data from the same biological sample. The latent space mapping module is used to input the raw data of single-cell multi-omics into a preset adversarial autoencoder network, and to map the raw data of single-cell multi-omics to a shared latent space through the encoder in the adversarial autoencoder network to obtain an initial latent representation. An orthogonal decoupling module is used to obtain a pre-constructed cell type prior knowledge matrix. Through a preset prior integration layer, the initial potential representation is orthogonally decoupled based on the cell type prior knowledge matrix, and the initial potential representation is divided into a prior-guided biological feature subspace and a technical noise subspace orthogonal to the biological feature subspace. The local adversarial alignment module is used to perform local adversarial alignment constraints within the technical noise subspace using the discriminator in the adversarial autoencoder network, and to perform gradient blocking on the biological feature subspace to eliminate batch effects. The data reconstruction module is used to reconstruct data based on the decoupled biological feature subspace and the technical noise subspace after local adversarial alignment constraints, through the decoder in the adversarial autoencoder network, and output fused single-cell data that eliminates technical noise and preserves the rare cell topology.

10. An electronic device, comprising: include: At least one processor; as well as, A memory that is communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the single-cell multi-omics data alignment method based on prior knowledge injection as described in any one of claims 1 to 8.

Images