Single cell gene characterization system and method based on deep learning

Abstract

The invention belongs to the technical field of cell data processing, and relates to a single cell gene characterization system and method based on deep learning. The system comprises a gene information input module, a data processing module, a context sensing interpolation module, a gene feature fusion module, a deep representation learning and joint optimization module and a cell typing and differentiation track inference module. The input module integrates a data source; the processing module preprocesses and enhances the original expression matrix; the interpolation module performs weighted interpolation by using a conditional Gaussian graph model; the gene feature fusion module generates a gene module activity matrix through 1 * 1 convolution and the like; the fusion module fuses the sample meta-information with the active matrix, the sequencing depth and the gene feature vector; the depth representation module captures high-dimensional dependence through a multi-layer Transform network, and generates cell high-order potential representation; the typing module outputs a cell population division and differentiation trajectory based on the deduced pseudo-time trajectory. The method has the beneficial effects that cell expression characteristics are recovered under high sparseness and noise, and a new way is provided for cell type identification and the like.

Description

Technical Field

[0001] This invention relates to the field of cell data processing technology, and in particular to a single-cell gene characterization system and method based on deep learning. Background Technology

[0002] The rapid development of single-cell sequencing technology has provided unprecedented opportunities for studying cellular heterogeneity, differentiation trajectories, and disease mechanisms. Technologies such as single-cell RNA sequencing (scRNA-seq) and single-cell ATAC sequencing (scATAC-seq) enable high-resolution analysis of gene expression, epigenetic status, and genomic characteristics of each cell. The application of these technologies has driven profound changes in oncology, immunology, developmental biology, and other fields, demonstrating enormous potential, particularly in precision medicine and personalized treatment.

[0003] However, single-cell transcriptome data itself has high dimensionality (number of genes G≈2×10). 4 Single-cell data exhibits several characteristics, including high sparsity (typically over 80%–95% of values ​​are zero), high noise levels (including technical noise such as dropout events, amplification bias, and sequencing depth differences, as well as biological noise such as gene expression randomness), and significant batch effects (systematic differences caused by different sequencing platforms, experimental batches, and tissue sources). These complexities pose significant challenges to the accurate analysis and reliable interpretation of single-cell data, specifically in the following aspects: 1. In single-cell transcriptome data, the expression value of the vast majority of genes is zero in each cell. These zero values ​​are not only a biological signal of gene non-expression, but also include missing values ​​caused by technical noise (such as dropout events, amplification bias, and sequencing depth differences). Traditional methods cannot effectively recover this missing information, affecting the accuracy of downstream analyses (such as cell type identification, gene regulatory network inference, etc.).

[0004] 2. Systematic differences arising from different sequencing platforms, experimental batches, and tissue sources often cause batch effects in data processing and analysis, affecting the integration and analysis of cross-experimental and cross-platform data. Existing batch correction methods mostly rely on linear assumptions or manual alignment, making it difficult to handle complex nonlinear batch structures.

[0005] 3. Single-cell data has extremely high dimensionality (the number of genes often exceeds 20,000), while the number of cells is usually only a few thousand to tens of thousands, leading to the "high-dimensional, low-sample" problem. This makes directly applying traditional machine learning methods face problems such as high computational cost, overfitting, and the curse of dimensionality.

[0006] 4. Although deep learning models have made significant progress in improving representation learning ability and clustering accuracy, the features of these models are often difficult to directly connect with specific biological processes (such as gene function, signaling pathways, etc.), which reduces the interpretability and clinical application value of the algorithms.

[0007] 5. With the rise of single-cell multi-omics technology, researchers can simultaneously obtain multidimensional data such as transcriptomics, chromatin accessibility, DNA methylation, and proteomics. However, these data differ significantly in measurement methods, noise characteristics, and statistical distributions, making it difficult for traditional methods to effectively perform cross-modal data fusion and feature alignment, thus affecting model performance and generalization ability.

[0008] To address the aforementioned issues, deep learning methods have been widely applied in single-cell data analysis in recent years, especially models based on variational autoencoders (VAEs) and generative adversarial networks (GANs) (such as scVI and scGAN), as well as models based on the Transformer architecture (such as scGPT and Geneformer). These methods have achieved significant progress in feature extraction, clustering, and representation learning. However, the following problems still exist: 1. Existing single-cell data analysis methods typically employ a step-by-step workflow, performing tasks such as data preprocessing, interpolation, clustering, and trajectory inference separately, making it difficult to fully utilize the complementary information between these tasks. The lack of a unified framework capable of handling multiple tasks simultaneously leads to error accumulation and performance limitations.

[0009] 2. Although some methods attempt to model the synergistic relationships between genes through co-expression networks or gene set scoring, these methods often fail to fully utilize gene functional module information in the core representation learning process of the model, resulting in insufficient interpretability and biological rationality of the model.

[0010] 3. The heterogeneity of multi-omics data poses challenges to existing methods, especially across multiple technology platforms or experimental batches. How to effectively integrate data from different omics and eliminate technological differences remains an unsolved problem.

[0011] Therefore, the current field of single-cell data analysis urgently needs a comprehensive method that can uniformly handle high sparsity, high noise, and strong batch effects, while simultaneously supporting gene functional module decoupling, differentiation trajectory modeling, comparative characterization learning, and sequencing depth adaptive correction. Summary of the Invention

[0012] Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a single-cell gene characterization system and method based on deep learning, which solves the technical problems of high sparsity, high noise, batch effect, high data dimensionality and feature redundancy, and lack of interpretability in existing single-cell data analysis.

[0013] Technical solution To achieve the above objectives, the main technical solutions adopted by the present invention include: In a first aspect, the present invention provides a single-cell gene representation system based on deep learning, comprising: a gene information input module, a data processing module, a context-aware interpolation module, a gene feature fusion module, a fusion module, and a deep representation learning and joint optimization module; The gene information input module is responsible for integrating the input data sources; The inputs include: the original single-cell expression matrix, the gene function feature matrix, sequencing depth information, sample metadata, and cell-level feature vectors; The data processing module is used to preprocess the original single-cell expression matrix and perform data augmentation operations on the preprocessed matrix. The context-aware interpolation module uses a conditional Gaussian graph model (GGM) to estimate the partial correlation coefficients between genes, and performs weighted interpolation on the processed matrix based on the top k related genes to obtain the interpolated single-cell expression matrix. The gene feature fusion module first fuses the gene functional feature matrix with the imputed single-cell expression matrix, and then activates it through 1×1 convolution and competitive activation mechanism to generate a gene module activity matrix. The fusion module fuses sample metadata, sequencing depth, and cell-level feature vectors with the gene module activity matrix to obtain a single-cell fusion matrix. The deep representation learning and joint optimization module uses a multi-layer stacked Transformer deep learning architecture to perform representation learning on the single-cell fusion matrix: a preliminary high-order latent representation of the cell is generated through a multi-head self-attention mechanism; this latent representation is constrained and enhanced through a multi-task joint optimization mechanism to output the final high-order latent representation of the cell as the cell representation.

[0014] Optionally, it also includes: a cell typing and differentiation trajectory inference module, which takes the higher-order latent representation of cells as input, performs pseudo-temporal trajectory inference on cells through the multi-head attention mechanism of the Transformer deep learning architecture, captures the continuous process and state changes of cell differentiation, and finally outputs the cell population division results and cell differentiation trajectory; The cell typing and differentiation trajectory inference module includes: a cell typing submodule and a differentiation trajectory inference submodule; The cell typing submodule inputs the higher-order latent representation of cells into the clustering adaptive layer. This layer calculates the similarity between cells through an optimized clustering algorithm and adopts an adaptive strategy based on internal evaluation indicators to automatically search and determine the optimal number of clusters, thereby determining the cell population division result and updating the cluster centers to divide biologically significant cell populations and complete cell type identification. The differentiation trajectory inference submodule predicts cell differentiation trajectories based on the Transformer deep learning architecture. This submodule includes two modes: Single-cell mode: directly takes the latent representation of a single cell as input and outputs its pseudo-time scalar through a regression network; Trajectory mode: A group of cells is arranged into a sequence in an initial order, input into a Transformer-based decoder, and its attention mechanism is used to refine the pseudo-temporal order of the entire sequence, outputting a smooth continuous differentiation trajectory.

[0015] In a second aspect, the present invention provides a single-cell gene characterization method based on a deep learning-based single-cell gene characterization system as described in any one of the first aspects above, comprising the following steps: Step 1: Obtain the input raw single-cell expression matrix, gene function feature matrix, sequencing depth information, sample metadata, and cell-level feature vector; Step 2: Perform data augmentation operations after preprocessing the original single-cell expression matrix; Step 3: Based on the processed matrix, use a conditional Gaussian graph model to iteratively estimate the partial correlation coefficients between genes; for each zero value in the matrix, perform weighted imputation based on the expression values ​​of the top-k related genes of the corresponding gene to obtain the imputed single-cell expression matrix; Step 4: Fuse the imputed expression matrix with the gene function feature matrix to obtain the gene feature fusion matrix; Step 5: Calculate the activation scores of each cell on multiple functional modules using 1×1 convolution and competitive activation mechanisms, and reweight and aggregate the gene feature fusion matrix using the module activation scores to generate a gene module activity matrix. Step 6: Fuse the sample metadata, sequencing depth, cell-level feature vectors, and the gene module activity matrix to obtain a single-cell fusion matrix; Step 7: Use a multi-layer stacked Transformer deep learning architecture to perform representation learning on the single-cell fusion matrix to generate a preliminary high-order latent representation of the cell; constrain and enhance this latent representation through a multi-task joint optimization mechanism, and output the final high-order latent representation of the cell as the cell representation.

[0016] Optionally, the original single-cell expression matrix, gene function feature matrix, sequencing depth information, sample metadata, and cell-level feature vector are specifically: The original single-cell expression matrix has a dimension of N×G1, where N represents the number of cells and G1 represents the number of genes; each element in the matrix is ​​a unique molecular identifier (UMI) count or original expression count commonly used in single-cell sequencing. The gene function feature matrix refers to the fixed feature representation corresponding to each gene in the gene function matrix, with a dimension of G2×F, where F represents the predefined gene feature dimension and G2 represents the predefined number of gene features; this matrix is ​​a static prior knowledge base obtained through pre-training. The sequencing depth information includes the total UMI number T for each cell and the standard depth S estimated by subsampling, which is used for depth normalization and perturbation enhancement during training. The sample metadata includes the technology platform, tissue source, experimental batch, and donor information; The cell-level feature vectors are generated from omics data, including chromatin accessibility features (ATAC-seq), DNA methylation features, and genetic variation features.

[0017] Optionally, step 2 specifically involves: performing basic quality control and library size normalization on the original single-cell expression matrix to obtain a preprocessed expression matrix; subsequently, applying library size perturbation and negative binary distribution noise to the preprocessed expression matrix to generate an enhanced expression matrix.

[0018] Optionally, in step 3, the partial correlation coefficient matrix between genes is estimated using GGM. For each gene expressed as zero in a cell, the k genes most related to it are found based on the partial correlation coefficient. The expression values ​​of the k genes are weighted by the partial correlation strength between the gene to be imputed and these k genes to obtain an estimate of the missing value. This process is iterated until the estimated value is stable, thereby obtaining the imputed single-cell expression matrix.

[0019] Optionally, step 5 specifically includes the following methods: Step 5.1: Using lightweight 1×1 convolutional layers, the functional feature vector of each gene is mapped to a higher-dimensional shared expression space to enhance its feature expression ability; Step 5.2: For each cell, calculate the matching degree between the expression patterns of all its mapped genes and the basis vectors of each module, and perform competitive activation using the softmax function to obtain the module activation score; Step 5.3: Reweight and aggregate the gene feature fusion matrix using module activation scores to finally generate the gene module activity matrix.

[0020] Optionally, the multi-task joint optimization mechanism in step 6 includes five task heads, specifically: representation reconstruction task, RDA prediction task, pseudo-temporal prediction task, contrastive learning task, and clustering adaptive task. The expression reconstruction task involves the decoder attempting to reconstruct the imputed expression matrix from the latent representation, using mean squared error loss to ensure that the representation retains basic gene expression information. The RDA prediction task is to predict the subsampling depth ratio for each cell, using mean squared error loss, so that the model can understand and correct for the impact of sequencing depth. The pseudo-time prediction task includes a regression head predicting pseudo-time values ​​for cells, using smoothed L1 loss to guide the model to capture continuous changes in cell state. The contrastive learning task enhances the discriminativeness and cross-batch consistency of representations by zooming in on different enhanced views of the same cell or representations of similar cells and zooming out on representations of different cells using InfoNCE loss. The clustering adaptive task involves simultaneously performing online clustering in the latent representation space, optimizing the loss function based on the silhouette coefficient, and promoting the formation of a cluster structure with clear boundaries in the representation space.

[0021] Optionally, it also includes step 8: processing the higher-order latent representation of the cells to output cell population segmentation results and cell differentiation trajectories; Step 8.1: Perform cluster analysis on the higher-order latent representation of the cells, determine the optimal number of clusters through an adaptive strategy, and obtain the cell typing results; Step 8.2: Based on the higher-order latent representation of the cell, perform pseudo-temporal prediction using a Transformer-based decoder to output the cell differentiation trajectory.

[0022] Beneficial effects The beneficial effects of this invention are: 1. By combining imputation, representation learning, clustering and differentiation trajectory inference, it can accurately recover cell expression characteristics in extremely sparse and noisy backgrounds, providing a new technical approach for cell type identification, state labeling and new cell population discovery.

[0023] 2. By integrating gene functional characteristics, partial correlation structures, and modular gene networks based on gene expression, the model can simultaneously possess statistical interpretability and biological rationality during the learning process, significantly improving the scientific rigor of cell characterization and classification results.

[0024] 3. By using data augmentation, GGM imputation, meta-information gated embedding, Transformer attention structure, and module context vectors, we can comprehensively mitigate sequencing depth differences and batch effects, and improve the stability and consistency of the embedding space.

[0025] 4. Through mechanisms such as pseudo-temporal modeling, adaptive clustering, and interpretable module activation, cell differentiation processes, state transitions, and population heterogeneity can be systematically analyzed, which has important reference value for both basic biological research and clinical translational research.

[0026] 5. The model adopts a multi-layer Transformer, gradient checkpoint, and high-dimensional module projection design, which can support large-scale single-cell data training and is suitable for complex sample scenarios such as cross-species, cross-platform, and cross-tissue. Attached Figure Description

[0027] Figure 1 A module structure diagram of a deep learning-based single-cell gene characterization system provided in an embodiment of the present invention; Figure 2 This is a flowchart of a single-cell gene characterization method based on deep learning provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the overall process of cross-modal alignment and multimodal fusion in an embodiment of the present invention; Figure 4 This is a schematic visualization of the original single-cell expression matrix in an embodiment of the present invention; Figure 5 This is a schematic visualization of the single-cell expression matrix after applying the context-aware interpolation strategy in an embodiment of the present invention. Figure 6 This is a schematic diagram of gene feature fusion and multimodal input in an embodiment of the present invention; Figure 7 This is a schematic diagram of the cell-module relationship based on module activation in an embodiment of the present invention; Figure 8 This is a schematic diagram of metadata embedding in UMAP in an embodiment of the present invention; Figure 9 This is a schematic diagram of the Transformer encoding and multi-task joint optimization structure in an embodiment of the present invention; Figure 10 This is a schematic diagram of the RDA task prediction trend in an embodiment of the present invention; Figure 11 This is a schematic diagram of the cell embedding spatial structure in an embodiment of the present invention; Figure 12 This is a schematic diagram of the pseudo-temporal differentiation trajectory structure in an embodiment of the present invention. Detailed Implementation

[0028] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0030] Firstly, in this embodiment, a single-cell gene characterization system based on deep learning, such as... Figure 1 As shown, it includes: a gene information input module, a data processing module, a context-aware interpolation module, a gene feature fusion module, a fusion module, a deep representation learning and joint optimization module, and a cell typing and differentiation trajectory inference module; The gene information input module is responsible for integrating input data sources. Inputs include: raw single-cell expression matrix, gene functional feature matrix, sequencing depth information, sample metadata, and cell-level feature vectors.

[0031] The data processing module performs basic quality control and library size normalization on the original single-cell expression matrix to obtain a preprocessed expression matrix. Subsequently, library size perturbation and negative binomial noise are applied to the preprocessed expression matrix to generate an enhanced expression matrix. These operations ensure that expression data from different cells are comparable, reduce the impact of depth differences, and improve the system's robustness to sequencing depth differences and technical noise.

[0032] The context-aware imputation module uses a conditional Gaussian graph model (GGM) to estimate the partial correlation coefficients between genes, and performs weighted imputation on the processed matrix based on the top k related genes to obtain the imputed single-cell expression matrix. The missing expression data is restored through an iterative process, which can effectively reduce the impact of data sparsity on model training.

[0033] The gene feature fusion module first fuses the gene functional feature matrix with the imputed single-cell expression matrix, and then activates it through 1×1 convolution and competitive activation mechanisms to generate a gene module activity matrix. This matrix effectively captures the synergistic expression relationships between genes and provides a biologically interpretable modular representation.

[0034] The fusion module fuses sample metadata, sequencing depth, and cell-level feature vectors with the gene module activity matrix in a controllable manner to obtain a single-cell fusion matrix.

[0035] The deep representation learning and joint optimization module employs a multi-layer stacked Transformer deep learning architecture to perform representation learning on the single-cell fusion matrix. Through its core multi-head self-attention mechanism, this network can effectively capture complex high-dimensional dependencies across genes, functional modules, and cells, thereby generating preliminary high-order latent representations of cells. To handle large-scale dataset training, gradient checkpointing is introduced at this stage to optimize memory usage and ensure efficient training of large models.

[0036] Building on this, the module further constrains and enhances this latent representation through a multi-task joint optimization mechanism to output the final high-order latent representation of the cell as a cell characterization.

[0037] Specifically, the model performs imputation and prediction tasks for expression values ​​or related variables through a regression decoder to maintain numerical accuracy; when RDA is enabled, the expression values ​​of random masks are reconstructed to improve the model's robustness to noise and missing data.

[0038] The cell typing and differentiation trajectory inference module takes the higher-order latent representation of cells as input and uses the multi-head attention mechanism of the Transformer deep learning architecture to infer pseudo-temporal trajectories of cells, capturing the continuous process and state changes of cell differentiation. During the training phase, its core supervised objective is the pseudo-temporal prediction task. Furthermore, this module introduces a contrastive learning projection head to enhance the discriminative ability of the latent space and cross-platform consistency, ultimately outputting cell representations that include cell population segmentation results and cell differentiation trajectories. It further includes a cell typing submodule and a differentiation trajectory inference submodule.

[0039] The cell typing submodule inputs the higher-order latent representation of cells into the adaptive clustering layer. This layer calculates the similarity between cells using an optimized clustering algorithm and employs an adaptive strategy based on internal evaluation metrics to automatically search for and determine the optimal number of clusters, thereby determining the cell population segmentation result. Simultaneously, it updates the cluster centers, segmenting biologically significant cell populations and completing cell type identification.

[0040] The differentiation trajectory inference submodule is based on the Transformer deep learning architecture. This submodule can operate in two modes, including: Single-cell mode: The potential representation of a single cell is taken as input and its pseudo-time scalar is output through a regression network.

[0041] Trajectory mode: A group of cells (such as cells within a cluster) are arranged into a sequence according to a certain initial order, input into a Transformer-based decoder, and its attention mechanism is used to refine the pseudo-temporal order of the entire sequence, outputting a smooth continuous trajectory.

[0042] In addition, this submodule introduces a contrastive learning projector that forces cells in similar developmental stages (even from different batches) to be close to each other in the projection space during pseudo-temporal learning, thereby significantly improving the consistency of trajectory inference across platforms and datasets.

[0043] The dimensions, network depth, specific iterations of enhancement and imputation, embedding dimensions and gating weights described in the deep learning-based single-cell gene characterization system provided in this embodiment can all be flexibly adjusted according to requirements; each module is independent and composable; it is suitable for multiple platforms (10x, Smart-seq, etc.), multiple tissues, different species or other highly sparse omics data.

[0044] Secondly, in this embodiment, the single-cell gene characterization method based on the aforementioned deep learning-based single-cell gene characterization system is used, such as... Figure 2 and Figure 3 As shown, it includes the following steps: The raw single-cell expression matrix has dimensions N×G1, where N represents the number of cells and G1 represents the number of genes. Each element in the matrix is ​​typically a unique molecular identifier (UMI) count or raw expression count, commonly used in single-cell sequencing. The matrix elements are UMI counts, recording the raw transcription level of each cell, and are highly sparse and noisy raw observation data.

[0045] The gene function feature matrix refers to the fixed feature representation corresponding to each gene in the gene function matrix, with a dimension of G2×F, where F represents the predefined gene feature dimension and G2 represents the predefined number of gene features. This matrix is ​​a static prior knowledge base obtained through pre-training, and gene alignment processing is performed on the original single-cell expression matrix based on this matrix. Each row in the matrix (corresponding to a gene) is a fixed feature vector. These features can be derived from the embedding representations of various biological knowledge sources, including gene ontology (GO) annotations, KEGG pathways, protein-protein interaction (PPI) networks, etc. Its function is to fuse the static functional attributes of genes with dynamic expression levels, providing the model with functional semantic information beyond expression levels, and assisting in modeling differences in techniques and tissues.

[0046] Table 1 illustrates the relationship between genes and functional features, i.e., the gene-functional feature matrix, where genes are the rows and functional feature dimensions are the columns. Each value in the matrix represents the representation value of the corresponding gene on a specific functional feature dimension. The functional features include the network prior feature dimension PPI_dim extracted based on the Protein-Protein Interaction (PPI) network, and the functional annotation feature dimension constructed based on Gene Ontology (GO). The GO functional features further include three aspects: biological process GO_BP, molecular function GO_MF, and cellular component GO_CC. The PPI network features and GO functional annotation features can be mapped to continuous numerical representations through feature encoding or embedding, and then further input into representation learning models such as variational autoencoders for joint modeling, thereby obtaining the latent feature representation VAE_combined_dim that integrates multiple prior information. The dimension or its statistical representation in VAE_combined_dim is used for illustration and is not limited to a specific number of dimensions.

[0047] Table 1 Example of a gene functional characteristic matrix

[0048] The sequencing depth information includes the total number of UMIs (T) per cell and the standard depth (S) estimated by subsampling, which is used for depth normalization and perturbation enhancement during training.

[0049] The sample metadata includes discrete background information such as technology platform, tissue origin, experimental batch, and donor information. This information will be encoded and then input into a dedicated embedding layer to be converted into a dense vector representation.

[0050] Table 2 presents the sample metadata associated with the cell expression data, including the ID of each cell, sequencing platform, tissue type, batch, and donor information. This metadata helps in understanding and controlling for potential technical biases and biological differences.

[0051] Table 2 Sample Element Information Table

[0052] The cell-level feature vectors are generated from omics data and include chromatin accessibility (ATAC-seq) features, DNA methylation features, and genetic variation features (such as SNP coding vectors). These multimodal features can be used to supplement the expression data with information at the regulatory level.

[0053] In ideal implementation, multi-omics fusion can be performed based on sequencing data from different omics sources within the same cell to achieve a detailed characterization of the cell state. However, this invention is not limited thereto. Due to the current level of development of sequencing technology, the single-cell multi-omics data that can be stably obtained at this stage mainly includes single-cell RNA sequencing data, chromatin accessibility sequencing data, and antibody-derived tag data, etc., but the method described in this invention is also applicable to situations where only one or more of these omics data are obtained.

[0054] In practice, when different omics data cannot be simultaneously derived from the same cell, cell-level feature vectors can be constructed based on omics data from the same tissue, sample, or cell type through mapping, aggregation, or inference, and then processed according to the feature fusion and representation learning process described in this invention.

[0055] Among them, the single-cell chromatin accessibility sequencing (scATAC-seq) data is divided into genome bins (10 kb is the sliding window division) and the accessibility feature vector is formed by counting the fragments in each bin. Table 3 shows the scATAC-seq feature matrix generated based on the genome bin method. These multimodal features can be used to supplement the information of expression data at the regulatory level during subsequent fusion.

[0056] Table 3. scATAC-seq feature matrix generated based on genome binning method

[0057] In this embodiment, the pbmc 3k classic single-cell sequencing dataset from peripheral blood mononuclear cells (PBMCs) is used as an example for illustration. The pbmc 3k data contains 2700 cells and 13714 genes. After preprocessing, it can be converted into a matrix where rows represent cells and columns represent genes. The input raw single-cell expression matrix is ​​a sparse matrix, and its structure is shown in Table 4.

[0058] Table 4. Single-cell expression matrix data structure

[0059] The table below uses the pbmc 3k single-cell sequencing dataset as an example to show the types of input data, their dimensions, sources, and descriptions used in the training of the deep learning-based single-cell gene characterization method. It reflects the method's fusion of multiple types of information, such as the original expression matrix, meta-information, prior features of gene function, and sequencing depth indicators, to help understand the overall data inflow structure.

[0060] Table 5. Data Input Composition and Characteristics (Taking pbmc 3k as an example)

[0061] Step 2: Perform data augmentation operations after preprocessing the original single-cell expression matrix; Specifically, basic quality control and library size normalization are performed on the original single-cell expression matrix to obtain the preprocessed expression matrix; then, library size perturbation and negative binomial distribution noise are applied to the preprocessed expression matrix to generate the enhanced expression matrix.

[0062] Preprocessing operations were performed on the raw single-cell expression matrix, including quality control, normalization, and gene alignment. The quality control refers to filtering cells or genes based on the statistical characteristics of the single-cell expression matrix. This includes excluding cells with excessively low total expression counts or those containing only a very small number of genes, thereby improving the stability of subsequent models. Specific thresholds can be flexibly set according to the data characteristics.

[0063] The standardization process aims to mitigate the impact of sequencing depth variations and can perform library size normalization, logarithmic normalization, or other common normalization transformations. This process ensures that expression levels across different cells are on a comparable scale, thereby improving the model's generalization ability.

[0064] The gene alignment refers to the process of aligning the single-cell expression matrix and the gene functional feature matrix when they come from different sources by matching the intersection of gene names or gene IDs. This ensures that the two are consistent in dimension and order, providing a unified data foundation for subsequent gene feature fusion.

[0065] Subsequently, data augmentation was performed on the preprocessed single-cell expression matrix, specifically including library size perturbation and the introduction of negative binomial distribution noise; In single-cell transcriptome analysis, data sparsity and technical noise are core bottlenecks hindering downstream biological discovery. For example, the pbmc 3k dataset has a 93.8% proportion of zero elements in its original single-cell expression matrix, meaning that the vast majority of genes are undetectable in the vast majority of cells. However, these zero elements do not all represent true biological silencing; rather, they are mixed with technical dropouts caused by capture efficiency, amplification bias, or insufficient sequencing depth (a phenomenon where technical limitations lead to gene expression that is actually undetected, manifesting as zero values ​​but not true biological silencing).

[0066] To address this, a dual strategy is introduced: enhancing robustness on the noise injection side and reconstructing the masked true representation on the signal recovery side, thus balancing the dual requirements of anti-interference and fidelity.

[0067] Specifically, based on the generative assumptions of scVI, negative binomial distribution noise and library size perturbation are applied simultaneously to the preprocessed single-cell expression matrix, including the following methods: First, discrete noise consistent with single-cell statistical characteristics is injected into each expression value through a negative binomial distribution NB to simulate random fluctuations during PCR amplification and capture. Subsequently, a Gaussian perturbation is applied to the cell-level library size, so that the deep learning-based single-cell gene characterization method can see extreme cases of excessive or insufficient depth during the training phase, thereby reducing overfitting to sequencing depth differences.

[0068] In this embodiment, the applied Gaussian perturbation is N(1, 0.2).

[0069] The above processing aims to provide structurally consistent and high-quality input data for subsequent interpolation, embedding, and inference functions.

[0070] Step 3: Based on the processed matrix, iteratively estimate the partial correlation coefficients between genes using a conditional Gaussian graph model. For each zero value in the matrix, weighted imputation is performed based on the expression values ​​of the top-k related genes of its corresponding gene. After multiple iterations, a denser single-cell expression matrix is ​​obtained after imputation.

[0071] Specifically, GGM is used to estimate the partial correlation coefficient between genes. The partial correlation coefficient eliminates the indirect influence of third-party genes and better reflects the direct regulatory or co-expression relationships between genes, thus providing a purer reflection of the direct regulatory relationships between genes.

[0072] The calculation of the partial correlation coefficient includes: First, using gene expression data from multiple cells within the current batch as a statistical sample and genes as random variables, a covariance matrix of gene expression is constructed. It is calculated according to the following formula: ; in, Let be the gene expression vector corresponding to the nth cell. This is a vector representing the average expression level of each gene within the current batch. The elements in the covariance matrix that reflect the degree of joint change in the expression of gene i and gene j are used to characterize the degree of joint change of the two genes in the cell population.

[0073] After obtaining the covariance matrix, inverting it yields the precision matrix. It is calculated according to the following formula: ; Based on the aforementioned precision matrix, the partial correlation coefficient between any two genes is further calculated according to the following formula: ; In the formula, This represents the partial correlation coefficient between the i-th gene and the j-th gene.

[0074] For each gene expressed at zero (dropout site) in each cell, the k most relevant genes (expressed non-zero in the same cell) are identified based on partial correlation coefficients. The expression values ​​of these k genes are then weighted and averaged to estimate the missing value, using the partial correlation strength between the gene to be inserted and these k genes. This average is calculated as follows: ; In the formula, This represents the interpolated expression value of the j-th gene in the c-th cell. This represents the set of k genes that have a high partial correlation with the stated gene. This represents the partial correlation coefficient between corresponding genes.

[0075] The above imputation process can be iterated multiple times based on the convergence of the covariance estimation. By repeatedly updating the covariance, a stable covariance estimation result can be obtained until the estimated value is stable, thus obtaining the imputed single-cell expression matrix.

[0076] The following tables show the original matrix and the matrix after context-aware imputation. The imputed data fills in the missing values ​​in the original matrix, restores the missing values ​​in the cell expression matrix, and preserves the overall trend of the data. (See Tables 6 and 7)

[0077] Table 6. Example of a matrix before context-aware interpolation (partial)

[0078] Table 7. Examples of matrices after context-aware interpolation (partial).

[0079] like Figure 4 As shown, this illustrates the distribution structure of zero and non-zero values ​​in a local sub-region within a single-cell expression matrix. Figure 5 As shown, after interpolation, the same submatrix shows a clear expression gradient, columns that were originally completely black show continuous low values, and previously isolated bright spots expand into band patterns, maintaining discernible difference bands in the heatmap row direction.

[0080] Step 4: Fuse the imputed expression matrix with the gene function feature matrix to obtain the gene feature fusion matrix.

[0081] like Figure 6As shown, in traditional single-cell embedding methods, genes are usually regarded as independent feature dimensions, making it difficult to characterize functional processes involving multiple genes. To address this, we explicitly model the collaborative relationships between genes through a learnable modular structure, transforming the original "gene-gene" dimensional structure into a hierarchical representation framework of "gene-module-cell". This achieves biological interpretability at the core layer of the model. Specifically, we use element-level weighted fusion based on a broadcast mechanism to fuse the imputed single-cell expression matrix with the gene functional feature matrix to obtain a gene feature fusion matrix.

[0082] Step 5: Calculate the activation scores of each cell across multiple functional modules using 1×1 convolution and competitive activation mechanisms. Then, reweight and aggregate the gene feature fusion matrix using these module activation scores to generate a gene module activity matrix. This includes the following methods:

[0083] Step 5.1: Using a lightweight 1×1 convolutional layer, the functional feature vector of each gene is mapped to a higher-dimensional shared expression space to enhance its feature expression ability.

[0084] The 1×1 convolution operation is equivalent to performing a linear mapping on the functional feature vector of each gene, which is calculated according to the following formula: ; In the formula, Let W represent the functional feature vector of the g-th gene, and let W and b be learnable mapping parameters. This represents the mapped gene feature representation. The weights of this mapping can be shared across all genes, ensuring consistency in the transformation of the same functional theme across different genes.

[0085] Step 5.2: For each cell, calculate the matching degree between the expression patterns of all its mapped genes and the basis vectors of each module, and perform competitive activation using the softmax function to obtain the module activation score.

[0086] Several learnable module basis vectors are maintained to represent possible functional modules. A competitive activation mechanism (softmax) is used to calculate the activation scores of multiple modules for each cell, resulting in a sparse module representation. Each cell tends to activate only a few functional modules most relevant to its current state, rather than averaging the activation of all modules. This simulates the biological phenomenon of pathway-specific activation within the cell.

[0087] Specifically, let the matching score of the c-th single cell on the m-th functional module be... The corresponding module activation score is calculated according to the following formula: ; In the formula, This represents the activation intensity of the c-th cell for the m-th functional module. The result is a sparse module activation score vector, indicating which functional modules the cell specifically activates.

[0088] Step 5.3: The gene feature fusion matrix is ​​reweighted and aggregated using module activation scores to generate the gene module activity matrix. This module activity matrix is ​​a low-dimensional, interpretable, compressed representation of cells in the functional module space. Table 8 shows the gene module activity matrix, illustrating the cell activity at the modules.

[0089] Table 8 Gene Module Activity Matrix

[0090] like Figure 7 The heatmap shown is a concept map of the processed single-cell data after imputation. The gene module activity matrix is ​​a compressed representation of cells in the functional module space, displaying the cooperative expression patterns of encoded genes. Compared to directly processing the raw expression of tens of thousands of genes, this gene module activity matrix provides a low-dimensional, compact, and semantically clear structured representation. The figure visually illustrates the differences in activation intensity across multiple functional module dimensions for different cell types in heatmap form, verifying that the competitive activation mechanism proposed in this invention can effectively capture and decouple gene expression patterns with clear biological significance.

[0091] This step not only utilizes the expression values ​​of the current sample but also integrates the inherent biological properties of genes. It can simultaneously capture the dynamic expression level and static functional semantics of genes. Through the above process, the originally high-dimensional, sparse, and noisy gene expression space is successfully transformed into a low-dimensional, sparse, and biologically meaningful module activation space.

[0092] Step 6: Fuse the sample metadata, sequencing depth, cell-level feature vectors, and the gene module activity matrix to obtain a single-cell fusion matrix.

[0093] To effectively eliminate batch effects, identify differences between different technology platforms, and introduce biological background information of cells into the modeling process, a meta-information encoder is used to encode cell meta-information into continuous embedding vectors that can participate in deep representation learning together with the gene module activity matrix.

[0094] Specifically, the encoder maps meta-information such as technology platform, tissue origin, and cell state into 64-dimensional continuous vector representations. These embedded vectors are weighted and fused using a sigmoid-based learnable gating mechanism to form an effective meta-information representation vector specific to the current cell. This vector, along with the gene module activity matrix, serves as input to the model, providing auxiliary information for distinguishing between different batches or experimental conditions. The weights of the gating mechanism are generated using a sigmoid activation function, ensuring their values ​​remain stable within the range of [0,1], allowing for flexible adjustment of the contribution of different meta-information. This method can effectively suppress batch bias while preserving biological signals and adaptively distinguish between technological differences and real biological differences.

[0095] Furthermore, cell-level feature vectors can be integrated through linear projection, feature splicing, or weighted fusion to supplement information from the regulatory and genetic layers while maintaining the original data structure. For ease of explanation, this embodiment uses gene expression data as an example.

[0096] like Figure 8 As shown, to help understand the conceptual structure of meta-information modeling, this figure illustrates the layout of meta-information vectors from different batches or platforms in the dimensionality reduction space, where color is used to distinguish different categories and spatial distance reflects potential differences.

[0097] After obtaining the meta-information representation vector, it is fused with the gene module activity matrix (2048 dimensions), sequencing depth metrics (including total UMI count T and subsampling depth S, a total of 2 dimensions), and cell-level feature vectors in the feature dimension to construct a unified representation matrix, namely the single-cell fusion matrix. This single-cell fusion matrix serves as the input to the Transformer encoder, constituting the basic input features of the entire deep representation learning backbone network, Transformer.

[0098] In specific implementation, the metadata embedding vector, sequencing depth index, and cell-level feature vector are respectively mapped to corresponding vector representations. It also generates corresponding weights for each type of information through a learnable sigmoid gating network. The weights range from (0,1), and k represents different types of metadata channels to dynamically adjust the contribution of each metadata source. This gating mechanism can adaptively suppress metadata channels with high noise or low correlation, thereby achieving model adaptability and robustness to cross-platform data.

[0099] Its fusion representation Calculate according to the following formula: ; The fusion will then be represented The single-cell fusion matrix is ​​obtained by fusing the gene module activity matrix with the gene module activity matrix.

[0100] Step 7: Use a multi-layer stacked Transformer deep learning architecture to perform representation learning on the single-cell fusion matrix to generate a preliminary high-order latent representation of the cell; constrain and enhance this latent representation through a multi-task joint optimization mechanism, and output the final high-order latent representation of the cell as the cell representation.

[0101] Specifically, to fully capture the high-order dependencies across genes, modules, and samples in single-cell expression data, a large-scale Transformer network, namely MegaTransformer, is used as the unified deep representation learning backbone network. This network receives the single-cell fusion matrix and maps it to a fixed high-dimensional hidden space through linear projection layers, thereby constructing a high-order latent representation of cells that is compatible with multi-layer Transformer stacked structures.

[0102] Among them, the large-scale Transformer network supports large-scale training and supports gradient checkpointing and mixed precision training.

[0103] The gradient checkpointing method significantly reduces GPU memory usage by recalculating some intermediate features during the backpropagation phase, enabling the model to be trained with larger batch sizes. Mixed-precision training reduces computational costs and accelerates training while maintaining sufficient numerical stability. These engineering-level optimizations effectively support training this invention on single-cell datasets of millions or even larger scales, improving the model's scalability and engineering feasibility.

[0104] The Transformer encoder outputs a 1024-dimensional higher-order latent representation of cells. This representation structurally integrates imputation expression, functional module patterns, meta-information differential sequencing depth correction, and cell-level feature vector results, serving as a unified input feature for cell typing and differentiation trajectory inference. Furthermore, due to the Transformer's powerful feature alignment and context modeling capabilities, the generated higher-order latent representation of cells exhibits strong discriminative power, interpretability, and batch stability.

[0105] In this embodiment, a 24-layer stacked Transformer encoder structure is adopted, with each layer containing two main sub-modules: multi-head self-attention and feed-forward network.

[0106] The multi-head self-attention mechanism is used to model the global dependencies between different feature dimensions, and it is calculated according to the following formula: ; In the formula, These represent the query vector, key vector, and value vector obtained from the input feature mapping, respectively, and d represents the feature dimension.

[0107] Multilayer Transformers can effectively model long-range dependency structures across genes, enabling global interactions between expression, module activation, and meta-information, thereby obtaining more robust, structured, and generalizable cellular potential representations.

[0108] Specifically, to simultaneously improve the expression reconstruction capability, representation robustness, pseudo-temporal modeling capability, and clustering structure quality of single-cell data, a multi-task joint optimization strategy is adopted to construct a multi-task learning framework consisting of five task heads. This framework uses the high-dimensional latent representation of MegaTransformer as a unified input and is optimized through joint training of multiple objectives. Specifically, the loss functions of all tasks are weighted and summed, and then backpropagated together to update the model parameters. This multi-task collaborative mechanism forces the Transformer encoder to generate a high-order latent representation of cells that simultaneously satisfies numerical reconstruction accuracy, trajectory continuity, intra-cluster compactness, and batch interference resistance. This enables deep learning-based single-cell gene characterization systems to achieve a balance between biological consistency, structural interpretability, and technical robustness, obtaining results for single-cell population segmentation and differentiation trajectory inference.

[0109] The five task heads mentioned here include: representation reconstruction task, RDA prediction task, pseudo-temporal prediction task, contrastive learning task, and clustering adaptation task; The Expression Reconstruction task aims to reconstruct the imputed single-cell expression matrix based on the latent representation output by the Transformer, thereby preserving the authenticity of gene-level expression and local variation patterns. This task employs a fully connected network to generate the reconstructed expression values ​​and uses mean squared error loss (MSE Loss) as the optimization objective. This effectively constrains the model's ability to maintain a good fit to the original expression structure, thus enhancing the stability of downstream tasks.

[0110] The RDA (Relative Depth Adjustment) prediction task is used to predict the subsampling ratio under different sequencing depths, fit technical biases, and simulate the impact of different sequencing depths on expression levels. During training, the deep learning-based single-cell gene characterization system outputs the predicted RDA value through the Transformer and calculates the MSE loss between it and the actual subsampling ratio.

[0111] The pseudotime estimation task aims to enable the model to capture the potential changes in cell state or developmental dynamics and differentiation trajectories. It uses the latent representation as input to the pseudotime regression network and optimizes the regression head through regression loss (smooth L1 loss), thereby enabling the model to learn the continuous state of cells in the potential developmental process.

[0112] The contrastive learning task aims to build more robust cell representations, making similar cells appear closer together in the latent space, while dissimilar cells appear farther apart. This task uses the InfoNCE loss function as the optimization objective, working in conjunction with data augmentation strategies to effectively improve the model's consistency across different batches and sampling conditions.

[0113] The Adaptive Clustering Task aims to directly supervise the structure of the latent space, enabling the model to dynamically update cluster centers and improve cluster structure quality during training. This task uses the silhouette score as the evaluation criterion for cluster structure and combines it with an optimizable distance constraint based on cluster centers to construct a prototype loss. During training, the number of clusters or cluster centers can be dynamically adjusted based on the feedback from the silhouette score regarding cluster structure quality, allowing the task to adaptively optimize the clustering pattern of the latent space and thus improve the performance of subsequent analyses.

[0114] During multi-task joint training, the total loss function is obtained by weighted summation of the losses from each task, and it is calculated according to the following formula: ; in, This represents the loss function corresponding to the i-th task. These are the corresponding weight parameters.

[0115] Based on this total loss function We jointly optimize the shared parameters of the MegaTransformer backbone network and each task head. This joint optimization structure enables the model to improve cell clustering accuracy, temporal continuity, and alignment between different batches while maintaining expression reconstruction capabilities, significantly enhancing the model's generalization ability and biological interpretability.

[0116] Table 9 shows the changes in the loss terms for different tasks during training, reflecting the process of the model gradually improving its performance during optimization. As training progresses, the various losses gradually decrease, indicating that the model has a significant optimization effect on each task.

[0117] Table 9. Examples of multi-task loss values ​​(training log snippets)

[0118] like Figure 9 As shown, the overall process of this multi-task training structure is as follows: the fused input vector is encoded by a Transformer and then fed into multiple task heads, finally summing the losses to complete end-to-end training. This structure not only improves the model's generalization ability but also enhances its adaptability and stability in different downstream analysis tasks.

[0119] As shown in Figure 10, this invention demonstrates the expected linear correspondence between the predicted sequencing depth ratio and the actual input depth under ideal training conditions. This figure illustrates the linear correspondence between the predicted subsampling depth and the actual input depth under ideal training convergence conditions. It is intended to aid in understanding the working mechanism of this invention and is not a result obtained from actual experiments. It should not be used to limit the final performance of the model.

[0120] Step 8: Process the higher-order latent representation of the cells and output the cell population segmentation results and cell differentiation trajectories.

[0121] To further verify the applicability and stability of the constructed deep single-cell characterization in downstream analysis tasks, cluster analysis and cell differentiation trajectory inference were carried out using a high-dimensional latent representation based on Transformer output. This method does not rely on the original high-noise expression matrix, but directly utilizes the enhanced, imputation, module decoupling and deep encoding of the cell's high-order latent representation, thereby significantly reducing the impact of technical noise and batch bias.

[0122] include: Step 8.1: Perform cluster analysis on the higher-order latent representation of the cells, determine the optimal number of clusters through an adaptive strategy, and obtain the cell typing results; Cluster analysis is performed on the higher-order latent representation of cells. First, the output of the Transformer encoder is processed by a feature aggregation module to generate a latent cell embedding representation for downstream analysis. This representation integrates local and global expression patterns. Subsequently, the optimized clustering algorithm K-Means is used to perform clustering and calculate silhouette coefficients at different preset numbers of clusters, automatically searching for and determining the optimal number of clusters K. Based on this, the cell population segmentation results are determined, the cluster centers are updated, and finally, biologically significant cell populations are identified, completing cell type identification.

[0123] Step 8.2: Based on the higher-order latent representation of the cell, perform pseudo-temporal prediction through a regression network or a Transformer-based decoder to output the cell differentiation trajectory.

[0124] Based on the aforementioned higher-order latent representation of the cell, pseudo-temporal prediction is performed using a decoder based on a Transformer architecture. This step includes two modes: Single-cell mode: directly takes the latent representation of a single cell as input and outputs its pseudo-time scalar through a regression network; Trajectory mode: A group of cells is arranged into a sequence in an initial order, input into a Transformer-based decoder, and its attention mechanism is used to refine the pseudo-temporal order of the entire sequence, outputting a smooth continuous differentiation trajectory.

[0125] Through the above steps, the final output includes cell typing results and their differentiation trajectories during continuous differentiation, providing direct support for downstream biological interpretation.

[0126] Table 10 shows the results of cell typing through cluster analysis. Each cell ID corresponds to its cluster label and cell type. The inference of these cluster labels and cell types provides an important basis for subsequent biological analysis.

[0127] Table 10 Example of cell typing results

[0128] In terms of 2D visualization, such as Figure 11 As shown, UMAP is used to perform dimensionality reduction projection on the latent representation to show the structural relationships of cells in the latent space. It can usually be observed that different cell clusters are separated in space.

[0129] Once the cell potential representation is generated, pseudotime modeling (PseudotimeTransformer) can be used to infer the cell's differentiation state or potential developmental trend. It can take cell embedding sequences organized by similarity or time order as input: in trajectory mode, multi-head attention can capture continuous changes in the embedding sequence and generate continuous pseudotime trajectories; in single-cell mode, it outputs a single-cell differentiation pseudotime scalar.

[0130] Table 11 shows the pseudo-time values ​​in the cell differentiation process obtained through pseudo-time inference. These pseudo-time values ​​help to understand the specific location of cells during differentiation.

[0131] Table 11 Examples of Pseudo-Time Trajectory Results

[0132] To avoid ambiguity, this embodiment defaults to using the higher-order output of the Transformer as a substitute input for more complex differentiation trajectory inference. The pseudo-temporal scalar output of pseudo-temporal modeling is typically normalized to a fixed interval to represent the relative position of cells during potential development.

[0133] As shown in Figure 12, pseudo-temporal modeling can also be combined with density estimation methods to analyze local density changes in pseudo-temporal space. It uses curves to illustrate the potential continuous changes of cells in the pseudo-temporal dimension, illustrating the structure and logical relationships of pseudo-temporal modeling, which is independent of specific cell types, specific developmental pathways, or any experimental data.

[0134] This embodiment presents a deep learning-based single-cell gene characterization method that constructs a complete single-cell data modeling workflow encompassing data augmentation and noise suppression, functional module-level interpretable feature construction, deep embedding learning, and multi-task downstream modeling. The method emphasizes modularity, scalability, and interpretability in its overall design, establishing a complete link in single-cell data processing from front-end quality improvement to high-level feature expression and functional modeling, without being limited to a single fixed parameter or specific implementation. This methodology is widely applicable to multi-platform, multi-species, and other highly sparse structured omics data, thus possessing significant versatility and expansion potential.

[0135] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed, implements the single-cell gene characterization method based on deep learning described in any of the second aspects above.

[0136] Fourthly, embodiments of the present invention provide a storage device, including a storage medium and a processor, wherein the storage medium stores a computer program, and when the program is executed by the processor, it implements the single-cell gene characterization method based on deep learning described in any of the second aspects above.

[0137] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention 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.

[0138] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

[0139] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A deep learning-based single-cell gene characterization system, characterized in that, include: The module includes a gene information input module, a data processing module, a context-aware interpolation module, a gene feature fusion module, a fusion module, and a deep representation learning and joint optimization module. The gene information input module is responsible for integrating the input data sources; The inputs include: the original single-cell expression matrix, the gene function feature matrix, sequencing depth information, sample metadata, and cell-level feature vectors; The data processing module is used to preprocess the original single-cell expression matrix and perform data augmentation operations on the preprocessed matrix. The context-aware interpolation module uses a conditional Gaussian graph model (GGM) to estimate the partial correlation coefficients between genes, and performs weighted interpolation on the processed matrix based on the top k related genes to obtain the interpolated single-cell expression matrix. The gene feature fusion module fuses the gene functional feature matrix with the imputed single-cell expression matrix, and then activates it through 1×1 convolution and competitive activation mechanism to generate a gene module activity matrix. The fusion module fuses sample metadata, sequencing depth, and cell-level feature vectors with the gene module activity matrix to obtain a single-cell fusion matrix. The deep representation learning and joint optimization module uses a multi-layer stacked Transformer deep learning architecture to perform representation learning on the single-cell fusion matrix: a preliminary high-order latent representation of the cell is generated through a multi-head self-attention mechanism; this latent representation is constrained and enhanced through a multi-task joint optimization mechanism to output the final high-order latent representation of the cell as the cell representation.

2. The deep learning-based single-cell gene characterization system according to claim 1, characterized in that, Also includes: The cell typing and differentiation trajectory inference module takes the higher-order latent representation of cells as input and uses the multi-head attention mechanism of the Transformer deep learning architecture to perform pseudo-temporal trajectory inference on cells, capture the continuous process and state changes of cell differentiation, and finally output the cell population division results and cell differentiation trajectory. The cell typing and differentiation trajectory inference module includes: a cell typing submodule and a differentiation trajectory inference submodule; The cell typing submodule inputs the higher-order latent representation of cells into the clustering adaptive layer. This layer calculates the similarity between cells through an optimized clustering algorithm and adopts an adaptive strategy based on internal evaluation indicators to automatically search and determine the optimal number of clusters, thereby determining the cell population division result and updating the cluster centers to divide biologically significant cell populations and complete cell type identification. The differentiation trajectory inference submodule predicts cell differentiation trajectories based on the Transformer deep learning architecture. This submodule includes two modes: Single-cell mode: directly takes the latent representation of a single cell as input and outputs its pseudo-time scalar through a regression network; Trajectory mode: A group of cells is arranged into a sequence in an initial order, input into a Transformer-based decoder, and its attention mechanism is used to refine the pseudo-temporal order of the entire sequence, outputting a smooth continuous differentiation trajectory.

3. A single-cell gene characterization method based on a deep learning-based single-cell gene characterization system as described in any one of claims 1-2, characterized in that, Includes the following steps: Step 1: Obtain the input raw single-cell expression matrix, gene function feature matrix, sequencing depth information, sample metadata, and cell-level feature vector; Step 2: Perform data augmentation operations after preprocessing the original single-cell expression matrix; Step 3: Based on the processed matrix, use a conditional Gaussian graph model to iteratively estimate the partial correlation coefficients between genes; for each zero value in the matrix, perform weighted imputation based on the expression values ​​of the top-k related genes of the corresponding gene to obtain the imputed single-cell expression matrix; Step 4: Fuse the imputed expression matrix with the gene function feature matrix to obtain the gene feature fusion matrix; Step 5: Calculate the activation scores of each cell on multiple functional modules using 1×1 convolution and competitive activation mechanisms, and reweight and aggregate the gene feature fusion matrix using the module activation scores to generate a gene module activity matrix. Step 6: Fuse the sample metadata, sequencing depth, cell-level feature vectors, and the gene module activity matrix to obtain a single-cell fusion matrix; Step 7: Use a multi-layer stacked Transformer deep learning architecture to perform representation learning on the single-cell fusion matrix to generate a preliminary high-order latent representation of the cell; constrain and enhance this latent representation through a multi-task joint optimization mechanism, and output the final high-order latent representation of the cell as the cell representation.

4. The single-cell gene characterization method based on deep learning according to claim 3, characterized in that, The original single-cell expression matrix, gene function feature matrix, sequencing depth information, sample metadata, and cell-level feature vector are specifically: The original single-cell expression matrix has a dimension of N×G1, where N represents the number of cells and G1 represents the number of genes; each element in the matrix is ​​a unique molecular identifier (UMI) count or original expression count commonly used in single-cell sequencing. The gene function feature matrix refers to the fixed feature representation corresponding to each gene in the gene function matrix, with a dimension of G2×F, where F represents the predefined gene feature dimension and G2 represents the predefined number of gene features; this matrix is ​​a static prior knowledge base obtained through pre-training. The sequencing depth information includes the total UMI number T for each cell and the standard depth S estimated by subsampling, which is used for depth normalization and perturbation enhancement during training. The sample metadata includes the technology platform, tissue source, experimental batch, and donor information; The cell-level feature vectors are generated from omics data, including chromatin accessibility features (ATAC-seq), DNA methylation features, and genetic variation features.

5. The single-cell gene characterization method based on deep learning according to claim 4, characterized in that, In step 2, specifically: basic quality control and library size normalization are performed on the original single-cell expression matrix to obtain the preprocessed expression matrix; then, library size perturbation and negative binomial distribution noise are applied to the preprocessed expression matrix to generate the enhanced expression matrix.

6. The single-cell gene characterization method based on deep learning according to claim 5, characterized in that, In step 3, the partial correlation coefficient matrix between genes is estimated using GGM. For each gene expressed as zero in a cell, the k genes most related to it are found based on the partial correlation coefficient. The expression values ​​of the k genes are weighted by the partial correlation strength between the gene to be imputed and these k genes to estimate the missing value. This process is repeated iteratively until the estimated value is stable, thus obtaining the imputed single-cell expression matrix.

7. The single-cell gene characterization method based on deep learning according to claim 6, characterized in that, Step 5 specifically includes the following methods: Step 5.1: Using lightweight 1×1 convolutional layers, the functional feature vector of each gene is mapped to a higher-dimensional shared expression space to enhance its feature expression ability; Step 5.2: For each cell, calculate the matching degree between the expression patterns of all its mapped genes and the basis vectors of each module, and perform competitive activation using the softmax function to obtain the module activation score; Step 5.3: Reweight and aggregate the gene feature fusion matrix using module activation scores to finally generate the gene module activity matrix.

8. The single-cell gene characterization method based on deep learning according to claim 7, characterized in that, The multi-task joint optimization mechanism in step 6 includes five task heads, specifically: representation reconstruction task, RDA prediction task, pseudo-temporal prediction task, contrastive learning task, and clustering adaptive task. The expression reconstruction task involves the decoder attempting to reconstruct the imputed expression matrix from the latent representation, using mean squared error loss to ensure that the representation retains basic gene expression information. The RDA prediction task is to predict the subsampling depth ratio for each cell, using mean squared error loss, so that the model can understand and correct for the impact of sequencing depth. The pseudo-time prediction task includes a regression head predicting pseudo-time values ​​for cells, using smoothed L1 loss to guide the model to capture continuous changes in cell state. The contrastive learning task enhances the discriminativeness and cross-batch consistency of representations by zooming in on different enhanced views of the same cell or representations of similar cells and zooming out on representations of different cells using InfoNCE loss. The clustering adaptive task involves simultaneously performing online clustering in the latent representation space, optimizing the loss function based on the silhouette coefficient, and promoting the formation of a cluster structure with clear boundaries in the representation space.

9. The single-cell gene characterization method based on deep learning according to claim 8, characterized in that, It also includes step 8: processing the higher-order latent representation of the cells and outputting the cell population segmentation results and cell differentiation trajectories; Step 8.1: Perform cluster analysis on the higher-order latent representation of the cells, determine the optimal number of clusters through an adaptive strategy, and obtain the cell typing results; Step 8.2: Based on the higher-order latent representation of the cell, perform pseudo-temporal prediction using a Transformer-based decoder to output the cell differentiation trajectory.

Images