An automatic analysis method and system for single-cell RNA sequencing data by fusing adaptive quality control optimization and cell intelligent annotation

Abstract

The application discloses a single-cell RNA sequencing data automatic analysis method and system fusing adaptive quality control optimization and cell intelligent annotation. The method receives single-cell RNA sequencing data and constructs a cell-gene expression matrix, establishes a candidate quality control parameter space containing multiple quality control indexes, adopts an adaptive search strategy to generate a candidate parameter combination and performs filtering, performs preset downstream analysis on the filtering result and calculates a comprehensive quality evaluation value, generates quality control post-expression data after determining a target quality control parameter, matches the quality control post-data with a target gene set of a cell intelligent annotation model through gene feature alignment to obtain aligned input data, inputs the cell intelligent annotation model to obtain a cell type prediction result, and generates a cell composition statistical result and an analysis result file.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of bioinformatics and artificial intelligence, specifically to an automated method and system for analyzing single-cell RNA sequencing data that integrates adaptive quality control optimization and intelligent cell annotation. This invention falls within the technical categories of single-cell transcriptome data analysis, machine learning model compression and knowledge distillation, and the construction of automated bioinformatics workflows. Background Technology

[0002] Single-cell RNA sequencing technology can characterize gene expression differences at single-cell resolution, and has become an important technical means to analyze cellular heterogeneity and complex biological processes. With the rapid increase in sequencing throughput, a single experiment can generate expression data from hundreds of thousands to millions of cells. A typical single-cell data analysis workflow includes steps such as quality control, normalization, feature selection, dimensionality reduction, clustering, and cell type annotation. These steps usually require professional bioinformatics personnel to perform step by step, resulting in a long analysis cycle and results that are greatly affected by the operator's experience.

[0003] In existing single-cell data analysis workflows, quality control typically relies on a set of empirical parameters, such as the minimum number of genes, the maximum number of genes, and the proportion of mitochondrial genes. These parameters often need to be manually set based on the specific characteristics of the dataset. Researchers usually judge the rationality of the filtering parameters by repeatedly observing the dimensionality reduction visualization results and conduct multiple rounds of experiments and corrections. The main limitations of existing technologies are: First, parameter setting depends on personal experience, and different researchers may provide different combinations of thresholds, lacking a unified optimization standard; second, the parameter selection process lacks a clear optimization modeling description, making the quality control process lack reproducibility and traceability; third, existing automatic filtering methods, such as those based on median absolute deviation, only focus on statistical anomaly detection and do not systematically link the selection of quality control parameters with the quality of downstream analysis, leading to a potential disconnect between the filtering strategy and the downstream analysis objectives.

[0004] Cell type annotation is a crucial step in single-cell analysis, bridging the gap between previous and subsequent steps. In recent years, foundational models pre-trained on large-scale single-cell data have achieved excellent performance in cell type classification tasks. However, these foundational models have a large number of parameters, require dedicated hardware for inference, suffer from high inference latency, are unsuitable for real-time analysis of large-scale cells, and are complex to deploy. Knowledge distillation techniques can transfer knowledge from large models to smaller models, but in the single-cell domain, existing distillation methods often employ single objectives, such as feature alignment or logical alignment only, failing to fully utilize the multi-layered knowledge of teacher models and lacking co-design with quality control processes.

[0005] More importantly, existing methods treat quality control and cell annotation as two independent steps, ignoring their cascading effects. The choice of quality control parameters alters the retained cell set, thus affecting the training data and final annotation accuracy of the annotation model. Currently, there is a lack of automated analysis systems that organically integrate quality control optimization with intelligent annotation. Summary of the Invention

[0006] The technical problems to be solved by this invention include: First, the selection of quality control parameters relies on human experience and lacks a systematic automated optimization framework; second, the quality control evaluation criteria are singular and do not fully utilize downstream analysis information such as dimensionality reduction and embedding structures; third, single-cell basic models have a large number of parameters, high inference costs, and are difficult to deploy; fourth, existing knowledge distillation methods have a single objective and do not fully utilize the potential of multi-task joint optimization; fifth, quality control and cell annotation are independent of each other and lack connection and synergy; sixth, there is a lack of an end-to-end automated analysis system that integrates quality control optimization and intelligent annotation.

[0007] This invention provides an automated analysis method for single-cell RNA sequencing data that integrates adaptive quality control optimization and intelligent cell annotation. The method receives single-cell RNA sequencing data, parses the data format, and constructs a standardized cell-gene expression matrix; establishes a candidate quality control parameter space based on multiple quality control indicators; generates candidate parameter combinations using an adaptive search strategy and filters the expression matrix; performs a preset downstream analysis process on the filtered expression matrix and calculates the data structure evaluation results; calculates a comprehensive quality evaluation value based on cell retention characteristics, expression distribution characteristics, and data structure evaluation results, and determines the target quality control parameter when a termination condition is met; generates quality-controlled expression data using the target quality control parameter; aligns the quality-controlled expression data with the target gene set of a cell intelligent annotation model using gene feature alignment; and inputs the aligned data into the cell intelligent annotation model to obtain cell type prediction results and cell composition statistical results. This invention also provides an automated analysis system, computer equipment, and computer-readable storage medium for implementing the above method.

[0008] The single-cell RNA sequencing data received in step S1 can be in comma-separated CSV format, h5ad format (AnnData format), HDF5 format, or 10XGenomics standard format. After parsing the data, a standardized cell-gene expression matrix X is constructed. This matrix is ​​an N-row, G-column two-dimensional matrix, where N is the number of cells and G is the number of genes. The i-th row of the matrix corresponds to the i-th cell, the j-th column corresponds to the j-th gene, and the matrix element Xi represents the expression count or normalized expression value of the j-th gene in the i-th cell. For sparse expression matrices, compressed sparse row (CSR) or compressed sparse column (CSC) format is used for storage to reduce memory usage. Simultaneously with matrix construction, cell barcodes are extracted as row indices, and gene names or gene identifiers are extracted as column indices.

[0009] For each cell, the following basic quality control statistics are calculated: gene detection count g_i, which is the number of elements with expression values ​​greater than zero in the matrix row corresponding to that cell; total count c_i, which is the sum of all elements in the matrix row corresponding to that cell; and mitochondrial gene expression percentage m_i, which is the sum of the elements in the column corresponding to mitochondrial genes in the matrix row corresponding to that cell, divided by the total value, and then multiplied by 100%. Mitochondrial genes are identified by the prefix of their gene names; for example, genes prefixed with MT- are mitochondrial genes.

[0010] Step S2 defines a candidate quality control parameter space. This parameter space contains five dimensions of threshold parameters, forming a candidate quality control parameter vector θ = (g_min, g_max, c_min, c_max, m_max), where: g_min is the minimum gene count threshold, used to filter low-quality cells or empty droplets with too few genes; g_max is the maximum gene count threshold, used to filter potentially two-cell or multi-cell droplets with abnormally high gene counts; c_min is the minimum total count threshold, used to filter poor-quality cells with excessively low total expression levels; c_max is the maximum total count threshold, used to filter abnormal cells with abnormally high total expression levels; and m_max is the maximum mitochondrial gene proportion threshold, used to filter damaged or apoptotic cells with excessively high mitochondrial gene proportions. Candidate values ​​for each dimension constitute a discrete candidate set. Invalid combinations of candidate parameters, such as combinations where g_min is greater than or equal to g_max or c_min is greater than or equal to c_max, are automatically excluded during the search.

[0011] In step S3, an adaptive search strategy is employed to search for candidate quality control parameter combinations in the parameter space. The search strategy can be at least one of grid search, greedy search, and simulated annealing. The grid search strategy traverses all valid candidate parameter combinations in the parameter space. For each combination, quality control filtering and a fixed downstream analysis process are executed sequentially, and a comprehensive score is calculated. Finally, the parameter combination with the highest score is selected. The greedy search strategy starts from the initial parameters and searches for candidate parameters with higher scores in the neighborhood of the current parameter. If a candidate parameter exists, the search continues; otherwise, it terminates.

[0012] Simulated annealing is the preferred search strategy in this invention. Starting from the initial parameter combination θ... Start from 0, set the initial temperature T_0 and the decay coefficient α. In each iteration, perform the following operations: randomly select a dimension from the parameter space, and in that dimension... Randomly select a candidate value different from the current value from the selected values ​​to form the neighborhood parameter combination θ. new. Regarding θ new execution quality control Filtration and downstream analysis were performed to calculate the overall quality evaluation value Q(θ). (new). The Metropolis criterion is used for acceptance judgment: if Q(θ) new) is greater than Q(θ) If the current is true, then accept it directly; otherwise, accept it with probability P=exp[(Q(θ)]. new)-Q(θ [current)) / T] Accept. After acceptance, update the current parameter combination and the current score. If the current score is higher than the historical best score, update the historical best parameter combination. After each iteration, the temperature decays exponentially: T (k+1)=α·T_k, where α is the attenuation coefficient, and its value ranges from 0 to 1. The search terminates when T drops below the minimum temperature T_min or reaches the preset maximum number of iterations, and the historical lowest value is output. The optimal parameter combination serves as the target quality control parameter θ. best.

[0013] The comprehensive quality evaluation value mentioned in step S5 is composed of a weighted combination of numerical structure score and embedded geometric structure score: Q(θ)=λ 1·S_num(θ)+λ 2. S_vis(θ). The numerical structure score S_num(θ) integrates the following clustering effectiveness indicators: Silhouette Score, Calinski-Harabasz Index (CH), Davies-Bouldin Index (DB), as well as small cluster penalty and cell retention rate reward terms. Each indicator is normalized before integration, mapping indicators of different dimensions to a unified zero-to-one interval. For indicators such as the DB index, where smaller values ​​indicate better quality, the direction is reversed by taking its complement. The embedding geometry score S_vis(θ) is calculated based on the geometric features of the low-dimensional embedding results, including at least two of the following: Cluster convex hull overlap, Cluster shape stretching, Inter-cluster minimum distance separation, and Outlier ratio.

[0014] When the termination condition is met, the historical optimal parameter combination is output as the target quality control parameter θ. best. In step S6 Using θThe `best` method filters the raw expression data: for each cell `i`, it retains cells that simultaneously satisfy `g_min≤g_i≤g_max`, `c_min≤c_i≤c_max`, and `m_i≤m_max`. After filtering, quality-controlled expression data is obtained, with a cell retention rate `r = N' / N`.

[0015] The gene feature alignment performed in step S7 is a key technical link between quality control and intelligent cell annotation. Specifically, it includes: extracting a list of gene names G_data from the column index of the expression data after quality control; standardizing the gene names, including removing leading and trailing spaces and unifying capitalization; for repeated gene names, taking the average expression value of the corresponding column of each repeated gene as the representative expression value of that gene; reading the target gene set G_model of the intelligent cell annotation model, which is a list of highly variable genes used during model training, with a length of G_model_len; performing string matching between G_data and G_model to determine the intersection gene set; for each gene in G_model, if it is in the intersection gene set, extracting the corresponding column from the data matrix; otherwise, filling the column with all zeros; arranging the genes in G_model sequentially, constructing an aligned matrix X_aligned with a dimension of N' rows multiplied by G_model_len columns.

[0016] The cell-level intelligent annotation model described in step S8 is a student model trained from the teacher model through knowledge distillation. The teacher model is a base model pre-trained on large-scale single-cell data. The student model is a lightweight neural network model. The architecture of the student model adopts a gene index embedding mechanism: for each gene g in the input matrix X_aligned, its scalar expression value x_g is mapped to a d-dimensional embedding space through a learnable projection vector w_g, and added to the learnable gene index embedding vector e_g^idx of that gene to form a gene-level token representation: h_g^(0)=x_g·w_g+e_g^idx. After the gene-level token sequence is context-encoded by a multi-layer Transformer encoder, the cell-level representation is obtained by pooling all gene tokens. The cell-level representation outputs the class probability distribution through a classifier.

[0017] During knowledge distillation training, the total loss function L_total is a weighted combination of the following six losses: L_total = L_cls + λ feat·L_feat+λ logit·L_logit+λ recon·L_recon+λmask·L_mask+λ_contrast·L_contrast. Where L_cls is the supervised classification cross-entropy loss; L_feat is the feature-level distillation loss, measured by mean squared error; L_logit is the logistic-level distillation loss, measured by KL divergence at temperature τ softening; L_recon is the reconstruction loss; L_mask is the masked gene prediction loss, which randomly masks approximately 15% of the input genes and predicts their original expression values; and L_contrast is the supervised contrastive learning loss.

[0018] During the inference phase, the gene feature-aligned input matrix X_aligned is fed into the trained student model row by row. The model outputs a probability distribution vector for each cell. The index corresponding to the maximum value in the probability distribution vector is taken and mapped to the cell type name through a label encoder, serving as the predicted type label for that cell. This maximum value is used as the prediction confidence: p_i = max_c P(c|x_i). The number of cells corresponding to each predicted cell type is counted. If the number of cells in the k-th type is n_k and the total number of cells is N', then the proportion of cells of that type is: r_k = n_k / N'.

[0019] The analysis result file generated in step S8 includes at least two of the following formats: HTML interactive report, CSV structured data table, and JSON result data. The analysis result file also includes a post-quality control data file (i.e., a target cell expression matrix in h5ad format) and a runtime log file.

[0020] This invention also provides an automated analysis system for single-cell RNA sequencing data that integrates adaptive quality control optimization and intelligent cell annotation. The system includes: a data input module, a quality control parameter optimization module, a quality control execution module, a gene feature alignment module, an intelligent cell annotation module, and a result output module. In one embodiment, the quality control parameter optimization module includes a candidate parameter generation unit, a data filtering unit, a downstream analysis unit, a comprehensive evaluation unit, and a parameter update unit. The gene feature alignment module includes a gene name standardization unit, a gene matching unit, a missing gene handling unit, and a gene order rearrangement unit. The intelligent cell annotation module includes a model input construction unit, a model inference unit, a label mapping unit, a confidence output unit, and a cell composition statistics unit.

[0021] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method. The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0022] Compared with existing technologies, this invention has the following advantages: First, it elevates the selection of quality control parameters from relying on manual experience to a systematic optimization model, automatically determining target quality control parameters through an adaptive search strategy, reducing the workload of repeated manual parameter tuning; Second, it combines numerical structure scoring and embedding geometric structure scoring to comprehensively evaluate the quality of quality control parameters from two dimensions: clustering effectiveness and embedding geometric features, making the evaluation results more comprehensive and objective than a single statistical indicator; Third, it achieves model compression through multi-task knowledge distillation, lowering the hardware deployment threshold; Fourth, it solves the problem of gene set mismatch between data and models through gene feature alignment technology, achieving automated connection between quality control and cell annotation; Fifth, the end-to-end automated analysis process is completed with one click from data input to result output, and multi-format analysis result files improve the readability and shareability of the analysis results. Attached Figure Description Figure 1 This is the overall architecture diagram of the automatic analysis system for single-cell RNA sequencing data of the present invention. Figure 2 This is the overall flowchart of the automatic analysis method for single-cell RNA sequencing data of the present invention. Figure 3 This is a flowchart of the adaptive quality control parameter optimization process of the present invention. Figure 4 This is a flowchart of the gene feature alignment and intelligent cell annotation of the present invention. Figure 5 This is a flowchart of the analysis results summary and output of this invention.

Claims

1. An automated method and system for analyzing single-cell RNA sequencing data by integrating adaptive quality control optimization and intelligent cell annotation. An automated analysis method for single-cell RNA sequencing data that integrates adaptive quality control optimization and intelligent cell annotation is characterized by, Includes the following steps: Step S1: Receive single-cell RNA sequencing data, parse the data format, calculate basic quality control statistics, and construct a standardized cell-gene expression matrix; Step S2: Establish a candidate quality control parameter space based on multiple quality control indicators. The parameter space includes a minimum gene number threshold, a maximum gene number threshold, a minimum total count threshold, a maximum total count threshold, and a maximum mitochondrial gene ratio threshold. Step S3: An adaptive search strategy is used to generate candidate parameter combinations in the parameter space, and the expression matrix is ​​filtered according to the candidate parameter combinations to obtain the filtered expression matrix. Step S4: Perform a preset downstream analysis process on the filtered expression matrix to obtain low-dimensional embedding results and clustering results, and calculate data structure evaluation indicators; Step S5: Calculate the comprehensive quality evaluation value based on cell retention characteristics, expression distribution characteristics, and the data structure evaluation index; update the candidate parameter combination based on the comprehensive quality evaluation value; and determine the target quality control parameter when the termination condition is met. Step S6: Filter the original expression data using the target quality control parameters to generate quality-controlled expression data; Step S7: Align the quality-controlled expression data with the target gene set of the cell intelligent annotation model using gene features to obtain aligned input data; Step S8: Input the aligned input data into the cell intelligent annotation model to obtain cell type prediction results and generate cell composition statistical results and analysis result files.

2. The method according to claim 1, characterized in that, The comprehensive quality evaluation value in step S5 is composed of a weighted combination of numerical structure score and embedded geometric structure score. The numerical structure score is calculated based on the clustering effectiveness index, and the embedded geometric structure score is calculated based on the geometric features of the low-dimensional embedding results.

3. The method according to claim 2, characterized in that, The embedded geometry score is based on at least two of the following geometric features of the low-dimensional embedding: cluster convex hull overlap rate, cluster shape stretching, minimum inter-cluster distance, and outlier ratio.

4. The method according to claim 1, characterized in that, The cell intelligent annotation model mentioned in step S8 is a student model trained from the teacher model through knowledge distillation; the loss function of knowledge distillation includes classification cross-entropy loss, feature-level distillation loss, logistic-level distillation loss, reconstruction loss, mask gene prediction loss and contrastive learning loss, and the total loss is a weighted combination of the above losses.

5. The method according to claim 6, characterized in that, The student model employs a gene index embedding mechanism: the scalar expression value of each gene is mapped to the embedding space through a learnable projection vector, and then added to the learnable gene index embedding vector of that gene to form a gene-level token representation; the cell-level representation is obtained by pooling all gene tokens.

6. The method according to claim 1, characterized in that, The termination conditions in step S5 include: the comprehensive quality evaluation value no longer increases in a preset number of consecutive iterations, the preset maximum number of iterations is reached, or the simulated annealing temperature drops to a preset minimum temperature.

7. An automated analysis system for single-cell RNA sequencing data that integrates adaptive quality control optimization and intelligent cell annotation, characterized in that, include: The data input module is used to parse single-cell expression data in various formats, construct a standardized cell-gene expression matrix, and calculate basic quality control statistics. The quality control parameter optimization module is used to establish a candidate quality control parameter space, generate candidate parameter combinations using an adaptive search strategy, filter the expression matrix based on the candidate parameter combinations, and perform a preset downstream analysis process and calculate the comprehensive quality evaluation value on the filtering results to determine the target quality control parameters. The quality control execution module is used to filter the original expression data using the target quality control parameters to generate quality-controlled expression data. The gene feature alignment module is used to align the quality-controlled expression data with the target gene set of the cell intelligent annotation model to obtain aligned input data. The cell intelligent annotation module is used to input the aligned input data into the cell intelligent annotation model to obtain cell type prediction results. The results output module is used to generate statistical results and analysis results files for cell composition.

8. The system according to claim 12, characterized in that, The quality control parameter optimization module includes a candidate parameter generation unit, a data filtering unit, a downstream analysis unit, a comprehensive evaluation unit, and a parameter update unit; the gene feature alignment module includes a gene name standardization unit, a gene matching unit, a missing gene processing unit, and a gene order rearrangement unit; the cell intelligent annotation module includes a model input construction unit, a model inference unit, a label mapping unit, a confidence output unit, and a cell composition statistics unit.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 11.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 11.

Images