High-resolution cell cycle stage prediction method based on single-cell transcriptome sequencing data
By constructing an integrated learning prediction model for single-cell transcriptome sequencing data, the problem of low-resolution cell cycle segmentation in traditional methods is solved, achieving high-resolution cell cycle prediction and providing more detailed cell cycle information and drug evaluation tools.
Patent Information
- Application Number
- CN202410028763.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-09
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-01-09
AI Technical Summary
Existing technologies struggle to achieve high-resolution cell cycle segmentation. Traditional methods can only provide low-resolution cell cycle information and lack methods to predict the cell cycle stage of a single or multiple cells using their characteristic gene expression profiles.
An ensemble learning prediction model based on single-cell transcriptome sequencing data was constructed. Through pseudo-time series construction, autoregressive testing, Fourier analysis, low-pass filtering, and TICC clustering, cell cycle characteristic genes were identified. Algorithms such as XGBoost, neural networks, SVM, and random forest were used for ensemble learning to achieve high-resolution cell cycle prediction.
It achieves higher resolution cell cycle segmentation, which can more accurately reveal the changes in cell state at different stages, providing a more comprehensive perspective for in-depth cell cycle research, and can more accurately assess the impact of drugs on the cell cycle, making it suitable for drug screening and cancer research.
Smart Images

Figure CN117809748B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cell cycle segmentation and prediction technology, and in particular relates to a cell cycle stage prediction method based on single-cell transcriptome sequencing data. Background Technology
[0002] The cell cycle is a highly regulated process that plays a crucial role in promoting cell growth, genome replication, and cell division. Traditionally, cell mitosis is divided into four (G1, S, G2, and M) or five (with the addition of G0) phases, and cell cycle and gene transcription regulation are studied based on this division. While this approach meets the needs of basic research to some extent, its coarse division makes it difficult to capture the more subtle dynamic changes within the cell cycle. A deeper understanding of the complex dynamics of the cell cycle and the achievement of a finer division of the cell cycle are essential for comprehending its progression.
[0003] Many traditional techniques, including cell synchronization, FUCCI experiments, fluorescence-activated cell sorting, and cell imaging, can only provide low-resolution information about the cell cycle process because they can only obtain cellular data at a limited number of time points during the cell cycle. With the development of single-cell RNA sequencing technology (scRNA-seq), researchers can study the dynamics of gene expression during the cell cycle at single-cell resolution, where each individual cell can be understood as a snapshot of the cell cycle process, corresponding to a single time point. Much work has been dedicated to reconstructing the transcriptional dynamics of genes during the cell cycle by sequencing cell snapshots from single-cell transcriptomes.
[0004] While these works have made significant progress in obtaining pseudo-chronological sequences of cells within the cell cycle, the method of dividing the cell cycle into four phases—G1, S, G2, and M—to represent different stages of the cell cycle provides only very limited resolution for understanding cell cycle progression. The G1 and G2 phases are merely defined as two interphases between the S and M phases, a coarse division that hinders researchers from gaining a deeper understanding of cell cycle progression. Cell cycle segmentation should reflect more details of cell cycle dynamics, thus a higher-resolution segmentation of the cell cycle is urgently needed. Furthermore, existing cell stage predictions often require in-depth dimensionality reduction and clustering analysis of a group of cells; currently, no method has been proposed that can predict the cell cycle stage of a cell using only cell cycle characteristic gene expression profiles of a single or multiple cells. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a cell cycle prediction method based on single-cell transcriptome sequencing data. The method of this invention constructs an integrated learning prediction model at different stage resolutions. This model can obtain predictions of the cell cycle sub-stages of single or multiple cells at different stages and even at higher stages by inputting only the cell cycle characteristic gene expression profiles of single or multiple cells.
[0006] The cell cycle prediction method based on single-cell transcriptome sequencing data in this invention is as follows:
[0007] 1. Collect and obtain several single-cell transcriptome datasets from cultured cell lines from the GEO database. Preprocess each single-cell transcriptome dataset and then construct pseudo-time series to obtain the pseudo-time series of each cell in the cell cycle. Based on the constructed pseudo-time series, rearrange the cells to obtain the temporal expression profiles of different genes. Obtain the cell cycle genes in each dataset through autoregression test. Select genes that appear in 2 / 3 or more of the total number of datasets as cell cycle feature genes.
[0008] The single-cell transcriptome dataset consists of multiple datasets composed of transcriptome data from different cell lines and cells subjected to different experimental treatments, and each dataset is required to contain more than 250 cells. The preprocessing of the single-cell transcriptome dataset involves removing cells with a sequencing depth of less than 10,000 and removing genes that are not expressed in 70% or more of the cells.
[0009] The pseudo-time series construction is based on existing known cell cycle genes from Seurat. Principal component analysis is performed on the expression profile of known cell cycle genes in each dataset. The scores of each cell on the first and second principal components represent the cell's progress in the cell cycle. The polar angle of the scatter points in the two-dimensional space corresponding to different cells is used as the measure of the pseudo-time series of the cell cycle.
[0010] 2. Based on the obtained cell cycle characteristic gene set, the cells in different datasets are reconstructed using pseudo-temporal sequences. The cells are rearranged based on the new pseudo-temporal sequences to obtain the temporal expression profiles of different genes. Fourier analysis and low-pass filtering are used to smooth and denoise the temporal expression profiles of cell cycle characteristic genes in each dataset to obtain the temporal transcription signals of cell cycle characteristic genes in each dataset.
[0011] The pseudo-time series construction is based on the cell cycle characteristic genes obtained in step 1, and the principal component analysis is the same as in step 1;
[0012] 3. The TICC clustering algorithm was used to divide the temporal transcription signals of cell cycle characteristic genes in each dataset into different number of phases to obtain the sub-phases in the cell cycle under different number of phases.
[0013] The number of different stages ranges from 3 to 20, each corresponding to a stage framework with different resolutions. These different resolution stage frameworks break down the cell cycle process to varying degrees. For example, when the number of stages is 9, stage C9.1 corresponds to the transition from the classical M stage to the G1 stage; stages C9.2 and C9.3 correspond to the early and late stages of the classical G1 stage; stage C9.4 corresponds to the transition from G1 to S stage; stages C9.5 and C9.6 correspond to the early and late stages of the classical S stage; stage C9.7 corresponds to the transition from S to G2 stage; and stages C9.8 and C9.9 correspond to the classical G2 stage and the M stage, respectively.
[0014] 4. Merge the transcription data of cell cycle characteristic genes from all datasets, standardize them, and divide them into training and testing sets. Use the expression data of cell cycle characteristic genes as features and the stage data of TICC clustering as labels. Use XGBoost (Limited Gradient Boosting), neural networks, SVM (Support Vector Machine), and random forest as base models, and use random forest as the meta-model to train and validate the ensemble learning model, and obtain the trained ensemble learning model under different number of stages.
[0015] 5. Based on the trained ensemble learning model, the expression data of cell cycle characteristic genes of single or multiple cells are used as input to the ensemble learning model to predict the sub-stage of the corresponding cell in the cell cycle.
[0016] The present invention also discloses a new naming method for cell cycle stages, wherein the stage name includes the total number of stages and the code of a specific stage. For example, "C6.2" represents the second sub-stage under the 6-stage framework, and "C9.5" represents the fifth sub-stage under the 9-stage framework.
[0017] Advantages and technical effects of the present invention:
[0018] (1) A robust set of cell cycle characteristic genes is provided: Through pseudo-time sequence construction and autoregression test, this invention successfully identified a more robust set of cell cycle genes, which provides a reliable basis for subsequent cell cycle staging studies and also provides a meaningful reference for other cell cycle studies.
[0019] (2) High-resolution cell cycle segmentation: Using Fourier analysis, low-pass filtering and TICC clustering, this invention achieves high-resolution segmentation of the cell cycle. Compared with traditional staging methods, this high-resolution segmentation can reveal the state changes of cells at different stages more accurately and in more detail, providing a more comprehensive perspective for a deeper understanding of the cell cycle.
[0020] (3) Simple and practical cell cycle stage predictor: This invention uses an integrated learning model to train cell cycle stage predictors at different stage resolutions; the stage predictor can predict the corresponding cell cycle sub-stage by using the expression data of cell cycle feature gene sets of single or multiple cells.
[0021] (4) High-resolution cell cycle segmentation can more accurately assess the effect of drugs on the cell cycle, which is crucial for drug screening and efficacy evaluation, and helps to select more effective drugs and reduce unnecessary side effects.
[0022] This invention can also be used to study the differentiation process of stem cells and determine the cell cycle distribution of stem cells at different developmental stages, thereby optimizing stem cell therapy and regenerative medicine applications. In the study of cancer cells, it can reveal the characteristics of cancer cells in different sub-phases of the cell cycle, providing strong support for a deeper understanding of cancer development mechanisms. By predicting the effects of drugs on different cell cycle stages, it can help discover potential therapeutic targets. Attached Figure Description
[0023] Figure 1 The process of constructing a pseudo-time series of single-cell cell cycle (Figure A) and the resulting image (Figure B);
[0024] Figure 2 The results of the screening for cell cycle characteristic genes;
[0025] Figure 3 This is a schematic diagram illustrating the functions of cell cycle characteristic genes during the cell cycle.
[0026] Figure 4 Figure A shows the temporal expression profile of cell cycle characteristic genes and the temporal transcription signal after Fourier analysis and low-pass filtering (Figure B).
[0027] Figure 5 A heatmap of naive mutual information between dataset pairs under different staging frameworks;
[0028] Figure 6 This is a schematic diagram illustrating the training and testing process of an ensemble learning model.
[0029] Figure 7 A schematic diagram showing the comparison results between the 6-stage and 9-stage frameworks and the classic staged framework of the FUCCI method;
[0030] Figure 8 A schematic diagram of transcriptional signal fluctuations of typical genes in the four cell cycles under different stage frameworks;
[0031] Figure 9Figure A shows the transcriptional signal fluctuations of four typical cell cycle genes within a 9-stage framework, and the temporal expression profiles of cell cycle characteristic genes that are significantly overexpressed in each sub-stage (Figure B).
[0032] Figure 10 The results show the predictions made using a trained ensemble learning model on single-cell transcriptome data from A549 cells. Detailed Implementation
[0033] The present invention will be further described in detail below through embodiments, but the scope of protection of the present invention is not limited to the content described. Unless otherwise specified, the methods in the embodiments are conventional methods.
[0034] Example 1: The cell cycle stage prediction method based on single-cell transcriptome sequencing data is as follows:
[0035] 1. To ensure the active mitosis of the relevant cells, single-cell transcriptome sequencing data from cultured cell lines were collected from the GEO database. Each dataset was required to contain more than 250 cells, resulting in 12 datasets covering various cell lines and experimental treatments. Details of each dataset are shown in the table below:
[0036] Table 1. 12 single-cell transcriptome datasets
[0037]
[0038] 2. To ensure data quality, the following quality control standards were used to preprocess the 12 datasets as follows:
[0039] (1) Remove cells with a sequencing depth of less than 10,000;
[0040] (2) Remove genes that are not expressed in 70% or more of the cells;
[0041] After processing, the gene expression matrix of cells in each dataset is obtained. The raw read counts are converted to counts per million (CPM) using the CPM function of the edgeR package, and the CPM data is transformed by a base-2 logarithmic function to prevent highly expressed genes from dominating.
[0042] 3. Pseudo-time series construction is performed on cells in each dataset.
[0043] Based on 97 known cell cycle genes provided by Seurat software, principal component analysis (PCA) was performed on the expression profiles of known cell cycle genes in each dataset. Cells showing significant deviations in principal component 1 (PC1) and principal component 2 (PC2) were then removed to eliminate potential confounding factors or outliers that might affect subsequent pseudo-time analysis. The scores of each cell on the first and second principal components were used to represent the cell cycle progression. Subsequently, the polar angles of the scatter points in the two-dimensional space corresponding to different cells were used as a measure of the pseudo-time series of the cell cycle. Figure 1 A), where the angle represents the progress of the cell cycle; thus, the pseudo-temporal sequence of cells in the cell cycle in each dataset is obtained;
[0044] The pseudo-time series obtained by PCA in this embodiment are compared with other state-of-the-art methods (such as Tricycle and reCAT). Across all datasets, the pseudo-time series constructed by different methods show excellent correlation. For example, in the HeLaCCL2_2019a dataset, the Pearson correlation coefficient between the pseudo-time series constructed by PCA and the Tricycle method is as high as 0.940. Figure 1 B); This indicates that the PCA of the present invention is one of the most effective methods for reconstructing scRNA-seq cellular pseudo-time sequences.
[0045] 4. Use the Ljung-Box test to identify cell cycle characteristic genes.
[0046] Based on the reconstructed pseudo-temporal sequence, cells in each dataset were rearranged to obtain temporal expression profiles of different genes. The Ljung-Box test was used to examine the autocorrelation of all gene expression data in each dataset. In multiple tests involving thousands of genes, the Benjamini-Hochberg multiple test correction method was used to estimate the false discovery rate (FDR) at different cutoff points, and genes with an FDR less than 0.05 were considered candidate cell cycle characteristic genes. Finally, the intersection of the cell cycle gene sets identified in different datasets was taken, and 152 genes identified as cell cycle genes in 8 or more datasets across 12 datasets were selected as cell cycle characteristic genes. Figure 2 Extensive literature review and analysis revealed that most cell cycle characteristic genes play crucial roles in cell growth signal transduction, DNA replication, microtubule formation, spindle formation, and sister chromatid separation. Figure 3 );
[0047] 5. Reconstructing the transcriptional waves of cell cycle characteristic genes
[0048] First, based on the cell cycle characteristic gene set identified in step 4, the pseudo-temporal sequence of cells in each dataset is reconstructed (using the same method as step 3). The cells are then rearranged based on the new pseudo-temporal sequence to obtain the temporal expression profile of cell cycle characteristic genes in each dataset.
[0049] Gene expression data are often affected by significant noise and sequencing errors, limiting the possibility of in-depth research into the potential transcriptional patterns of different genes. This invention employs Fourier transform and low-pass filtering to denoise and smooth the transcriptional levels of periodic genes, aiming to recover the temporal transcriptional signals of cell cycle characteristic genes. For each cell cycle characteristic gene, its expression level during the cell cycle can be decomposed into component waves with different frequencies and amplitudes. High-frequency waves correspond to noise and outliers, while low-frequency waves represent the general pattern of periodic transcription, i.e., the temporal transcriptional signal.
[0050] To improve computational efficiency, the FFT function from the R package `stats` was used to perform a fast Fourier transform on the expression data of cell cycle characteristic genes. Since the energy in the frequency domain after the Fourier transform is concentrated in the low-frequency region, a low-pass filter was used to remove these high-frequency components. The `ifftf` function from the R package `pracma` was then used to perform an inverse Fourier transform on the filtered result, ultimately obtaining the time-series transcriptional signals of cell cycle characteristic genes in each dataset. Figure 4 The temporal transcription signal of each gene can be viewed as a smooth curve that changes with pseudo-time. This smooth curve reflects the transcriptional dynamics of cell cycle characteristic genes to a certain extent. The above process is encapsulated in an R function called fftf. The specific R script can be obtained on GitHub (https: / / github.com / FengXu1218 / HRCC).
[0051] 6. TICC clustering of cells based on reconstructed transcriptional signals
[0052] To characterize cells at different stages of the cell cycle through gene transcription at different time points, the TICC algorithm was used to perform cluster analysis on cells from different datasets. TICC is a multivariate time series clustering algorithm based on the Topplitz inverse covariance matrix. The TICC algorithm has two important characteristics: multivariate clustering and time series clustering. In this embodiment, the expression profile of each cell cycle characteristic gene can be considered as a feature, and the expression level of each gene can be considered as a time series that changes with pseudo-time series. Therefore, the expression data of cell cycle characteristic genes after pseudo-time series rearrangement can be regarded as multivariate time series data, making the TICC algorithm more applicable when processing existing data. The Python program TICC (https: / / github.com / davidhallac / TICC) was used to perform cluster analysis on the time-series transcription signals of cell cycle characteristic genes in each dataset, and the number of clusters was set from 3 to 20.
[0053] 7. Evaluation of staging frameworks at different resolutions based on naive mutual information
[0054] To determine the advantages of a high-resolution staging framework, we developed a naive mutual information (MIS) model. Naive mutual information (MMI) serves as a measure of consistency in cell cycle segmentation across different datasets. The more closely related the transcription of cell cycle characteristic genes is to each other across different datasets, the greater the information content. In other words, a staging framework with more naive mutual information can provide more information about the overall progression of the cell cycle. The definition of naive mutual information is as follows:
[0055] For two datasets A and B with the same genes, we first divide their respective cells into k sub-stages. Assume the i-th stage of dataset A (A... i The peak gene count is m. i The j-th period of dataset B (B j The peak number of genes is n. j If i = 1, 2, ..., k, j = 1, 2, ..., k, then the total number of genes w involved at this time is:
[0056]
[0057] For gene pair g i' and g j' Where i' = 1, 2...152; j' = 1, 2...152, this gene pair in A i and B j The probability of a transcription peak during this period is
[0058]
[0059] in,
[0060]
[0061] Because of the inherent connection between the definition of φ and classical mutual information, we call φ naive mutual information.
[0062] Assumption Then the above probability can be simplified to:
[0063]
[0064] The expected probability of each gene pair peaking in the same sub-period is:
[0065]
[0066] because
[0067]
[0068] Therefore, the expected value can be simplified to...
[0069]
[0070] Further transformation yields,
[0071]
[0072] We use φ 2 As an indicator of the interdependence of transcriptional signals between datasets A and B, it can be obtained by directly calculating the number of gene pairs that peak in the same sub-period. When the total number of datasets is K, we can calculate...
[0073]
[0074] Where p and q represent indices of different datasets, This represents the average degree of interdependence of transcriptional signals across all datasets.
[0075] Analysis shows that the mutual information of the aforementioned stage frameworks varies with the number of stages; for example, the hiPSC and hNPC datasets (hiPSC_2019, hNPC_2021) have the most information in the 19-stage framework, while the hiPSC and HeLa datasets (hiPSC_2019, HeLaCCL2_2019a) have the most information in the 10-stage framework. Notably, the information content increases with the number of stages, reaching its maximum at nine or more stages. These results indicate that a framework dividing the cell cycle into nine or more stages better represents the transcriptional processes within the cell cycle. Figure 5 ).
[0076] 8. Construction of Phase Predictors under Different Phase Prediction Frameworks
[0077] (1) The expression data of cell cycle characteristic genes from 12 datasets were merged to obtain the gene expression profiles of 13,754 cells. The data were standardized using the R function scale to obtain the standardized expression matrix of cell cycle characteristic genes. Subsequently, the data were randomly divided into a training set (70%, 9,630 cells) and a test set (30%, 4,124 cells) in a 7:3 ratio.
[0078] (2) A stacked ensemble learning strategy was adopted, integrating four machine learning algorithms: XGBoost, Neural Networks (NNs), Support Vector Machines (SVM), and Random Forests (RF). By combining and balancing different models, we constructed a more robust and accurate ensemble learning model. The ensemble learning model uses the above four machine learning models as base models and Random Forest as the meta-model; it uses the expression data of 152 cell cycle characteristic genes as input features, and the TICC stage result of each cell as the true label of the model; on the training set, 5-fold cross-validation was performed on the four base models respectively, and then the validation set results of the above four base models were integrated. The integrated result was used as the training set of the ensemble learning meta-model to obtain the trained ensemble learning models under different stages. Figure 6 );
[0079] Meanwhile, XGBoost, neural networks (NNs), support vector machines (SVM), and random forests (RF) were used to train different single prediction models in a 3- to 10-stage framework, respectively.
[0080] (3) The different prediction models trained in step (2) were tested using a test set. The test results showed that the accuracy of the ensemble learning model was significantly improved compared to the single prediction model. For example, under the 9-stage framework, the accuracy of the ensemble learning model reached 81%, exceeding the individual prediction results of the other four base models (Table 2). The results indicate that the trained ensemble learning classifier can use the expression profiles of cell cycle characteristic genes of single or multiple cells to determine the cell cycle stage of various cells. The trained prediction model can be used online at www.hrcc-predictor.fun / hrcc or www.compgenelab.info / hrcc.
[0081] Table 2 Prediction accuracy of single and ensemble models under different staging frameworks
[0082]
[0083]
[0084] 9. New phased framework naming system
[0085] We provide a naming system for different staging frameworks to encode sub-periods under different frameworks, including the total number of periods and the code for a specific period. For example, "C6.2" represents the second sub-period under the 6-period framework.
[0086] We further compared our proposed 6-stage and 9-stage frameworks with the classic staged framework of the FUCCI method. In the hESC dataset (hESC_2015), the traditional FUCCI method divides cells into G1, S, and G2 phases, with 91 cells in G1, 70 cells in S, and 74 cells in G2. In our 6-stage framework, most of the G1 phase cells are further subdivided into C6.1 (49 cells) and C6.2 (29 cells); most of the S phase cells are subdivided into C6.3 (39 cells) and C6.4 (28 cells); and most of the G2 phase cells are further subdivided into C6.5 (25 cells) and C6.6 (37 cells). Figure 7 -A). A comparison between the more refined 9-stage framework and the FUCCI stage framework provides us with further insights and considerations. For example, in both C9.4 and C9.7 sub-stages, there are small numbers of cells from adjacent FUCCI stages, which may correspond to the transition between different classical stages (...). Figure 7 -B). In addition, we observed the same result in the hU2OS dataset ( Figure 7 (C and D). Our research indicates that a more comprehensive transcriptional staging framework provides more information about cell cycle progression.
[0087] 10. Biological significance of staging frameworks at different resolutions
[0088] To demonstrate the correspondence between different stage resolution frameworks and the classical cell cycle, we selected the transcriptional signals of four typical cell cycle genes—CCNE1, PCNA, CCNA2, and CCNB1—during the entire cell cycle as references. Based on the intensity of the transcriptional signals of these four typical genes at different time points, their corresponding classical cell cycle stages can be inferred. Figure 8 ).
[0089] For example, within the 9-phase framework, in phase C9.1, the transcription of CCNE1 and PCNA reaches its lowest level in the cell cycle. Simultaneously, the expression of CCNB1 and some cell cycle characteristic genes in phase M persists but decreases. This indicates the transition from classical M phase to G1 phase. Subsequently, in phases C9.2 and C9.3, the expression of cell cycle characteristic genes in phase M significantly decreases, while the expression of CCNE1, PCNA, and many cell cycle characteristic genes involved in DNA replication in phase S begins to increase; therefore, phases C9.2 and C9.3 correspond to the early and late stages of classical G1 phase, respectively. Then, in phase C9.4, the expression of CCNE1 and some DNA replication-related cell cycle characteristic genes reaches its maximum, indicating that genome replication gradually becomes more active during this period. Subsequently, in phase C9.5, the expression of CCNE1 gradually decreases, while the expression of PCNA and other cell cycle characteristic genes in phase S continuously increases and reaches a peak. This stage corresponds to the peak of DNA replication. In phase C9.6, PCNA expression decreases as CCNA2 expression gradually increases, indicating the approaching end of S phase and the gradual transition of cells into G2 phase. When CCNA2 expression surpasses PCNA, cells enter phase C9.7. At this stage, genes related to spindle formation, nuclear division, and sister chromatid separation begin to show high expression, marking the entry of cells into classic G2 phase. CCNA2 peaks in phase C9.8, and CCNB1 expression also gradually increases during this period. Simultaneously, genes involved in chromosome formation and sister chromatid separation also reach peak expression, indicating that cells have entered the mid-to-late stages of classic G2 phase. When CCNB1 expression surpasses CCNA2, cells enter phase C9.9. At this stage, the expression signals of both CCNB1 and CCNA2 begin to decline, indicating the beginning of the transition into classic late G2 and M phases. Figure 9 ).
[0090] 11. Application of high-resolution cell cycle stage predictors
[0091] To further validate the effectiveness of the ensemble learning model trained in step 8, we applied the trained 9-stage model to the recently released single-cell transcriptome dataset of A549 cells. This dataset includes cells treated with 6 Gy γ-rays for 2 hours (IR_2h), 6 hours (IR_6h), and control cells. The results showed that the distribution of cells in the IR-2h group across each sub-stage was very similar to that of the control group. However, in the IR-6h group, the number of cells in stages C9.6–C9.9 was significantly higher than that of the control cells. Figure 10A). In the control group and the IR-2h group, the overall proportion of cells in the C9.6-C9.9 phase was only 28.54% and 24.66%, respectively, while in the IR-6h group, the corresponding proportion significantly increased to 48.79%. Analysis of the aforementioned four classic cell cycle genes showed that during the C9.6-C9.9 phase, the expression levels of CCNA2 and CCNB1 continuously increased, while the expression levels of CCNE1 and PCNA continuously decreased. Figure 10 B). This is consistent with previous analysis, namely that C9.6-C9.9 correspond to different stages of the classic G2 phase (B). Figure 9 Furthermore, previous studies have shown that radiation may trigger G1 / S and G2 / M cell cycle arrest in cultured cells. The presence of the G2 / M phase checkpoint may lead to a significant proportion of cells in IR-6h cells being arrested in the G2 phase. Figure 10 A). The only minor difference observed between the IR-2h group and the control group may be due to the short duration of radiation exposure. These results demonstrate that our cell cycle classifier provides an effective tool for high-resolution staged cell cycle analysis, offering a clearer perspective for in-depth research into biological processes within the cell cycle.
Claims
1. A method for predicting cell cycle stages based on single-cell transcriptome sequencing data, characterized in that, The steps are as follows: (1) Collect several single-cell transcriptome datasets, preprocess each single-cell transcriptome dataset, and then construct pseudo-time series to obtain the pseudo-time series of each cell in the cell cycle. Based on the constructed pseudo-time series, rearrange the cells to obtain the temporal expression profiles of different genes. Obtain the cell cycle genes in each dataset through autoregression test, and select genes that appear in 2 / 3 or more of the total number of datasets as cell cycle characteristic genes. (2) Based on the obtained cell cycle feature gene set, the cells in different datasets are reconstructed using pseudo-temporal sequences. The cells are rearranged based on the new pseudo-temporal sequences to obtain the temporal expression profiles of different genes. Fourier analysis and low-pass filtering are used to smooth and denoise the temporal expression profiles of cell cycle feature genes in each dataset to obtain the temporal transcription signals of cell cycle feature genes in each dataset. (3) The TICC clustering algorithm was used to divide the temporal transcription signals of cell cycle characteristic genes in each dataset into different number of phases to obtain the sub-phases in the cell cycle under different number of phases. (4) Merge the expression data of cell cycle characteristic genes in all datasets and standardize them. Use the expression data of cell cycle characteristic genes as features and the stage data of TICC clustering as labels. Use XGBoost, neural network, SVM and random forest as base models and random forest as meta-model to train and validate the ensemble learning model and obtain the trained ensemble learning model under different number of stages. (5) Based on the trained ensemble learning model, the transcriptional data of cell cycle characteristic genes of single or multiple cells are used as input to the ensemble learning model to predict the sub-stages of the corresponding cells in the cell cycle. The pseudo-time sequence construction in step (1) is based on known cell cycle genes; the pseudo-time sequence construction in step (2) is based on cell cycle characteristic genes obtained in step (1); Specifically, principal component analysis is performed on the cell cycle gene expression profile of cells in each dataset. The scores of each cell on the first and second principal components represent the cell's progress in the cell cycle. The polar angles of the scatter points in the two-dimensional space corresponding to different cells are used as a measure of the pseudo-time series of the cell cycle.
2. The cell cycle stage prediction method based on single-cell transcriptome sequencing data according to claim 1, characterized in that: A single-cell transcriptome dataset consists of multiple datasets composed of transcriptome data from different cell lines and cells subjected to different experimental treatments. Preprocessing of a single-cell transcriptome dataset involves removing cells with a sequencing depth of less than 10,000 and removing genes that are not expressed in 70% or more of the cells.
3. The cell cycle stage prediction method based on single-cell transcriptome sequencing data according to claim 1, characterized in that: The number of different phases in step (3) is 3-20.
4. The cell cycle stage prediction method based on single-cell transcriptome sequencing data according to claim 3, characterized in that: Based on different number of stages, the cell cycle stage is named Cx.y, where x is the number of stages and y is the sub-stage under that number of stages.
5. The cell cycle stage prediction method based on single-cell transcriptome sequencing data according to claim 4, characterized in that: When the number of phases is 9, phase C9.1 corresponds to the transition from the classical M phase to the G1 phase in the cell cycle; phases C9.2 and C9.3 correspond to the early and late stages of the classical G1 phase; phase C9.4 corresponds to the transition from G1 to S phase; phases C9.5 and C9.6 correspond to the early and late stages of the classical S phase; phase C9.7 corresponds to the transition from S to G2 phase; and phases C9.8 and C9.9 correspond to the classical G2 and M phases, respectively.
Citation Information
Patent Citations
Method for evaluating cell aging degree based on transcriptome data and machine learning strategy
CN113838531A
Method for detecting mesenchymal stem cells with different tissue sources, generations and culture conditions based on single cell sequencing
CN116884484A
