Timestamp correction method based on time sequence scRNA data

Abstract

The invention provides a time stamp calibration technology named as ScPace. The time stamp calibration technology is used for processing time sequence scRNA-seq data. According to the method, the noise indication hidden variables are introduced to identify and correct mistakenly marked cells, so that the accuracy and reliability of timestamp automatic annotation and pseudo time analysis supervision are improved. The technical process comprises the steps of PCA dimension reduction preprocessing, noise cell recognition and correction, and selective re-classification or deletion of error cells. Different from a traditional probability sampling method, the ScPace realizes higher-precision timestamp calibration based on a hidden variable mechanism, has higher robustness, is suitable for various types of time sequence scRNA-seq data, and remarkably improves downstream analysis performance.

Description

Technical field:

[0001] The present invention relates to a timestamp correction method for time-series scRNA-seq data, specifically using noise-indicating latent variables for correction, and belongs to the field of biomedical information processing technology. Technical background:

[0002] The timestamps in time series scRNA-seq data are usually derived from the time of sample collection. However, due to the influence of various factors, timestamps are often accompanied by label noise, which may cause cells in the dataset to be mislabeled, resulting in inaccurate labels and thus compromising the quality of the data. There are many reasons for time series labeling errors. For example, when different heterogeneous cell types are present at the same time point, classification errors may occur, causing cells to be incorrectly assigned to adjacent time points. In addition, in some cells, due to the presence of different maturation states, a timestamp may overlap with cells in other maturation states, making the cell state difficult to distinguish from other timestamps, which further increases the complexity of the sequential relationship between timestamps. These factors work together to lead to a decline in the performance of automatic timestamp annotation.

[0003] To address this issue, ScReclassify uses a computational approach to reclassify mislabeled cell types through semi-supervised Adasampling, which can also be applied to timestamp calibration. However, the current mechanism for selecting potentially mislabeled cells relies on probabilistic sampling, which may result in the selection of unimportant samples during the sample selection process. In addition, ScReclassify does not fully account for the complex relationships inherent in time series scRNA-seq data.

[0004] Therefore, we hypothesize that traditional timestamp automatic annotation methods rely heavily on high-quality labels, but when applied to real-world datasets, these labels can hinder performance, leading to suboptimal results. Current solutions to this problem mostly rely on probabilistic sampling, which can result in the selection of unimportant samples. To address these challenges, we developed a novel timestamp calibration method, named ScPace. Unlike previous methods that use semi-supervised Adasampling for cell type label correction, this method introduces a latent variable that serves as an indicator for selecting potentially mislabeled samples, enabling efficient removal or reclassification of these samples based on specific requirements. This method is inspired by self-paced learning, a machine learning training paradigm that aims to mimic the human cognitive process of learning by gradually mastering more difficult samples, starting with easier ones. Self-paced learning has been widely used in various fields, such as multimodal image classification, domain adaptation, imbalanced sampling, and functional brain network cleanup. In particular, previous studies have shown that imbalanced sampling using self-paced learning achieves better performance than random imbalanced sampling methods, emphasizing the retention of high-quality samples while discarding low-quality ones, thereby constructing a more balanced sample set. Summary of the invention:

[0005] ScPace

[0006] We describe each proposed module in detail. Inspired by self-paced learning, we introduce noise indicator latent variables, such as Figure 1 Specifically, this method consists of the following three main components:

[0007] (1) Use PCA or KernelPCA for data preprocessing

[0008] (2) Noise indicator variables, which are used to specify which cell samples are potentially mislabeled through a two-stage optimization.

[0009] (3) Based on the output of the noise indicator variable, determine whether to reclassify or delete the designated cells.

[0010] Step 1: Data preprocessing

[0011] This paper provides the option to perform dimensionality reduction on both the training and test sets; however, this preprocessing is optional. Here, we implement principal component analysis (PCA) and kernel PCA. We consider the original dimensionality of the time series scRNA-seq dataset to be m × n. We determine d as the number of principal components (PCs) needed to capture at least 70% of the overall variability in the dataset. After obtaining d, dimensionality reduction is defined in the following equation, transforming the dataset to m′ × n. Note that in kernelPCA, the dimensionality is set to the default value of m′ = 20.

[0012]

[0013] Step 2: Train OvO multi-classification SVM

[0014] In order to obtain the optimal w and b in the SVM dual space, we need to maximize the equation on the parameter α. Therefore, we get the following optimal solution with the noise indicator variable:

[0015] Chemical problems:

[0016]

[0017] After solving the optimization problem, we can get the optimal α according to the KKT condition.

[0018] Consider that we have the optimal solution

[0019]

[0020] Note that the optimal parameter b * is obtained from the support vector, and the optimal parameter w * is obtained from the above equation. Therefore, after obtaining the optimal solution, we can get the decision boundary corresponding to its decision function with latent variables:

[0021]

[0022] Step 3: Calculate the decision value of the OvR shape based on the training results

[0023] According to the algorithm Figure 16 As shown, we can get the voting matrix The sum and confidence matrix S are then monotonically transformed according to the following equation:

[0024]

[0025] here a ij∈S, i = 1, 2, ..., n, j = 1, 2. Therefore, the final decision value converted to One vs Rest shape is the sum of the voting matrix and the converted total confidence matrix.

[0026]

[0027] Step 4: Optimize noise-indicating latent variables

[0028] After we obtain the decision value of the N-dimensional shape of each sample, we use the multi-class hinge loss to calculate the loss value of each sample. Therefore, considering n samples and is the corresponding label y of the i-th sample i The predicted decision value of , and The loss value of each sample is as follows:

[0029]

[0030] After computing the loss value for each sample, we can optimize the latent variable using the following modified version to take into account which samples are clean or noisy:

[0031]

[0032] The above formula is equivalent to the following:

[0033]

[0034] Step 5: Repeat steps 1 to 3

[0035] according to Figure 17 The present invention performs two-stage continuous optimization

[0036] Innovation

[0037] The following are the innovative features of the present invention:

[0038] 1. Unlike previous methods that calibrate cell type labels using semi-supervised Adasampling, our method introduces a latent variable indicator to identify samples that may be mislabeled. This indicator allows us to flexibly remove or reclassify these samples based on specific needs, thereby improving the accuracy and overall performance of timestamp calibration.

[0039] 2. In addition to automatic timestamp annotation, supervised pseudo-time analysis based on timestamp labels relies heavily on timestamp quality. Therefore, by using our proposed method to timestamp-calibrate raw time series scRNA-seq data, we significantly improved analysis performance. This correction not only improves timestamp accuracy but also provides more reliable data support for subsequent downstream pseudo-time analysis, thus making a positive contribution.

[0040] 3. The present invention demonstrates enhanced robustness when handling unbalanced and high-dimensional nonlinear datasets. This demonstrates that the present invention can be effectively applied to various time series scRNA-seq datasets, enabling precise timestamp calibration and thus improving the stability and accuracy of data analysis.

[0041] Result Analysis

[0042] Quantitative results

[0043] In order to verify the quantitative performance of the present invention, we used four public data sets and three simulation data sets, and the relevant statistical results are shown in Table 1. We compared the models using three indicators, Accuracy, Balanced accuracy and Correction score, on the data sets, through cross-validation and artificial noise labeling, and the results are shown in Figure 3 , Figure 4 , Figure 5 , Figure 6 These results demonstrate that our model remains robust even in the presence of strongly noisy labels.

[0044]

[0045] Table 1

[0046] Sensitivity analysis

[0047] In order to verify the impact of different parameters on the performance of the present invention, the present invention conducted sensitivity analysis on four public data sets and three simulated data sets, and evaluated them with Accuracy. The relevant parameters are shown in Table 4 and Table 5. Through cross-validation and artificial noise labeling, the results are shown in Figure 7 , Figure 8 , Figure 9 , Figure 10 It can be concluded that except for the hyperparameter C, the performance of the hyperparameters including λ, Iter, and p do not show significant performance changes. In contrast, we observe that the performance of the hyperparameter C varies significantly on most datasets.

[0048]

[0049] Table 2

[0050]

[0051] Table 3

[0052] Enhancement of supervised quasi-time series

[0053] The present invention reclassified the simulated data set with artificial labeling errors and the original real time series scRNA-seq data set (excluding artificial labeling errors) through three calibration methods as the starting point of the analysis. After completing the reclassification of the original noisy simulated data and the real time series scRNA-seq data using the three calibration methods, we applied the timestamps reclassified by the two methods to Psupertime respectively, and calculated the average training accuracy, test accuracy and pseudo-time value of each cell. In order to evaluate the performance of calculating pseudo-time, we quantified it by calculating the Spearman and Kendall correlation between the calculated pseudo-time value and the calibrated timestamp. Finally, we applied the pseudo-random number to the original timestamp for comparison to test whether timestamp calibration can significantly improve the performance. The results are shown in Figure 11 , Figure 12 , Table 2, Table 3. The results show that the present invention exhibits superior performance on different datasets.

[0054]

[0055] Table 4

[0056]

[0057] Table 5

[0058] Case Studies

[0059] Identification of different maturation states during hepatoblast differentiation

[0060] We further evaluated the cells detected by the present invention in the GSE90047 dataset, which includes two different developmental trajectories: one from hepatoblasts to cholangiocytes and the other from hepatoblasts to hepatocytes. Our analysis showed that most of the hepatoblasts detected by the present invention at early time points (such as E10.5 and E11.5) were noisy samples, indicating that these timestamps may be erroneous. However, we noticed that there was a certain degree of overlap in the cholangiocytes observed along the hepatoblast-to-cholangiocyte trajectory at E14.5 and E15.5 time points, which may be due to the presence of timestamp noise.

[0061] Further analysis revealed that timestamp noise between E12.5 and E13.5 was present in both trajectories, with the noise being particularly pronounced on the hepatoblast-to-hepatocyte pathway. In contrast, the hepatoblast-to-cholangiocyte pathway exhibited less of a marker noise impact. This phenomenon suggests that timestamp noise in the GSE90047 dataset may be due to varying maturation at specific time points.

[0062] exist Figure 13 In the figure, red circles highlight potentially mislabeled cells (these cells are denoted as v = 0). Our results show that our method can successfully detect the majority of overlapping cells in the GSE90047 dataset, further demonstrating its ability to effectively identify labeling noise. This finding demonstrates that our method can accurately distinguish and correct these errors in the face of timestamp noise, thereby improving the accuracy of data analysis.

[0063] Given the challenges posed by these noisy timestamp labels, we believe that simply removing detected samples from the GSE90047 dataset is not an optimal solution. Instead, we advocate a reclassification approach to calibrate cells to more reliable timestamp labels. This approach is ideal because the observed label noise may not be due to technical issues but rather reflects biological variability. By employing reclassification techniques, we not only enhance the integrity of our analysis but also preserve valuable information in the dataset, thereby providing more accurate biological interpretations for subsequent studies.

[0064] Characterizing the regenerative function of cardiomyocytes after heart failure

[0065] The present invention conducted a preliminary experiment to show the developmental state of cardiomyocytes after myocardial infarction. The datasets are from the GEO database, numbered GSE126128, GSE131181, and GSE132658, containing 2404 cells and 6 time points (E7.75, E8.25, E9.25, E10.5, E13.5, and E16.5). The data were preprocessed using the Scanpy software package, and the dataset was divided into a training set (75%) and a validation set (25%). According to Figure 14 In the confusion matrix of the training results shown, we observe a certain degree of ambiguity between E8.25 and E9.25. This may be due to the smaller developmental changes between these two time points, making it more difficult to distinguish between the two time points.

[0066] For the test set, we collected 2082 cells from the National Center for Biotechnology Information Sequence Read Archive (NCBI) with accession number PRJNA595398 from the BioProject. Figure 15The test set prediction results shown in the figure show that the concentration of cardiomyocytes after injury shifts to earlier stages, with a significant peak increase in cell concentration at embryonic day 13.5, indicating a significant increase in the number of cardiomyocytes during recovery. This phenomenon may be closely related to the regenerative mechanism during myocardial repair, indicating that cardiomyocytes gradually recover and transition to a more mature state.

[0067] Furthermore, our results suggest that cardiomyocyte density changes relatively little between embryonic days 10.5 and 13.5, suggesting that this period may represent a barrier to cardiomyocyte dedifferentiation. Specifically, cardiomyocytes at this stage may experience a temporary arrest, preventing them from transitioning smoothly to a fully mature state. This phenomenon may be related to the reprogramming process of cardiomyocytes, particularly after myocardial infarction, when cells may fail to fully dedifferentiate or reprogram.

[0068] Overall, our findings demonstrate that cardiomyocytes undergo an incomplete genomic reprogramming process during post-MI repair. During this process, the number of immature cardiomyocytes gradually increases over time, indicating that cardiomyocytes not only change in number but also exhibit a degree of heterogeneity in their functional recovery. Our findings provide new insights into the mechanisms of post-MI repair, particularly regarding potential pathways for cardiomyocyte dedifferentiation and regeneration.

[0069] Runtime

[0070] To evaluate the computational cost of our invention, we used a GPU equipped with NVIDIA GeForce RTX 3080Ti and 12th Gen Core TM Training was performed on a research server with an i7-12700KF CPU (125Gi). The computation time for timestamp calibration with swapped and random mislabeling is shown in Tables 6 and 7, and the computation time for a real time series dataset is shown in Table 8. The results show that compared to ScReclassify, which introduces PCA in the initial step, the computation time for our method and CL is significantly longer when no dimensionality reduction is used. However, we found that after applying dimensionality reduction to our method, its computation time is even shorter than that of ScReclassify.

[0071]

[0072] Table 6

[0073]

[0074] Table 7

[0075]

[0076] Table 8 Description of the drawings:

[0077] (1) Figure 1 The main model architecture includes data preprocessing, training ScPace, and performing two timestamp calibrations.

[0078] (2) Figure 2 Conceptual image of swapping error labels and random error labels

[0079] (3) Figure 3 Cross-validation results with different mislabeling rates on simulated datasets (AC) Cross-validation results with exchange mislabeling (DF) Cross-validation results with random mislabeling (GI) PCA dimensionality reduction visualization of the dataset simulated by Splatter, visualization of the simulated datasets simulated by Sim1 and Sim2 are class balanced, while Sim3 is unbalanced.

[0080] (4) Figure 4 Cross-validation results with different mislabeling rates on real time series scRNA-seq datasets (AD), cross-validation results with exchanged label errors (EH), and mutual validation results with random label errors.

[0081] (5) Figure 5 is the correction score (AC) of cross-validation with different mislabeling rates on the simulated dataset. The correction score result of exchanged mislabeling (DF) is the correction score result of random mislabeling.

[0082] (6) Figure 6 It is the correction score of cross validation with different mislabeling rates on the real dataset (AD), the correction score result of exchanged mislabeling (EH), and the correction score result of random mislabeling.

[0083] (7) Figure 7 This is a sensitivity analysis of the four hyperparameters (C, \lambda, iter, p) on the simulated dataset. Each dataset contains two rows, the first row represents the cross-validation sensitivity analysis of swap label errors, and the second row represents the mutual validation sensitivity analysis of random label errors.

[0084] (8) Figure 8 It is a sensitivity analysis of the four hyperparameters (C, \lambda, iter, p) on the real data set. Each data set contains two rows, the first row represents the cross-validation sensitivity analysis of swap label errors, and the second row represents the mutual validation sensitivity analysis of random label errors.

[0085] (9) Figure 9Sensitivity analysis results of using different dimensionality reduction techniques on simulated datasets (AC) Cross-validation results of using different dimensionality reduction techniques on exchanged mislabeling (DF) Cross-validation results of using different dimensionality reduction techniques on random mislabeling.

[0086] (10) Figure 10 Sensitivity analysis results of different dimensionality reduction techniques on real datasets (AD) Cross-validation results of different dimensionality reduction techniques on exchanged mislabeling (EH) Cross-validation results of different dimensionality reduction techniques on random mislabeling.

[0087] (11) Figure 11 Enhanced results of supervised pseudo-time analysis using two types of mislabeling and two types of metrics for reclassification (AC) Enhanced results of supervised pseudo-time analysis with swapped mislabeling using Kendall correlation (DF) Enhanced results of supervised pseudo-time analysis with swapped mislabeling using Spearman correlation (GI) Enhanced results of supervised pseudo-time analysis with random mislabeling using Kendall correlation (JL) Enhanced results of supervised pseudo-time analysis with random mislabeling using Spearman correlation

[0088] (12) Figure 12 Enhanced results of supervised pseudo-time analysis using two types of mislabeling and two types of metrics for removal (AC) Enhanced results of supervised pseudo-time analysis with swapped mislabeling using Kendall correlation (DF) Enhanced results of supervised pseudo-time analysis with swapped mislabeling using Spearman correlation (GI) Enhanced results of supervised pseudo-time analysis with random mislabeling using Kendall correlation (JL) Enhanced results of supervised pseudo-time analysis with random mislabeling using Spearman correlation

[0089] (13) Figure 13 (A) TSNE dimensionality reduction visualization on GSE90047 (B) ScPace detection of potential mislabeled cells (v=0) in GSE90047.

[0090] (14) Figure 14 (A) Confusion matrix generated from the validation set and (B) PCA visualization of cardiomyocytes collected at embryonic day 6.

[0091] (15) Figure 15 is the prediction result on the (AC) test set. We can clearly observe that the peak occurs at embryonic day 13.5.

[0092] (16) Figure 16 It is an algorithm flow chart that converts the OvO decision function of SVM into OvR form

[0093] (17) Figure 17 This is the algorithm flow chart of the entire ScPace

[0094] (18) Table 1 shows the statistical results of the simulated dataset and the real-time scRNA-seq dataset. Each row corresponds to a different dataset. Sim1 and Sim2 are class-balanced datasets with different trajectories, linearity, and bifurcation. Sim3 corresponds to an unbalanced dataset on a bifurcated path. Genes and cells are listed according to the preprocessed dataset.

[0095] (19) Table 2 shows the enhancement effect of supervised pseudo-time analysis based on reclassification on the original real-time sequence scRNA-seq dataset

[0096] (20) Table 3 shows the performance improvement results of the supervised pseudo-time analysis on the original noise dataset and the ScPace on the deleted dataset using the real sequence scRNA-seq dataset.

[0097] (21) Table 4 shows the hyperparameters used for grid search adjustment during cross-validation experiments.

[0098] (22) Table 5 shows the hyperparameters used for grid search tuning, while performing timestamp calibration and supervised pseudo-time analysis.

[0099] (23) Table 6 shows the calculation time of timestamp calibration of simulated time series scRNA-seq with exchanged mislabeling.

[0100] (24) Table 7 shows the calculation time of timestamp calibration for the simulated time series scRNA-seq with random mislabeling.

[0101] (25) Table 8 shows the computation time for timestamp calibration on real time series scRNA-seq.

Claims

1. A timestamp correction method based on time-series scRNA-seq data, characterized by: The following steps are included: Step 1: Input time-series scRNA-seq data and give the timestamp of each cell; Step 2: Based on the multi-classification SVM model, the training model introduces noise indicator latent variables to determine whether the timestamp of each cell is mislabeled; Step three: Determine whether to reclassify or delete the latent variables based on their training results to achieve the corresponding correction effect.

2. The time stamp correction method for time series scRNA-seq according to claim 1, characterized in that: The mechanism of introducing noise-indicating latent variables can select samples in a more confident manner, rather than randomly extracting samples marked as errors by probability.

3. The time stamp correction method for time series scRNA-seq according to claim 1, characterized in that: The method can be embedded in other downstream model tasks or enhance the performance of related models, such as supervised pseudo-time series.

4. The time stamp correction method for time series scRNA-seq according to claim 1, characterized in that: The method can be applied in various real-world scenarios, such as revealing the regeneration of myocardial cells after injury.

5. The time stamp correction method for time series scRNA-seq according to claim 1, characterized in that The proposed method is highly robust to high-dimensional and imbalanced time-series scRNA-seq data. Therefore, this suggests that the proposed method can be effectively applied to various time-series scRNA-seq datasets for timestamp calibration.

6. The time stamp correction method for time series scRNA-seq according to claim 1, characterized in that :The method was trained and tested on multiple public datasets and simulated datasets, including GSE90047, GSE67310, GSE98664, and E-MTAB-3929, to verify its correction performance in scRNA-seq at different time series.

Images