Method and system for evaluating cell differentiation stage based on Raman spectrum

Abstract

The present invention provides a method of assessing the stage of cell differentiation. The method comprises the following steps: (a) providing a spectral data set of a single cell, wherein the spectral data set comprises Raman spectral data of the single cell at multiple differentiation stages; (b) performing data preprocessing on the Raman spectrum data in the spectrum data set to obtain a preprocessed spectrum data set; and (c) taking the preprocessed spectral data set as a training data set, and constructing the classification model for evaluating the cell differentiation stage through a machine learning algorithm. The invention also provides a classification model for evaluating a cell differentiation stage, and a method and a system for evaluating a differentiation stage of donor cells based on the model. The method can be used for quality control of cell culture passage, is especially suitable for high-accuracy and rapid evaluation of a donor-crossing cell stage, and has an excellent practical application prospect.

Description

Technical Field

[0001] The present application relates to the field of biometric data analysis, and more particularly to a method and system for evaluating cell differentiation stages based on Raman spectroscopy. More particularly, the present invention provides a machine learning model for evaluating cell differentiation stages and a method for evaluating cell differentiation stages using the machine learning model. Background Art

[0002] Following small molecule drugs and macromolecule drugs, cell therapy has become a new means of treating diseases in recent years. There is a very important issue that needs to be resolved before cells are used in clinical practice: cells with complex physiological changes require strict quality control and safety assessment. The main methods for evaluating cell quality focus on detecting mRNA and protein levels related to cell function. However, these methods require a large number of samples and take a long time, with large differences between batches. The results reflect a trend in the detection group and cannot be achieved at the single cell level. It is also difficult to capture subtle changes in the complex physiological processes of cell proliferation and differentiation, and the sensitivity is low for quality control. Using spectral technology, single cell information can be collected in a short period of time, requiring very small amounts of samples and high repeatability.

[0003] However, it is a huge challenge to evaluate the stage of cell differentiation due to subtle differences between generations of cell differentiation, heterogeneity of donors, spectral noise, etc. Currently, there is no method to use large sample libraries to perform pattern recognition of cell functions, exclude individual differences, and accurately evaluate the cell status of samples from unknown donors.

[0004] Therefore, there is an urgent need to develop a method and system that can evaluate the differentiation stage of donor cells with high accuracy, simplicity and rapidity. Summary of the invention

[0005] The object of the present invention is to provide a method and system (or device) that can evaluate the differentiation stage of donor cells with high accuracy, simplicity and rapidity.

[0006] In a first aspect of the present invention, there is provided a method for establishing a classification model for assessing the cell differentiation stage, the method comprising:

[0007] (a) providing a single-cell spectral dataset, the spectral dataset comprising Raman spectral data of single cells at n differentiation stages from m donors; wherein m is an integer ≥3, n is an integer ≥2; the number of cells at each differentiation stage is at least p, and p is an integer ≥100;

[0008] (b) performing data preprocessing on the Raman spectral data in the spectral data set in step (a) to obtain a preprocessed spectral data set; and

[0009] (c) Using the preprocessed spectral data set as a training data set, and constructing the classification model for evaluating the cell differentiation stage through a machine learning algorithm.

[0010] In another preferred embodiment, the data preprocessing is selected from the following group: spike removal processing, baseline correction processing, data normalization processing, or a combination thereof.

[0011] In another preferred example, the data preprocessing includes spike removal processing, baseline correction processing, and data normalization processing.

[0012] In another preferred embodiment, the data preprocessing further includes: spectral line correlation screening processing.

[0013] In another preferred embodiment, the spectral line correlation screening process refers to eliminating spectral lines with correlation lower than 0.9.

[0014] In another preferred embodiment, p is an integer ≥500, preferably p is an integer ≥1000, and more preferably p is an integer ≥2000.

[0015] In another preferred embodiment, the single cell comes from multiple single cells in a culture container.

[0016] In another preferred embodiment, the single cell comes from a plurality of single cells in a culture stage in a culture container.

[0017] In another preferred embodiment, the n differentiation stages include primary generation, passage 1 to passage t, and t is a positive integer of 2-100.

[0018] In another preferred embodiment, the method comprises: using a test data set to verify the classification model constructed in step (c) for assessing the cell differentiation stage.

[0019] In another preferred example, the training data set is selected from spectral data sets of cells at n differentiation stages from i donors, wherein i is a positive integer ≥1.

[0020] In another preferred embodiment, the verification includes: evaluating the performance of the model.

[0021] In another preferred embodiment, evaluating the performance of the model includes evaluating an indicator selected from the following group: precision, recall, specificity, and accuracy.

[0022] In another preferred embodiment, the Raman spectrum includes a wave number range of 300-1800 cm -1 Raman spectrum.

[0023] In another preferred embodiment, the Raman spectrum includes a wave number range of 320-1780cm -1Raman spectrum.

[0024] In another preferred embodiment, the Raman spectrum in step (a) is a partial or complete Raman spectrum curve (or Raman spectrum curve data) within the wavenumber range.

[0025] In another preferred embodiment, in step (b), the method further includes: performing quality control evaluation on the preprocessed spectral data set.

[0026] In another preferred embodiment, the quality control assessment includes assessing cellular heterogeneity of Raman spectral data in the spectral dataset.

[0027] In another preferred embodiment, the cell heterogeneity is selected from the following group: cell heterogeneity at different differentiation stages and cell heterogeneity from multiple donors.

[0028] In another preferred example, the cell heterogeneity assessment is performed by calculating the centroid distance.

[0029] In another preferred embodiment, the cell heterogeneity assessment comprises:

[0030] (S1) Dimensionality reduction processing;

[0031] (S2) Calculation of centroid distance; wherein the centroid distance is calculated using the following formula:

[0032]

[0033] Among them, (x i ,y i ) is the coordinate of the Raman spectrum of cell i after t-SNE dimension reduction, are the coordinates of the center of mass.

[0034] In another preferred example, the dimensionality reduction method includes stochastic neighbor embedding (t-Distributed Stochastic Neighbor Embedding, t-SNE).

[0035] In another preferred embodiment, the machine learning algorithm for constructing the cell classification model is selected from the following group: support vector machine, random forest, logistic regression, naive Bayes, neural network.

[0036] In another preferred embodiment, step (b) further comprises: performing characteristic peak difference evaluation on the preprocessed spectral data set.

[0037] In another preferred embodiment, the characteristic peak difference assessment includes identifying differential characteristic peaks in data sets of cells at different differentiation stages.

[0038] In another preferred embodiment, the characteristic peak difference evaluation includes:

[0039] (i) Plotting of average spectrum;

[0040] (ii) sliding window processing;

[0041] (iii) summing and pooling processing;

[0042] (iv) Statistical tests to screen the treatments.

[0043] In another preferred embodiment, the statistical test screening process refers to screening characteristic peaks with p value < 0.05 and effect value > 0.2.

[0044] In another preferred example, the characteristic peak includes a characteristic peak from a biomarker.

[0045] In another preferred embodiment, the biomarker is selected from the group consisting of lipids, amino acids, carbohydrates, proteins, or a combination thereof.

[0046] In another preferred embodiment, the biomarker is selected from the following group: glycogen, tyrosine, polysaccharide, phenylalanine, phospholipid, hydroxyproline, collagen, tryptophan, lipid, amide III, carbonyl in lipid, or a combination thereof.

[0047] In another preferred embodiment, the characteristic peak of the biomarker is selected from the following group: 480±5cm -1 、831±5cm -1 , 840±5 to 860±5cm -1 、933±5cm -1 、1003±5cm -1 、1080±5cm -1 、1172±5cm -1 、1206±5cm -1 、1265±5cm -1 、1300±5cm -1 、1337±5cm -1 、1440±5cm -1 、1660±5cm -1 、1744±5cm -1 , or a combination thereof.

[0048] In another preferred embodiment, the method further comprises: evaluating the per-plate accuracy used in the test, wherein the per-plate accuracy refers to the number of correctly evaluated cells in a single culture dish / the total number of cells in a single culture dish.

[0049] In another preferred embodiment, the method further comprises: adopting a sample size plan to determine the optimal number of donors and the optimal data volume of the training data set.

[0050] In another preferred embodiment, m≥4, preferably ≥5.

[0051] In another preferred embodiment, n≥3, more preferably ≥4.

[0052] In another preferred embodiment, the single cells belong to the same type or kind of cells.

[0053] In another preferred embodiment, the cells are selected from the group consisting of: hepatocytes, pancreatic cells, cardiomyocytes, nerve cells, muscle cells, intestinal cells, and adipocytes.

[0054] In another preferred embodiment, the cells are hepatocytes.

[0055] In another preferred embodiment, the differentiation stage of the cells is selected from: primary, 1st to 20th generations (preferably 1st to 15th generations, more preferably 1st to 10th generations).

[0056] In another preferred embodiment, the differentiation stage of the cells is selected from: primary, first, second, third, fourth, fifth, sixth, seventh, eighth and ninth generations.

[0057] In another preferred embodiment, the differentiation stage of the cells is selected from: primary, first generation, fourth generation, and ninth generation.

[0058] In another preferred embodiment, in the differentiation stage of the cells, two adjacent generations are separated by 2, 3 or 4 generations (except that the primary generation and the first generation are separated by 1 generation).

[0059] In another preferred embodiment, the cells are selected from the group consisting of: hepatocytes, pancreatic cells, cardiomyocytes, nerve cells, muscle cells, intestinal cells, and adipocytes.

[0060] In another preferred embodiment, the cells are hepatocytes.

[0061] A second aspect of the present invention provides a method for evaluating the stage of cell differentiation, the method comprising:

[0062] (s1) providing Raman spectral data from donor cells; the number of cells of the donor is at least q, wherein q is a positive integer ≥50;

[0063] (s2) inputting the Raman spectroscopy data into a classification model constructed using the method as claimed in claim 1;

[0064] (s3) evaluating the Raman spectroscopy data using the classification model to obtain an evaluation result of the differentiation stage of the donor's cells; and

[0065] (s4) Outputting the evaluation result.

[0066] In another preferred embodiment, q is a positive integer ≥80, preferably q is a positive integer ≥100, and more preferably q is a positive integer ≥200.

[0067] In another preferred embodiment, the method is non-diagnostic and non-therapeutic.

[0068] A third aspect of the present invention provides a system for evaluating the differentiation stage of donor cells, characterized in that the system comprises:

[0069] (M1) an input module, the input module being configured to input Raman spectrum data of cells of a donor;

[0070] (M2) a cell differentiation stage evaluation module, wherein the evaluation module is configured to evaluate the donor's cells based on a classification model for evaluating the cell differentiation stage, thereby obtaining an evaluation result of the differentiation of the donor's cells; wherein the classification model for evaluating the cell differentiation stage is based on the Raman spectrum of a single cell, and the classification model for evaluating the cell differentiation stage is constructed by a machine learning algorithm;

[0071] Alternatively, the classification model for evaluating the cell differentiation stage is a classification model for evaluating the cell differentiation stage constructed by the construction method as claimed in claim 1; and

[0072] (M3) An output module, configured to output the evaluation result.

[0073] In another preferred embodiment, the input module is selected from the following group: a Raman spectrometer, a scanner, a keyboard, a tablet computer (PAD), a smart phone, or a combination thereof.

[0074] In another preferred embodiment, the cell differentiation stage assessment module comprises a processor and a storage device, wherein the Raman spectrum data is stored in the storage device; and / or

[0075] The output modules include a display, a printer, a tablet computer (PAD), and a smart phone.

[0076] In another preferred embodiment, the cells are selected from the group consisting of: hepatocytes, pancreatic cells, cardiomyocytes, nerve cells, muscle cells, intestinal cells, and adipocytes.

[0077] In another preferred embodiment, the cells are hepatocytes.

[0078] In another preferred embodiment, the Raman spectrum includes a wave number range of 300-1800 cm -1 Raman spectrum.

[0079] In another preferred embodiment, the Raman spectrum includes a wave number range of 320-1780cm -1Raman spectrum.

[0080] In another preferred embodiment, the Raman spectrum in step (s1) or (M1) is a partial or complete Raman spectrum curve (or Raman spectrum curve data) within the wavenumber range.

[0081] It should be understood that within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features specifically described below (such as embodiments) can be combined with each other to form a new or preferred technical solution. Due to space limitations, they will not be described one by one here. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A general flow chart showing the single-donor cell recognition model and the cross-donor cell recognition model.

[0083] Figure 2 Showing the identification process of single donor cells.

[0084] Figure 3 The identification process across donor cells is shown.

[0085] Figure 4 Raman spectra of primary human hepatocytes (PHH) and proliferating hepatocytes (passage 1 (P1), 4 (P4), and 9 (P9)) are shown. (A) Schematic diagram of the study, including Raman spectra measurement, data preprocessing, and classification model construction. (B) Average Raman spectra obtained from PHH, P1, P4, and P9 from six different donors. (CD) Raman spectra were visualized by t-SNE to obtain two-dimensional differentiation images, where colors represent different cell stages (C) and donor origins (D). (EF) Comparison of Raman band intensities between different cell stages from six donors was performed to quantify various biomolecular components, including (E) 1208cm -1 (Hydroxyproline) and (F) 1440cm -1 (Lipids), from left to right, PHH, P1, P4, and P9 generations, respectively. Student's t-test with Bonferroni correction was used. Asterisks indicate the level of statistical significance test: *P < 0.05, **P < 0.01, ns not significant.

[0086] Figure 5Figure 2 shows the differential changes in Raman peaks between primary and proliferating hepatocytes. (A) Average Raman spectra acquired from six donors, representing PHH, P1, P4, P9. Spectra are annotated with Raman bands indicating differences in cell stages. (B) Heatmap of aggregated results of differentially changing peaks between two stages, where the color indicates the number of donors that produced significant results. (C-D) Raman band intensities between different cell stages from six donors were quantified to understand various biomolecular components, including (C) 480cm -1 (glycogen), (D) 933cm -1 (ν(CC) skeleton) collagen skeleton). Student's t-test with Bonferroni correction was used. Asterisks indicate the level of statistical significance test: *P < 0.05, **P < 0.01, ns not significant.

[0087] Figure 6 Intra-donor cell stage classification is shown. (A) Intra-stage and (B) inter-stage heterogeneity is quantified using centroid distance. Intra-stage refers to cells within the same cell group, e.g., PHH. Inter-stage refers to cells between two different cell groups, e.g., PHH and P1. (CF) Performance of classification models trained on cells from each donor individually. (C) Accuracy, (D) Precision, (E) Recall, and (F) Specificity. Error bars represent 95% confidence intervals.

[0088] Figure 7 The classification model across donor cell stages is shown. (A) Quantification of cross-donor heterogeneity by centroid distance. (B) Proportion of data variance explained by cell stage and donor origin. (CF) Performance of the classification model trained on cells from five donors and tested on another donor. (C) Accuracy, (D) Precision, (E) Recall, and (F) Specificity. Error bars represent 95% confidence intervals.

[0089] Figure 8 Sample size planning showing the experimental design. (A) Fraction of correctly and incorrectly assessed cells for a given dish. (BM) Three sample size planning strategies were performed to show how the single dish accuracy changes as a function of the number of (BE) training donors, (FI) training cells, and (JM) test cells. The minimum sample size is at the inflection point of the rising curve. DETAILED DESCRIPTION

[0090] After extensive and in-depth research, the inventors have developed for the first time a method that can accurately evaluate the differentiation stage of donor cells by analyzing the Raman spectral data of cells based on machine learning. Specifically, the inventors unexpectedly discovered that before constructing a classification model for cell differentiation stages, the data sets of different cell differentiation stages from different donors were analyzed and compared, and the differences between cell stages and between different donors were fully evaluated. Based on the differences, a model that can accurately evaluate the differentiation type of cells across donors was designed, and the sampling amount is small and the operation is simple. The present invention was completed on this basis.

[0091] The following is a description of the process for evaluating the differentiation type of cells in the present invention.

[0092] The overall flow chart of the present invention is shown in Figure 1 .like Figure 1 As shown, the evaluation process of the present invention can be divided into the main steps of data collection, data processing and model building, etc. The following will explain them respectively.

[0093] Data Collection

[0094] In the present invention, although other cell spectral data acquisition methods that meet the requirements of the present invention may also be used, Raman spectral signals of cells at different differentiation stages measured by a Raman spectrometer are mainly used as training data and test data.

[0095] Raman spectroscopy is a non-destructive analytical technique based on the interaction between light and the internal chemical bonds of materials. Raman spectroscopy can provide detailed information on the sample's chemical structure, phase and morphology, crystallinity, and molecular interactions.

[0096] In the present invention, conventional techniques in the art are used to collect single-cell Raman spectroscopy data at different cell stages from different donor cells.

[0097] Data processing

[0098] In the present invention, the collected Raman spectrum data are processed as follows: data preprocessing, quantitative analysis (data visualization), and characteristic peak identification.

[0099] Data preprocessing: remove poor quality data, such as hole burning effect, equipment false signal, defocus, impurity signal; eliminate data spikes caused by particles generated by high-energy cosmic particles on CCD or complementary metal oxide semiconductor detectors; baseline correction, remove baseline offset caused by substrate or fluorescence signal; data normalization, eliminate the influence of excitation intensity fluctuations or focus changes.

[0100] Quantitative analysis (data visualization): In the present invention, before establishing the cell classification model, in order to perform differential analysis on the heterogeneity between different donor cells, the centroid distance was selected to quantify the heterogeneity. Due to the curse of high dimensionality, the Raman spectral features were projected into two-dimensional space through t-Distributed Stochastic Neighbor Embedding (t-SNE) for distance calculation. The specific calculation formula is as follows:

[0101]

[0102] Where (x i ,y i ) is the coordinate of the Raman spectrum of cell i after t-SNE dimension reduction, is the coordinate of the center of mass. The difference in cell state is calculated by the distance between the center of mass of cells in the same state, and the calculation formula is as follows where n s is the number of cells at a given cell stage, Calculated as The heterogeneity between cell states was calculated using the distance between the centroids of different cell stages. in and Represents the number of cells in stage A and stage B, and is the distance between the centroids of cells at stage A and stage B, respectively. The heterogeneity across donors is the distance between the centroids of cells at the same stage but from different donors, expressed as Where n is the number of terms to be summed, D = {D k}(k=1,2,…), are the coordinates of the center of mass of donor k.

[0103] The degree to which the donor and cell stage explained the overall data variance was obtained by variance decomposition analysis, in which donor and cell stage were used as explanatory variables, while the internal centroid distance of the full data set was used as the response variable.

[0104] Characteristic peak identification: In the present invention, in order to obtain the characteristic peaks of differences between cell generations, the Raman spectrum is subjected to sliding window and summing pooling processing. The spectral signals of different windows are analyzed by Student T test with Bonferroni correction to determine the difference peaks.

[0105] Identification of cell differentiation stages

[0106] The inventors' research shows that identifying cell differentiation stages can be classified as a multi-classification problem.

[0107] Support vector machine (SVM) aims to find a hyperplane that can separate positive and negative samples from a high-dimensional space. Since SVM is used for binary classification tasks, it needs to be transformed using certain strategies for multi-classification tasks. Two typical strategies for converting binary classification tasks into multi-classification tasks are One-vs-rest (OvR) and One-vs-one (OvO). OvR builds a model for each category, with samples of the corresponding category treated as positive examples and remaining samples as negative examples. For N categories, N models will be obtained, and the category with the highest score will be used as the prediction result; OvO builds models by combining different categories in pairs. For N categories, N(N-1) / 2 models will be obtained, and the category with the most votes will be used as the prediction result.

[0108] When evaluating the model, the accuracy, precision, recall, and specificity are calculated for each category, and the performance evaluation results of multiple categories are summarized through macro-average and micro-average; the overall accuracy is also calculated.

[0109]

[0110]

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119] TP i ,FP i ,TN i and FN i is the number of true positives, false positives, true negatives, and false negatives of category i; N is the number of categories.

[0120] Single donor classification model: The flowchart of the single donor classification model of the present invention is shown in Figure 2.like Figure 2 As shown, in the single-donor internal classification task, the Raman spectra of the given donor cells were used to evaluate the model performance by nested cross-validation. The inner loop is responsible for tuning the hyperparameters (i.e., cost and smoothing parameter) through grid search, and the outer loop is responsible for model evaluation.

[0121] Cross-donor classification model: The flowchart of the cross-donor classification model of the present invention is shown in Figure 3 .like Figure 3 As shown in Figure 2, in the cross-donor classification task, the data from a given donor is retained as an independent test set. Cross-validation is applied to find the best hyperparameter configuration using the rest of the data in the donor pool. The performance is finally evaluated on the data from the retained donor.

[0122] Sample size planning (SSP): In order to guide clinical use, a systematic sample size plan (SSP) is used to estimate the impact of sample size on the classification model. SSP aims to determine the minimum number of measurements required for statistical analysis. For biological Raman spectroscopy, it is very important to maintain efficiency and reliability. A learning curve (LC)-based method is usually used, which measures the performance change of the model as the amount of data increases. In the present invention, there are three types of SSP for cross-donor tasks, with single-disc accuracy as the model indicator, and the formula is as follows:

[0123]

[0124] Among them, n i represents the number of correctly predicted cells in plate i, N i Represents the total number of cells in plate i.

[0125] Cell differentiation stage assessment method and system

[0126] The present invention provides a method for establishing a method for evaluating a cell differentiation stage, a method for evaluating a cell differentiation stage, and a device.

[0127] A typical method for assessing the cell differentiation stage is as described in the first aspect of the present invention, the method comprising:

[0128] (a) providing a spectral dataset of a single cell, the spectral dataset comprising Raman spectral data from a single cell at multiple differentiation stages;

[0129] (b) performing data preprocessing on the Raman spectral data in the spectral data set in step (a) to obtain a preprocessed spectral data set; and

[0130] (c) Using the preprocessed spectral data set as a training data set, and constructing the classification model for evaluating the cell differentiation stage through a machine learning algorithm.

[0131] A typical method for assessing the cell differentiation stage is as described in the second aspect of the present invention, the method comprising:

[0132] (s1) providing Raman spectroscopy data from donor cells;

[0133] (s2) inputting the Raman spectrum data into a classification model constructed by the method described in the first invention of the present invention;

[0134] (s3) evaluating the Raman spectroscopy data using the classification model to obtain an evaluation result of the differentiation stage of the donor's cells; and

[0135] (s4) Outputting the evaluation result.

[0136] A typical system (or device) for evaluating the differentiation stage of donor cells comprises:

[0137] (M1) an input module, the input module being configured to input Raman spectrum data of cells of a donor;

[0138] (M2) a cell differentiation stage evaluation module, wherein the evaluation module is configured to evaluate the donor's cells based on a classification model for evaluating the cell differentiation stage, thereby obtaining an evaluation result of the differentiation of the donor's cells; wherein the classification model for evaluating the cell differentiation stage is based on the Raman spectrum of a single cell, and the classification model for evaluating the cell differentiation stage is constructed by a machine learning algorithm;

[0139] Alternatively, the classification model for evaluating the cell differentiation stage is a classification model for evaluating the cell differentiation stage constructed by the method of the first invention of the present invention; and

[0140] (M3) An output module, configured to output the evaluation result.

[0141] In the present invention, the Raman spectrum data of cells can be input manually or automatically. Typically, the input module is selected from the following group: a Raman spectrum collector, a scanner, a keyboard, a tablet computer (PAD), a smart phone, or a combination thereof.

[0142] Preferably, in the present invention, the cell differentiation stage assessment module comprises a processor and a storage device, wherein the storage device stores Raman spectroscopy data.

[0143] In the present invention, representative output modules include (but are not limited to): a display, a printer, a tablet computer (PAD), and a smart phone.

[0144] The system for evaluating the differentiation stage of donor cells of the present invention can be in the form of an integrated device or a distributed device. For example, the input module can be independent, and the collected or input Raman spectrum data can be transmitted to the local processing module by wire or wireless means, or uploaded to a non-local processing module (for example, a server in a remote center) by wireless or communication means, thereby realizing remote evaluation.

[0145] In one embodiment, after the remote processing module has identified the Raman spectrum, the evaluation result can be wirelessly transmitted back to an output device connected to the network, such as a tablet computer (PAD) or a smart phone, thereby achieving rapid evaluation.

[0146] In a preferred example of the present invention, when a Raman spectrometer or other equipment is used and images are uploaded through the input terminal, there is basically no need or no need to rely on manual feature reading, thereby greatly liberating manpower, and greatly reducing or even eliminating the requirements for the level of Raman spectroscopy knowledge mastered by the user, and also helping to improve the accuracy of the evaluation.

[0147] Compared with the prior art, the main advantages of the present invention include:

[0148] (1) The method of the present invention has high accuracy in evaluating cross-donor cell stages and is an evaluation method that is independent of donors and has broader practical application prospects;

[0149] (2) The present invention can realize direct sampling and detection of single-disk cells. Compared with the detection of single cells, it has strong operability in large-scale production and manufacturing and has good industrial application prospects;

[0150] (3) The present invention utilizes spectroscopy technology to achieve rapid, non-invasive, and label-free cell assessment;

[0151] (4) The present invention can reflect the states of various biological molecules including sugars, nucleic acids, proteins, and lipids, and can provide a more detailed description of cell states and a higher sensitivity in distinguishing cell differentiation stages.

[0152] The present invention will be further described below in conjunction with specific examples. It should be understood that these examples are intended to illustrate the present invention only and are not intended to limit the scope of the present invention. The experimental methods in the following examples where specific conditions are not specified are usually performed under conventional conditions, such as those described in Sambrook et al., Molecular Cloning: A Laboratory Manual (New York: Cold Spring Harbor Laboratory Press, 1989), or under conditions recommended by the manufacturer. Unless otherwise indicated, percentages and parts are weight percentages and weight parts.

[0153] Example 1

[0154] 1.1 Cell culture

[0155] Primary human hepatocytes (PHH) were purchased from Corning (lot numbers: 416 and 393B) and Novabiosis (lot numbers: 005, 2011025-01, 2016789-01, and 1922211-01). PHH were cultured in OptiCulture Hepatocyte Medium (Novabiosis).

[0156] Proliferating human hepatocytes (ProliHHs) were induced according to the previously reported method (Zhang et al., 2018, 23(6), pp. 806-819; Ma et al., Stem Cell Res Ther (2021) 12:555). In brief, to generate ProliHHs generation 1 (P1), PHHs were treated with HM under hypoxic cell culture conditions (2% O2) for seven days. Using a similar method, generation 4 (P4) and generation 9 (P9) were obtained.

[0157] PHH, P1, P4, and P9 were seeded on 8-well chamber Raman scattering microslides for 24 hours. The cells were washed with PBS and then fixed in 4% paraformaldehyde for 15 minutes. To collect Raman spectra, the cells were washed with sterile water and evaporated naturally.

[0158] 1.2 Raman spectroscopy data acquisition

[0159] Raman spectroscopy was performed using a WITec alpha 300R Raman imaging microscope. In brief, the spectra were performed at an excitation wavelength of 532 nm and 100-fold air interference (NA0.9). To improve the signal-to-noise ratio of the spectra, a 1800 line / mm grating was used, with an integration time of 10 s and a laser power of 9 mW. The fingerprint region of the Raman spectrum (from 300 to 1800 cm -1). Ten Raman spectra were randomly detected for each cell. At least 60 cells were collected for each donor / batch. The spectral data were preprocessed as follows: quality control, calculating the correlation between each spectral line and the average spectrum, and removing the spectral lines with correlation less than 0.9; peak removal: by comparing with the average spectrum, detect points with abnormal intensity, and then use the detection value of the adjacent point to replace the intensity of the peak position; baseline correction: baseline correction by polynomial fitting; data normalization: finally normalize the data by vector normalization.

[0160] PHH from six donors (lots: 416, 393B, 005, 2011025-01, 2016789-01, and 1922211-01) were used to induce ProliHH. PHH, generation 1 (P1), generation 4 (P4), and generation 9 (P9) of ProliHH were collected to obtain Raman spectra. The fingerprint region used was 320 cm -1 Up to 1800cm -1 , consisting of 2214 Raman bands ( Figure 4 B), covering most vibration modes of biomolecules. At least 600 spectra of PHH, P1, P4 and P9 were obtained for each donor PHH, P1, P4 and P9, and more than 14,000 spectra were obtained after quality control.

[0161] Based on previous studies on donor 005 (Ma et al., Stem Cell Res Ther (2021) 12:555), during the transition from PHH to ProliHH, the Raman spectrum showed a peak at 1206 cm -1 (hydroxyproline) and 1440cm -1 Here, other donors were also evaluated for these changes ( Figure 4 EF). 1206cm -1 The intensity of the hydroxyproline band increased significantly from PHH to P9 ( Figure 4 E). During the proliferation of ProliHHs, 1440 cm -1 The Raman band intensity of lipids is significantly reduced ( Figure 4 F). This shows that the differentially shifted peaks are consistent across different donors.

[0162] 1.3 Characteristic peak identification

[0163] The Raman spectra were processed by sliding window and summing pooling. The window size was 20 sampling points and the step was 5 sampling points. The spectral signals in different windows were analyzed by Student T test with Bonferroni correction. If the p value was less than 0.05 and the effect value was greater than 0.2, the peak was determined to be a difference peak.

[0164] The summary results of different donors are as follows Figure 5 As shown in B. After artificial peak position assignment, several Raman bands with corresponding biological significance were identified ( Figure 5 A), including 480cm -1 (glycogen), 831cm -1 (tyrosine), 840–860 cm -1 (polysaccharide), 933cm -1 (v(CC) skeleton of collagen skeleton), 1003cm -1 (phenylalanine), 1080cm -1 (amide II, typical phospholipid), 1172 cm -1 (CH in the plane bending mode of tyrosine), 1206cm -1 (Hydroxyproline), 1265cm -1 (α-helix, collagen, tryptophan), 1300cm -1 (lipid), 1337cm -1 (amide III), 1440 cm -1 (CH2 and CH3 form vibration) lipids), 1660cm -1 (amide I of protein) and 1744 cm -1 (Carbonyl features of lipid spectra). Hepatocytes maintain glucose homeostasis through glycogen synthase and degradation, and 480 cm -1 The peak at P1 decreased from P1 to P9, which means that glycogen synthase is downregulated during cell dedifferentiation ( Figure 5 C) In addition, it was observed that from PHH to P9, there was a -1 The band at Figure 5 D).

[0165] 1.4 tSNE and Data Heterogeneity Measures

[0166] Further visualization results of Raman spectrum t-SNE are shown in Figure 4 CD. In terms of cell stages, PHH showed greater homogeneity and differences compared with other stages ( Figure 4 C). At the same time, there was little variation between donors ( Figure 4 D). This indicates that Raman spectroscopy can distinguish different cell stages, especially PHH and ProliHH; moreover, the differences caused by donor origin are controllable.

[0167] Centroid distances were used to quantify intra- and inter-stage heterogeneity. Figure 6As shown in A, the observed centroid distances within PHHs are smaller than those in other cell states, indicating that PHHs have less intra-stage heterogeneity. At the same time, the heterogeneity between PHHs and ProliHHs is significantly higher than the heterogeneity between different stages of ProliHHs ( Figure 6 B) These results suggest that during the transition from PHH to ProliHH, the growth rates and states of cells become more diverse.

[0168] 1.5 Machine learning to build a single donor classification model

[0169] Machine learning models were constructed to classify different cell types from the same donor. Since this is a multi-class task, OvR and OvO SVM models were applied and evaluated using 10 replicates of nested cross validation. OvO produced higher performance than OvR, indicating that OvO can better capture subtle differences between cell types (Table 1).

[0170] Table 1. Comparison of the accuracy of one-vs-one (OvO) and one-vs-rest (OvR) strategies in multi-classification SVM models

[0171]

[0172] Note: 95% confidence intervals are shown in brackets.

[0173] Figure 6 CF shows the performance of the OvO SVM model in each donor. The overall accuracy of the models across different donors ranged from 0.87 to 0.90. Since the number of cells in multiple datasets was roughly balanced, the micro-average and macro-average across multiple categories for each evaluation metric were almost identical. The macro-precision and micro-precision ranged from 0.87 to 0.90, the macro-recall and micro-recall ranged from 0.87 to 0.90, and the macro-specificity and micro-specificity ranged from 0.96 to 0.97. It is worth noting that there were fluctuations in performance across different cell types, with PHHs consistently achieving superior precision, recall, and sensitivity compared to other cell types.

[0174] 1.6 Machine Learning to Build a Cross-Donor Model

[0175] A practical application scenario in cell culture is to use Raman spectroscopy to determine the cell type of a new donor by using a classification model built on previously collected donors. Cross-donor classification is more challenging than the intra-donor task because the model must not only distinguish between different cell types but also reduce the batch effect and individual heterogeneity of the donors. The distances to the centroids of different cell types were quantified and it was found that P9 showed higher cross-donor heterogeneity ( Figure 7A). As cell culture time increases, the growth rate of cells may become more inconsistent, and cells collected on the same day may be a mixture of different passages. In addition, ANOVA was used to examine the contribution of cell type and donor origin to the variance in the data ( Figure 7 B). Notably, cell type accounted for 47.47% of the total variance, while donor origin accounted for only 0.83%. This sets the stage for cross-donor classification, as the variance between cell types exceeds that of donor origin. Therefore, cross-donor classification was evaluated, leaving one donor for independent testing and the rest for model training. Figure 7 As shown in CF, the accuracy of classification across donors decreased, ranging from 0.73 to 0.78. Precision, recall, and specificity also decreased slightly. In terms of cell type, PHH tended to achieve higher performance.

[0176] Due to donor heterogeneity, experimental environment, etc., cross-donor classification tasks are usually more difficult than single-donor classification tasks. As expected, the performance of each cell of the cross-donor classification model degrades. In practical scenarios, the actual requirement is to use Raman spectra to infer the cell stage from randomly selected cells in a given culture dish, rather than accurately infer the stage of each cell. Figure 8 A shows the results of the fraction of correctly and incorrectly assessed cells for a given culture dish. The distance between the positive and negative examples reflects the credibility of the model built. There is a clear distinction between them, especially for donors 005, 416, 2016789-01 and 393B, indicating that the cross-donor classification model is suitable for determining PHH and ProliHH passages of new donors. Therefore, the per-plate accuracy is defined, which is the percentage of cells correctly classified using the cross-donor classification model. When the data set is exhausted (where all cells from five donors are used to train the model to evaluate all cells from the remaining donors), the per-plate accuracy is 0.76.

[0177] 1.7 Sample size planning

[0178] Train the SSP for the number of donors and infer how many donors are needed in the training set to obtain good generalization performance to unseen donors. With k training donors and m total donors, there is possible iterative combinations. Then, SSP of the number of training cells was performed. The training cells were stratified and sampled from 5 donors, ranging from approximately 120 to 10,800. Each sampling process was repeated 10 times. Third, SSP of the number of test cells was performed. The test cells sampled from the test donors ranged from approximately 25 to 2,300. Each sampling process was repeated 10 times.

[0179] First, the lower limit of the donor library is confirmed. As the number of donors increases, the single-disk accuracy and model stability will improve ( Figure 8 B). Three donors would be the optimal choice to achieve the best performance in all cell types. Second, the training cells were downsampled to estimate the minimum number of training cells. In the downsampling, the classification accuracy of all cell types showed a clear trend of improvement. When the data size approached 9,600 cells (i.e., about 480 cells from four cell types from five donors), the learning curve reached a plateau, which means that the performance of the model will no longer increase significantly with more data ( Figure 8 FI). Finally, the minimum number of cells to test was determined. The learning curve began to converge around 200 cells, indicating that about 200 cells were sufficient to measure the overall phase of a dish that had never been seen before ( Figure 8 JM).

[0180] All documents mentioned in the present invention are cited as references in this application, just as each document is cited as reference individually. In addition, it should be understood that after reading the above teachings of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the claims attached to this application.

Claims

1. A method for establishing a classification model for evaluating cell differentiation stages, characterized in that The method comprises: (a) providing a single-cell spectral dataset, the spectral dataset comprising Raman spectral data of single cells at n differentiation stages from m donors; wherein m is an integer ≥3, n is an integer ≥2; the number of cells at each differentiation stage is at least p, and p is an integer ≥100; (b) performing data preprocessing on the Raman spectral data in the spectral data set in step (a) to obtain a preprocessed spectral data set; and (c) Using the preprocessed spectral data set as a training data set, and constructing the classification model for evaluating the cell differentiation stage through a machine learning algorithm.

2. The method according to claim 1, characterized in that The data preprocessing is selected from the following group: spike removal processing, baseline correction processing, data normalization processing, or a combination thereof.

3. The method according to claim 1, characterized in that p is an integer ≥500, preferably p is an integer ≥1000, more preferably p is an integer ≥2000.

4. The method according to claim 1, characterized in that The n differentiation stages include primary generation, passage 1 to passage t, where t is a positive integer of 2-100.

5. The method according to claim 1, characterized in that The method comprises: validating the classification model for assessing the cell differentiation stage constructed in step (c) using a test data set.

6. The method according to claim 1, characterized in that The Raman spectrum includes a wave number range of 300-1800 cm -1 Raman spectrum.

7. The method according to claim 1, characterized in that In step (b), it also includes: performing quality control evaluation on the preprocessed spectral data set.

8. The method according to claim 1, characterized in that The step (b) further includes: performing characteristic peak difference evaluation on the preprocessed spectral data set.

9. A method for evaluating the stage of cell differentiation, characterized in that The method comprises: (s1) providing Raman spectral data from donor cells; the number of cells of the donor is at least q, wherein q is a positive integer ≥50; (s2) inputting the Raman spectroscopy data into a classification model constructed using the method as claimed in claim 1; (s3) evaluating the Raman spectroscopy data using the classification model to obtain an evaluation result of the differentiation stage of the donor's cells; and (s4) Outputting the evaluation result.

10. A system for assessing the differentiation stage of donor cells, characterized in that The system comprises: (M1) an input module, the input module being configured to input Raman spectrum data of cells of a donor; (M2) a cell differentiation stage evaluation module, wherein the evaluation module is configured to evaluate the donor's cells based on a classification model for evaluating the cell differentiation stage, thereby obtaining an evaluation result of the differentiation of the donor's cells; wherein the classification model for evaluating the cell differentiation stage is based on the Raman spectrum of a single cell, and the classification model for evaluating the cell differentiation stage is constructed by a machine learning algorithm; Alternatively, the classification model for evaluating the cell differentiation stage is a classification model for evaluating the cell differentiation stage constructed by the construction method as claimed in claim 1; and (M3) An output module, configured to output the evaluation result.

Images