Mesenchymal stem cell adipogenic differentiation prediction model, construction method and application

Abstract

The application discloses a mesenchymal stem cell adipogenic differentiation prediction model, a construction method and application, and belongs to the field of biomedical engineering and cell therapy technology. The prediction model is a model constructed based on a long short-term memory network (LSTM), input of the model includes relative expression amount time series data of characteristic genes of the mesenchymal stem cells at multiple continuous time points in an early stage of adipogenic induction differentiation, and output of the model is a quantitative prediction value representing terminal adipogenic differentiation potential; the characteristic genes are one or more of FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN and PPARgamma. The prediction model can be used not only for quality control of MSC cell preparations, batch release detection, but also for screening of donors or tissue sources with high adipogenic potential, and optimization of adipogenic induction differentiation process, and has a wide application prospect.

Description

Technical Field

[0001] This invention belongs to the field of biomedical engineering and cell therapy technology, and particularly relates to a predictive model for adipogenic differentiation of mesenchymal stem cells, its construction method, and its application. Background Technology

[0002] Mesenchymal stem cells (MSCs) have shown great potential in regenerative medicine and cell therapy due to their multi-lineage differentiation potential, immune regulation, and tissue repair capabilities. However, the function of MSCs exhibits significant heterogeneity, stemming from differences in tissue origin, donor, passage number, and culture batch, which seriously affects the quality consistency of cell therapy products, the predictability of clinical efficacy, and treatment safety.

[0003] Currently, the "gold standard" method for assessing the adipogenic differentiation capacity of MSCs is to perform Oil Red O staining after 14-21 days of in vitro induction, and then perform qualitative or semi-quantitative evaluation by observing lipid droplets under a microscope or measuring the absorbance of the eluent. This method has the following significant drawbacks: Long cycle: It requires an induction culture of up to 21 days, which is difficult to meet the timeliness requirements for cell product release testing.

[0004] The process is cumbersome and highly subjective: staining results are easily affected by operation, field of view selection, and human judgment, resulting in low standardization.

[0005] Endpoint evaluation: It is impossible to predict cell fate and potential in the early stages of differentiation, which is not conducive to identifying problems and optimizing processes in the early stages of production.

[0006] Therefore, there is an urgent need to develop a new method for the early evaluation of the adipogenic differentiation potential of MSCs that is rapid, objective, quantitative, and standardized. Summary of the Invention

[0007] In view of this, the present invention discloses a predictive model for adipogenic differentiation of mesenchymal stem cells, its construction method and application.

[0008] The present invention adopts the following technical solution: A predictive model for adipogenic differentiation of mesenchymal stem cells is provided. The predictive model is based on a long short-term memory network (LSTM). The input of the predictive model includes time-series data on the relative expression levels of characteristic genes of the mesenchymal stem cells at multiple consecutive time points in the early stage of adipogenic differentiation. The output of the predictive model is a quantitative predicted value characterizing the terminal adipogenic differentiation potential. The multiple consecutive time points include the time points from day 0 before adipogenic differentiation to day 7 after induction.

[0009] Furthermore, the characteristic gene is one or more of FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, and PPARγ.

[0010] Furthermore, the multiple consecutive time points are time points among day 0, day 1, day 3, and day 7.

[0011] A method for constructing the above-mentioned prediction model, the method comprising the following steps: 1.1 Obtain mesenchymal stem cell training samples from multiple sources and in multiple batches; 1.2 All training samples underwent in vitro adipogenic differentiation induction; 1.3 Detect the relative expression levels of the feature genes of each training sample at multiple set time points to obtain time-series data of relative expression levels, which are used as input features of the model; 1.4 The absorbance value of Oil Red O staining for each training sample at the induction endpoint was detected and used as the output label of the model; 1.5 Construct a training dataset using the input features and output labels, and train it using an LSTM network until the model converges to obtain the prediction model.

[0012] Application of the above-mentioned prediction model in predicting the adipogenic differentiation potential of mesenchymal stem cells.

[0013] Furthermore, the method of application includes the following steps: 2.1 In vitro adipogenic differentiation induction of the test mesenchymal stem cells; 2.2 At multiple set time points during induced differentiation, the relative expression levels of the characteristic genes in the test cells were detected to obtain time-series data of the relative expression levels. The set time points included day 0, day 1, day 3, and day 7 before induction. 2.3 Input the relative expression level time series data into the prediction model; 2.4 Obtain the quantitative predicted value output by the prediction model.

[0014] Furthermore, the set time point mentioned in 2.2 is a time point among day 0, day 1, and day 3.

[0015] A kit for assessing the adipogenic differentiation potential of mesenchymal stem cells, the kit comprising RT-qPCR detection reagents for detecting the aforementioned characteristic genes.

[0016] Furthermore, the kit also includes a computer-readable medium or a remotely invoked model interface storing the prediction model of any one of claims 1-3.

[0017] The above-mentioned predictive model is applied in the quality control, batch release testing, donor screening, or differentiation process optimization of mesenchymal stem cells.

[0018] The present invention has the following significant advantages: Significantly shorten the evaluation cycle: The traditional 21-day endpoint testing cycle is greatly shortened to 7 days, or even to 3 days, which greatly improves the evaluation efficiency and makes it possible to release cell products quickly and control the process.

[0019] High prediction accuracy: The MSC-FateNet model (LSTM) constructed in this invention achieves excellent performance on the validation set (R0). 2 =0.9551, RMSE=0.0440), which can quantitatively predict terminal adipogenic differentiation potential with high accuracy.

[0020] Objective, quantitative, and standardized: Gene expression data and model predictions obtained through RT-qPCR replace the subjective judgment-dependent Oil Red O staining morphological observation, resulting in more objective and quantitative evaluation results that facilitate standardized comparisons between different laboratories and batches.

[0021] Capturing dynamic processes: The model utilizes the temporal expression information of eight characteristic genes in the early stages of differentiation, which aligns with the dynamic biological process of cell differentiation. This approach is more informative and advantageous than methods using static data from a single time point, significantly improving the accuracy of model predictions.

[0022] The invention has broad application prospects: It can be used not only for quality control and batch release testing of MSCs cell preparations, but also for screening donors or tissue sources with high adipogenic potential, as well as optimizing adipogenic differentiation processes. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram illustrating the construction and application process of the MSC-FateNet model of this invention; Figure 2 This is a schematic diagram comparing the prediction performance of datasets at different time points in this invention across three deep learning models; Figure 3 This is a schematic diagram of the Oil Red O staining results of three independent verification samples after 21 days of in vitro adipogenic induction. Detailed Implementation

[0025] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] It should be understood that the described embodiments are merely some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the reagents or methods involved in the following embodiments are all conventional reagents or experimental methods in the art.

[0027] Example 1

[0028] A predictive model for adipogenic differentiation of mesenchymal stem cells is provided. The predictive model is based on a long short-term memory network (LSTM). The input of the predictive model includes time-series data on the relative expression levels of characteristic genes of the mesenchymal stem cells at multiple consecutive time points in the early stage of adipogenic differentiation. The output of the predictive model is a quantitative predicted value characterizing the terminal adipogenic differentiation potential. The multiple consecutive time points include the time points from day 0 before adipogenic differentiation to day 7 after induction.

[0029] Furthermore, the characteristic gene is one or more of FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, and PPARγ.

[0030] The full names of the proteins or genes corresponding to the characteristic genes FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, and PPARγ are as follows: FABP4 (Fatty Acid Binding Protein 4), SCD (Stearoyl-CoA Desaturase), G0S2 (G0 / G1 Switch Gene 2), SRPX (Sushi Repeat Containing Protein X-linked), PDK4 (Pyruvate Dehydrogenase Kinase 4), SAT1 (Spermidine / Spermine N1-Acetyltransferase 1), FASN (Fatty Acid Synthase), and PPARγ (Peroxisome Proliferator-Activated Receptor). Gamma (peroxisome proliferator-activated receptor γ).

[0031] Furthermore, the multiple consecutive time points are time points among day 0, day 1, day 3, and day 7.

[0032] A method for constructing the above-mentioned prediction model, the method comprising the following steps: 1.1 Obtain mesenchymal stem cell training samples from multiple sources and in multiple batches; 1.2 All training samples underwent in vitro adipogenic differentiation induction; 1.3 Detect the relative expression levels of the feature genes of each training sample at multiple set time points to obtain time-series data of relative expression levels, which are used as input features of the model; 1.4 The absorbance value of Oil Red O staining for each training sample at the induction endpoint was detected and used as the output label of the model; 1.5 Construct a training dataset using the input features and output labels, and train it using an LSTM network until the model converges to obtain the prediction model.

[0033] Application of the above-mentioned prediction model in predicting the adipogenic differentiation potential of mesenchymal stem cells.

[0034] Furthermore, the method of application includes the following steps: 2.1 In vitro adipogenic differentiation induction of the test mesenchymal stem cells; 2.2 At multiple set time points during induced differentiation, the relative expression levels of the characteristic genes in the test cells were detected to obtain time-series data of the relative expression levels. The set time points included day 0, day 1, day 3, and day 7 before induction. 2.3 Input the relative expression level time series data into the prediction model; 2.4 Obtain the quantitative predicted value output by the prediction model.

[0035] Furthermore, the set time point mentioned in 2.2 is a time point among day 0, day 1, and day 3.

[0036] A kit for assessing the adipogenic differentiation potential of mesenchymal stem cells, the kit comprising RT-qPCR detection reagents for detecting the aforementioned characteristic genes.

[0037] Furthermore, the kit also includes a computer-readable medium or a remotely invoked model interface storing the prediction model of any one of claims 1-3.

[0038] The above-mentioned predictive model is applied in the quality control, batch release testing, donor screening, or differentiation process optimization of mesenchymal stem cells.

[0039] Example 2

[0040] An early prediction model for the adipogenic differentiation potential of mesenchymal stem cells, wherein the model is an MSC-FateNet model constructed based on a long short-term memory network (LSTM); Its input features are the relative expression levels of a specific set of adipogenic-related genes detected at multiple consecutive time points (at least before induction and day 7) in the early stage of adipogenic differentiation of mesenchymal stem cells; its output is a predicted value (e.g., predicted Oil Red O staining absorbance value) that can quantitatively characterize the terminal adipogenic differentiation potential of the mesenchymal stem cells.

[0041] The method for constructing the prediction model includes the following steps: 1. Screening of the feature gene set: 1.1 Obtain single-cell transcriptome sequencing data of bone marrow-derived MSCs before adipogenic induction and on day 7 after induction.

[0042] 1.2 Genes that are highly expressed specifically in adipogenic differentiation branches and show a significant positive correlation with pseudotime were screened by differential expression analysis and lineage-specific trajectory analysis.

[0043] 1.3 The core feature gene set was finally determined, including: FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, and PPARγ.

[0044] 2. Construction of the training dataset: 2.1 Collect MSCs samples from multiple sources (such as umbilical cord, fat, and dental pulp), multiple batches, and different passages.

[0045] 2.2 In vitro adipogenic differentiation was induced in each sample.

[0046] 2.3 Cells were collected at multiple time points in the early stage of induction (e.g., days 0, 1, 3, and 7), RNA was extracted, and the relative expression levels of the characteristic gene set were detected by RT-qPCR as input features of the model.

[0047] 2.4 At the induction endpoint (e.g., day 21), differentiated cells were stained with Oil Red O and eluted with isopropanol, and the absorbance (e.g., OD490) was measured as an output label for quantitative characterization of adipogenic differentiation potential.

[0048] 3. Model training and optimization: 3.1 Construct training and validation datasets using the input features and output labels.

[0049] 3.2 Construct an LSTM regression model, using gene expression time series data from multiple early time points as input and the endpoint Oil Red O absorbance value as output for training.

[0050] 3.3 By comparing the prediction performance of different time windows (such as 0-3 days, 0-7 days, and 0-14 days of data) and different models (such as LSTM, GRU, CNN, and random forest), the optimal model structure was determined to be LSTM, and the optimal prediction time windows were the 0th, 1st, 3rd, and 7th days after induction.

[0051] 3.4 After training, an LSTM model for predicting the adipogenic differentiation potential of MSCs, namely MSC-FateNet, is obtained.

[0052] A method for predicting the adipogenic differentiation potential of mesenchymal stem cells using the above model includes the following steps: 1. Obtain the mesenchymal stem cell sample to be tested.

[0053] 2. The test sample was subjected to in vitro adipogenic differentiation.

[0054] 3. Collect cells at early time points during induced differentiation (days 0, 1, 3, and 7).

[0055] 4. Detect the relative expression levels of the characteristic gene set (FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, PPARγ) in cells.

[0056] 5. Input the obtained gene expression time series data from multiple time points into the pre-trained MSC-FateNet model. When the gene expression time series data from multiple time points simultaneously includes the expression time series data of the feature gene sets FABP4, SCD, and G0S2, the prediction accuracy of the model can be significantly improved.

[0057] The model outputs a predicted value, which corresponds to the predicted final Oil Red O staining absorbance value. The level of this value represents the level of adipogenic differentiation potential of the sample to be tested.

[0058] like Figure 1 As shown, Figure 1 This is a schematic diagram illustrating the construction and application process of the MSC-FateNet model of this invention.

[0059] A kit for evaluating the adipogenic differentiation potential of mesenchymal stem cells, the kit comprising: RT-qPCR detection reagents (e.g., specific primer pairs) for detecting characteristic gene sets (FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, PPARγ).

[0060] Optionally, the kit may also include a storage medium or right to use the MSC-FateNet model, as well as a standard operating procedure instruction manual.

[0061] Preferred technical solution of the present invention: The mesenchymal stem cells mentioned in 1 are derived from sources including but not limited to umbilical cord, bone marrow, adipose tissue, and dental pulp.

[0062] The preferred early prediction time points are days 0, 1, 3, and 7 after induced differentiation.

[0063] The optimal input time window for the LSTM model described in section 3 is the time series data on days 0, 1, 3, and 7 after induction.

[0064] The predictive performance of the LSTM model described in section 4 can reach: coefficient of determination R0 2 ≥ 0.95, RMSE ≤ 0.05.

[0065] Example 3

[0066] Feature gene screening and dataset construction Feature gene screening: Single-cell transcriptome data of human bone marrow MSCs at days 0 and 7 of adipogenic induction were obtained from public databases.

[0067] Use toolkits such as OmicVerse for data quality control, dimensionality reduction, clustering, and trajectory inference based on pyVIA.

[0068] Gene expression dynamics in the differentiation trajectory of MSCs into the adipogenic lineage were analyzed. Genes that were highly positively correlated with pseudotime (corr>0.4) and lowly correlated in the osteogenic branch (corr<0.3) were screened. Based on the literature, the characteristic gene set was finally determined to be: FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, PPARγ.

[0069] Sample collection and processing: Seven MSCs from three different tissue sources (umbilical cord, fat, and dental pulp) were collected and amplified at different passages (e.g., P5, P8, P11).

[0070] All samples underwent in vitro adipogenic differentiation.

[0071] Data Acquisition: Input data: Cells were collected on days 0, 1, 3, and 7 of induction, RNA was extracted, and the relative expression levels of the above 8 characteristic genes were detected by RT-qPCR (with GAPDH as an internal control). The expression values ​​of 8 genes were obtained for each sample at each time point.

[0072] Label data: On day 21 of induction, differentiated cells were stained with Oil Red O, eluted with isopropanol, and the absorbance (OD490) was measured at 490 nm. A continuous value in the range of 0.2-1.2 was obtained for each sample as a quantitative label of its adipogenic differentiation potential.

[0073] Example 4

[0074] Model training and selection Data partitioning: The dataset of the above 7 MSCs was divided into a training set (5 strains) and a validation set (2 strains).

[0075] Model building: Construct three deep regression models: LSTM, GRU, and CNN.

[0076] The model input consists of the expression values ​​of eight genes for each sample at four time points: 0, 1, 3, and 7 days (i.e., a 4×8 time series matrix).

[0077] The model output is a scalar value, namely the predicted 21-day oil red OOD value.

[0078] Meanwhile, traditional machine learning models (SVM, XGBoost, RF, DT) were used for comparison based on data from single time points (days 3, 7, and 14).

[0079] Model training and evaluation: The Adam optimizer and mean squared error (MSE) are used as the loss function for training.

[0080] Compute R on the validation set 2 The performance of the model is evaluated using metrics such as RMSE and MAE.

[0081] result: Traditional models suffer from overfitting and poor generalization ability.

[0082] Among the three deep learning models, the LSTM model performed best. When the input data was from days 0, 1, 3, and 7, the validation set performance of the LSTM model reached: R0 2 =0.9551, RMSE=0.0440, MAE=0.0277. The predicted values ​​are in high agreement with the actual Oil Red O staining OD values. However, the performance is lower when using only 0, 1, and 3-day data (R² = 0.9551, RMSE = 0.0440, MAE = 0.0277). 2 =0.6830), and performance did not improve further when using 14 days of data (R = 0.6830). 2 =0.5618). Performance comparisons of the regression models are shown in Table 1. The prediction performance of the datasets at different time points in the three deep learning models is compared as follows: Figure 2 As shown.

[0083] Table 1 Performance comparison of deep regression models

[0084] Conclusion: The optimal model is LSTM, and the best early prediction time windows are days 0, 1, 3, and 7 after induction. This optimized LSTM model is named MSC-FateNet.

[0085] Example 5

[0086] Independent Sample Validation Verification process: Three additional umbilical cord MSCs from different donor sources were obtained.

[0087] Adipogenic induction, RT-qPCR detection (days 0, 1, 3, and 7), and Oil Red O quantitative detection (day 21) were repeated in Example 1.

[0088] Gene expression data from days 0, 1, 3, and 7 were input into the trained MSC-FateNet model to obtain predicted OD values.

[0089] The verification results are shown in Table 2: Table 2 Prediction performance of the model of this invention in independent validation samples.

[0090] The model's predicted OD values ​​for the three independent samples were very close to the actual detected OD values ​​(e.g., actual 0.6142 vs. predicted 0.6670; actual 0.0908 vs. predicted 0.1006; actual 0.5072 vs. predicted 0.5386), with small prediction errors, demonstrating that the model has good generalization ability and practical application value. Figure 3 This is a schematic diagram of the Oil Red O staining results of three independent verification samples after 21 days of in vitro adipogenic induction.

[0091] Example 6

[0092] A predictive model for adipogenic differentiation of mesenchymal stem cells is provided. The predictive model is a model constructed based on a long short-term memory network (LSTM). The input of the predictive model includes: time-series data of the relative expression levels of characteristic genes of the mesenchymal stem cells at multiple consecutive time points in the early stage of adipogenic differentiation. The output of the predictive model is a quantitative predicted value characterizing the terminal adipogenic differentiation potential. The multiple consecutive time points include the time points from day 0 before adipogenic differentiation to day 7 after induction.

[0093] Furthermore, the characteristic genes are one or more of FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, and PPARγ.

[0094] Furthermore, the multiple consecutive time points are time points among day 0, day 1, day 3, and day 7.

[0095] A method for constructing the above-mentioned prediction model, the method comprising the following steps: 1.1 Obtain mesenchymal stem cell training samples from multiple sources and in multiple batches; 1.2 All training samples underwent in vitro adipogenic differentiation induction; 1.3 The relative expression levels of the characteristic genes in each training sample at multiple set time points were detected to obtain relative expression time-series data, which were used as input features for the model. During the model training process, the relative expression time-series data of different combinations of characteristic gene sets were examined. The prediction model construction results showed that including the expression time-series data of the characteristic gene sets FABP4, SCD, G0S2, and SRPX significantly improved the model training effect and prediction accuracy. Furthermore, the model construction process also found that including CEBPA as a characteristic gene also improved the model's prediction stability and accuracy. CEBPA stands for CCAAT enhancer binding protein alpha, which corresponds to a CCAAT enhancer binding protein α.

[0096] 1.4 The absorbance value of Oil Red O staining for each training sample at the induction endpoint was detected and used as the output label of the model; 1.5 Construct a training dataset using the input features and output labels, and train it using an LSTM network until the model converges to obtain the prediction model.

[0097] The embodiments of the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A predictive model for adipogenic differentiation of mesenchymal stem cells, characterized in that, The prediction model is a model constructed based on a long short-term memory network. The input of the prediction model includes: time-series data of the relative expression levels of characteristic genes of mesenchymal stem cells at multiple consecutive time points in the early stage of adipogenic differentiation. The output of the prediction model is a quantitative prediction value characterizing the terminal adipogenic differentiation potential. The multiple consecutive time points include the time points from day 0 before adipogenic differentiation to day 7 after induction.

2. The prediction model according to claim 1, characterized in that, The characteristic gene is one or more of FABP4, SCD, G0S2, SRPX, PDK4, SAT1, FASN, and PPARγ.

3. The prediction model according to claim 2, characterized in that, The multiple consecutive time points are time points among day 0, day 1, day 3, and day 7.

4. A method for constructing a prediction model as described in any one of claims 1-3, characterized in that, The method includes the following steps: 1.1 Obtain mesenchymal stem cell training samples from multiple sources and in multiple batches; 1.2 All training samples underwent in vitro adipogenic differentiation induction; 1.3 Detect the relative expression levels of the feature genes of each training sample at multiple set time points to obtain time-series data of relative expression levels, which are used as input features of the model; 1.4 The absorbance value of Oil Red O staining for each training sample at the induction endpoint was detected and used as the output label of the model; 1.5 Construct a training dataset using the input features and output labels, and train it using an LSTM network until the model converges to obtain the prediction model.

5. The application of the prediction model as described in any one of claims 1-3 in predicting the adipogenic differentiation potential of mesenchymal stem cells.

6. The application of the prediction model according to claim 5, characterized in that, The method of application includes the following steps: 2.1 In vitro adipogenic differentiation induction of the test mesenchymal stem cells; 2.2 At multiple set time points during induced differentiation, the relative expression levels of the characteristic genes in the test cells were detected to obtain time-series data of the relative expression levels. The set time points included day 0, day 1, day 3, and day 7 before induction. 2.3 Input the relative expression level time series data into the prediction model; 2.4 Obtain the quantitative predicted value output by the prediction model.

7. The method according to claim 6, characterized in that, 2.2 The set time points are the time points of day 0, day 1, and day 3.

8. A kit for evaluating the adipogenic differentiation potential of mesenchymal stem cells, characterized in that, The kit contains RT-qPCR detection reagents for detecting the characteristic gene of claim 1.

9. The reagent kit according to claim 8, characterized in that, The kit also includes a computer-readable medium or a remotely invoked model interface storing the prediction model of any one of claims 1-3.

10. The application of the prediction model according to any one of claims 1-3 in the quality control, batch release testing, donor screening, or differentiation process optimization of the mesenchymal stem cells.

Images