Machine Learning-based Prediction of Stemness and Differentiation Trajectories in Mesenchymal and Pluripotent Stem Cells Using Stemformatics Datasets
Zhadyra Zhalgasbaeva
A
BSTRACT
Background:
Stemness maintenance and controlled differentiation are central challenges in regenerative medicine. Predicting whether stem cell samples remain stem-like or progress toward lineage commitment can support stem cell-based therapy, disease modeling, and tissue engineering. Curated transcriptomic resources such as Stemformatics provide suitable datasets for computational modeling of stem cell states.
Objective:
This project aimed to develop a machine learning-based framework for predicting stemness and differentiation states in pluripotent and mesenchymal stem cell systems using Stemformatics transcriptomic datasets. The study also aimed to compare induced pluripotent stem cell (iPSC) and mesenchymal stromal cell (MSC) differentiation dynamics and identify genes and pathways associated with stemness loss and lineage commitment.
Methods:
Two Stemformatics datasets were analyzed: Ang (2016), representing iPSC differentiation toward cardiac lineage, and Dani (2012), representing MSC-related osteogenic differentiation. Expression matrices were preprocessed separately because the datasets were generated using different transcriptomic platforms. Biological labels were assigned based on cell type annotations. Random Forest, Support Vector Machine (SVM), and Logistic Regression models were trained to classify stem-like and differentiated states. Model performance was evaluated using accuracy, F1-score, receiver operating characteristic–area under the curve (ROC-AUC), and confusion matrix analysis. Stemness scores were calculated as the predicted probability of the stem-like class. Principal component analysis (PCA), ROC curves, stemness score distribution plots, differentiation trajectory plots, feature selection, and functional enrichment analysis were used to interpret differentiation dynamics and regulatory genes.
Results:
After preprocessing, the iPSC dataset contained 17 samples and 22,938 genes, whereas the MSC dataset contained 13 samples and 21,743 genes, with no missing values remaining after filtering. PCA showed separation of samples according to differentiation state. In the iPSC dataset, stem, early differentiation, and late differentiation samples formed distinct groups. In the MSC dataset, stem-like MSC and differentiated osteogenic samples were also distinguishable. Machine learning models showed strong classification performance for stem-like versus differentiated states. Random Forest, SVM, and Logistic Regression achieved ROC-AUC values of 1.000 in both the datasets, with MSC classification accuracy of 0.923 and F1-score of 0.857. Stemness score and trajectory analyses showed decreasing stemness during differentiation. Feature selection and functional annotation identified genes and pathways associated with stemness maintenance, cardiac differentiation, osteogenic differentiation, and lineage commitment.
Conclusion:
The findings demonstrate that machine learning can be used to predict stemness and differentiation states from curated Stemformatics transcriptomic datasets and can support comparative analysis of iPSC and MSC differentiation dynamics.