Deep learning–based quantitative prediction of MSC adipogenic potency from early transcriptional feature

Mesenchymal stem cells (MSCs) are increasingly explored for regenerative medicine and cell-based therapiess, yet substantial heterogeneity across tissue sources and manufacturing batches complicates potency testing and quality control. Current adipogenic assessment largely relies on long-cycle endpoint assays, limiting early decision-making during production. In this study, we performed single-cell transcriptomic profiling of bone marrow–derived MSCs (BM-hMSCs) before induction and after 7 days of osteogenic or adipogenic induction to identify adipogenic-associated candidate genes. We quantified early induction kinetics of these genes by real-time PCR (RT-qPCR) and integrating multi time point RT-qPCR features from MSCs line of different tissue origins with quantitative Oil Red O absorbance as a quantitative phenotypic target. We then developed MSC-FateNet (adipogenic-specific version) to predict adipogenic potential. In our preliminary dataset, the LSTM-based MSC-FateNet model achieved an R² of 0.9551, a root-mean-square error (RMSE) of 0.0440, and a mean absolute error (MAE) of 0.0277. However, these results were obtained from only seven mesenchymal stem cell lines and require further validation in larger independent datasets. This model supports predictive analysis using longitudinal gene expression profiles collected on days 0, 1, 3, and 7 after induction, reducing the evaluation period from 21 days to 7 days and thus outperforming traditional endpoint staining methods. When using only the data from days 0, 1, and 3 as input, the predictive performance drops markedly (R² = 0.683), indicating that expression data from the first 7 days are critical for achieving robust and reliable prediction. MSC-FateNet represents a proof-of-concept rapid, quantitative framework for early adipogenic potency assessment. With further validation on larger and more diverse MSC samples, it may potentially support quality control of MSC-based products. Currently, the model is limited to adipogenic differentiation and has not been validated for osteogenic or chondrogenic lineages.