Label-free morphology-based prediction of multiple differentiation potentials of human mesenchymal stem cells for early evaluation of intact cells

Precise quantification of cellular potential of stem cells, such as human bone marrow-derived mesenchymal stem cells (hBMSCs), is important for achieving stable and effective outcomes in clinical stem cell therapy. Here, we report a method for image-based prediction of the multiple differentiation potentials of hBMSCs. This method has four major advantages: (1) the cells used for potential prediction are fully intact, and therefore directly usable for clinical applications; (2) predictions of potentials are generated before differentiation cultures are initiated; (3) prediction of multiple potentials can be provided simultaneously for each sample; and (4) predictions of potentials yield quantitative values that correlate strongly with the experimental data. Our results show that the collapse of hBMSC differentiation potentials, triggered by in vitro expansion, can be quantitatively predicted far in advance by predicting multiple potentials, multi-lineage differentiation potentials (osteogenic, adipogenic, and chondrogenic) and population doubling potential using morphological features apparent during the first 4 days of expansion culture. In order to understand how such morphological features can be effective for advance predictions, we measured gene-expression profiles of the same early undifferentiated cells. Both senescence-related genes (p16 and p21) and cytoskeleton-related genes (PTK2, CD146, and CD49) already correlated to the decrease of potentials at this stage. To objectively compare the performance of morphology and gene expression for such early prediction, we tested a range of models using various combinations of features. Such comparison of predictive performances revealed that morphological features performed better overall than gene-expression profiles, balancing the predictive accuracy with the effort required for model construction. This benchmark list of various prediction models not only identifies the best morphological feature conversion method for objective potential prediction, but should also allow clinicians to choose the most practical morphology-based prediction method for their own purposes.