In these comparisons, we addressed three technical questions, with the aim of identifying the most practical scheme for obtaining such cell-quality prediction models in clinical facilities. First, can morphology-based prediction Cabozantinib methods be expanded to the prediction of multiple differentiation potentials? Second, is morphological information of greater use than gene-expression information in predicting the qualities of hBMSCs? Third, how far can we optimize model performance by selecting the appropriate conversion and combination of information from the time-course morphological features? To our great surprise, considering the current lack of comparable evaluation methods, most of the examined prediction models using only morphological features showed practically useful performance in multiple predictions. Even with the Model 9, the multilineage potential prediction was available. Practically, potential II can be predicted with high accuracy using only morphological data from the first 4 days of culture. Both potentials I and III could also be predicted with reasonable accuracy from the early morphological data. In addition to differentiation rates, future PDT following repeated passages can also be predicted with high accuracy using only morphological features. These results strongly indicate that it will be possible to develop practical methods for cell assessment that are multiple, rapid, cheap, non-invasive, and significantly more effective than conventional staining-based assessment techniques. Our models’ performance indicate that such novel predictive methods will enjoy LY2835219 CDK inhibitor several advantages: non-invasiveness, i.e., avoiding damage to patients’ cells; synchronism, repeated quality evaluation throughout the culture period for all patients; and multivalent consideration of the same sample, i.e., multiple quality assessments can be performed with the same sample, which is not possible when using data obtained by destructive methods such as fluorescently labeled imaging analysis. The quantitative predictions made possible by these methods will permit prior evaluation of cellular fate, which will in turn facilitate scheduling of cell-therapy operations in the clinic. As shown in Fig. 3, most of the transition events in hBMSC potentials were abrupt, and would be nearly impossible to estimate the future linearly from the present result plots. Therefore, conventional cellassessment techniques could never outperform quantitative prediction methods for hBMSC quality assessment. Our results thus provide a successful example of the use of machine-learning models to model biological information and generate output that can overcome a major practical problem in clinical cell therapy. Taken together with the non-linear correlation of conventional marker gene-expression levels with passage numbers and the predictive performance of models, we concluded that morphological data 
from the early stage of culture are more useful than measurements of conventional markers in forecasting future quality disruptions. In some cases, gene-expression measurement enhanced morphological predictions, when an early gene marker such as SPP1 occasionally function as extreme early osteogenesis predictor. However, differentiation gene markers are not always promising to function as extreme early predictor in the undifferentiation stage.