For prediction of potential III, a shape factor on day 1, but not throughout the whole period of expansion culture, was weighted. This can be interpreted to mean that very early roundness of hBMSCs can indicate the potential for chondrogenesis. For prediction of differentiation potential before induction, morphogenic markers are sufficient because these data contain both time-course and multi-parametric information. However, in microarray experiments, we identified four genes related to osteogenic differentiation, three genes related to adipogenic differentiation, and one gene related to chondrogenic differentiation. From this background, these gene profiles were not sufficient to explain the difference in differentiation potential. As it happens, in prediction of potential I, PIK3CA and SPP1 were useful markers even before differentiation induction; however, for the other two cell types, there were no critical marker genes. For prediction of potential IV, both gene-expression profiles and morphological features were weighted. Cell cycle–related genes such as TP53 and CDKN1A, actin-related genes such as ACTA2 and IQGAP1, and the cellular senescence–related gene CDKN2 were heavily weighted in the best prediction models. Such gene selection reflects reports that indicate a correlation between senescence and cytoskeleton gene cascades in hBMSCs. The weighted morphological features reflect the common culture sense of regular size and slenderness during the expansion process are markers of active hBMSCs. To confirm the utility of our proposed method in clinics, its adaptive performance in the context of the cellular diversity is an important criterion. Cellular diversity derived from patient diversity should be the first concern; we have previously investigated a new modeling scenario, designated as the ”ongoing patient scenario”, as one strategy for adapting to such diversity. In this scenario, we proved that inclusion of new patients’ own morphological features in prediction models can greatly enhance the prediction accuracy for new patients. Thus, this proposed method avoids attempting to adapt to all cellular variations arising due to patient diversity, which is ideal but impossible, and instead seeks to ‘re-train’ the prediction model upon the arrival of each new patient, thereby allowing the system to adapt to new morphologies. Such re-training is Ruxolitinib feasible with our proposed regression models, and it is also feasible at various times in the clinic, because there are various opportunities to obtain images of primary cells before making critical potential predictions. Therefore, in this study, we attempted to optimize the adaptive performance of our model in the context of cellular diversity arising due to culture processes. In contrast to patient diversity, such culture process–derived diversity may expand during the cell production process, and may therefore require daily monitoring with non-invasive methods. Our results showed that even an extraordinarily diverse group of samples could be modeled feasibly using only morphological features. In addition, because our model performance is a LY2109761 summary of detailed crossvalidation results combined with our ongoing patient scenario, we believe that our results provide a reliable performance benchmark reflecting robustness. Throughout this work, we have successfully demonstrated models for predicting multiple future qualities of hBMSCs.