In particular, by interpreting the parameter usages chosen through automatic exploration of the best LASSO model formula, it was possible to clearly detect the correlation of cytoskeleton-related and senescence-related genes with the decrease of potentials. Such correlation was supported by previous studies showing that the TGF-b signaling cascade links cellular quality collapse with Dabrafenib morphological changes. Compared to the successful predictions of adipogenic differentiation potential and PDT, the predictive performance for osteogenic and chondrogenic differentiation could not be increased by altering the modeling techniques. We believe that the main reason for this limitation on performance did not reflect a shortcoming of our method. Machine-learning performance relies heavily on the quality of training data. In this work, we are uncertain of the quality of our ‘teacher signal’ data, i.e., the converted data from conventional MLN4924 Metabolic Enzyme/Protease inhibitor staining assays for evaluations of differentiation potential. In practice, the staining technique is usually used as only one aspect of differentiation confirmation, but is not commonly used for strict quantitative analysis. A growing body of evidence describes the quantitative use of immunohistochemical staining results in high-content analysis, analogous to the way in which we converted the staining results into numerical values. However, our machine-learning model results show that only Oil red, but not other staining values, results in excellent performances. In addition we have identified a few critical sources of experimental noise that might partially explain the comparatively poorer performance of osteogenic and chondrogenic differentiation. For example, in osteogenic staining with Alizarin red, small parts of stained cells tend to be ripped from the plate during the washing process, resulting in larger deviations within some wells. Similarly, in chondrogenic staining, the pellet size and its slice position greatly affect the staining level, resulting in larger deviations within some samples. Hence, the large difference between prediction models suggests that reproducibility and signal-to-noise ratio of staining results must be carefully examined in order for modeling to be effective. In other words, if one can introduce more stable staining, the prediction models should perform better. Our expression data regarding other types of genes lead us to expect that effective gene combinations could be defined as new quality assessment markers. The interpretation of weights of LASSO regression models can provide insights regarding essential parameters that contributed to successful predictions. For prediction of potential I, a combination of morphological features from the whole pre-differentiation period, together with expression of the cytoskeleton-related genes and the early osteogenic marker SPP1, were weighted. The decision to weight these genes reflected previous reports of interactions between osteogenic marker genes and cytoskeleton genes. Together, the selected morphological features support the commonly observed flat and expanded cellular morphology of hBMSCs, known as an indication of bone differentiation. This result can be interpreted to mean that the continuous evolution of hBMSCs during the expansion period toward a jagged morphology is the signature of adipogenic.