Background: Within the extensively developed therapeutic application of mesenchymal stem cells (MSCs), allogenic immunomodulatory therapy is among the promising categories. Although donor selection is a critical early process that can maximize the production yield, determining the promising candidate is challenging owing to the lack of effective biomarkers and variations of cell sources. In this study, we developed the morphology-based non-invasive prediction models for two quality attributes, the T-cell proliferation inhibitory potency and growth rate.
Methods: Eleven lots of mixing bone marrow-derived and adipose-derived MSCs were analyzed. Their morphological profiles and growth rates were quantified by image processing by acquiring 6 h interval time-course phase-contrast microscopic image acquisition. T-cell proliferation inhibitory potency was measured by employing flow cytometry for counting the proliferation rate of peripheral blood mononuclear cells (PBMCs) co-cultured with MSCs. Subsequently, the morphological profile comprising 32 parameters describing the time-course transition of cell population distribution was used for explanatory parameters to construct T-cell proliferation inhibitory potency classification and growth rate prediction models. For constructing prediction models, the effect of machine learning methods, parameter types, and time-course window size of morphological profiles were examined to identify those providing the best performance.
Results: Unsupervised morphology-based visualization enabled the identification of anomaly lots lacking T-cell proliferation inhibitory potencies. The best performing machine learning models exhibited high performances of predictions (accuracy > 0.95 for classifying risky lots, and RMSE < 1.50 for predicting growth rate) using only the first 4 days of morphological profiles. A comparison of morphological parameter types showed that the accumulated time-course information of morphological heterogeneity in cell populations is important for predicting the potencies.
Conclusions: To enable more consistent cell manufacturing of allogenic MSC-based therapeutic products, this study indicated that early non-invasive morphology-based prediction can facilitate the lot selection process for effective cell bank establishment. It was also found that morphological heterogeneity description is important for such potency prediction. Furthermore, performances of the morphology-based prediction models trained with data consisting of origin-different MSCs demonstrated the effectiveness of sharing morphological data between different types of MSCs, thereby complementing the data limitation issue in the morphology-based quality prediction concept.
Keywords: Cell manufacturing; Mesenchymal stem cells; Morphological analysis; Quality control; T-cell proliferation inhibitory potency.
© 2022. The Author(s).