Senescence screening of mesenchymal stromal cells by morphology-based deep learning senescence analysis

Mesenchymal stromal cells (MSCs) therapy has emerged as a new approach for regenerative medicine. However, with the expansion of MSCs, passaging-induced in vitro senescence inevitably leads to impaired cellular quality and reduced their regenerative capacity, thereby limiting therapeutic efficacy. Until now, there is no quick and effective method monitoring the senescence of clinic-grade MSCs to ensure their quality before clinic use. Here, we introduced a morphology-based deep learning senescence analysis system to identify the passage number of MSCs to evaluate the senescence state of MSCs from pre-trained ResNeXt network. We used some specific molecular markers to identify aged MSCs, which exhibited flat and hypertrophic senescent-like morphology, decreased proliferation ability and immunomodulatory function. Especially, we used the expression of Ki67, RPS6, and R-galactosidase (R-gal) to identify senescent cells with increasing cell generations. Simultaneously, we used the ResNeXt network as the classification network, which could analyze the images of human umbilical cord mesenchymal stromal cells (HUCMSCs) and human bone marrow mesenchymal stromal cells (HBMMSCs) automatically. We found that culture generation could induce in vitro natural senescence of HUCMSCs, exhibiting flat and hypertrophic morphology, decreased proliferation ability and immunomodulatory function, increased aging markers including Ki67(-), pRPS6(+), SA-R-gal(+), P16, P21, and p53 expressions. Thus the number of cell passage represents the aging degree of cells. Subsequently, we successfully evaluate the senescence state of MSCs from pre-trained ResNeXt network optimized for classifying the natural cellular senescence. Our system identified the passage number and types of MSCs with high accuracy.