A feature learning model identifies predictive attributes of mesenchymal stromal cell efficacy for osteoarthritis therapy
Abstract
Human mesenchymal stromal cells (hMSCs) have shown inconsistent clinical benefit, in part because donor-to-donor variability obscures the cellular attributes that drive therapeutic potency. To address this challenge, we screened hMSCs from multiple donors and integrated in vitro secretome profiling with efficacy data in an in vivo preclinical osteoarthritis model; we identified cellular paracrine signatures that distinguish more-therapeutic from less-therapeutic hMSCs. We trained a feature learning model relating secretome to in vivo structural outcomes and, in a prospective proof-of-concept study, used it to classify additional donors, with in vivo outcomes directionally consistent with predictions. Transcriptomic and phospho-protein analyses showed elevated JNK pathway activation in more-therapeutic hMSCs. Pharmacologic JNK inhibition using a small molecule inhibitor SP600125 in more-therapeutic cells reduced therapeutic cytokine secretion, while metformin treatment reported to increase phosphorylation, shifted the secretome of less-therapeutic hMSCs towards a more-therapeutic profile. Together, these findings define potency-associated attributes of hMSC efficacy and establish a prospective predictive framework for donor screening and potency optimization. This approach addresses a major barrier in scalable cell therapies and advances precision hMSC-based treatments.