Morphology-based noninvasive early prediction of serial-passage potency enhances the selection of clone-derived high-potency cell bank from mesenchymal stem cells

Rapidly expanding clones (RECs) are one of the single-cell-derived mesenchymal stem cell clones sorted from human bone marrow mononuclear cells (BMMCs), which possess advantageous features. The RECs exhibit long-lasting proliferation potency that allows more than 10 repeated serial passages in vitro, considerably benefiting the manufacturing process of allogenic MSC-based therapeutic products. Although RECs aid the preparation of large-variation clone libraries for a greedy selection of better-quality clones, such a selection is only possible by establishing multiple-candidate cell banks for quality comparisons. Thus, there is a high demand for a novel method that can predict "low-risk and high-potency clones" early and in a feasible manner given the excessive cost and effort required to maintain such an establishment.

We successfully developed a noninvasive morphology-based machine learning model to enhance the efficiency of establishing cell banks with single-cell-derived RECs for quantitatively predicting the future serial-passage potencies of clones. Conventional methods that can make noninvasive and quantitative predictions without wasting precious cells in the early stage are lacking; the proposed method will provide a more efficient and robust cell bank establishment process for allogenic therapeutic product manufacturing.

LNGFR and Thy-1 co-positive cells from BMMCs were single-cell-sorted into 96-well plates, and only fast-growing clones that reached confluency in 2 weeks were picked up and passaged as RECs. Fifteen RECs were prepared as passage 3 (P3) cryostock as the primary cell bank. From this cryostock, RECs were passaged until their proliferation limitation; their serial-passage limitation numbers were labeled as serial-passage potencies. At the P1 stage, phase-contrast microscopic images were obtained over 6-90 h to identify time-course changes of 24 morphological descriptors describing cell population information. Machine learning models were constructed using the morphological descriptors for predicting serial-passage potencies. The time window and field-of-view-number effects were evaluated to identify the most efficient image data usage condition for realizing high-performance serial-passage potency models.

Serial-passage test results indicated variations of 7-13-repeated serial-passage potencies within RECs. Such potency values were predicted quantitatively with high performance (RMSE < 1.0) from P1 morphological profiles using a LASSO model. The earliest and minimum effort predictions require 6-30 h with 40 FOVs and 6-90 h with 15 FOVs, respectively.