Image-based discrimination of the early stages of mesenchymal stem cell differentiation

Abstract

Mesenchymal stem cells (MSCs) are self-renewing, multipotent cells, which can be used in cellular and tissue therapeutics. MSCs cell number can be expanded in vitro, but premature differentiation results in reduced cell number and compromised therapeutic efficacies. Current techniques fail to discriminate the "stem-like" population from early stages (12 h) of differentiated MSC population. Here, we imaged nuclear structure and actin architecture using immunofluorescence and used deep learning-based computer vision technology to discriminate the early stages (6-12 h) of MSC differentiation. Convolutional neural network models trained by nucleus and actin images have high accuracy in reporting MSC differentiation; nuclear images alone can identify early stages of differentiation. Concurrently, we show that chromatin fluidity and heterochromatin levels or localization change during early MSC differentiation. This study quantifies changes in cell architecture during early MSC differentiation and describes a novel image-based diagnostic tool that could be widely used in MSC culture, expansion and utilization.

FIGURE 1:

Morphological markers lipid droplet and cytoskeleton were stained with Nile red and rhodamine phalloidin, respectively. (A) Lipid droplets were stained with Nile red (red) and nuclei were stained with Hoechst 33342 (blue) at different timepoints during MSC adipogenesis. (B) F-actin was stained with rhodamine phalloidin (red) and nuclei were stained with Hoechst 33342 (blue) at different timepoints during MSC adipogenesis. The scale bar is 20 μm.

FIGURE 2:

Heterochromatin markers H3K9me3 and H3K27me3 were assessed with widefield immunocytochemistry. (A) Example immunostaining images of heterochromatin markers H3K9me3 (first row, green) and H3K27me3 (second row, green) in MSC adipogenesis at different timepoints. The scale bar is 20 μm. (B) Quantification of mean fluorescent intensity (MFI) for H3K9me3 stained cells (n = 62). Statistical analysis was performed with the Kruskal–Wallis test. ns, not significant; *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001. (C) Quantification of MFI from H3K27me3 stained cells (n = 62). Statistical analysis was performed with one-way ANOVA, P > 0.05, ns; *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001.

FIGURE 3:

Chromatin mobility was decreased at 24-h after adipogenic differentiation. (A) Representative images of cell nuclei expressing mCherry2-TRF1. From left to right, the top row represents the mCherry, Hoechst, and merged channels, respectively. The bottom row represents the removal of nucleus movement, the possibility map of “real” puncta within the cell nucleus (Red, yellow, and green represent the least, intermediate, and highest possibility, respectively) and the trajectory of puncta movements over time. (B) Comparisons of CD for control (Time = 0 h) and 24-h adipo are displayed on a log-log scale. N = 12. (C) Parameters associated with chromatin fluidity decrease are shown for control and 24-h adipo (reduced fluidity is assessed from chromatin mobility). Parameters for force generated by the actin-myosin were statistically similar between control and 24-h adipo. ns, not significant; *, P< 0.05.

FIGURE 4:

Detection of MSC differentiation stages using deep learning classification and regression. (A) Classification models were trained to predict the timepoints of segmented cell images in the training dataset, and the accuracy measured for the test set. Nuclear images are more informative than actin images for training a small network with low learnable parameters. Combining actin and nuclei provides a small accuracy boost of ∼2%. Brightfield images do significantly better than random (20%) but not as well as nuclear images. Larger models with more learnable parameters (MobileNet) perform similarly to small networks. Results are also shown for traditional image analysis using nuclear area or texture. (B) Using just the nucleus dataset, a regression model was trained to predict the time post differentiation using the training set, and treatment time was predicted for each nucleus in the test set. Results are for 271-462 cells per timepoint (see Materials and Methods).

FIGURE 5:

Machine learning pipeline for predicting and characterizing differentiation time periods of mesenchymal stem cells. (A) Input F-actin and chromatin images are segmented into regions containing individual nuclei, matched with the corresponding F-actin region. Datasets containing nuclei only, actin only, and combined nuclei and actin images are created Brightfield dataset is also created with the same region as the nuclei. (B) All datasets are processed with a set of CNN approaches.