Abstract
Mesenchymal stromal cells (MSCs) are multipotent cells that have great potential for regenerative medicine, tissue repair, and immunotherapy. Unfortunately, the outcomes of MSC-based research and therapies can be highly inconsistent and difficult to reproduce, largely due to the inherently significant heterogeneity in MSCs, which has not been well investigated. To quantify cell heterogeneity, a standard approach is to measure marker expression on the protein level via immunochemistry assays. Performing such measurements non-invasively and at scale has remained challenging as conventional methods such as flow cytometry and immunofluorescence microscopy typically require cell fixation and laborious sample preparation. Here, we developed an artificial intelligence (AI)-based method that converts transmitted light microscopy images of MSCs into quantitative measurements of protein expression levels. By training a U-Net+ conditional generative adversarial network (cGAN) model that accurately (mean $r_s = 0.77$) predicts expression of 8 MSC-specific markers, we showed that expression of surface markers provides a heterogeneity characterization that is complementary to conventional cell-level morphological analyses. Using this label-free imaging method, we also observed a multi-marker temporal-spatial fluctuation of protein distributions in live MSCs. These demonstrations suggest that our AI-based microscopy can be utilized to perform quantitative, non-invasive, single-cell, and multi-marker characterizations of heterogeneous live MSC culture. Our method provides a foundational step toward the instant integrative assessment of MSC properties, which is critical for high-throughput screening and quality control in cellular therapies.
Introduction
Mesenchymal stromal cells (MSCs) are often referred to as the "Swiss Army knives" of regenerative medicine due to their ability to repair tissues and modulate immune responses. However, MSCs are notoriously heterogeneous; cells from the same donor or batch can exhibit vastly different behaviors, leading to inconsistent clinical outcomes. Understanding this diversity is critical for the success of MSC-based therapies.
Traditionally, scientists measure surface proteins using immunofluorescence, a process that requires "fixing" (killing) the cells and staining them with antibodies. This destructive process prevents the monitoring of live cell dynamics over time and renders the analyzed cells unusable for therapy. While non-invasive morphological analysis is possible, cell shape alone does not always correlate strongly with biological function.
Methods: AI-Driven Virtual Staining
To address these challenges, researchers developed a deep learning pipeline for image-to-image translation. The system utilizes a U-Net generator paired with a conditional Generative Adversarial Network (cGAN) discriminator. The U-Net captures fine structural details from phase-contrast microscopy, while the cGAN ensures the "virtually stained" output is statistically indistinguishable from real fluorescent images.
The model was trained on a dataset of 600–1500 pairs of phase-contrast and immunofluorescent images for each of eight markers, including standard MSC markers (CD105, CD90, CD73) and markers for migration (CD29), vascular commitment (CD146), and immunomodulation (CD106, STRO-1).
Results: Quantitative Accuracy and Dynamic Heterogeneity
The model achieved a mean Pearson correlation coefficient ($r_s$) of 0.77 across all eight markers, demonstrating high fidelity to ground-truth chemical stains. Beyond binary presence, the AI accurately predicted sub-cellular localization, distinguishing between membrane and perinuclear regions.
Multi-Marker Integration
Because each marker is trained individually, the researchers can stack predictions to create a composite proteomic profile from a single light image. This avoids the "spectral overlap" issues common in traditional multi-color fluorescence.
Key Insights
- Orthogonal Information: Hierarchical clustering revealed that protein expression provides information complementary to morphology. For instance, while some markers correlate with aspect ratio, others do not correlate with cell area, suggesting that "looking" at a cell's shape is insufficient for a full biological assessment.
- Stochastic Fluctuations: By applying the model to 48-hour time-lapse videos, the authors observed that protein levels fluctuate over time with a correlation timescale of approximately 40 minutes. This highlights the importance of continuous monitoring over single-snapshot assessments.
Discussion and Future Work
This AI-based platform allows for the real-time, simultaneous tracking of multiple cellular properties without the need for destructive fixatives or expensive dyes. While the "black box" nature of convolutional neural networks remains a challenge for interpretability, the system's robustness—maintaining performance even with low signal-to-noise ratios—makes it a promising tool for industrial quality control.
Future efforts will focus on accounting for donor-to-donor heterogeneity and integrating automated cell segmentation to scale the platform for high-throughput screening in cellular therapies.
Summary updated