Abstract: Imaging flow cytometry combines the high-throughput capabilities of conventional flow cytometry with single-cell imaging. Here we demonstrate label-free prediction of DNA content and quantification of the mitotic cell cycle phases by applying supervised machine learning to morphological features extracted from brightfield and the typically ignored darkfield images of cells from an imaging flow cytometer. This method facilitates non-destructive monitoring of cells avoiding potentially confounding effects of fluorescent stains while maximizing available fluorescence channels. The method is effective in cell cycle analysis for mammalian cells, both fixed and live, and accurately assesses the impact of a cell cycle mitotic phase blocking agent. As the same method is effective in predicting the DNA content of fission yeast, it is likely to have a broad application to other cell types.
Introduction
To understand how cells grow, divide, or respond to drugs, scientists must determine where a cell is in its cycle (G1, S, G2, or Mitosis). Traditionally, this requires staining cells with fluorescent dyes that bind to DNA. However, these stains come with a "biological tax": they can be toxic to the cells, often require cell fixation (killing the cells), and consume limited detection channels on flow cytometers.
This research presents a significant leap in cell monitoring by combining high-throughput imaging with machine learning. The researchers demonstrate that a cell's internal state can be accurately predicted using only brightfield and darkfield images, bypassing the need for traditional fluorescent dyes.
Methodology
The team utilized Imaging Flow Cytometry (IFC), a technology that captures high-resolution images of thousands of individual cells as they zoom through a fluidic channel. The workflow involves three main stages: data acquisition, feature extraction, and machine learning classification.
Feature Extraction and Machine Learning
Using the open-source software CellProfiler, the researchers extracted 213 distinct morphological features from both brightfield (standard light) and darkfield (scattered light) images. These features include size, shape, texture, and granularity.
To process this data, they employed two specific machine learning algorithms:
- LSboosting (Regression): Used to predict the continuous value of DNA content.
- RUSboosting (Classification): Used to identify specific mitotic phases. This algorithm is particularly effective at handling "class imbalance," ensuring that rare mitotic cells are not ignored in favor of more common G1 or G2 cells.
Key Results and Insights
The study validated this label-free method across several biological contexts, including fixed and live human Jurkat cells and fission yeast (Schizosaccharomyces pombe).
- High Accuracy: In fixed Jurkat cells, the correlation between predicted DNA content and actual stained DNA content was remarkably high ($r = 0.896$).
- Mitotic Precision: The model achieved a 100% true positive rate for identifying cells in the anaphase and telophase stages of mitosis.
- Live-Cell Success: Even in live cells, the model maintained a strong correlation ($r = 0.786$) and successfully detected the impact of Nocodazole, a drug that blocks cells in the mitotic phase.
- The Power of Darkfield: While often ignored, the darkfield channel proved to be a goldmine for texture and granularity data, which is essential for distinguishing subtle cell states.
Conclusion
The physical structure of a cell is so tightly coupled to its DNA content that machine learning can learn to see the latter by looking at the former. This "morphology as a proxy" approach is non-destructive, cost-effective, and generalizable across species. While the method currently requires an initial annotated dataset for training and does not yet support real-time physical sorting, it paves the way for a new generation of stain-free cytometers capable of identifying complex phenotypes like cell death or differentiation.
Summary updated