Predicting cell cycle stage from 3D single-cell nuclear-stained images.

Area of Science:

• Cell Biology
• Biomedical Research
• Machine Learning Applications

Background:

• Cell cycle dynamics are crucial for understanding eukaryotic cell proliferation, differentiation, and regeneration.
• Current cell cycle profiling methods are complex and can interfere with experimental outcomes.
• Efficient cell cycle annotation is vital for research in development, health, aging, and disease.

Purpose of the Study:

• To develop an efficient, cost-effective machine learning workflow (CellCycleNet) for cell cycle staging from microscopy data.
• To simplify cell cycle annotation with minimal experimenter intervention.
• To enable accurate cell cycle phase prediction using only DAPI staining in fixed interphase cells.

Main Methods:

• Developed CellCycleNet, a machine learning workflow utilizing DAPI-stained nuclei in fixed interphase cells.
• Trained support vector machine (SVM) and deep neural network models using the Fucci2a reporter system as ground truth.
• Utilized two benchmarking image datasets for model training and validation.
• Compared performance using 3D image features versus 2D projections.

Main Results:

• CellCycleNet accurately predicts cell cycle phase (G1 vs. S/G2) using only DAPI staining.
• The workflow outperforms state-of-the-art SVM models.
• Simultaneous training on multiple datasets yielded the highest classification accuracy (AUROC improvement of 0.08-0.09).
• The model demonstrated excellent generalization across different microscopes (AUROC of 0.95).
• 3D image features significantly improved classification performance over 2D projections.

Conclusions:

• CellCycleNet offers a simplified and accurate method for cell cycle profiling from microscopy images.
• The machine learning approach reduces experimental complexity and cost.
• The developed models and data are released as a community resource to advance cell cycle research.