Automated comet assay analysis using YOLOv5-based deep learning.

Area of Science:

• Biomolecular Sciences
• Computational Biology
• Toxicology

Background:

• The comet assay is crucial for single-cell DNA damage detection.
• Manual scoring is slow, subjective, and limits scalability for high-throughput studies.
• Current methods face challenges in standardization and efficiency.

Purpose of the Study:

• To develop and validate a deep learning system for automated comet assay image classification.
• To overcome the limitations of manual and semi-automated scoring methods.
• To enhance accuracy, reproducibility, and processing speed in DNA damage analysis.

Main Methods:

• A YOLOv5 object detection model was trained on 875 expert-annotated comet assay images.
• Hyperparameter tuning and data augmentation were employed for performance optimization.
• Model evaluation utilized mAP, precision, recall, and confusion matrix analysis.

Main Results:

• The YOLOv5 model achieved a high mean Average Precision (mAP@0.5) of 0.98 and recall over 0.8.
• The system demonstrated robust learning, generalization capacity, and high classification accuracy.
• Despite challenges with overlapping comets and class imbalance, the model showed improved scalability and speed.

Conclusions:

• The YOLOv5-based system provides a scalable and accurate automated solution for comet assay analysis.
• This AI approach significantly increases throughput and reduces human error in genotoxicity testing.
• Future work will address overlapping structures and real-world laboratory integration.