COLLEMBOT: AI-based counting of Collembola for OECD 232 Tests.

Area of Science:

• Ecotoxicology
• Environmental Science
• Computational Biology

Background:

• Ecotoxicological tests with soil organisms are crucial for chemical risk assessment in terrestrial ecosystems.
• Current OECD 232 reproduction tests for Collembola rely on manual counting, which is labor-intensive, time-consuming, and prone to bias, limiting data availability.

Purpose of the Study:

• To develop an automated counting tool, COLLEMBOT, for Collembola reproduction tests.
• To integrate COLLEMBOT into existing OECD workflows without protocol modifications.
• To improve the efficiency and reproducibility of ecotoxicological data generation.

Main Methods:

• Development of COLLEMBOT, an automated counting tool utilizing a YOLOv11 convolutional neural network.
• Training the model on 3207 high-resolution images from multiple laboratories.
• Validation using 22 independent datasets (1704 images) from diverse international locations and standard soil types.

Main Results:

• Automated counts demonstrated strong agreement with manual counts (R² = 0.79-0.99).
• Dose-response curves and effect concentrations (EC10, EC50) derived from automated and manual counts showed minimal differences, remaining within regulatory limits.
• Processing time was reduced by approximately 97%, from 137 hours for manual counting to under 3 hours for automated counting.

Conclusions:

• COLLEMBOT provides a reliable and efficient automated solution for counting Collembola in ecotoxicological studies.
• The tool significantly reduces labor and improves reproducibility, enabling broader hazard data generation for collembola.
• Public availability of the code and workflow encourages adoption and community-driven development for enhanced chemical risk assessment.