COLLEMBOT. AI-based counting of Collembola for OECD 232 tests

Abstract

Ecotoxicological tests with soil organisms, such as the collembola Folsomia candida, are essential for assessing chemical risks in terrestrial ecosystems. However, the current Organization for Economic Co-operation and Development (OECD) 232 reproduction tests rely on manual counting of juvenile and adult Collembola, a process that is costly, labor-intensive, time-consuming and prone to operator bias. These limitations restrict data availability and hinder robust risk assessments. We therefore developed COLLEMBOT, an automated counting tool based on a YOLOv11 convolutional neural network, designed to integrate seamlessly into OECD workflows without protocol modifications. The model was trained on high-resolution images (n = 3207) from multiple laboratories and validated using 22 independent datasets (n = 1704 images) from Amsterdam (Netherlands), Basel (Switzerland), Bayreuth (Germany), Coimbra (Portugal) and Aarhus (Denmark). Datasets consisted of relevant standard soils (OECD artificial soils with 2.5%, 5% and 10% sphagnum peat; LUFA 2.2) and the springtail Folsomia candida. Automated counts showed strong agreement with manual counts (R² = 0.79–0.99). Dose-response curves derived from automated and manual counts strongly overlapped and effect concentrations (EC10 and EC50) differed minimally (Median %Δ 6.2 ± 23 and EC10–EC90 R2 ≥ 0.977), remaining within acceptable limits for regulatory risk assessment and confirming reliability. Time efficiency improved significantly: a test with approximately 300 images and up to 1,500 individuals per image was processed in less than 3 hr, compared to approximately 137 hr needed for manual counting, a reduction of approximately 97%. By reducing labor and improving reproducibility, COLLEMBOT enables broader hazard data generation for collembola, supporting science-based chemical risk assessment. The code and workflow are publicly available to facilitate adoption and community-driven development.