Related Experiment Video
Updated: Jul 31, 2025
08:49
Cultivating a Three-dimensional Reconstructed Human Epidermis at a Large Scale
Published on: May 28, 2021
11.3K
The Good, The Bad, and The Perplexing: Structural Alerts and Read-Across for Predicting Skin Sensitization Using
Emily Golden1, Daniel C Ukaegbu1, Peter Ranslow2
1Center for Alternatives to Animal Testing (CAAT), Johns Hopkins Bloomberg School of Public Health, Baltimore, Maryland 21205, United States.
Chemical Research in Toxicology
|May 1, 2023
Summary
Computational tools for skin sensitization showed lower accuracy on an expanded dataset, especially with more non-sensitizing chemicals. Combining methods may improve predictions, but a deeper understanding of sensitization mechanisms is needed.
Area of Science:
- Toxicology
- Computational Chemistry
- Dermatology
Background:
- Previous computational tools for skin sensitization achieved 70-80% accuracy using human data.
- An expanded dataset from the NICEATM human skin sensitization database was utilized for further analysis.
Purpose of the Study:
- To re-evaluate the performance of computational tools for skin sensitization using an expanded dataset.
- To analyze mispredictions and identify limitations of current models.
Main Methods:
- Evaluated Toxtree, OECD QSAR Toolbox, VEGA's CAESAR, and a k-nearest-neighbor (kNN) approach on 1355 chemicals.
- Analyzed model performance, focusing on balanced accuracy and misprediction patterns.
- Investigated the impact of metabolic simulation and structural alerts.
Main Results:
- Balanced accuracies were lower than previously reported: 63% (Toxtree), 65% (OECD QSAR Toolbox), 46% (VEGA), and 59% (kNN).
- The higher proportion of non-sensitizing chemicals (∼70%) likely contributed to reduced accuracy.
- 287 chemicals (20%) were mispredicted by both Toxtree and OECD QSAR Toolbox, predominantly as false positives.
Conclusions:
- Current computational tools show limitations, particularly with non-sensitizing chemicals and mechanisms not captured by structural alerts.
- A kNN approach identified different mispredictions, suggesting complementary information.
- Combining structural alerts, QSAR, and read-across shows promise, but a deeper understanding of skin sensitization mechanisms is crucial for improving predictive accuracy.
