1. A defined endpoint | |
Clarify the endpoint of a test system for the predictive model (predict Ames test results, chromosomal aberration test results, not predict genotoxicity or mutagenicity). | |
2. An unambiguous algorithm | |
Clarify the types of models (rule-based and statistical-based) and the methods (algorithms, descriptors, etc.) used to build the models, and ensure their transparency. However, in the case of models for commercial purposes, this information is often not necessarily disclosed. | |
3. A defined domain of applicability | |
Since the predictability of QSAR depends on the training set used to build the model, the types of chemicals that can make highly accurate predictions are limited. Therefore, clarify the limits of the chemical structure to which the QSAR model can be applied (Clarification of Out of Domain). | |
4. Appropriate measures of goodness-of–fit, robustness and predictivity | |
The fitness and robustness of the predictive model should be evaluated using an internal training set. Also, its predictability should be determined using an appropriate external dataset. | |
5. A mechanistic interpretation, if possible | |
If possible, show the mechanical association between the model descriptor and the prediction endpoint. If it can be interpreted by mechanisms, it can be part of the scope of Principle-3. |