Transparency is fundamental for models for regulatory purposes. CAESAR's models are transparent at the maximum level:

• All biological values
• Chemical structures
• Values of chemical descriptors and fragments
• Algorithms developed within CAESAR
• Procedures

will be publicly available at the CAESAR website.

In order to maximise the reproducibility, which is fundamental for models for regulatory purposes, we

• checked if different tautomers gave different predictive results (this check is not typically done in QSAR modelling),
• used chemical descriptors based on bi-dimensional structures (the use of tri-dimensional structures typically requires manual optimisation of the conformation, which introduces variability),
• will make publicly available the modelling algorithms we developed, defining and fixing all model parameters.

For the model validation we used a wide series of statistical checks. We also used external test sets, to verify that the model performs correctly on new compounds. For the internal tests of QSAR models we adopted the Tropsha parameters.

The experimental data used within CAESAR to build up the QSAR models have been produced according to procedures suitable for REACH. When possible, data have been compared from different sources. The experimental variability has been characterised.

Furthermore, careful check has been done on the data, including the chemical information. At least two different groups within the CAESAR consortium have independently and individually checked all structures. A number of errors have been identified in the data, even if taken from good sources. The presence of chiral centers and tautomers has been checked. The influence on model results on the basis of the use of different tautomers has been checked.

The model output is suitable for REACH, for the different uses (risk assessment or classification and labelling). Attention has been given to the thresholds identified within REACH (for instance for bioconcentrative compounds).

Chemical descriptors and fragments have been calculated using different programs, such as DRAGON, CODESSA, MDL, Pallas, and more.

Models have been developed according to the state-of-the-art of modern algorithms, exploring many possibilities, such as linear multivariate regressions, neural network, GMDH, fuzzy logic, etc. In some cases, several models have been produced, and combined into a single model, to improve reliability.

Local and general models have been developed.

Particular attention has been given to reduce/avoid false negatives.