Prediction of physico-chemical properties of amines based on QSPR models
Résumé
The new EU regulation REACH requires the evaluation of the physico-chemical properties of a large number of existing substances (143 000 pre-registered substances in 2008) in order to allow their use before 2018. Due to the number of substances and properties, the timing, the economic costs, the feasibility at the R&D level and the risks for the manipulator, in particular for the characterization of the dangerous physico-chemical properties (explosibility, flammability), the complete gathering of the data solely based on experimental measurements is not realistic. Thus, the use of alternative predictive methods was recommended in the framework of REACH for the evaluation of the properties of substances. In this context, the French PREDIMOL (molecular modeling prediction of physico-chemical properties of products) project funded by ANR (National Research Agency) has started in November 2010 for 3 years. This project is conducted by INERIS associated with several public and private partners. Its objective is to demonstrate that molecular modeling, notably through QSPR (Quantitative Structure-Property Relationships), is a credible alternative approach to experimental characterization to access, in a reliable and fast manner, to the whole range of physico-chemical properties of substances required by EU-REACH’s regulation (annexes VII and IX) and for the industry in terms of property-screening method. The present contribution focuses on the use of QSPR models to predict the physico-chemical hazardous properties of amines (boiling point, flash point, flammability limits, self-ignition temperature). At first, existing models applicable to amines have been identified, stressing onto performances and limits of each model. Then, an inventory of existing experimental data from literature has also been established. Based on these data, existing models were tested upon availability of algorithm and tools. New models have also been developed and validated on a series of internal and external validation tests. In particular, models were developed for the flash point of 66 amines based on multilinear regressions using only computed descriptors with a mean absolute error of 10K on an external validation. A critical objective of this project is the regulatory validation of models by European chemical instances in order that predictive data could be used for registration. Hence, models will be submitted to the JRC comity for acceptance or to existing tools (like QSAR toolbox) for integration.