|
The hybrid models are algorithms to predict the toxicity of pesticides. These models were efficinetly developed by using several QSAR and classification algorithms on
- all the five endpoints (trout, daphnia, oral and dietary quail, bee), by using 2D descriptors;
- the first three endpoint (trout, daphnia, oral quail), by employing 3D descriptors
Generally, good and interesting results were obtained. The best classification models allowed to predict the test set compounds with satisfactory scores of 70-75%, for all endpoints. Moreover, good cross-validation values of 65-70% were derived. These values underline that the best classification models developed are robust and predictive, also those related with data sets including a weak number of compound, like the quails and bee series. Finally, for most models, the best predicted classes were the most toxic ones; this result is very important for real applications of these models, because predicting as being toxic a compound which is not (false positive) is preferable to predicting the inverse.
QSAR and classification algorithms
The main target of the models derived is the integration of the best algorithms obtained as the basis for a hybrid system software to be used for predictive purposes.
The informatics architecture (object oriented, OO) for the software package (Demetra Tool) that implements the hybrid neuro-fuzzy system is based on a neuro-fuzzy approach, to be generalizable and able to manage both the "learning from data" paradigm (from data sets algorithms are built) and the "learning from experts" paradigm (from algorithms or expert rules software modules are built and integrated with data). In particular 3 methods of integration have been considered: statistical, fuzzy aggregation and connectionist integration. Moreover an terative testing and development cycle has been used until the final integration of all the algorithms was achieved..
Integration of the models
|
|
Last Updated ( Monday, 23 October 2006 )
|
Home 
Models 
