Abstract
A data set containing acute toxicity values (96-h LC50) of 69 substituted benzenes for fathead minnow (Pimephales promelas) was investigated with two Quantitative Structure-Activity Relationship (QSAR) models, either using or not using molecular descriptors, respectively. Recursive Neural Networks (RNN) derive a QSAR by direct treatment of the molecular structure, described through an appropriate graphical tool (variable-size labeled rooted ordered trees) by defining suitable representation rules. The input trees are encoded by an adaptive process able to learn, by tuning its free parameters, from a given set of structure-activity training examples. Owing to the use of a flexible encoding approach, the model is target invariant and does not need a priori definition of molecular descriptors. The results obtained in this study were analyzed together with those of a model based on molecular descriptors, i.e. a Multiple Linear Regression (MLR) model using CROatian MultiRegression selection of descriptors (CROMRsel). The comparison revealed interesting similarities that could lead to the development of a combined approach, exploiting the complementary characteristics of the two approaches.
Original language | English |
---|---|
Pages (from-to) | 1005-1021 |
Number of pages | 17 |
Journal | MATCH: COMMUNICATIONS IN MATHEMATICAL AND IN COMPUTER CHEMISTRY |
Volume | 70 |
Issue number | 3 |
Publication status | Published - 2013 |
MoE publication type | A1 Journal article-refereed |
Keywords
- RECURSIVE NEURAL-NETWORKS
- GLASS-TRANSITION TEMPERATURE
- PREDICTION
- QSPR
- REPRESENTATIONS
- TOXICOLOGY
- POLYMERS