Related Experiment Videos
Nonlinear multivariate regression outperforms several concisely designed neural networks on three QSPR data sets
Lucic1, Amic, Trinajstic
1Rugjer Boskovic Institute, Zagreb, Croatia. lucic@faust.irb.hr
Journal of Chemical Information and Computer Sciences
|April 13, 2000
Summary
Nonlinear multiregression (MR) models outperform neural networks (NNs) in quantitative structure-activity/property relationship (QSAR/QSPR) modeling. These simpler MR models offer better predictive accuracy and interpretability than complex NNs.
Area of Science:
- * Computational Chemistry
- * Cheminformatics
- * Quantitative Structure-Property Relationships (QSPR)
Background:
- * Neural networks (NNs) are powerful nonlinear techniques for QSAR and QSPR but are often complex with numerous parameters.
- * Previous research indicated simpler nonlinear multiregression (MR) models outperformed robust NNs.
- * This study investigates if nonlinear MR models also surpass concisely designed NN models.
Purpose of the Study:
- * To compare the performance of nonlinear multiregression (MR) models against concisely designed neural network (NN) models in QSAR/QSPR.
- * To demonstrate a method for developing nonlinear MR models with superior predictive capabilities.
- * To highlight the advantages of simpler, more interpretable models.
Main Methods:
- * Generation of nonlinear MR models by incorporating 2-fold and 3-fold cross-products of initial descriptors.
- * Application of descriptor selection techniques (CROMRsel and CROMRiisel) to identify key descriptors.
- * Evaluation of model performance using cross-validation (CV) and prediction metrics (standard error, correlation coefficient).
Main Results:
- * For alkane boiling points, a 20-descriptor MR model achieved a CV standard error of 2.88 K, outperforming NNs (3.60 K).
- * For chemical shifts, 15- and 9-descriptor MR models yielded CV standard errors of 0.89 and 1.19 ppm, respectively, superior to NNs (1.42 ppm).
- * Sublimation enthalpy modeling showed a CV correlation coefficient of 0.97 for a 4-descriptor MR model versus 0.93 for NNs.
Conclusions:
- * Nonlinear MR models demonstrate superior, unambiguous fitted, cross-validated, and predictive performance compared to NN models.
- * Nonlinear MR models are significantly simpler than NNs, facilitating clearer functional relationship interpretation.
- * This study establishes a viable approach for developing highly effective and interpretable nonlinear QSAR/QSPR models.