Pre-processing of three-way data by pulse-coupled neural networks—an imaging approach

A new method for pre-processing three-dimensional data to model quantitative structure-retention relationships (QSRR) is presented. The pre-processing of three-dimensional images of molecules is done with a pulse-coupled neural network (PCNN). The PCNN is capable of transforming an image to a short time series representation of the molecule, which is more suitable for QSRR modelling with partial least squares than the original data. The method was developed and tested on a steroid data set of 24 compounds with reversed-phase high-performance liquid chromatographic retention data. The QSRR models are stable with respect to the parameters of the PCNN. Test set correlations (q2) of 0.95 and cross-validated r2 of about 0.95 are readily obtained.