Gradient liquid chromatographic retention time prediction for suspect screening applications: A critical assessment of a generalised artificial neural network-based approach across 10 multi-residue reversed-phase analytical methods

Abstract

For the first time, the performance of a generalised artificial neural network (ANN) approach for the prediction of 2492 chromatographic retention times (tR) is presented for a total of 1117 chemically diverse compounds present in a range of complex matrices and across 10 gradient reversed-phase liquid chromatography-(high resolution) mass spectrometry methods. Probabilistic, generalised regression, radial basis function as well as 2- and 3-layer multilayer perceptron-type neural networks were investigated to determine the most robust and accurate model for this purpose. Multi-layer perceptrons most frequently yielded the best correlations in 8 out of 10 methods. Averaged correlations of predicted versus measured tR across all methods were R(2)=0.918, 0.924 and 0.898 for the training, verification and test sets respectively. Predictions of blind test compounds (n=8-84 cases) resulted in an average absolute accuracy of 1.02±0.54min for all methods. Within this variation, absolute accuracy was observed to marginally improve for shorter runtimes, but was found to be relatively consistent with respect to analyte retention ranges (~5%). Finally, optimised and replicated network dependency on molecular descriptor data is presented and critically discussed across all methods. Overall, ANNs were considered especially suitable for suspects screening applications and could potentially be utilised in bracketed-type analyses in combination with high resolution mass spectrometry.

DNA methylation-based forensic age prediction using artificial neural networks and next generation sequencing

“…For the blind test set in particular, the 75th percentile of all 231 case errors lay within 6.3 years. This performance is consistent with other ANN-based applications from our research group which revealed a 3–5% average inaccuracy across predictions [58] and with a recent study reporting a percentage of prediction error of 6.3% [38].…”

The First Attempt at Non-Linear in Silico Prediction of Sampling Rates for Polar Organic Chemical Integrative Samplers (POCIS)

“…Furthermore, given that models developed herein are derived from a very limited number of training cases, any new reported R s data generated by similar methods to those used herein will likely enable better generalizability in the future, as was observed with retention time predictions in reversed-phase liquid chromatography. 27 …”

Recent applications of chemometrics in one‐ and two‐dimensional chromatography

“…A second argument against ANNs is that, at least in the examples discussed, the molecular structure had to be known to a certain extent. Attempts have been made to apply ANNs for unknown samples [215][216][217]. In that case, the models obtained are not related to any physicochemical interactions that occur within the column.…”