19


by re-equilibration to 3 % B for 3 min. The column temperature was 35ºC, the injection volume was 2 µL and the flow rate was 0.4 mL/min.


2.2 Descriptor generation, feature selection and retention time prediction


Figure 2. (a) Range of data for each descriptor used in the optimised tR prediction model and (b) the


coverage of measured tR of all 653 compounds across the 12 min gradient runtime. For molecular descriptor abbreviations, see section 2.2.


small molecules including pharmaceuticals, metabolites, pesticides, herbicides and industrial chemicals [6-9]. This technique involves the use of computer algorithms that can learn to predict tR


by finding trends


in compound structures, properties and functions. Recently, we published work using artificial neural networks to predict retention of 1,117 chemically diverse compounds across ten reversed-phase chromatography (RPLC) methods for a range of different applications and sample types [10]. Predictions were generally very good and with an average inaccuracy of 1.02 ±0.54 min across all methods. Longer runtimes generally yielded more inaccuracy, but it was found that inaccuracy was relative and within ~3-5 % of the retention ranges of all analytes measured. Therefore, faster separations in general enabled better accuracy. However, all of these methods involved separations performed on either C18


or C8 media. Other


RPLC phases exist, such as aromatic, polar embedded and hybrid phases, which can offer alternative selectivity to C18


, especially


for isomers and more polar compounds. To our knowledge, no machine learning-based prediction models have yet been developed for other RPLC media.


Fast, selective and highly sensitive methods for an array of different compounds have recently been developed using biphenyl columns and have gained in popularity as an alternative to C18


. Shimadzu Corporation


recently published a method for 646 pesticides using this phase [11]. The aim of this work was to train and evaluate a machine learning model to predict tR


models in this method. Particularly relevant to pesticides, the aromatic character in the stationary phase can improve separation of these compounds and offers an excellent alternative to C18


. Predictions of tR


For all compounds, canonical simplified molecular input line entry system (SMILES) strings were generated using Chemspider freeware (Royal Society of Chemistry, UK). Molecular descriptors were generated using two licenced software packages ACD Labs Percepta for logD only (Advanced Chemistry Development Laboratories, ON, Canada) and Dragon version 7 for all other descriptors (Kode Chemoinformatics s.r.l., Pisa, Italy). For prediction of pesticide tR


, on several


different RPLC media in this way offers the possibility for rapid shortlisting of candidates when performing suspect screening on environmental samples.


2. Experimental 2.1 Retention time datasets


Retention data for 653 compounds were generated using a biphenyl column (Restek Raptor 100 x 2.1 mm, 2.7 µm) configured to a Shimadzu LCMS-8060 LC-MS/MS instrument in polarity switching mode and with MRM data for up to three transitions per compound. Multistep gradient elution was using mobile phase reservoirs containing water (mobile phase A) and methanol (mobile phase B) with both containing a buffer of 2 mM ammonium formate and 0.002% formic acid in each. A 0.002% formic acid concentration resulted in a higher signal intensity in MS/MS, particularly for negative ion mode, and this ion signal response was consistent within and between batch analyses. This approach has been reproduced within food safety applications but also within drugs of abuse testing [12]. The % relative standard deviation (%RSD) of n=100 replicate injections of a spiked apple matrix at 50 µg/L was previously measured at an average of 0.12 %.


for


application to these compound types given the size of the dataset available for training


The gradient was as follows: 3-10% B over 1 min; 10-55% B for 2 min; 55-100% B over 7.5 min; held at 100% B for 1.5 min followed


n=16 molecular descriptors were based on our previous work [10] including unsaturation index (Ui), hydrophilic factor (Hy), Ghose– Crippen logP (AlogP), Moriguchi logP (MlogP), number of benzene-like rings (nBnz), number of double and triple bonds (nDB/nTB), number of 4–9 membered rings (nR04-nR09), number of carbons (nC), number of oxygens (nO) and logD (calculated at pH 5.4). These descriptors were sub-selected from a larger set of >200 user-curated constitutional, topological and physicochemical descriptors deemed relevant to reversed-phase LC mechanisms in the Network Designer Tool in the neural network simulator package, Trajan v6.0.


2.3 Machine learning, optimisation and procedures


All artificial neural network modelling was performed using Trajan v6.0 software (Trajan Software Ltd, Lincolnshire, UK). The intelligent problem solver tool was used to optimise a suitable neural network and architecture in several steps, each comprising 15 min intervals for training. Briefly, the network type was first selected from a range of different types, including probabilistic neural networks (PNNs), generalised regression neural networks (GRNNs), radial basis function (RBFs), as well as three- and four-layer multilayer perceptrons (MLPs). Here, the MLPs were the best choice for tR


prediction, and


are a type of feed-forward, non-linear model that comprises of: (a) an input layer (i.e., molecular descriptors); (b) one or two hidden layers which each contain an optimised number nodes which are


Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32  |  Page 33  |  Page 34  |  Page 35  |  Page 36  |  Page 37  |  Page 38  |  Page 39  |  Page 40  |  Page 41  |  Page 42  |  Page 43  |  Page 44  |  Page 45  |  Page 46  |  Page 47  |  Page 48  |  Page 49  |  Page 50  |  Page 51  |  Page 52  |  Page 53  |  Page 54  |  Page 55  |  Page 56  |  Page 57  |  Page 58  |  Page 59  |  Page 60  |  Page 61  |  Page 62  |  Page 63  |  Page 64  |  Page 65  |  Page 66  |  Page 67  |  Page 68