20 August / September 2019
media contain large electron clouds which promote enhanced π-π and dipole-induced dipole interactions in addition to van der Waals interactions. These early eluting compounds contained no phenyl rings, but all contained at least one double bond or displayed some level of aromaticity, and enough to result in significant retention from the void. The full selection of 653 compounds covered a wide range of polarities (e.g., -1.63≤AlogP≤7.88). A total of n=460 compounds had between one and four benzene-like rings and n=565 had between one and seven double bonds. However, correlation of AlogP with tR
was lower than expected at R2 =0.5496. Many
compounds were partly or fully ionised under these slightly acidic mobile phase conditions (pH=5.4), with logD for all compounds between -2.75 (diuron) and 7.99 (acequinocyl). This descriptor was correlated to a larger extent with tR
(R2 = 0.6279) than
AlogP and better took into account the ionised portion of all compounds under mobile phase pH conditions. However,
Figure 3. (a) Predicted versus measured tR using a single 16-5-1 MLP model and its associated residual errors
(b), (c) predicted versus measured tR using an ensemble of four MLPs and its associated residual errors (d). Data split into training data (n=457); verification and blind test data (n=98 each).
interconnected and weighted; and (c) the output layer which collates the information from the hidden layer(s) and generates tR
via
an activation function. The proportioning of all datasets was set at 70:15:15 across training, verification and blind test cases (optimised). Cases were randomly re- assigned for every neural network type investigated. The best model type was retained based on the lowest predicted errors obtained and with consistency across the three datasets. The best model overall was then replicated exactly via the custom network designer (n=6) and evaluated for correlation coefficient (R2
), slope, intercept,
residual error, as well as overall accuracy, precision and sensitivity to molecular descriptor data.
3. Results and Discussion
3.1 Compound selection and retention behaviour on biphenyl media
The biphenyl column used here offers complimentary selectivity to C18
especially for aromatics and polar
compounds and especially when used with methanolic mobile phases. According to the manufacturer, it is suitable for fast separations and increases the retention of early eluting species to minimise matrix suppression when used with mass spectrometry. The column is packed with superficially porous particles with a surface area of 130 m2
/g which offered high efficiency in this case.
Retention data for all 653 compounds covered most of the gradient separation space (Figure 1 and Figure 2(b)), which was considered desirable to allow models to learn more fully from quantitative structure-activity relationship (QSAR) data at each timepoint. Higher retention of polar compounds was observed meaning that model predictive accuracy for any new compounds eluting early could be less reliable. For example, and following elution of the first compound, aminopyralid at 0.806 min, the next compound to elute was methamidophos at 1.758 min. Only 9 compounds in total eluted within the first 3 min, after which the remaining compounds eluted in more rapid succession up to etofenprox at 10.367 min (Figure 2). Biphenyl
it alone could not be used to predict tR reliably. In the main, nC was high across the board in comparison to nO, potentially resulting in preferential retention via van der Waals forces over dipole-induced dipole interactions. In previous works involving tR
prediction on C18 media, a prioritised
list of 16 molecular descriptors enabled reliable models to be built for >1,117 drugs, pesticides and industrial chemicals. However, here some of these descriptors yielded no data in the main. These were nTB, nR04-05 and nR07-09. Nonetheless, these were retained in the model as a small number of compounds did possess some of these features and it was decided to test the generalisability of the C18
model to another reversed-phase medium as is.
3.2 Performance of the optimised model
During optimisation, it was quickly apparent that MLPs performed best. This was in line with previous models for C8
or C18 media
[7, 10]. The best neural network-type model had a 16-5-1 MLP architecture (Figure 3(a)). Fewer layers and nodes was desirable so that the model could be more easily interpreted and to enhance its stability for generalised application. An R2
>0.85 was
achieved for all three datasets, including the blind test data. Overall, excellent consistency was also observed between the training, verification and blind tests datasets showing that the model was not over-trained. All three yielded a mean
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
