Maria Anna Maggi

Cross-column retention prediction in reversed-phase high-performance liquid chromatography by artificial neural network modelling

Linear solvation energy relationships (LSERs) are commonly applied to model the effect of solute structure on the retention of analytes in reversed-phase high-performance liquid chromatography (RP-HPLC). Standard LSER approaches can be used, in principle, to predict RP-HPLC behaviour of unknown analytes under fixed separation condition. However, as solute structure is the only source of variability described by the model, a LSER established for a given column/eluent pair cannot be transferred to external separation conditions. In the present investigation, we attempt cross-column prediction by combining in the same model usual LSER molecular descriptors with observed retentions of selected solutes within the calibration set, adopted to represent the stationary phase features. A multi-layer artificial neural network (ANN) is used as regression tool to model the combined effect of solute structure and column on retention. This model is generated and validated using literature retention data of 34 solutes collected on 15 different RP-HPLC columns at a fixed eluent composition (acetonitrile-water 30:70, v/v). The calibration set is designed by selecting 25 solutes and 11 columns able to represent the variability of the chemical structure of the investigated compounds and dissimilarity of the stationary phases of the data set, respectively. The final predictive performance of the optimised ANN model is tested on the four columns excluded from calibration. Retention of the 25 solutes used to train the network and that of the nine unknown molecules on the external stationary phases is comparably well predicted.

Optimisation by response surface methodology of microextraction by packed sorbent of non steroidal anti-inflammatory drugs and ultra-high performance liquid chromatography analysis of dialyzed samples

A procedure based on microextraction by packed sorbent (MEPS) followed by ultra-high performance liquid chromatography (UHPLC) with photodiode array (PDA) detection has been developed for the analysis of seven selected non steroidal anti-inflammatory drugs (NSAIDs) in human dialysates. The influence on MEPS efficiency of pH of the sample, pH of the washing solvent and methanol content in the hydro-alcoholic elution mixture has been investigated by response surface methodology based on a Box-Behnken design of experiments. Among the above factors, pH of sample is the variable that mostly influences MEPS recovery. UHPLC separation of the NSAIDs was completed within less than 4min under isocratic elution conditions on a Fortis SpeedCore C18 column (150×4.6mm I.D., 2.6μm) using acetonitrile-phosphate buffer as the mobile phase. Calibration curves of the NSAIDs were linear over the concentration range 0.025-15μg/mL, with correlation coefficients≥0.998. Intra- and inter-assay relative standard deviations were <8% and recovery values ranged from 94% to 100% for the quality control samples. The results reveal that the developed MEPS/PDA-UHPLC method exhibits a good accuracy and precision and is well suited for the rapid analysis of human dialysate from patients treated with the selected NSAIDs.

Cross-column prediction of gas-chromatographic retention of polybrominated diphenyl ethers

In this paper, we predict the retention of polybrominated diphenyl ethers (PBDEs) in capillary gas-chromatography (GC) within a useful range of separation conditions. In a first stage of this study, quantitative structure-retention relationships (QSRRs) of PBDEs in six stationary phases with different polarity are established. The single-column QSRR models are generated using the retention data of 126 PBDE congeners by multilinear regression (MLR) coupled to genetic algorithm variable selection applied to a large set of theoretical molecular descriptors of different classes. A quite accurate fitting of experimental retentions is obtained for each of the six GC columns adopting five molecular descriptors. In a further step of this work six molecular descriptors were extracted within the set of molecular descriptors (17 variables) involved in the various single-column QSRRs. The selected molecular descriptors are combined with observed retentions of ten representative PBDEs, adopted as descriptors of the GC system. These quantities are considered as the independent variables of a multiple-column retention model able to simultaneously relate GC retention to PBDE molecular structure and kind of column. The quantitative structure/column-retention relationship is established using a multi-layer artificial neural network (ANN) as regression tool. To optimise the ANN model, a validation set is generated by selecting two out of the six calibration columns. Splitting of columns between training and validation sets, as well as selection of PBDE congeners to be used as column descriptors, is performed with the help of a principal component analysis on the retention data. Cross-column predictive performance of the final model is tested on a large external set consisting of retention data of 180 PBDEs collected in four separation conditions different from those considered in model calibration (different columns and/or temperature program).

Multi-variable retention modelling in reversed-phase high-performance liquid chromatography based on the solvation method: A comparison between curvilinear and artificial neural network regression

The linear solvation energy relationships (LSERs) have been widely used in the last decades for description and prediction of retention in reversed-phase high-performance liquid chromatography (RP-HPLC). LSERs are usually applied to model the effect of solute structure on the RP-HPLC retention at a fixed separation condition. Some authors by combining LSER with known empirical relationships relating retention with mobile phase composition of binary eluents (ϕ) have proposed a predictive model able to simultaneously relate RP-HPLC retention to both solute LSER descriptors and mobile phase composition. The resulting relationship can be established for a given column/organic modifier combination by curvilinear regression aimed at defining 18 model coefficients. In this study, we compare predictive performance of such approach and that of artificial neural network (ANN) regression in which the five solute LSER descriptors and ϕ are directly considered as the network inputs. To this purpose we analyse literature retention data of 31 molecules of different types collected on five reversed-phase columns either in water-acetonitrile and water-methanol mobile phase, the organic modifier content ranging between 20 and 70% (v/v). For each column/organic modifier combination both a curvilinear and an ANN-based model is built using data referred to 25 solutes, while the alternative models are later tested on the remaining six solutes excluded from calibration. Further, we compare capability of curvilinear and ANN regression after including into the respective models also variability related with the stationary phase, represented by the average retention of calibration solutes extrapolated at pure water as the mobile phase. The results of this investigation demonstrate that regardless of the kind of column and organic modifier ANN regression, as compared with curvilinear modelling, provides lower prediction errors and these are more uniformly distributed over the investigated retention range.

Artificial neural network modelling of retention of pesticides in various octadecylsiloxane-bonded reversed-phase columns and water-acetonitrile mobile phase.

Previously, retention of 26 pesticides in the reversed-phase column Gemini (Phenomenex) and water-acetonitrile mobile phase was modelled using a feed-forward artificial neural network (ANN) learned by error back-propagation, accounting for both the effect of solute structure and mobile phase composition. To this end, logK(ow) of solutes and four quantum chemical molecular descriptors (the dipole moment, the mean polarizability, the anisotropy of the polarizability and an hydrogen-bonding descriptor based on the atomic charges located on the acid and basic functional groups) and acetonitrile % (v/v) in the eluent (%ACN) were used as ANN inputs. The above ANN-based approach is here tested on further five similar octadecylsiloxane-bonded columns in water-acetonitrile mobile phase within the %ACN range 30-70%. A quite good predictive performance evaluated on three external solutes in the whole %ACN range is observed, prediction errors being lower than +/-0.1 log k units or slightly higher although still within +/-0.15 log k units. On the other hand, multilinear regression used in place of ANN provides a more diffuse and non-uniform residual distribution for all the investigated columns. ANN multiple-column retention prediction is attempted by adding to the above variables a column descriptor defined as the average retention of calibration solutes extrapolated to 100% water. This more general model is built using 16 solutes and five 5-microm columns in calibration, while its predictive performance is tested on the remaining 10 compounds. Under these conditions, prediction errors are generally within +/-0.2 log k units regardless of the kind of column. The possibility of cross-column prediction is evaluated by column leave-one-out cross-validation within the five 5-microm stationary phases and on a 4-microm external column. This analysis reveals that accuracy of retention prediction for unknown solutes in unknown columns is acceptable provided that the external column is not very dissimilar to those used in calibration.

Investigation by response surface methodology of the combined effect of pH and composition of water-methanol mixtures on the stability of curcuminoids.

Response surface methodology, coupled to a full factorial three-level experimental design, was applied to investigate the combined influence of pH (between 7.0 and 8.6) and composition of methanol-water mixtures (between 30 and 70% v/v of methanol content) on the stability of curcumin and its analogues demethoxycurcumin and bisdemethoxycurcumin. The response plots revealed that addition of methanol noticeably improved the stability of curcuminoids, this effect being both pH- and structure-dependent. In the central point of the experimental domain, half-life times of curcumin, demethoxycurcumin and bisdemethoxycurcumin were 3.8±0.2, 27±2 and 251±17h, respectively. Stability of curcuminoids increased at lower pH and higher methanol content and decreased in the opposite vertex of the experimental domain. These results can be interpreted by assuming that addition of methanol to water produces a different variation of pH of the medium and apparent pKa values of the ionisable groups of curcuminoids.

Modelling of UPLC behaviour of acylcarnitines by quantitative structure-retention relationships

In the present work, the retention time (RT) of acylcarnitines, collected by ultra-performance liquid-chromatography after formation of butyl esters, is modelled by quantitative structure-retention relationship (QSRR) method. The investigated set consists of free carnitine and 46 different acylcarnitines, including the isomers commonly monitored in screening metabolic disorders. To describe the structure of (butylated) acylcarnitines, a large number of computational molecular descriptors generated by software Dragon are subjected to variable selection methods aimed at identifying a small informative subset. The QSRR model is established using two different approaches: the multi linear regression (MLR) combined with a genetic algorithm (GA) variable selection and the partial least square (PLS) regression after iterative stepwise elimination (ISE) of useless descriptors. Predictive performance of both models is evaluated using an external set consisting of 10 representative acylcarnitines, and, successively, by repeated random data partitions between the calibration and prediction sets. Finally, a principal component analysis (PCA) is performed on the model variables to facilitate the interpretation of the established QSRRs. A PLS model based on seven latent variables extracted from 20 molecular descriptors selected by ISE permits to calculate/predict the retention time of acylcarnitine with accuracy better than 5%, whereas a 6-dimensional model identified by GA-MLR provides a slightly worse performance.

Cross-column prediction of gas-chromatographic retention indices of saturated esters

We combine computational molecular descriptors and variables related with the gas-chromatographic stationary phase into a comprehensive model able to predict the retention of solutes in external columns. To explore the quality of various approaches based on alternative column descriptors, we analyse the Kováts retention indices (RIs) of 90 saturated esters collected with seven columns of different polarity (SE-30, OV-7, DC-710, OV-25, XE-60, OV-225 and Silar-5CP). Cross-column retention prediction is evaluated on an internal validation set consisting of data of 40 selected esters collected with each of the seven columns, sequentially excluded from calibration. The molecular descriptors are identified by a genetic algorithm variable selection method applied to a large set of non-empirical structural quantities aimed at finding the best multi-linear quantitative structure-retention relationship (QSRR) for the column OV-25 having intermediate polarity. To describe the columns, we consider the sum of the first five McReynolds phase constants and, alternatively, the coefficients of the corresponding QSRRs. Moreover, the mean RI value for the subset of esters used in QSRR calibration or RIs of a few selected compounds are used as column descriptors. For each combination of solute and column descriptors, the retention model is generated both by multi-linear regression and artificial neural network regression.

Optimisation of temperature-programmed gas chromatographic separation of organochloride pesticides by response surface methodology.

A response surface methodology (RSM) approach is applied to optimise the temperature-programme gas-chromatographic separation of 16 organochloride pesticides, including 12 compounds identified as highly toxic chemicals by the Stockholm Convention on Persistent Organic Pollutants. A three-parameter relationship describing both linear and curve temperature programmes is derived adapting a model previously used in literature to describe concentration gradients in liquid chromatography with binary eluents. To investigate the influence of the three temperature profile descriptors (the starting temperature, the gradient duration and a shape parameter), a three-level full-factorial design of experiments is used to identify suitable combinations of the above variables spanning over a useful domain. Resolutions of adjacent peaks are the responses modelled by RSM using two alternative methods: a multi-layer artificial network (ANN) and usual polynomial regression. The proposed ANN-based approach permits to model simultaneously the resolutions of all the consecutive analyte pairs as a function of the temperature profile descriptors. Four critical pairs giving partially overlapped peaks are identified and multiresponse optimisation is carried out by analysing the surface plot of a global resolution defined as the average of the resolutions of the critical pairs. Descriptive/predictive performance and applicability of the ANN and polynomial RSM methods are compared and discussed.

Quantitative structure–retention relationships of cannabimimetic aminoalkilindole derivatives and their metabolites

Development of chromatographic analyses of synthetic cannabinoids is complicated by the lack of commercial reference standards, especially for new analogues introduced in the clandestine market to bypass legal controls and for their metabolites. In the present work, we explore the possibility of predicting the retention behaviour of the cannabimimetic aminoalkilindoles and their urinary metabolites in high-performance liquid-chromatography using a quantitative structure-retention relationship (QSRR) generated by multilinear regression. To represent the structure of the 43 investigated analytes, 617 computational molecular descriptors are subjected to genetic algorithm variable selection aimed at identifying a small but informative subset. Predictive performance of the QSRR model is evaluated on an external set consisting of 10 representative compounds, including both drugs and their metabolites, and, successively by a Monte Carlo validation method. The best QSRR model, based on six molecular descriptors, exhibits a promising predictive performance and robustness.