Application of multivariate chemometrics tools for spectrophotometric determination of naphazoline HCl, pheniramine maleate and three official impurities in their eye drops

This work is concerned with exploiting the power of chemometrics in the assay and purity determination of naphazoline HCl (NZ) and pheniramine maleate (PN) in their combined eye drops. Partial least squares (PLS) and artificial neural network (ANN) were the chosen models for that purpose where three selected official impurities, namely; NZ impurity B and PN impurities A and B, were successfully determined. The quantitative determinations of studied components were assessed by percentage recoveries, standard errors of prediction as well as root mean square errors of prediction. The developed models were constructed in the ranges of 5.0–13.0 μg mL−1 for NZ, 10.0–60.0 μg mL−1 for PN, 1.0–5.0 μg mL−1 for NZ impurity B and 2.0–14.0 μg mL−1 for two PN impurities. The proposed models could determine NZ and PN with respective detection limits of 0.447 and 1.750 μg mL−1 for PLS, and 0.494 and 2.093 μg mL−1 for ANN. The two established models were compared favorably with official methods where no significant difference observed.

www.nature.com/scientificreports/chromatographic (HPLC) technique.In addition, two impurities, A and B, are reported in its BP monograph 14,15 .A spectrophotometric method based on complexation with ferric ion has been published for its simultaneous determination with colorpheniramine maleate 23 .
For the treatment of inflammatory eye disorders and allergic conjunctivitis, the approach of using NZ and PN together in combined eye drops is found to be more useful than using each drug alone 24 .Literature review reveals that NZ and PN have been quantified together using HPLC [25][26][27][28] , capillary electrophoresis 29 and thin layer chromatography 30 .Two of these chromatographic methods have been published by our research group where their determination along with three reported official impurities, NZ impurity B, PN impurity A and PN impurity B, was achieved 28,30 .In spite of the applicability of the chromatographic technique in analyzing such complex mixtures, it is still considered challenging.This is attributed to the requirements for the successive sample pretreatment steps as well as the selection of suitable mobile phase and stationary phase to attain the best separation and system suitability parameters.Besides all of this, the time needed for optimization, the expensive tools required, and the hazardous organic reagents utilized are considered potential obstacles to that technique 31 .On the other hand, spectrophotometric techniques can overcome these previously mentioned drawbacks as it is fast, cheap, simple and time saving 32 .To this end, our aim is to simultaneously analyze NZ, PN, and three official impurities, namely; NZ impurity B, PN impurity A and PN impurity B, using simple chemometrics-assisted spectrophotometric methods.Calibration mixtures are prepared based on five-level five-factor design.Partial least squares (PLS) and Artificial neural network (ANN) regressions are the chosen models for the determination of the five cited compounds (Fig. 1) in marketed dosage form and laboratory prepared mixtures.

Results and discussion
Among different analytical techniques used for drugs determination in their pharmaceutical formulations, spectrophotometry is the widely applicable one.This is attributed to simplicity and rapidity of spectro-analytical methods which has no need for neither sophisticated apparatus nor chemical pretreatment as other chromatographic methods 9 .Due to presence of sever overlapping between the spectra of the five studied components, Fig. 2. Multivariate spectrophotometric are the technique of choice to resolve this overlap.It was found that readings below 250.0 nm showed high noise which may affect the obtained result, otherwise readings above 300.0nm gave almost zero absorption leading to invaluable information, thus the range between 250.0 and 300.0 nm was the range of choice in calculation.Brereton five-level calibration design was followed in order to prepare different mixtures of the five cited drugs 33 .25 mixtures were prepared and divided into two groups.15 mixtures for the calibration group whereas the remaining 10 mixtures were utilized for the validation one.

Partial least squares model
During construction of PLS model, mean centering of all spectral data was adopted.After applying the leaveone-out as a cross-validation tool, latent variables' number was obtained 6,34 .The plot relating this number to root mean square error of calibration (RMSEC) revealed that six was the optimum number to be utilized, Fig. 3.It is worth noting that numerous models for calibration were tried and built whereas obtaining least noise as well as satisfactory recovery results were the parameters for selecting the optimum one 35 .

Artificial neural network model
Three transferable layers (input, hidden and output) are usually incorporated in ANN model 8,12 In our model, the 251 points of spectral data were used as input neurons 11,35 .On the other hand, 5 neurons related to number of drugs determined by such model comprised the output layer.For the hidden layer, various numbers were  tried whereas RMSEC was calculated in each time.It was found that the lowest error (< 0.1 for the five drugs) was gained upon using 4 neurons in this layer with no further enhancement upon increasing their number.It is worth noting that pure-line transfer function, 0.1-learning rate as well as 50-epochs were also utilized.
The prediction's capability of the two proposed models was verified by the aid of the validation group mixtures.For each drug, average recovery percentages as well as relative standard deviations were tabulated, Table 1.It is worth noting that the concentration ranges of the two drugs were chosen to facilitate their direct determination in the challenging eye drops ratio.Linearity parameters, limits of detection and quantification were obtained, Table 2.As shown, our proposed models could detect NZ and PN impurities at limits of about ≈1% and ≈2%, respectively, of their parent drugs' highest calibration concentration.Moreover and in order to test the presence of possible over-fitting, root mean square error of prediction (RMSEP) and standard error of prediction (SEP) were computed as well.As shown in Table 2, their close values assured over-fitting absence in our models.

Eye drops application
Successful determination of NZ and PN in naphcon-A drops was attained by the proposed multivariate chemometrics models.As shown in Table 3, good recoveries were obtained and the models' validity was further assured through recoveries of the added standards.

Statistical analysis
Student's t-test as well as F-one were statistically applied to compare obtained results with that acquired by BP methods of NZ and PN analyses.Table 4 summarizes the outputs of this comparison along with average recoveries, standard deviations, variances, and number of observations utilized in those tests.It is worth noting that the lower t and F values obtained relative to theoretical ones assured the statistical absence of significant differences.

Comparison with reported chromatographic methods
Comparing our proposed models with the reported chromatographic methods for the determination of NZ, PN, and the three selected impurities was conducted.The solvents used, the linearity ranges as well as the obtained LOD values were the chosen items for this comparison, Table 5.The table proved the superiority of these proposed PLS and ANN spectrophotometric models in terms of the simplicity of solvents required.Lower LOD values were also obtained compared to the reported HPLC method.

Conclusion
In this work NZ, PN along with three selected official impurities are determined in their mixtures and eye drops pharmaceutical dosage form using accurate and simple chemometrics assisted spectro-analytical models.The proposed models are considered time-saving and cost-effective comparing with the reported chromatographic methods.The two proposed models show the advantage of the low values for prediction error.PLS supposes that the error is scattered equally between the two matrices; concentration and spectral response.As a result, robust results were given by this model through removing the absorbance as well as concentration data noises simultaneously.On the other hand, the ANN model's predicting ability was exploited for obtaining more precise results during studied drugs' quantification.

Instrument
A dual-beam Shimadzu, UV 1601 spectrophotometer, Kyoto, Japan.Spectra Scanning was conducted at 200.0-400.0nm range with 0.2 nm intervals using 1.00-cm quartz cuvettes.The utilized Matlab ® software (7.0.1) was integrated with a PLS Toolbox 2.1.

Standards
The studied drugs were supplied by Eva-pharma Co. (Egypt).Purities were checked as per BP methods to be 100.12%for NZ, and 99.58% for PN 15 .The impurities were purchased from Alfa Aesar Co. (Germany).Potencies were certified to be 99.00%,100.30% and 99.70% for NZ impurity B, PN impurity A and PN impurity B, respectively.

Reagents
Methanol of analytical-grade was used (Alpha, Egypt).

Solutions
Five standard solutions, of 1.0 mg mL −1 concentration, were prepared separately using methanol.

Procedures
Construction of the calibration models Spectral features of the five studied substances were determined through their normalized spectra using methanol as a blank.Twenty-five mixtures containing different concentrations of the five cited compounds were prepared following a five-levels five-factors experimental design to reach concentration ranges of 5.0-13.0μg mL −1 for   Mean ± SD 100.0 ± 1.9 100.0 ± 1.4 Mean ± SD 101.0 ± 1.9 100.2 ± 1.9 Mean ± SD 99.3 ± 1.9 98.7 ± 1.9 www.nature.com/scientificreports/NZ, 10.0-60.0μg mL −1 for PN, 1.0-5.0μg mL −1 for NZ impurity B and 2.0-14.0μg mL −1 for two PN impurities.Scanning of the prepared solutions were then conducted at 200.0-400.0nm range.The range from 250.0 to 300.0 nm was used for further calculations.The data were analyzed using Matlab ® , where multivariate calibration models were built by using fifteen mixtures from the previously prepared solutions.

Validation of the calibrated models
The remaining ten mixtures from the previously prepared solutions were used for the validation purpose.PLS and ANN obtained parameters were used for quantification of the five cited compounds.

Application to naphcon-A drops
A 1.0-mL solution was transferred from naphcon-A drops to a 50-mL flask, and 25.0 mL methanol was introduced.Sonication was then applied for 10.0 min.A final concentration of 5.0 µg mL −1 NZ and 60.0 µg mL −1 PN was obtained after completing the volume with methanol.The prepared solutions were scanned and concentrations were predicted by the developed models.

Figure 1 .
Figure 1.Chemical structures of the five cited components.

Figure 3 .
Figure 3. Root mean square error of calibration (RMSEC) versus the number of latent variables used to construct the PLS calibration for the assay of NZ, PN, NZ impurity B, PN impurity A and B in their mixtures.

Table 1 .
Prediction recoveries of the validation set samples by the constructed PLS and ANN models.

Table 2 .
Regression parameters of the validation sets calculated for each developed model.

Table 3 .
Determination of NZ, PN in their dosage form and application of standard addition technique using the proposed PLS and ANN models.a Average determinations of four eye drop dosage form.

Table 4 .
Statistical comparison between the results obtained by the proposed PLS and ANN models and the official BP method of analysis of NZ, PN. a NZ was analyzed using stationary phase octylsilyl silica gel and mobile phase of sodium octane sulphonate in a mixture of 5.0 mL of glacial acetic acid, 300.0 mL of acetonitrile and 700.0 mL of water, flow rate 1.0 mL min −1 and detection at 280.0 nm.PN was analyzed using stationary phase C18 with gradient mobile phase of sodium heptane sulphonate pH 2.5 and acetonitrile, flow rate 1.0 mL min −1 and detection at 264.0 nm.b These values represent the corresponding tabulated values of t and F at p = 0.05.