Letrozole-loaded nanoparticles were successfully prepared using the emulsification–solvent evaporation method outlined. Since Placket–Burman designs are resolution 4 designs, only main effects of the selected variables were analyzed and factors that had significant main effects on the responses were selected for further studies during optimization. SEM image (Fig. ) showed that the prepared nanoparticles were well-formed and non-porous with some agglomeration, which was expected with PLGA.
Pareto ranking analysis (Fig. ) showed that the amount of drug (X1
) was the most significant factor controlling the EE, followed by the organic to aqueous phase volume ratio (X5
) and the type of organic solvent (X6
), as depicted by the length of the bars and their p
values being less than the a priori
value of 0.05. Other factors such as polymer loading (X2
), stirring rate (X3
), emulsifier concentration (X4
), and homogenization time (X7
) also affected EE but their effects were not statistically significant (p
0.05). A quantile–quantile plot showed a linear correlation between the observed and predicted values of the EE with R2
0.92 (Fig. ). Further analysis by ANOVA confirmed that the model was statistically significant in its prediction of EE (Prob
0.0463). The Prob
is the observed significance probability (p
value) of obtaining a greater F
value by chance alone if the specified model fits no better than the overall mean response.
Generally, NIR spectra of solid powdered samples recorded in the reflectance mode are affected by variations within spectral groups due to varying particle size distributions (19
). These variations result in dissimilar packing densities, intensity differences, light scattering effects, path length variations, and ultimately baseline shifts in the spectra (20
). Instrumental effects such as random noise, changes in lamp intensity, and detector response may also cause variations within spectral groups and can adversely affect the robustness and reliability of the multivariate calibration model to be developed (5
). In an attempt to eliminate, reduce, or standardize the effects of these variations on our multivariate model, we applied three different preprocessing methods, namely multiplicative scatter correction, standard normal variate, and Savitzky–Golay second derivative transformation with third-order polynomial using 7 filter points, to study the influence each pretreatment method on the robustness of our regression model.
It was observed that spectra for both MSC and SNV looked similar in shape (Fig. ). Compared to the original spectra, both MSC and SNV were able to remove a large part of the variance between the spectra without distorting the spectral features. This observation has been reported by other authors that MSC- and SNV-transformed spectra are closely related and that the difference in prediction ability between the two pretreatment methods is very small (21
). Both MSC and SNV are normalization methods that effectively minimize or eliminate variations due to path length and baseline offsets, thus improving the linearity of the relationship between the constituents and the spectral values. MSC gives an estimation of the relation of the scatter of each sample with respect to the scatter of a mean spectrum which is calculated from all the spectra in a defined data set by a least squares regression. Using this mean spectrum, the same least squares regression is performed on every spectrum to minimize the variations due to amplified or additive effects (20
). SNV, on the other hand, is a row-oriented correction method which corrects for each individual spectrum at each wavelength. Each spectrum is mean centered and divided by its standard deviation, so that the new spectra are centered in zero and their standard deviations are one, with a common scale for all spectra (23
Fig. 5 Raw NIR absorbance spectra, MSC-treated, SNV-treated, and Savitzky–Golay second derivative-treated spectra of the 12 formulations of PB design. MSC and SNV pretreatment resulted in the removal of a large part of the variance among spectra without (more ...)
For the Savitzky–Golay second derivative-treated spectra, it was observed that spectral peak and troughs were more prominent than the original, untreated spectra and were also different from SNV- and MSC-treated spectra. Savitzky–Golay second derivative transformation with third-order polynomial applied smoothing and differentiation to the spectra to reduce random noise and enhance spectral resolution. This resulted in peaks which were sharp and not overlapping.
Principal component analysis (PCA) was then carried out on each of the pretreated spectra. Figures , , and show the score plots of the second principal component (PC 2) against the first principal component (PC 1) for the pretreated spectra. It was observed that for both MSC-corrected and SNV-corrected spectra, PC 1 and PC 2 accounted for 92 % of the total variance in the data, with PC 3 accounting for the remaining 8 %.
Fig. 6 Principal component analysis score plot of PC1 against PC2 for MSC-treated NIR spectra showed a clear pattern of clusters along PC 1 axis. From left to right, the order of the sequence coincided with an increasing order of drug loading in the nanoparticle (more ...)
Fig. 7 Principal component analysis score plot of PC1 against PC2 for SNV-treated NIR spectra showed a clear pattern of clusters along PC 1 axis. From left to right, the order off the sequence coincided with an increasing order of drug loading in the nanoparticle (more ...)
Fig. 8 Principal component analysis score plot for PC1 and PC2 for Savitzky–Golay second derivative-transformed NIR spectra. Second derivative transformation resulted in the splitting of overlapping peaks and the introduction of peaks and troughs. As (more ...)
The MSC- and SNV-treated score plots showed clear patterns of clusters along the PC 1 axes. From left to right, the sequence of the score plot was formulation 11, 10, 1, 6, 7, 9, 2, 4, 8, 5, 3, and 12. This sequence coincided with increasing order of letrozole content and decreasing order of PLGA in the formulation as determined by the destructive RP-HPLC method (Table ). Thus, it could be inferred from these score plots that PC1, which explained 70 % of the total variance, primarily correlated with the amount of letrozole within the nanoparticles while PC 2, which explained 22 % of the total variance, correlated with the amount of PLGA in the nanoparticles.
For spectra that were treated with third-order polynomial Savitzky–Golay second derivative with 7 filter points, a look at the score plot (Fig. ) showed no clear pattern even though PC 1 and PC 2 accounted for 99 % of the data variance. Moreover, the scores along both PC 1 and PC 2 axes were very small compared with the scores from the SNV- and MSC-treated spectra. It is reported that derivatization generally results in splitting of overlapping peaks and results in decrease in sensitivity due to the introduction of new spectral peaks (21
). However, examination of the loading plots for the Savitzky–Golay second derivative-transformed spectra (Fig. ) showed that for PC1, there were no prominent peaks between 1,100 and 1,600 nm while PC2 and PC 3 showed both positive and negative peaks between 1,100 and 1,600 nm. By comparing these three loadings with the second derivative spectra of pure letrozole and PLGA, the following assignments could be made: PC1 loading vector showed peaks at 1,624, 1,640, 1,868, 2,138, 2,170, 2,280, and 2,456 nm which could be attributed to the letrozole component of the nanoparticulate system while PC 2 showed peaks at 1,172, 1,332, 1,654, 1,694, 2,138, 2,228, 2,272, and 2,450 nm which could be attributed to the PLGA component of the system. This showed that PC1 correlated with the amount of letrozole in the nanoparticles while PC 2 correlated with the PLGA content.
Fig. 9 Principal component loading vectors of the three principal components and second derivative spectra of the pure components. The spectrum of PC 1, which accounted for 91 % of the total variance was similar to the spectrum of pure letrozole while (more ...)
To quantitatively predict the drug content in the nanoparticulate system, the influence of two standard regression methods, PCR and PLS regression, on the pretreated samples were also examined. These two regression methods were selected because they are the most commonly used regression methods and have been found to be very accurate in terms of model prediction (23
). Tables and show the results of the PCR and PLS regressions for the different pretreatment methods. In both PCR and PLS regression, spectra that were pretreated using Savitzky–Golay second derivative gave the best predictive ability based on the correlation coefficients, RMSEC, RMSEP, SEC, and SEP values for both calibration and prediction models. Thus, spectral pretreatment with Savitzky–Golay second derivative transformation was the best method in effectively minimizing the variations in the raw spectra of letrozole-loaded nanoparticles. PLS models for the Savitzky–Golay second derivative-transformed spectra showed better prediction ability than the corresponding PCR model as depicted by higher correlation coefficients (0.991 for both calibration and prediction models for PLS compared with 0.988 and 0.990 for calibration and prediction for PCR), lower root mean square errors of calibration and prediction (RMSEC and RMSEP) and lower standard errors of calibration and prediction (SEC and SEP; Tables and ). Moreover, there was a smaller difference between RMSEC and RMSEP for the PLS models pretreated with Savitzky–Golay second derivative transformation (RMSEC
0.747 %; RMSEP
0.786 %) compared with the RMSEC and RMSEP for the PCR models pretreated with the same method (RMSEC
0.849 %; RMSEP
0.740 %). It is reported that multivariate models that result in big differences between RMSEC and RMSEP usually yield prediction models that are not very robust and may fail when tested with an independent validation set (18
). Figure shows the relationship between the actual and predicted drug loadings for the PLS calibration and validation models that were pretreated by the Savitzky–Golay second derivative transformation. Very low RMSEC and RMSEP (0.747 and 0.786 %), SEC and SEP (0.758 and 0.589 %) and high calibration and prediction correlation coefficients between measured and predicted drug loadings indicate that NIR spectroscopy has good predictive potential for drug loading in nanoparticles.
Principal Component Regression of MSC, SNV, and Savitzky–Golay Second Derivative-Transformed NIR Spectra for Calibration and Prediction of Letrozole and PLGA
PLS Regression of MSC, SNV, and Savitzky–Golay Second Derivative-Transformed NIR Spectra for Calibration and Prediction of Letrozole and PLGA
Fig. 10 Calibration and validation plots of measured versus predicted drug loading for spectra pretreated by Savitzky–Golay second derivative transformation. Very low RMSEC and RMSEP; SEC and SEP and high calibration and prediction correlation coefficients (more ...)
The difference between PLS regression and PCR is that in PCR, the three principal components were derived from principal component analysis and were used to perform regression on the drug loading whereas in PLS regression, the PLS components or latents variables were derived by comparing both spectral and target property information to find the direction of greatest variability (5
). In other words, whereas in PCR, we knew from PCA that the first principal component (PC 1) which represented the largest variation in the spectra correlated with the amount of letrozole in the nanoparticulate system, in PLS regression, the first PLS component represented the most relevant variations showing the best correlation with the drug loading in the system. The first two PLS factors, PLS 1 and PLS 2, for the Savitzky–Golay second derivative-transformed spectra explained 76 and 19 %, respectively, of the total variance in the spectra while PLS 3 accounted for 5 % of the variance.
Spectral and chemical information contained in the PLS model were estimated by correlating the PLS loading vectors with the second derivative spectra of the individual components of the nanoparticulate system (Fig. ). PLS 1 loading vector showed effective smoothing with no peaks between 1,100–1,600 nm but showed positive and negative peaks at 1,624, 1,640, 1,868, 2,138, 2,170, 2,280, and 2,456 nm which could be attributed to the letrozole component of the system while PLS 2 showed peaks at 1,172, 1,332, 1,654, 1,694, 2,138, 2,228, 2,272, and 2,450 nm corresponding to the PLGA component of the system. PLS 3 may be due to physical variations such as differences in particle size distribution.
Fig. 11 Loading vectors of the PLS factors and second derivative spectra of letrozole and PLGA. The spectrum of PLS 1, which accounted for 76 % of the total variance was similar to the spectrum of pure letrozole while PLS 2, which accounted for 19 % (more ...)