All the determinations were performed in duplicate and the results were expressed as the mean ± standard deviation. NIR spectroscopy was obtained by Matrix-I FT-NIR spectrometer (Bruker Optics, Ettlingen, German) equipped with an integrating sphere in the sampling area. OPUS spectroscopy software (v.6.5 Bruker Optics, Ettlingen, Germany) was used for instrumental control and spectral acquisition. Sample
was poured into 50 mm rotating cup on holder and scanned over the spectra range 4000–12,500 cm− 1 (800–2500 nm) at 1 nm interval. The spectrum of each sample was find more the average of 64 scans with the resolution ratio of 16 cm− 1. All acquisitions of the sample spectrum were performed in triplicate. Prior to modeling, the original data were smoothed using the Savitzky–Golay (9 points) algorithm to avoid noise enhancement [20]. To optimize the models, the available data preprocessing methods were performed on the data using mathematical transformation method such as vector normalization, multiplicative scattering correction, the first derivative + vector normalization and the first derivative + multiplicative scattering correction. Limiting wavenumber region was used to decrease the spectral noise [13]. Partial least squares (PLS) algorithm was used to obtain the fundamental relation between the spectral data and corresponding chemical values. The reliability LGK-974 solubility dmso of prediction model
was tested by leave-one-sample-out cross validation and external validation. All models were originally based on a calibration set (203 samples) and a validation set (41 samples). Therefore, the choice of the calibration and validation sets ensured a large representative range and a good uniformity of gradient distribution. Various statistics, such as the coefficient
of correlation (r2), the coefficient of determination (R2), the root mean square error (RMSE) and residual predictive deviation (RPD), were computed by OPUS 6.5 to judge the quality of models. The coefficient PtdIns(3,4)P2 of determination (R2) indicates the percentage of variance present in the chemical values, which was reproduced in the prediction. The root mean square error in cross-validation (RMSECV) gives an average of the uncertainty that can be expected for the predicted values. The root mean square error of prediction in test set validation (RMSEP) was also computed. The residual prediction deviation (RPD), defined as the ratio between the standard deviation of the values and the standard error of performance, indicated the predictive capacity. The prediction accuracy of models was regarded as excellent or good when RPD was above 2.5. The models could be applied for a rough prediction when RPD ranged from 2.0 to 2.5. Reliable PLS model should have high value of r2, R2 and RPD and low value of RMSECV [20] and [21]. For preventing PLS model from over-fitting, the max rank value was determinate at ten. Two-step clustering analysis was performed by SPSS (Version 13.