Abstract: To meet the different functional needs for soybean protein isolate (SPI) in different food applications, this study utilized infrared spectroscopy to rapidly analyze 70 SPI samples subjected to different pH treatments, and explored the effect of pH changes on the secondary structure content of SPI. Mean centralization (MC), multivariate scattering correction (MSC), standard normal variate transformation, and normalization were used for infrared data preprocessing. Feature wavebands were identified based on two-dimensional correlation infrared spectra, and predictive modeling of the secondary structure content of SPI against pH was performed using partial least squares (PLS) and arithmetic optimization algorithm-random forests (AOA-RF). The results showed that the relative standard deviations of the α-helix, β-sheet, β-turn, and random coil prediction models developed by the combined use of MC and MSC were 1.29%, 1.60%, 1.37%, and 7.28%, respectively, indicating their combination to be the best spectral pre-processing method. The optimal model for predicting α-helix and β-sheet contents was AOA-RF (characteristic wavebands), with calibration determination coefficients of 0.935 0 and 0.926 6 and prediction determination coefficients of 0.856 8 and 0.870 1, respectively. The optimal model for predicting β-turn and random coil contents was PLS (characteristic wavebands), with calibration determination coefficients of 0.915 4 and 0.881 7 and prediction determination coefficients of 0.891 3 and 0.784 3, respectively. The results of this study provide a theoretical basis for product quality detection and processing condition control in industrial settings.

Key words: two-dimensional correlation infrared spectroscopy; soybean protein isolate; secondary structure; pH changes; predictive modeling; rapid analysis