Spectroscopic quantitation of tetrazolium formazan in nano-toxicity assay with interval-based partial least squares regression and genetic algorithm

• Partial least square regression for tetrazolium-based nano-toxicity assay.
• Interval PLS, backward interval PLS and synergy interval PLS methods were compared.
• siPLS and biPLS models are simpler than full-spectrum PLS models.
• Genetic algorithm based siPLS regression comes up with lowest prediction error.

Spectrophotometric quantitation of formazan in tetrazolium-based nanoparticle (NP) toxicity assay requires a robust calibration model immune to optical interference. For the first time, variant of partial least squares (PLS) regression models, such as, full-spectrum (250–700 nm) PLS, interval PLS (iPLS), backward interval PLS (biPLS), and synergy interval PLS (siPLS) models have been adopted for formazan quantitation. Models were evaluated based on root mean square error of cross-validation (RMSECV), and prediction (RMSEP). The spectral variables in optimal iPLS, biPLS and siPLS models, as well as variables retained above a selection frequency threshold (for all intervals), were further refined in a genetic algorithm (GA). The results suggest that the optimal biPLS (140 variables, 5 LVs, RMSECV: 0.4438, RMSEP: 0.2936) and siPLS (88 variables, 5 LVs, RMSECV: 0.4401, RMSEP: 0.316) models were superior either to the full-spectrum PLS (4 LVs, RMSECV: 0.9674, RMSEP: 0.4618) or traditional single wavelength calibration (414 nm, RMSECV: 2.0864, RMSEP: 2.1628). Minimum RMSEP (0.2976) was observed when GA was performed on spectral variables retained (above a threshold frequency) from the cumulative frequency distribution of all siPLS models. Finally, applicability of the selected PLS regression models in real NP toxicity assay is demonstrated.