Enhancing understanding of the prediction behavior of the iterative optimization technology (IOT) algorithm using model diagnostics.

Spectroscopy-based process analytical technology (PAT) has become the method of choice in the pharmaceutical industry for its non-destructive and rapid monitoring of critical quality attributes (CQA) of both intermediate and final drug products. When coupled with appropriate chemometric models, complex spectroscopic signals can be translated into meaningful process understanding. However, a significant amount of calibration data with corresponding reference values is often required to train traditional chemometric models, such as Partial Least Squares (PLS) regression, for optimum prediction accuracy. The scarcity of active pharmaceutical ingredients (APIs) and the time and effort required for collecting reference values limit the application of PAT at early drug product development stages. An alternative lean chemometric approach, iterative optimization technology (IOT), can be utilized in a calibration-free manner. Unlike the traditional regression approach in PLS, IOT formulates spectral interpretation as an optimization problem, using numerical solvers to estimate mixture compositions from known pure component spectra with pre-determined constraints. The consistency and reliability of the solver are critical for achieving accurate predictions and model robustness. In this study, a solver parameter, Lagrange Multiplier (LM), was studied along with principal component analysis (PCA) to understand solver performance and predictive behavior of IOT algorithm, particularly for the low-concentration components within the formulation. Furthermore, the integration of model diagnostics enhances confidence in predictive performance. Signal-to-noise ratio (SNR) was explored to characterize the predictive behavior of IOT, particularly in scenarios involving low-concentration components in formulation.