Applications of Infrared Spectroscopy in Monitoring Solvent Distillation during Early-Phase Pharmaceutical Process Development─Lean Chemometrics to Address Temperature and Matrix Effects

Abstract

Solvent exchanges are a common step in the synthetic chemical crystallization process during pharmaceutical development. In this study, we explore the application of in-line mid-infrared (IR) spectroscopy-based Process Analytical Technology (PAT) tools to determine the solvent content during distillation in early-phase development in the presence of an active pharmaceutical ingredient (API). During solvent swaps, the composition of solvents and temperature in the reactor can change drastically, generating nonideal solvent mixtures, and interference from the API signal creates additional challenges in modeling these systems. IR spectroscopy can be a useful tool in monitoring these changes, and lean chemometric models built off pure solvent spectra present an opportunity to capture the effects of temperature and API addition without requiring extensive calibration samples. An example solvent system was used for the study, where 2-methyltetrahydrofuran was replaced with heptane through a binary solvent exchange across a temperature range of 10–80 °C. Various solvent combinations were prepared at different ratios for a calibrated partial least-squares (PLS) model. Lean chemometric algorithms, namely classical least-squares and prediction-augmented classical least-squares (PACLS), with generalized least-squares weighting (GLSW) preprocessing, were applied to monitor these systems via minimizing the calibration burden. Lean model performance and traditional calibrated PLS performance were compared and evaluated by the root mean squared error of prediction for solvent exchanges between binary solvent mixture runs and additional runs with the inclusion of a surrogate API (benzophenone). Results demonstrate that spectral variation influenced by the temperature and API signal on the predictive abilities of multivariate models can effectively be mitigated through a two-pronged GLSW and PACLS lean modeling approach. This combination modeling approach appears to be competitive with PLS, without the additional time and resource constraints of preparing extensive training data, while also circumventing matrix dependence.