Process analytical technology (PAT) has seen increased application in the pharmaceutical manufacturing industry following the introduction of the PAT guidance from FDA (FDA, 2004). Its flexibility in implementation to various unit operations promotes process monitoring and control by providing data and associated feedback in a timely manner. Systematic and mechanistic relationships between critical process parameters and critical quality attributes are established through analysis of intermediate materials (ICH , 2009). Effective and timely data collection from PAT is especially critical for continuous manufacturing (CM) processes, resulting in significant efforts in PAT implementation (ICH, 2021).

Near-infrared (NIR) spectroscopy is a popular PAT tool with both the pharmaceutical manufacturers and regulatory agencies. It is particularly valuable for its rapid collection times and minimal sample preparation requirements (Roggo, 2020, Ward, 2013, Hetrick, 2017). These characteristics are favorable for PAT application and facilitate interfacing of NIR with CM processes. A major application of NIR is blend potency monitoring in both upstream and downstream unit operations to ensure appropriate active pharmaceutical ingredient (API) potency and blend homogeneity (El-Hagrasy et al., 2006, El-Hagrasy and Drennen, 2006). Partial least squares (PLS) regression has become the preferred quantitative model for predicting API potency from NIR spectra. While PLS models show desirable predictive ability, the calibration burden is sometimes considered untenable. A PLS model requires a calibration set that captures the expected variance encountered during the manufacturing process. Typical calibration sets employ complex designs which often consume large quantities of experimental time and materials (Bondi, 2012). Reducing the calibration burden for models accompanying PAT tools has become an active interest across the pharmaceutical industry.

Pure component modeling, particularly iterative optimization technology (IOT), has been suggested as a strategy for minimizing calibration burden for quantitative prediction models. Models utilizing IOT have been demonstrated as effective for API potency prediction in pharmaceutical powder systems (Rish et al., 2022). The individual pure component spectra are the only model input for pure component models, drastically reducing the material demand compared to a traditional calibration set. The significant reduction in the calibration burden required to develop IOT models shows potential for minimization of the overall calibration burden for in-line PAT methods. IOT was selected over classical least squares (CLS) as the preferred pure component approach due to previous work showing its superior performance over CLS regression (Gupta, 2021, Rish et al., 2022). Pure component models achieve a material sparing approach to model calibration while maintaining sufficient prediction performance, motivating further investigation throughout pharmaceutical research.

A fundamental assumption of IOT is that the measured response in the mixture data is linear with respect to the concentration of the pure components (Muteki, 2013). An important corollary to this assumption is that the pure component spectral collection conditions are similar to the collection conditions for the mixture spectra. This criterion is challenging when one or more pure components cannot be processed independently, which is often true for CM processes. Additionally, pure components processed independently may not provide the best representation of the mixture spectra, resulting in subpar prediction performance. The focus of pure component collection conditions should be to maximize the representation of the mixture spectra rather than match the collection conditions.

Pure component spectral collections under varying conditions facilitate the construction of a pure component spectral library form which combinations of pure components comprise individual spectral sets. The spectral sets facilitate an assessment of the most representative pure component spectra for the prediction of mixture spectra. The primary challenge is the selection of representative pure component spectral sets, since there is no calibration set or mixture samples used as IOT model inputs. Introducing a small number of mixture samples facilitates pure component spectra selection while maintaining the reduced calibration burden associated with pure component approaches.

Strategies which reduce the material consumption for collection of pure component spectra are advantageous when collection in the production scale is unfavorable or unavailable. Differences in spectral shape (such as baseline or intensity) between production scale and pure component collection conditions can arise even when the conditions are similar, prompting the use of spectral preprocessing to align NIR spectral features between different spectral datasets (Blanco, 1997, De Maesschalck, 1998). Identification of suitable preprocessing techniques often requires mixture spectra with reference samples, presenting a similar challenge to the selection of a representative pure component spectral set.

A small set of mixture samples focused on the variance in the dependent variable is an alternative to a full calibration set for the purpose of identifying suitable preprocessing techniques and selecting representative pure component spectral sets. This is referred to as a “development set” and captures the primary variance of interest in the mixture samples while avoiding the size and complexity of a calibration set. Model performance metrics from IOT prediction of the development set can guide the selection of representative pure component spectra while maintaining a material sparing approach to model calibration. In this work, the development set was generated with a small-scale simulator designed to imitate the conditions of a production scale system with drastically reduced material requirements. This system has been used previously to create PLS calibration sets (Alam, 2021). Fig. 1 shows the workflow for applying the development set during IOT model calibration.

In this study, a development set containing ten spectra at six API potency steps (80–120 %) informed an IOT algorithm for prediction of API potency via in-line NIR spectra. The API potency was predicted for the development set using spectral sets consisting of combinations of stagnant and dynamic pure components, and the model performance metrics identified appropriate spectral sets for predicting API potency test sets generated in a portable continuous miniature modular (PCMM) manufacturing platform (Alam, 2021). The development set facilitated a material sparing selection of the best pure component spectral sets for the IOT model, effectively demonstrating a method for improving the prediction performance of pure component models in PAT applications.