(M1330-09-59) Development of Iterative Optimization Technology Algorithms with a Small-Scale Feed Frame Simulator: Stagnant vs. Dynamic Pure Component Spectra

Purpose: Process analytical technology (PAT) is widely used in the pharmaceutical industry for its ability to detect critical quality attributes in-line. One of the most common PAT tools is near-infrared (NIR) spectroscopy. The NIR data, however, requires advanced mathematical modeling to extract the relevant information. Partial least squares (PLS) models are preferred throughout the industry for predicting API content from NIR data, but their calibration demand is potentially burdensome to manufacturers. Pure component modeling approaches seek to reduce the calibration burden by minimizing or removing the calibration data requirement. Iterative optimization technology (IOT) algorithms have been demonstrated as viable pure component models for predicting API content.¹ One fundamental consideration assumption for IOT is the similarity of collection conditions for the pure components and the mixture samples.² In this work, pure components collected under stagnant and dynamic conditions are applied with IOT for the prediction of API content in a feed-frame simulator. A development set comprised of six API potency step samples was used to inform the selection of preprocessing and pure component spectral set. Spectral sets were selected using various model performance metrics. Overall, this work demonstrates the applicability of a development set for maintaining a material sparing pure component modeling effort while generating appropriate in-line API content predictions.

Methods: A five-component model formulation used Saccharin (Mutchler Inc.) as the model API, with Microcrystalline Cellulose (MCC, PH 102, FMC Corporation), Dicalcium Phosphate Anhydrous (DCP, Innophos, Inc), Sodium Starch Glycolate (SSG, Roquette) and Magnesium Stearate (MgSt, Mallinckrodt Inc) as excipients. The SentroPAT FO spectrometer with the SentroProbe NIR probe (Sentronic, Dresden, Germany) was used to collect the NIR spectra. A custom sample holder presented the stagnant pure components, while the small-scale simulator housed the SentroProbe for dynamic pure component and development set spectra collection. The five pure component spectra were combined with one another to create fifteen spectral sets. The development set was comprised of mixture spectra from six API potency steps. Several model metrics (root mean square error of prediction for the test set (RMSEP), root mean square error of prediction for the development set (RMSED), correlation coefficient (R2), absolute value of prediction bias, intercept and slope of prediction, and Euclidean norm of the IOT metric squared spectral residual (SSR)) were used to identify the best spectral set for each preprocessing treatment. Test set prediction performance highlighted the utility of metrics other than RMSED for indicating test set performance.

Results: Standard normal variate (SNV) preprocessing and Savistky-Golay (SavGol) smoothing with first derivative (SG1) and no derivative plus SNV (SG0 + SNV) were applied to the spectra. The model metrics mentioned in the methods section were calculated for the IOT model predicting the development set API content using each pure component spectral set under each preprocessing technique. The metrics were scaled from 0 to 1 for visualization purposes, seen in the heatmaps (figure 1). The lack of clarity regarding the best spectral sets prompted a statistical investigation to determine the best indicators of test set RMSEP. Each development set model metric was correlated with test set RMSEP under each preprocessing treatment. The results in table 1 highlight a key finding: the RMSED does not always directly inform the RMSEP. This is contradictory to the general assumption in PLS models where RMSE of calibration (RMSEC) and cross-validation (RMSECV) directly indicate the RMSEP. The significant correlations ( >0.70) are bolded. The correlated metrics were used in combination with the development set to select the best pure component spectral set for each preprocessing technique. The results are shown in table 2. The primary comparison is between the spectra set selected with RMSED and the set selected with the correlated metric for each preprocessing technique. Overall, the correlated metrics selected pure components which provided superior predictions to those selected with RMSED.

Conclusion: This work advances the application of IOT to consider two important situations in model development. First, the collection conditions of the pure component spectra can significantly impact the prediction performance based on the spectral shape change observed under varying collection conditions. Second, the metrics used to inform the selection of pure component spectra are dependent on preprocessing and should be carefully considered rather than assuming the representative nature of RMSE. Future work on this topic includes testing model transferability for the selection metrics by assessing prediction performance on alternate datasets using combinations of stagnant and dynamic pure component spectra.

References: 1. Rish, A.J., Henson, S. R., Alam, M., Liu, Y., Drennen, J. K., & Anderson, C. A., Comparison between pure component modeling approaches for monitoring pharmaceutical powder blends with near-infrared spectroscopy in continuous manufacturing schemes. The AAPS Journal, 2022. 24(4): p. 1-10.
2. Muteki, K., et al., Mixture component prediction using iterative optimization technology (calibration-free/minimum approach). Industrial & Engineering Chemistry Research, 2013. 52(35): p. 12258-12268.

Acknowledgements: The authors wish to thank Pfizer Inc. for funding this work.

Figure 1: Heatmaps of the development set metrics scaled from 0-1. The lack of observable superiority of any individual metric indicates the need for a statistical investigation.

Table 1: Development set metrics and their correlations with test set RMSEP under each preprocessing treatment. Significant correlations are bold. Metrics were selected based on the sign and magnitude of the correlations.

Table 2: RMSEP for the test sets for each preprocessing treatment. For each preprocessing treatment, RMSEP for pure component selection via RMSED and correlated metrics are shown to demonstrate the improvement in prediction performance.