(T1230-11-61) Reaching an Optimum Calibration-Free Solution: The Role of Numerical Solver Parameters on IOT Prediction of Active Ingredients in Pharmaceutical Powder Blends

Purpose: Spectroscopic data along with chemometrics models is commonly used to achieve process understanding in pharmaceutical manufacturing. Patrial least squares regression (PLS) is a widely used chemometric approach for process monitoring and, it requires substantial calibration data. Another approach to modeling spectroscopic data is classical least squares regression (CLS); this has an advantage of a greatly reduced calibration burden and functions as a pure component-based approach. Iterative optimization technology (IOT) is a different pure component-based calibration-free approach. Often, IOT applications are limited to simple systems like solutions and powder mixtures. This study aims to extend the applicability of IOT algorithm for process monitoring of a pharmaceutical powder blend containing a co-processed granular excipient. With the utilization of variable selection and available diagnostics, IOT demonstrated better prediction than the classical least squares regression (CLS) model while maintaining the advantage of no, or minimal mixture samples being required for implementation. This work also tested the role of solver parameters in prediction performances of the algorithm. With careful selection of solver parameters, the study demonstrated consistent and interpretable IOT predictions across different solvers.

Methods: A four components formulation containing two APIs: theophylline monohydrate (THY) and acetaminophen (APAP), and two excipients: Ludipress (contains 93% α-lactose monohydrate, 3.5% povidone, and 3.5% crospovidone) and magnesium stearate (MgST) was used for this study. A full-factorial mixture design with one central point was utilized to vary the ludipress amount and the THY:APAP ratio. Magnesium stearate % w/w was kept unchanged in all samples. Feed frame fill level and paddle speed were changed to provide additional variance to the system. The study design is illustrated in figure 1. All components except MgSt were blended in a V-blender for 15 minutes at 15 rpm followed by addition of MgSt into the mixture and blended for 5 minutes at 15 rpm. Samples were then transferred to the feed-frame simulator and NIR scans of range of 900-1700 nm were collected using MicroNIR PAT-U (VIAVI Solutions Inc., Scottsdale, Arizona, USA). Spectral data for pure components were collected in a small, inverted container. The data analysis was facilitated by the use of MATLAB (R2023a, MathWorks, USA).

Results: The sample dataset was divided into two groups based on the APAP concentration in the powder blends. Samples with < 5% w/w of APAP content are marked as ‘APAP low’ whereas the samples with APAP content higher than this value are considered ‘APAP high’ samples. IOT predictions showed distinctly different results between the two groups. The components with lower concentration levels were predicted poorly by both CLS and IOT. In particular, MgSt predictions demonstrated consistently large errors in all samples. MgSt prediction can be improved by putting a narrow constraint on the MgSt % w/w (for example from 0.5%-1.5% range).^[1] IOT predictions for THY, APAP, and Ludipress were improved by utilizing a variable selection method called spectral window wavelength mapper (SWAM) reported by Rish et al.^[2] The wavelength selection operation was done after preprocessing the raw spectral data with Savitzky-Golay (5 points, 2nd order, 1st derivative). These results are summarized in figure 2. Previous literature on the IOT algorithm has reported using different numerical solvers to obtain solutions.^[3] The effects of different solvers parameters were found critical for obtaining consistent realistic predictions from the IOT algorithm. Among these parameters, the initial point for the solver is found to be critical for the optimal solution. Since IOT approaches the optimization problem iteratively, presence of saddle points in feasibility region can yield different results if approached from different initial points.^[4] Other solver parameters that are important for IOT algorithm include, but are not limited to, tolerance for objective function value, maximum number of iterations, convergence criteria, mechanism to handle saddle points, etc. which needs to be carefully selected to obtain consistent outputs from IOT algorithm. The constraints used for the IOT algorithm also have a direct effect on the solution. Usually, a constraint of total mass fraction of all components summed to one is used (sum(r) = 1). For example, a continuous manufacturing process may have an NIR probe that is not in contact with the powder stream. In this case, the total mass fraction constraint of (r) = 1 may lead to unrealistic predictions. Figure 3 illustrates such an example, where IOT provided false prediction in absence of sample in front of the probe. Such false predictions should be associated with diagnostics that indicate inappropriate data for a prediction, such as SSR and NAS-T.^[5] Relaxation of constraints (e.g., (r) ≤ 1 facilitates more realistic predictions in this case.

Conclusion: IOT was implemented to predict APIs and co-processed excipients in powder blends in feed-frame simulator. The role of solver parameters is found to be critical for consistent IOT prediction. Constraint relaxation demonstrated interpretable prediction results.

References: 1. Muteki et al., 2013. Mixture component prediction using iterative optimization technology (calibration-free/minimum approach). Industrial & Engineering Chemistry Research, 52(35), 12258-12268.
2. Rish et al., 2023. Application of a wavelength angle mapper for variable selection in iterative optimization technology predictions of drug content in pharmaceutical powder mixtures. International Journal of Pharmaceutics, 643, 123261.
3. Shi et al., 2019. Estimation of mass‐based composition in powder mixtures using Extended Iterative Optimization Technology (EIOT). AIChE Journal, 65(1), 87-98.
4. Coelho et al., 2006. Combining of chaotic differential evolution and quadratic programming for economic dispatch optimization with valve-point effect. IEEE Transactions on power systems, 21(2), 989-996.
5. Rish et al., 2023. Diagnostic development using net analyte signal for pure component modeling approaches. Chemometrics and Intelligent Laboratory Systems, 243, 105007.

Acknowledgements: This work was funded by Genentech, Inc. (South San Francisco, CA, USA).


Figure 1: (A) Mixture design points for the formulation, and (B) table showing physical variant levels.


Figure 2: IOT prediction comparison between no wavelength selection (A) and wavelength selection with SWAM (B).


Figure 3: IOT prediction of unusual (blank) sample prediction. Showing (A) smear of powder on probe tip, (B) empty feed-frame, no sample in front of NIR probe, (C) comparison between regular vs blank sample spectral shape, (D) IOT predictions disparities at different constraint levels.