DEVELOPMENT OF QUALITY BY DESIGN QBD GUIDANCE ELEMENTS

DEVELOPMENT OF QUALITY BY DESIGN (QBD) GUIDANCE ELEMENTS ON DESIGN SPACE SPECIFICATIONS ACROSS SCALES WITH STABILITY CONSIDERATIONS Blending Benoit Igne, Sameer Talwar, Brian Zacour, Carl Anderson, James Drennen Duquesne University Center for Pharmaceutical Technology

Objectives • Multi-sensor blend monitoring – 2 NIR sensors on a V-blender – Global decision criterion • Development of efficient calibration strategies – Limited sampling – Alternative calibration algorithm • Multi-component based end-point criteria

Instrumentation • 3. 5 quarts stainless-steel V-blender • 2 Near-infrared sensors (Spectral. Probe, Thermo. Fisher) • Real-time data collection and blend homogeneity monitoring

Formulation • Fluid bed dried granules (72%) were blended with extra-granular excipients – MCC (11. 3 %) – Starch (6. 8 %) – HPC (4. 5 %) – Crospovidone (2. 5 %) – Poloxamer (1. 25 %) – Talc (1 %) – Magnesium Stearate (0. 8%)

Modeling • Models based on an “efficient calibration” approach using Classical Least Squares (CLS) • 3 to 4 design points: 0, 100%, nominal(s) (target(s)) concentration(s) • Idea: – Take advantage of pure component spectra – Limit sample handling

Blend point • Root Mean Square Error to the Nominal Value Weighted, cumulative, pooled standard deviation that takes into account the deviation of the predicted concentration of the major components of a mixture to their target concentration, over a given number of rotations.

Workflow Router Decision

Calibration performances • Calibration performances (Instr. 1 and 2) RMSEC (%) = RMSECnom (%) = API MCC HPC Starch 1. 40 0. 95 1. 08 0. 70 1. 66 1. 09 1. 25 0. 75 RMSEC (%) = RMSECnom (%) = API MCC HPC Starch 1. 59 1. 13 0. 78 1. 13 1. 91 1. 32 0. 87 1. 27

Sensor 1 output

Pooled RMSNV • Combination of RMSNVs from both sensors for decision making Blend end-point

Scale up considerations • Significant challenges – Different blender shape – Different NIR sensors • Only 1 available • Only access to predictions • CLS modeling with efficient approach – Ready in 1 day – Powder properties comparable to those observed at small scale

Conclusions • The use of multiple sensors helped understand powder behaviors and limit risks of under or over blending • Efficient modeling techniques allowed for a simpler method implementation at scale up • The multi-component blend point statistic helped stop the blend consistently, even when properties of the granules were altered (different design of experiments), by relying on other components