Characterization of Near Infrared and Raman Spectroscopy for In-line

Subscriber access provided by Bethel University

Article

Characterization of Near Infrared and Raman Spectroscopy for In-line Monitoring of Low-Drug Load Formulation in Continuous Manufacturing Process Zachary D. Harms, Zhenqi Shi, Rajesh A. Kulkarni, and David P. Myers Anal. Chem., Just Accepted Manuscript • DOI: 10.1021/acs.analchem.8b05002 • Publication Date (Web): 29 May 2019 Downloaded from http://pubs.acs.org on June 4, 2019

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

Characterization of Near Infrared and Raman Spectroscopy for Inline Monitoring of Low-Drug Load Formulation in Continuous Manufacturing Process Zachary D. Harms,* Zhenqi Shi, Rajesh A. Kulkarni, and David P. Myers Small Molecule Design and Development, Eli Lilly and Company, Indianapolis IN, 46285. ABSTRACT: Reflectance spectroscopy is an excellent candidate for process analytical technology (PAT) applications in continuous manufacturing of pharmaceutical tablets. Spectroscopic methods provide a real-time, non-destructive measurement of active pharmaceutical ingredient (API) concentration in order to ensure product quality and uniformity. Of particular challenge are powder blends with low drug loads (< 5% w/w) where measurement signal-to-noise, and in turn precision, limit ability of the method. We evaluate both near infrared (NIR) and Raman spectroscopy for use in PAT applications by measuring pharmaceutical blends of varying active ingredient concentration. Both spectrometers are equipped with a fiber-optically coupled probe head for non-contact measurement of powder blends. A mockup of the interface between the spectrometer and powders within the feed frame of a rotary tablet press is used to simulate the movement of powder blends from the mixer to the press. A port on the feed frame allows measurement by NIR or Raman spectroscopy of the blends just before tablet compression. For our model compound, Raman spectroscopy is shown to have a lower limit-of-detection and less day-to-day variability than NIR spectroscopy. Raman spectroscopy was chosen as the PAT platform to support process development and working distance and spot size were both optimized for use in the feed-frame of a tablet press. Sufficient limit-of-detection was achieved for monitoring active pharmaceutical ingredient concentration (API) down to 1% w/w during a semi-continuous manufacture of tablets. An innovative optimization-based model (EIOT) was used to trend API concentration and demonstrated that the process could be capable of detecting out-of-trend material.

The issuance of the PAT guideline by the FDA1 in 2004 combined with the trend of continuous manufacturing of drug product (CM DP) within the pharmaceutical industry has propelled the use of PAT tools in commercial manufacturing in the recent years. In a batch process, it is often the case that PAT effort in development, intended to support process development via quality-by-design (QbD) studies and subsequent scale-up/transfer to a commercial manufacturing site, ceases its journey during the tech transfer because the use of PAT tools improves process understanding and reveals/enables the final control strategy. Thus, historically only a small fraction of PAT tools have been transferred and deployed for routine batch manufacturing. For continuous processing, the use of spectroscopy-based PAT tools have enabled characterization of system dynamics,2 empowered us to differentiate and understand the root cause for common cause variabilities,3,4 and ultimately to develop multi-layered control strategies to address such variabilities.5 Moreover, given pharmaceutical industry’s limited experience on CM, a PAT tool used throughout the development cycle is likely to be deployed onto the manufacturing floor in order to not only serve as the verification tool to assure critical quality attributes meet their specifications, but also to derisk any carry-over from special cause variabilities either independently or via a combination with other process monitoring tools, such as multivariate statistical process control (MSPC). Spectroscopy-based PAT tools for routine manufacturing not only supports real-time decision making, but also enables real-time release testing (RTRt). In addition, continued use of spectroscopic PAT tools and MSPC for CM DP could provide justification to switch from traditional 3-batch process validation to continuous process verification, which could result in significant resource and material savings. In CM DP, real-time monitoring at the feed-frame of the tablet press offers many advantages.6-12 For a continuous direct compression line, the feed-frame was reported to be the final mixing element.4 As a consequence, the blend homogeneity determined upstream before the feed-frame may not be representative of the content uniformity of individual solid oral dosage forms. Moreover, the feed-frame is the last location before powder is compressed into tablets. The real-time measurement of API concentration at the feed-frame that is outside acceptable ranges triggers immediate tablet rejection by the press. Therefore, developing and deploying the use of near infrared (NIR) ACS Paragon Plus Environment

1

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 2 of 16

spectroscopy measurements at the feed-frame for the CM DP has seen increased adoption by both industrial and academic practitioners.4 With more understanding and experience accumulated on the use of near infrared (NIR) spectroscopy at feedframe for CM DP, the detection of API content in low drug-load blends (i.e., < 5%, w/w) has been realized as a challenge in the field. Raman spectroscopy is one such alternative that can be used in place of NIR spectroscopy for monitoring powder blending processes,13 though concerns exist on its sensitivity relative to NIR spectroscopy.14 Prior reports have demonstrated that Raman spectroscopy can be used for feedback control in a twin-screw blending process15 and for monitoring blending of 1% API in a batch process.16 Recently, the use of Raman spectroscopy at the feed-frame was reported to be capable of measuring the API content in the final blend with a target amount of 8% API (w/w).17 Given these reasons, this paper documents our recent efforts at Lilly to conduct a comparison on the use of NIR and Raman spectroscopy at the feed-frame for real-time process monitoring for CM DP. Three drug loads (1%, 3%, and 8%) of an investigational API were evaluated on an off-line device (i.e., feed-frame table)4 by both NIR and Raman spectroscopy. In this paper, the initial feasibility of Raman spectroscopy was not only characterized by optimizing working distance across different optic lenses, but also compared against that of NIR spectroscopy in terms of its sensitivity, within and between day precision, and robustness against process conditions (i.e., paddle speed). In this report, the use of Raman spectroscopy to perform real-time monitoring on a semicontinuous manufacturing run of the 1% drug-load blend illustrates the potential of Raman spectroscopy to serve as a potential solution to measure API content in low drug-load blend for CM DP. EXPERIMENTAL SECTION Materials. One kg each of nominal blends of three different drug loads (1%, 3%, and 8% (w/w) active pharmaceutical ingredient) of a proprietary drug molecule, and the associated placebo blends were prepared using a Turbula blender (Glen Mills Inc., Clifton, NJ) for analysis by NIR and Raman spectroscopy on the feed-frame table. The formulation contained lactose monohydrate, microcrystalline cellulose, croscarmellose sodium and sodium stearyl fumarate as the excipients. When the API concentration changed across three nominal blends, the lactose amount was proportionally adjusted to balance the blend composition. Spectral data collection on the feed-frame table. Both NIR and Raman spectroscopy were used in the reflectance mode and spectral data were collected on an off-line device (i.e., feed-frame table). A more detailed description of the feed-frame table can be found elsewhere.4 In brief, the off-line feed-frame device consisted of a feed-frame from a commercial tablet press (Korsch XL200, Berlin, Germany, used in CM DP rig) installed on a table with a motor to drive the feed-frame paddles at a desired speed, revolutions-per-minute (RPM), and also with a slide gate at the end of the feed-frame to regulate the flow of powder. The intended purpose is to mimic the dynamics of moving powder off-line in order to collect representative spectra using a small (≤ 1 kg) batch of blended powders and avoid using the whole CM DP rig. The spectroscopic measurement is made on the moving powder immediately before material would be directed to the dies of the tablet press. The NIR spectrometer and probe was a Prozess 611 NIR Monitoring System (Prozess Technologie, St. Louis, MO) consisting of a tungstenhalogen lamp for illumination and a NIR diode-array detector with a nominal wavelength range of 1100 – 2100 nm with approximately 5 nm wavelength spacing. The integration time was 11 ms with 100 averages for a total time of 1.1 s. Spectral collection frequency was set to every two seconds. The Raman spectrometer was a Raman Rxn2 analyzer (Kaiser Optical Systems, Inc, Ann Arbor, MI) equipped with a PhAT probe and had an excitation wavelength of 785 nm. Three different optic lenses with varied working distances were compared that had spot sizes of 3 mm, 4.7 mm, and 6 mm. The 3 mm optic was used in the PCA analysis and in all comparisons to the NIR data. The nominal wavenumber range was 150 – 1890 cm-1 with 1 cm-1 spacing. The exposure time was 1 second with no averaging. Spectral collection frequency was every three seconds, given the overhead to read out and store the spectral signals. Spectral collections via both spectroscopic tools were conducted across two paddle speed conditions including 15 RPM and 30 RPM to simulate the throughput that would typically be observed in a continuous manufacturing process. Slide gate setting was fixed throughout data collection, per an early report4 that it was not found to be as significant as paddle speed to contribute to spectral variability. The spectral collection took place across multiple days, in which NIR and Raman data were collected across five and three days, respectively. Raman data collection and analysis on CM DP process equipment for the 1% drug load tablet. The Raman spectrometer and probe equipped with the 3 mm optic lens was used to perform real-time monitoring on a semicontinuous manufacture of a 1% drug load tablet. Data collection frequency on the process data was every three seconds. The 1% drug load manufacture was a semi-continuous manufacturing process in which pre-blended ACS Paragon Plus Environment

2

Page 3 of 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


material was dispensed into the hopper of a tablet press. The tablet press was the same as used in Lilly’s CM process. Throughout the run, core tablets were sampled for HPLC analysis of API content (assay) and expressed as percent nominal concentration, which is the measured API content divided by expected API content after adjusting corresponding tablet weight. Chemometric and EIOT analysis. For the data collected on the off-line feed-frame table, appropriate spectral preprocessing was conducted on both NIRS and Raman data followed by principal component analysis to characterize the spectral variability in score space caused by API concentration, paddle speed, and within/between day precision. For the Raman data collected on the CM DP for the 1% drug-load formulation, extended iterative optimization technique (EIOT) was used to translate the Raman spectra to predicted API concentration. More detailed info of EIOT can be found elsewhere.18 In brief, EIOT is a novel spectral analysis technique based upon optimization under a constrained environment to decompose spectral data by strictly following Lambert-Beer’s law. The unique advantage of EIOT compared to traditional chemometric analysis is that it only requires very lean/limited training data. EIOT was designed to handle spectroscopy-based process monitoring in early R&D when API supply is too limited to perform PLS calibration. The EIOT analysis was composed of two steps. First, the spectral data collected on the feed-frame table for the placebo blend and the 1% blends were used as the training data (DM) in order to extract the estimated pure component spectra (𝐒′) for API and the counter-diluent via Eq. (1). Since the major diluent was used in the placebo blend to counter the loss of API, the nominal blend compositions for the API and the summation of excipient were used in order not to introduce constant mass fraction column into the inverse operation in Eq. (1). Standard normal variate (SNV) was used as the spectral preprocessing. (1)

𝐒' = (𝐂T𝐂)

―1 T

𝐂 𝐃𝐌

Dm represents the training spectral data and C represents the mass fraction data from the experiments on the feedframe table for both placebo and 1% drug load blend at the 30 RPM paddle speed. Spectral data collected at only one paddle speed were used in the training data in order to accurately capture estimated pure component spectra for individual chemical species without potential physical interference caused by paddle speed. Prior experience on EIOT illustrated that mixing both chemical and physical variability in the training data prohibits the accurate representation of the pure component species.18 Second, the estimated pure component spectra (𝐒′) were used to estimate mass fraction for a new sample spectrum collected during the continuous run under constrained environment using non-linear programming via Eq. (2). Spectral signal from 720 cm-1 to 1720 cm-1 were used for EIOT analysis. Standard normal variate (SNV) was also used as the spectral preprocessing for the analysis. min (𝜺𝑇𝜺) (2) subject to 𝜺 = 𝑑𝑠𝑎𝑚𝑝𝑙𝑒 ― 𝑑𝑠𝑎𝑚𝑝𝑙𝑒 𝑑𝑠𝑎𝑚𝑝𝑙𝑒=𝑟𝑠𝑎𝑚𝑝𝑙𝑒 ∙ 𝑺 ∑𝑛

𝑟 𝑖=1 𝑖

=1

0 ≤ 𝑟𝑖 ≤ 1 𝑑𝑠𝑎𝑚𝑝𝑙𝑒 represents a new sample spectrum with an unknown mass fraction on n components. Here n = 2, the API and the sum of the excipients. 𝑟𝑠𝑎𝑚𝑝𝑙𝑒 represents the estimated mass fraction of API and the sum of excipients for the new sample spectrum, in which the mass fraction was subjected to two constraints of being non-negative and the summation to be unity. The algorithm EIOT itself is capable to be expanded to take into consideration of those non-chemical interferences, such as powder density and paddle speed in the feed frame table.18 Given the non-observable impact of paddle speed (see the Results and Discussion section) on Raman data, the capability of the algorithm handling non-chemical interference was not shown here. All relevant calculations were executed using either MATLAB 2017b or MATLAB 2018a (The Mathworks Inc., Natick, MA). The PLS Toolbox 8.5.2 (Eigenvector Research, Inc., Manson, WA) used for PCA. An in-house code was written for the analysis on EIOT.

ACS Paragon Plus Environment

3


Page 4 of 16

RESULTS AND DISCUSSION Raman and NIR Spectra. Representative NIR and Raman spectra are shown in Figure 1 for the pharmaceutical blend samples. Spectra have been pre-processed by SNV normalization to aid in visualization of the API absorbance. The wavelength regions in the figure are truncated to the regions in which the active compound absorbs and also correspond to the wavelength regions used in the final PCA models in Figures 2 and 4. The complete NIR and Raman spectra of both the blend samples and API are given in Supporting Information, Figures S1 and S2. In 1a, NIR spectra show that the active compound absorbs at 1135 nm as demonstrated by the light blue trace corresponding to the 8% drug load blend. However, there is little-to-no differentiation between the 1%, 3%, and placebo blends. The Raman spectra for these samples are shown in Figure 1b, where the active compound scatters at 1610 to 1640 cm-1. The slope in the baseline of the Raman spectra is due to the fluorescence of the excipients in the blend. In contrast to the NIR data, the peaks in the Raman spectra have a much higher signal-to-noise ratio and all blends are clearly distinguishable from the placebo blend, including the 1% drug load blend. These Raman peaks are most likely attributed to C=C or C=O bonds in the API due to their position and strength at these low levels.19 Principal component analysis (PCA) is used to compare the NIR and Raman data. In both data sets, a two principle component (PC) model was sufficient to capture the majority (98.79% for NIR and 84.93% for Raman) of the variance. NIR analysis by PCA. In the NIR spectra, there are two wavelength regions that show high API absorbance and minimal interference from excipient absorbance, 1100 to 1300 nm and 1600 to 1750 nm as shown in Figure S1. PCA models were built from both regions and compared to models of the entire wavelength range. These wavelength selections and a variety of spectral pre-processing conditions were examined as summarized in Supporting Information, Table S1. As demonstrated previously for NIR datasets, SNV normalization alone or in combination with 1st or 2nd derivative filtering within a narrow region of the spectrum provided low RMSEC while minimizing sensitivity to effects of excipient variability.4 Here, 2nd derivative pre-processing alone resulted in the lowest RMSEC for each wavelength region tested. In all models, there was significant day-to-day variability as shown in four example scores plots in Figure S3. The wavelength range of 1104 to 1296 nm and 2nd derivative preprocessing resulted in the lowest RMSEC of 4.69 x10-6 and was chosen for the final model. The scores and loadings plots of the final NIR PCA model are shown in Figure 2. 94.96% of variance is in PC1 and 3.83% of the variance is in PC2. The scores plots are color-coded by rate-per-minute (RPM) of the feed frame paddle in (a), day in (b), and concentration in (c). The model is largely insensitive to paddle speed, as seen in the overlap of the scores plots for the 15 and 30 RPM data in 2a. For API concentration, there is some separation in the scores along PC2 that trend with API concentration, however there is significant day-to-day variability along PC1. The difficulty in building a PCA model sensitive to API concentration is likely due to the poor signal-to-noise of the API in the sample spectra (Figure 1). The PC2 loadings are overlaid with -1 * the 2nd derivative API spectrum in Figure 2d. The 2nd derivative is taken of the API spectrum to match the pre-processing of the data in the PCA model, and the spectrum is also multiplied by -1 to as the scores along PC2 in 2c trend negative as API concentration increases. The PC2 loadings and API spectrum show alignment with the API absorbance at 1135 nm. Data were collected over five days and the variability in the measurement from day to day was visible in the scores plots as shown in Figure 2b. To quantify day-to-day variability, the resolution between scores distributions between days three and four was studied for each drug load. First, the average width of each distribution was calculated. For each drug load within a given day, the length of the vector connecting the coordinates of each score to the distribution center is calculated, and the average vector length for day three and day four scores distributions is given by 𝑤 in Eq. 3 below. For each drug load, the length of the vector connecting the center of day three’s scores plots to the center of day four’s scores plots is l. The equation for the resolution of the two distributions, R, is defined in Eq. 3 as, (3)

𝑅 = 𝑙/𝑤

For the NIR data, the placebo had the greatest separation between days three and four at 50.7 ± 12.7. The 3% drug load was at 23.1 ± 9.5 and the 8% drug load was at 2.9 ± 1.6. Day-to-day variability decreases as drug load increases. However, even at the highest drug load, the high day-to-day variability is problematic in that it makes prediction of API concentration of an unknown sample difficult. Within-day variability for days three and four was also determined. The RSD in the average distribution widths, 𝑤, is used as an indication of within-day variability. For the 8% drug load, day three had an average RSD of 48% and day four had an average RSD of 69%. The RSD for the placebo and 3% drug load is in Table S3. Due to the scale of the scores plots, it is difficult to visualize the ACS Paragon Plus Environment

4

Page 5 of 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


shape of each individual distribution, so for the 8% drug load data a magnified view of the scores distributions for days three and four is shown in Figure S4. Raman Analysis by PCA. In the Raman spectra, there are several sharp peaks in the API spectra across the wavelength range of 1700 to 200 cm-1 as shown in Figure S2. However, fluorescence from excipients is pronounced at low wavenumbers. As with the NIR data, several wavelength selections and spectral pre-processing operations were surveyed and the results are summarized in Table S2. SNV normalization followed by 2nd derivative preprocessing provided the lowest RMSEC for each wavelength range. For wavelength selection, it was determined that as the spectral window decreased, the RMSEC increased but the day-to-day variability as observed in the scores plots decreased. Scores plots from three models are shown in Figure 3, the entire wavelength region (1890 to 150 cm-1), 1700 to 1200 cm-1, and 1700 to 1575 cm-1. As the wavelength selection decreases, the separations in scores from different days’ experiments also decreases, with near complete overlap for the smallest spectral window. RMSEC increases from 2.07E-4 for the model in (a) to 1.19E-3 to the model in (c). Minimizing day-to-day variability is more important here than a modest decrease in RMSEC. The region of 1700 to 1575 cm-1 has two peaks 1610 and 1640 cm-1 (Figure 1b) corresponding to the API and also has a minimum of background fluorescence. These data suggest that the variability in background fluorescence are driving some of the day-to-day variability in the Raman measurements. SNV normalization followed by 2nd derivative on the wavelength range of 1700 to 1575 cm-1 was selected to build the final PCA model for the Raman data analysis and the scores plots are shown in Figure 4. Here, 81.75% of variance is in PC1 and 3.18% of the variance is in PC2. Again, the scores plots are color-coded by RPM of the feed frame paddle in (a), day in (b), and concentration in (c). As with the NIR PCA model, no effect was observed with regards to feed frame paddle speed (4a). At 1 s per spectrum, the variability associated with the moving paddle is likely minimized. In contrast to the NIR model, most of the concentration variability is captured in the first PC as shown in 4c. The Raman signal provides acceptable signal down to the 1% drug load, and there is clear definition between the placebo, 1%, 3%, and 8% drug loads in the scores plot. The PC1 loadings are overlaid with -1 * the 2nd derivative API spectrum in 2d and align with the two peaks at 1610 and 1640 cm-1. There is less day-to-day variability (4b) in the Raman measurement than in the NIR measurement. Equation 3 is used to quantify day-to-day variability through the separation between scores plots between days one and two in Figure 4b. The placebo scores had a separation of 0.63 ± 0.40, the 3% drug load was 0.55 ± 0.33, and the 8% drug load was 0.99 ± 0.58. These separation factors indicate that the distributions are not resolved and that the day-today variability associated with the measurement is minimized. Average RSD for 𝑤 for the 8% drug load was 64% for day one and 54% for day 2. These values are comparable to the NIR values and the sample size is insufficient to show a difference in between-day precision. In the comparison between NIR and Raman, consideration should also be given to the amount of material analyzed in each respective measurement. Acquisition times for the NIR and Raman were 1.1 and 1 s, respectively. Spot sizes were 5 mm for the NIR and 3 to 6 mm for Raman. Powder density and feed-frame paddle speed are constant between both experiments. Due to the similar spot size and acquisition time, the volume of powder scanned per measurement is comparable for both measurements. Raman development. In order to optimize the signal, the effect of spot size and working distance between the optic and the sample were investigated. Three optics were studied, with spot-size diameters of 3, 4.7, and 6 mm. All three optics were largely insensitive to small changes (~2 – 4 cm) in working distance, demonstrating excellent depth-of-field which is important in studying the dynamic powder flows in the feed frame. In Figure S5, the signal amplitude as a function of working distance is shown for each optic. Optimized working distances of 10.2, 15.2, and 20.3 cm were chosen for the 3, 4.7, and 6 mm optics, respectively. Figure 5 shows the comparison of three optics of varying spot size. Working distance for each optic was set at the optimized position and the peak at 1610 cm-1 is monitored for the 8% blend sample. As shown in 5a, the raw counts in signal intensity correlate to diameter, with counts increasing as diameter decreases. Laser power was held constant at 400 mW (180 mW measured through a fiber optic at the sample) and the increasing signal is most likely due to the increased power density or irradiance (power/cm2) as the laser is focused over a smaller spot size with the 3 mm optic having the highest raw intensity signal. The magnitude of the fluorescent background also increases with decreasing optic diameter, but this effect can be removed by SNV pre-processing. Figure 5b shows the raw signal after it has been normalized by SNV pre-processing. Peak heights are all identical in the processed spectra, however the baseline noise decreases with decreasing spot size. The relationship between spot size and SNR is shown in Figure 6. The standard deviation of the SNV-normalized baseline decreases with decreasing spot size. Conversely, the SNR increases with decreasing spot size, with the 3 ACS Paragon Plus Environment

5


Page 6 of 16

mm optic having the highest SNR at 38 for the peak at 1610 cm-1. As peak height is constant for the normalized spectra, the decreasing noise is driving the increase in SNR. High SNR and low RSD are needed so that disturbances of 5-10% from the target level can be detected during continuous manufacturing of drug product and nonconforming material can be rejected. Raman monitoring of 1% drug load manufacture. NIR was shown to be incapable of measuring the analyte, so Raman was chosen to monitor the mixing homogeneity at the feed frame during development manufacture campaign of the pharmaceutical compound. The 3 mm optic demonstrated the best SNR under the SNV-normalized spectral pretreatment, and was chosen to monitor a semi-continuous run of 1% drug load tablet. The estimated pure component spectra by EIOT are compared to experimental spectra in Figure 7b. The most unique scattering bands around 1600 cm-1 (i.e., the double peaks) were well captured by Eq. (1) indicating the validity of EIOT to capture concentration variability of API. Throughout the semi-continuous run, the Raman spectra were collected every 3 s during a 25 min run time (Figure 7a) and the stratified core tablets from six locations were collected and 3 tablets from each location were analyzed by HPLC for % nominal concentration, which is the measured amount of API in each tablet divided by the expected amount. HPLC data are also weight corrected, meaning the data are normalized by tablet weight so that the data represent the amount of API in the blend and do not fluctuate with tablet weight differences. The estimated mass fractions by EIOT were normalized by the target API concentration to obtain % theoretical concentration and was compared against weight-corrected nominal concentration on tablets analyzed using HPLC. Agreement between the weight-corrected HPLC and the EIOT prediction is good, with average values of 97.9 ± 0.7% and 98.9 ± 1.1% respectively. These data demonstrate that both accuracy and precision of the EIOT prediction are comparable to conventional HPLC testing. CONCLUSIONS NIR and Raman spectroscopy are valuable and complementary tools for monitoring continuous drug product manufacturing. Choice of NIR versus Raman spectroscopy will depend on both the API properties as well as the excipient matrix. In this study, Raman spectroscopy has demonstrated sufficient limit-of-detection for monitoring drug load down to 1% w/w. Raman was also shown to have less day-to-day variability, and has the advantage of greater chemical specificity than NIR. EIOT is also shown to be successful in predicting drug load for the Raman data set. EIOT represents a leaner approach to model building as compared to a full PLS prediction and is a valuable tool in early development in which time and resources are constrained. Further investigation is needed to build a quantitative model from 80 to 120% of nominal concentration to study the sensitivity of Raman spectroscopy and to understand the impact of varying excipient matrices (both in level of each excipient and in lot-to-lot raw material differences). Faster acquisition times and larger spot sizes are both advantageous in content uniformity determinations in CM. Application of currently available Raman platforms into the GMP environment of pharmaceutical tablet production is also under consideration.


6

Page 7 of 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 1. NIR and Raman spectra. NIR spectra (a) and Raman spectra (b) of placebo and sample blends. Inset in (a) shows API NIR absorbance peak at 1135 nm. All spectra have been normalized by SNV pre-processing.


7


Page 8 of 16

Figure 2. Principal component analysis (PCA) of NIR spectra. Scores on PC 2 versus PC 1 are color-coded by RPM (a), day (b), and active concentration (c). Dashed line represents 95% confidence level. The loadings for PC2 and -1 * the second derivative of the API spectrum are shown in (d).


8

Page 9 of 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 3. Raman PCA scores plots for three models. All three models are pre-processed by SNV then 2nd derivative, but differ in the wavelength region used to build the model. In a) all wavelengths are used (Model Number 11 in Table S2), b) 1700 to 1200 cm-1 (Model Number 14), and c) 1700 to 1575 cm-1 (Model Number 5, final model used in analysis). RMSEC values are 2.07E-4 for (a), 3.93E-4 for (b) and 1.19E-3 for (c). All data are color-coded by API concentration in weight percent.


9


Page 10 of 16

Figure 4. Principal component analysis (PCA) of Raman spectra. Scores on PC 2 versus PC 1 are color-coded by RPM (a), day (b), and active concentration (c). Dashed line represents 95% confidence level. The loadings for PC1 and -1 * the second derivative of the API spectrum are shown in (d).


10

Page 11 of 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 5. Raman spectra of sample blends as a function of spot size. Raw spectra of a Raman peak at 1610 cm-1 for 3, 4.7, and 6 mm optics are shown in (a) and the SNV normalized spectra of the same region are shown in (b). Spectra in (b) are y-offset to aid visualization.


11


Page 12 of 16

Figure 6. Effect of Raman spot size on signal-to-noise ratio (SNR). Standard deviation of SNV normalized Raman spectra (a) decreases and SNR (b) increases as spot size decreases.


12

Page 13 of 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 7. EIOT prediction. In (a) EIOT is used to predict active concentration for a 1% target blend concentration during a batch manufacture and compared against the weight-corrected HPLC measurement. Values are expressed in percent nominal concentration (target is 1% w/w). In (b) the estimated API and placebo spectra from the EIOT model are compared to measured API and placebo spectra. Both estimated and measured spectra were after SNV preprocessing.


13


Page 14 of 16

ASSOCIATED CONTENT Supporting Information The supporting Information is available free of charge on the ACS Publications website. NIR and Raman spectra of API and blend samples, NIR PCA model building and scores plots, Raman PCA model building, Tables of RMSEC and RMSECV for PCA models, Table of within-day variability calculations and Raman working distance optimization

AUTHOR INFORMATION *E-mail: [email protected]

ACKNOWLEDGEMENTS The authors thank Tony Cooper, Aaron Garrett, Robert Glenn Rupard, James Hermiller, Bryan Castle, and Ahmad Almaya for technical guidance and helpful discussions related to the development of this manuscript.

REFERENCES

(1) Services, U. S. D. o. H. a. H. Guidance for Industry: PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance. FDA, http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/ucm0703 05.pdf, 2004. (2) García‐Muñoz, S.; Butterbaugh, A.; Leavesley, I.; Manley, L. M.; Slade, D.; Bermingham, S. A flowsheet model for the development of a continuous process for pharmaceutical tablets: An industrial perspective. AIChE J 2018, 64, 511-525. (3) Hanson, J. Control of a system of loss-in-weight feeders for drug product continuous manufacturing. Powder Technol 2018, 331, 236-243. (4) Hetrick, E. M.; Shi, Z. Q.; Barnes, L. E.; Garrett, A. W.; Rupard, R. G.; Kramer, T. T.; Cooper, T. M.; Myers, D. P.; Castle, B. C. Development of Near Infrared Spectroscopy-based Process Monitoring Methodology for Pharmaceutical Continuous Manufacturing Using an Offline Calibration Approach. Anal Chem 2017, 89, 9175-9183. (5) Almaya, A.; De Belder, L.; Meyer, R.; Nagapudi, K.; Lin, H. H.; Leavesley, I.; Jayanth, J.; Bajwa, G.; DiNunzio, J.; Tantuccio, A.; Blackwood, D.; Abebe, A. Control Strategies for Drug Product Continuous Direct Compression-State of Control, Product Collection Strategies, and Startup/Shutdown Operations for the Production of Clinical Trial Materials and Commercial Products. J Pharm Sci 2017, 106, 930-943. (6) Liu, Y.; Blackwood, D. Sample Presentation in Rotary Tablet Press Feed Frame Monitoring by Near Infrared Spectroscopy. Am Pharm Rev 2012, http://www.americanpharmaceuticalreview.com/FeaturedArticles/116357-Sample-Presentation-in-Rotary-Tablet-Press-Feed-Frame-Monitoring-by-NearInfrared-Spectroscopy/. (7) Mateo-Ortiz, D.; Colon, Y.; Romañach, R. J.; Méndez, R. Analysis of powder phenomena inside a Fette 3090 feed frame using in-line NIR spectroscopy. J Pharm Biomed Anal 2014, 100, 40-49. (8) Šašić, S.; Blackwood, D. O.; Liu, A.; Ward, H. W.; Clarke, H. Detailed analysis of the online nearinfrared spectra of pharmaceutical blend in a rotary tablet press feed frame. J Pharm Biomed Anal 2015, 103, 73-79. (9) Ward, H. W.; Blackwood, D. O.; Polizzi, M.; Clarke, H. Monitoring blend potency in a tablet press feed frame using near infrared spectroscopy. J Pharm Biomed Anal 2013, 80, 18-23. (10) Y., L.; Blackwood, D. Sample Presentation in Rotary Tablet Press Feed Frame Monitoring by Near Infrared Spectroscopy. Am Pharm Rev 2012, http://www.americanpharmaceuticalreview.com/FeaturedArticles/116357-Sample-Presentation-in-Rotary-Tablet-Press-Feed-Frame-Monitoring-by-NearInfrared-Spectroscopy/. ACS Paragon Plus Environment

14

Page 15 of 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


(11) Sierra-Vega, N. O.; Sánchez-Paternina, A.; Maldonado, N.; Cárdenas, V.; Romañach, R. J.; Méndez, R. In line monitoring of the powder flow behavior and drug content in a Fette 3090 feed frame at different operating conditions using Near Infrared spectroscopy. J Pharm Biomed Anal 2018, 154, 384-396. (12) De Leersnyder, F.; Peeters, E.; Djalabi, H.; Vanhoorne, V.; Van Snick, B.; Hong, K.; Hammond, S.; Liu, A. Y.; Ziemons, E.; Vervaet, C.; De Beer, T. Development and validation of an in-line NIR spectroscopic method for continuous blend potency determination in the feed frame of a tablet press. J Pharm Biomed Anal 2018, 151, 274-283. (13) De Beer, T. R. M.; Bodson, C.; Dejaegher, B.; Walczak, B.; Vercruysse, P.; Burggraeve, A.; Lemos, A.; Delattre, L.; Heyden, Y. V.; Remon, J. P.; Vervaet, C.; Baeyens, W. R. G. Raman spectroscopy as a process analytical technology (PAT) tool for the in-line monitoring and understanding of a powder blending process. J Pharm Biomed Anal 2008, 48, 772-779. (14) Allan, P.; Bellamy, L. J.; Nordon, A.; Littlejohn, D.; Andrews, J.; Dallin, P. In situ monitoring of powder blending by non-invasive Raman spectrometry with wide area illumination. J Pharm Biomed Anal 2013, 76, 28-35. (15) Nagy, B.; Farkas, A.; Gyürkés, M.; Komaromy-Hiller, S.; Démuth, B.; Szabó, B.; Nusser, D.; Borbás, E.; Marosi, G.; Nagy, Z. K. In-line Raman spectroscopic monitoring and feedback control of a continuous twin-screw pharmaceutical powder blending and tableting process. Int J Pharm 2017, 530, 21-29. (16) Hausman, D. S.; Cambron, R. T.; Sakr, A. Application of Raman spectroscopy for on-line monitoring of low dose blend uniformity. Int J Pharm 2005, 298, 80-90. (17) Li, Y.; Anderson, C. A.; Drennen, J. K.; Airiau, C.; Igne, B. Method Development and Validation of an In Line Process Analytical Technology Method for Blend Monitoring in the Tablet Feed Frame Using Raman Spectroscopy. Anal Chem 2018, 90, 8436-8444. (18) Shi, Z.; Hermiller, J.; Garcia-Muñoz, S. Estimation of mass-based composition in powder mixtures using Extended Iterative Optimization Technology (EIOT). AIChE J 2019, 65, 87-98. (19) Socrates, G. Infrared and Raman Characteristic Group Frequencies: Tablets and Charts, 3rd ed.; John Wiley & Sons Ltd.: West Sussex, England, 2001.


15


Page 16 of 16

For TOC only


16

Characterization of Near Infrared and Raman Spectroscopy for In-line

Recommend Documents