- Email: [email protected]
Determination and differentiation of surface and bound water in drug substances by near infrared spectroscopy
Determination and Differentiation of Surface and Bound Water in Drug Substances by Near Infrared Spectroscopy GEORGE X. ZHOU, ZHIHONG GE, JASON DORWART, BILL IZZO, JOSEPH KUKURA, GARY BICKER, JEAN WYVRATT Merck and Co., Inc., Merck Research Laboratories, P.O. Box 2000, Rahway, New Jersey 07065
Received 22 October 2002; revised 13 December 2002; accepted 13 December 2002
ABSTRACT: Near infrared spectroscopy (NIRS) was utilized to determine the water content during the drying of a drug substance with Karl Fisher titration as the reference measurement. NIRS calibration models were built with Partial Least Squares (PLS) regression based on spectral region of 1822–1948 nm for samples with 1–40% (w/w) water from five batches. A standard error of prediction (SEP) of 1.85% (w/w) was obtained from validation of the model with additional batches. A second NIRS calibration model using PLS was constructed in the region of 1–10% (w/w) water with samples from the same five calibration batches. This calibration model improved the accuracy of the prediction in the critical region around the end point of drying and provided a standard error of prediction 0.42% (w/w). In addition, direct spectral analyses with Principal Component Analysis (PCA) and peak ratios were applied to distinguish between surface (unbound) water and bound water incorporated into the crystal lattice of the drug substance. Direct spectral analysis indicated the existence of significant numerical and graphical differences between samples containing both surface and bound water, and those containing bound water only. Applying this method to monitor an actual drying process with the graphical assistance of spectral analysis, the drying process can be controlled, and the end point of drying clearly determined to ensure the desired hydrate form of the product. This study has demonstrated the in-line monitoring capability of NIRS to differentiate the surface and bound water simultaneously in addition to the determination of the water level. ß 2003 Wiley-Liss, Inc. and the American Pharmaceutical Association J Pharm Sci 92:1058–1065, 2003
Keywords: near-infrared spectroscopy; surface water; free water; bound water; water determination; on-line monitoring
INTRODUCTION Conventional methods for water determination in pharmaceutical drug products are based on weight loss on drying and Karl Fischer titration.1,2 Other methods such as gas chromatography are alternative choices that provide comparable results.3 However, each method has pros and cons in terms of accuracy, speed, ease of operation,
Correspondence to: George X. Zhou (Telephone: 732-5942277; Fax: 732-594-3887; E-mail: [email protected]) and Bill Izzo (Telephone: 732-594-8431; E-mail: [email protected]) Journal of Pharmaceutical Sciences, Vol. 92, 1058–1065 (2003) ß 2003 Wiley-Liss, Inc. and the American Pharmaceutical Association
1058
and compatibility with certain analytes. For example, Karl Fischer titration is a routine assay in pharmaceutical laboratories, but generally it is labor intensive, time consuming, and often uses toxic reagents. Determining low moisture levels is often difficult requiring control of ambient moisture to achieve reproducible results. In addition, the physical properties of the sample, such as particle size, which affects the dissolution time of solids during Karl Fisher titration, could result in substantial variability in measurement. If large particles are present, it may need a longer time to dissolve entrapped water in the titrator. Above all, all of these techniques are off-line methods that cannot provide real-time water levels during the drying of a drug substance.
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
SURFACE AND BOUND WATER IN DRUG SUBSTANCES
Drying of a drug substance is typically the last step in its production prior to formulation. In some instances, the drying step can become the critical step in the production from either a quality or time cycle perspective. This is certainly true for drug substances exhibiting several hydrate forms. Maintaining an intermediate hydrate form of a drug substance can be challenging. Therefore, inline monitoring of drying processes can assist in both the development and optimization of the drying process as well as the production of quality bulk drug substance. Near infrared spectroscopy (NIRS) has found diverse applications for rapid and nondestructive analysis.4 It is especially suitable for moisture determination because water has strong absorption bands in the near infrared region that provide the sensitivity needed for accurate determinations.1–6 NIRS methods for water determination have been demonstrated to be not only rapid and nondestructive, with no sample pretreatment, but also highly accurate.5,6 Several previous studies have used Karl Fischer (KF) titration as the reference method for NIRS development.2,6 NIRS methods built with GC reference methods have also been reported to provide accurate prediction of water values.5 During a fluidized bed granulation, NIRS was successfully applied for in-process control of a placebo formulation to measure granule moisture content.11 These applications have made NIRS an attractive candidate for the development of an in-line monitoring method for moisture. Water in organic materials can be divided into free (surface) and bound water based on how it exists. In the study of water sorption and desorption of an amorphous raffinose, Hogan and Buchton12 observed and tentatively assigned the penta-, tetra-, tri-, and dihydrate forms in crystalline with a NIR peak around 1440 nm related to OH. The nature of water binding in stratum corneum on skins was studied by Walling and Dabney13 with NIRS and regression equations for free water and bound water were derived. In addition Pyper etc. reported14 a microwave method to measure bound and free water in organic materials. Even though bound and free (surface) water has been studied, none of the work to date has reported the application of NIRS for in-line monitoring of a drying process with concomitant distinction between bound and free water levels. In this study, the drying process of a drug substance M1 has been investigated using NIRS, with KF titration as reference, to in-line monitor the water content. This compound can be isolated
1059
in several hydrated forms with different amounts of bound water, trihydrate 6.8% (w/w), dihydrate 5.6% (w/w), monohydrate 3.7% (w/w), and dehydrate 1.5% (w/w), which were identified by X-ray powder diffractometry (XRPD), NMR, and DSC. The desired hydrate form has been designated as Form I, a trihydrate, and the drying process has been design to preserve the final bulk compound as Form I. During the drying, process control is required to determine not only when the drying step is complete but also when to switch over to the humid nitrogen sweep to preserve the desired form of the compound, and when to stop constant agitation to reduce shear-induced dehydration at low water levels. With KF off-line analysis, one has to interrupt the drying process by releasing the vacuum, pull a sample, and wait for the results of KF analysis. Even though the desired water level has been achieved based on KF analysis, the sample could contain additional surface water and might be a different hydrate because KF analysis cannot differentiate surface water from bound water. Therefore, NIRS has also been evaluated for differentiating between free surface water and bound water, serving as a quality control tool to help achieve the desired hydrate form.
EXPERIMENTAL Apparatus and Materials Diffuse reflectance spectra were obtained with a Foss NIRSystems 6500 spectrophotometer (Foss NIRSystems, Silver Spring, MD) equipped with an in-line probe. Karl Fischer titration was performed with a 701 KF Titrino (L-43406) (Metrohm, Switzerland), and the KF titrant is a single solution Karl Fisher Reagent (stabilized) diluted with KF reagent, both from Fisher Scientific, Pittsburgh, PA. The drying process was monitored in an in-house designed 1-liter filterdryer interfaced with a NIRS probe, and with connections to nitrogen, vacuum, water bath, and mechanical stirrer (Scheme 1). Drug substance M1 under development at Merck and Company was used. Drying Process The drying process involves drying a batch with an approximately 50% (w/w) initial water content, inside an agitated filter dryer, under full vacuum, JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
1060
ZHOU ET AL.
Unscrambler Chemometrics Software (ver. 7.5) from Camo Inc., Norway, was used to perform the Principal Component Analysis for direct spectral analysis in this study. Volumetric Karl-Fischer Titration
Scheme 1. Drying setup for the in-line monitoring.
The volumetric KF titration was performed by charging 30–120 mg of the drug substance solid into the titration vessel. The percent (w/w) moisture level (Cw%) of the sample was calculated by the following equation: Cw% ¼
constant agitation and heat (358C jacket temperature) down to a KF of 25% water. At this point the filter dryer is operated using a humid nitrogen sweep, intermittent agitation, and 358C jacket temperature to dry the wet cake down to the specification of 5–8% (w/w) water and less than 0.5% (w/w) n-propanol and preserve the desired form of the compound. The n-propanol is removed relatively fast, leaving the bulk of the drying time cycle to remove the free surface water.
Spectrum Collection and Data Analysis NIR spectra were collected for each drying sample prior to KF titration. Each spectrum was the average of 32 scans over the range of 1100 to 2500 nm. A reference spectrum of an internal white ceramic was collected automatically for every sample spectrum. The software package Vision (ver. 2.22) accompanying the NIRSystems 6500 analyzer was used to collect the spectra. Calibration models were built with partial least squares (PLS) regression on PLSplus/IQ, an addon program of GRAMS/32 (Galactic Industries Corporation, Salem, NH). During the calibration process, full crossvalidation was applied with a calibration data set, and the standard error of crossvalidation (SECV) was obtained for each calibration model. The number of factors providing the lowest SECV5 was selected as the optimal and the corresponding model was used to predict a separate validation data set. During validation, the standard error of prediction (SEP)5 was calculated by predicting an independent data set. In addition, spectral pretreatment has been reported to improve the quality of the spectra and thus improve the performance of calibration models.8 In this study, the baseline shift was minimized using second derivative pretreatment. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
V T 100 W
where V is the volume of titrant in mL; T is the titer of reagent in mg water per mL; and W is the weight of the sample in mg. The amount of solid powder is adjusted according to its moisture level. In addition, it was found that grinding samples prior to analysis resulted in more representative KF values, especially for samples with large particles at high water levels. For this compound, large particles need longer time to dissolve entrapped water in the titrator and thus introduce additional variation in water measurement due to titration drifting.
RESULTS AND DISCUSSION Spectral Features In the near infrared region, water has large absorption bands with peak maxima around 1420 (overtones) and 1920 nm (combination bands). These absorption bands are very strong, especially the combination bands of OH, and the exact position and width of these bands vary slightly, depending on the chemical and physical environment. Significant changes in spectral features of the sample studied were observed at varying moisture levels. Figure 1 shows the effect of water on the raw reflectance spectra of representative samples, and Figure 2 shows the corresponding second derivative spectra. In Figure 1, the absorbance around 1920 nm increases when water level increases from 1.5 to 33.5% (w/w). These variations in the raw absorbance spectra are retained in the corresponding second derivative spectra (Figure 2). By visual inspection of these representative spectra, it is clear that a quantitative relationship exists between the magnitude of absorption and the moisture level.
SURFACE AND BOUND WATER IN DRUG SUBSTANCES
1061
samples with bound water from those with both bound and surface water. Calibration
Figure 1. NIR spectra of M1 samples with different amount of surface and bound water.
In addition, the NIR peaks at 1904 and 1936 nm vary as water level changes in different samples (Figure 2). Peak 1904 predominately attributes to surface water, while peak 1936 largely corresponds to bound water. In Figure 2, for samples containing only bound water, peak 1936 shrinks as the sample changes from trihydrate to dihydrate, monohydrate, and dehydrate, but there are no changes at 1906 nm. However, the opposite phenomenon is observed for samples at high water level, containing both surface and bound water. As water levels change from 8.7 to 9.3, 16.9, 22.9, and 33.5% (w/w) water the peak at 1904 nm increases correspondingly, while the peak around 1936 nm undergoes almost no changes because similar amount of bound water is present in these samples. These characteristic features clearly separate the
Figure 2. Second derivative spectra of M1 samples with different levels of water with eight data points averaging.
In this work, partial least squares (PLS) regression was selected to construct NIRS calibration models. PLS regression is a factor-based analysis that could enhance S/N ratio. During the process, the variables of the original data sets (spectra and concentrations) are reconstructed with new variables (factors). These factors are believed to better describe the analytical variations in the data sets using few dimensions.7 If experimental conditions are properly chosen, PLS will provide a model that minimizes the effect of minor sources of variance and illustrates the quality of our overall calibration process. The spectral range and the number of PLS factors are the two most crucial parameters in the PLS regression process.7 Based on the absorption features in Figures 1 and 2, several subranges within the 1350–1500 and 1822–1950 nm were investigated. The first region focuses on the 1420 nm absorption bands, while the second one covers the 1920 nm absorption bands. Calibration models were tested by using various combinations of spectral ranges and different number of factors. Initial experiments were carried out on M1 samples with varying levels of water from eight batches to demonstrate the feasibility of NIRS for this application. In total, 94 M1 samples were divided into two data sets, calibration and prediction. There were 62 samples from five batches in the calibration set and 32 samples from the remaining three prediction batches. The former was used to build the calibration models with KF as the reference method while the latter was used to evaluate the performance of these calibration models. For the early stage of drying, an optimum NIRS calibration model for a wide water range of 1–40% (w/w) was built with PLS regression. This model was constructed on the second derivative spectra with four PLS factors in the spectral range of 1822–1948 nm. The corresponding SECV was 1.81% (w/w) water. Regression analysis of the concentration correlation (water by KF versus by NIRS) of this model resulted in a slope of 0.981 and a y-intercept of 0.319% (w/w). Using this calibration model to predict the other three batches, a SEP of 1.85% (w/w) was attained. The large error for this model was attributed primarily to the KF JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
1062
ZHOU ET AL.
analysis, as these samples contained relatively large particles at high water content. To accurately determine the end point of the drying process [6–8% (w/w)], a second NIRS calibration model based on the second derivative spectra with four PLS factors in the range of 1822–1948 nm for samples with 1–10% (w/w) water was developed. This model provided a SECV of 0.43% (w/w). Regression analysis of the concentration correlation of this model resulted in a slope of 0.991 and a y-intercept of 0.062% (w/w). Applying this model to predict the validation batches, the corresponding SEP was 0.42% (w/w) (see Table 1), which was within the desired level of accuracy for control of the drying process. The performance of this model far exceeded that of the model covering the entire (wide) concentration range. This improvement was partially attributed to sample characteristics, namely reduced particle size at low water level, which ultimately improved the precision of the reference KF analysis. Direct Spectral Discriminant Analysis In addition to the quantitative determination of the water level, graphically monitoring the drying process could provide rapid interpretation of the water content without the intensive processing of calibration models. This is of particular importance when targeting an intermediate hydrate form that rapidly transitions to an undesired
Table 1. Prediction Results of External Batches by a Calibration Model of PLS in Concentration Range of 1–10% (w/w)
Sample 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
KF % (w/w)
NIRS % (w/w)
Difference % (w/w)
6.37 6.77 7.00 8.50 6.63 6.82 6.91 6.65 9.63 6.47 7.27 6.91 8.87 9.08 8.86 7.17
6.82 7.08 7.11 8.76 6.95 6.92 7.22 7.20 9.64 6.26 7.16 7.29 9.05 9.06 8.97 6.69
0.45 0.31 0.11 0.26 0.32 0.10 0.31 0.55 0.01 0.21 0.11 0.38 0.18 0.03 0.11 0.48
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
lower hydrate. Such ‘‘overdrying’’ may require extensive rework. To this end, Principal Component Analysis (PCA) was used to examine the spectra of all samples. In PCA, the spectral data set is decomposed into its most common variations (factors or loadings), generating a small subset of well defined numbers (scores) for each sample representing the amount of each variation present in the spectrum.9 The scores can be used for discrimination because they provide an accurate description of the entire set of spectra. In this study, the spectral range of 1890–1950 nm and second derivative pretreatment with 16 data point averaging were used in PCA analysis. Figure 3 shows the first three loadings (factors) of these samples. Compared to Figure 2, factor one (PC1) describes mainly the dried drug substance as observed for dehydrate [1.5% (w/w)], and PC2 mainly characterizes the surface water with a maximum around 1904 nm. PC3 is largely related to bound water with a maximum around 1940 nm. When the scores of PC2 are plotted against PC3 in Figure 4, the samples fall into two distinct groups, one with bound water and the other with both surface and bound water. This figure clearly illustrates that samples with water level less than 8% (w/w) are statistically and graphically different from those over 9% (w/w). At a high water level [above 9% (w/w)], samples contain abundant free surface water whose absorption overwhelms that of bound water, resulting in a low PC2 (high surface water level) and low PC3 in the left corner of Figure 4a. As water level decreases this impact of surface water absorption on that of bound water decreases, and the sample
Figure 3. Loadings of PCA analysis on the spectra of M1 samples in 1890–1950 nm with second derivative pretreatment.
SURFACE AND BOUND WATER IN DRUG SUBSTANCES
1063
water. In this approach, the spectral range of 1890–1950 nm and second derivative pretreatment with 6 to 20 data point averaging were evaluated. The number of consecutive data points used for second derivative calculation affected the performance of this approach with 20 data point averaging yielding significant overlapping of the bound and unbound water bands. Eight-point averaging provided the best separation of NIR bands of surface and bound water (Figure 2); however, 16 data point averaging (Figure 5a) was selected for optimum performance yielding adequate graphical difference between surface and bound water with robust performance. In the region of 1890–1950 nm, the difference can be observed between samples with both bound and surface water, and those with no surface water (Figure 4a). The peak at 1904 nm largely
Figure 4. Score plot of PCA for all M1 samples from eight batches with different water levels in 1890– 1950 nm with second derivative, (a). PC2 versus PC3 and (b) PC3 versus water level.
reaches a high PC3 around 8–9% (w/w) water with minimal amount surface water but substantial amount of bound water. As the water level decreases further, bound water begins to decrease sharply within a narrow range of water content. At a water level below 6.8% (w/w) in Figure 4b where PC3 was plotted against the water level in the samples, samples progressed from trihydrate, to dihydrate, to monohydrate, and eventually became a dehydrated form of the drug substance M1 as drying continued. This corresponds to the changes of peak 1936 nm in Figure 2. Between water levels of 9 and 6.8% (w/w), samples transit from the surface water region to bound water region with trihydrate at the bottom of Figure 4. This relationship could be used to determine if trihydrate has been reached during the drying process. An alternate semiquantitative procedure involving simple peak ratios was investigated to differentiate samples with and without surface
Figure 5. (a) Second derivative spectra of M1 samples with different levels of water with 16 data points averaging; (b) peak ratio versus samples with different amount of water based on all M1 samples from eight batches. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
1064
ZHOU ET AL.
corresponds to surface water while the peak at 1936 nm corresponds primarily to bound water. Because the second derivative pretreatment was based on 16 data point averaging and the spectral resolution was 2 nm, this procedure superimposed some surface water information on the peak at 1932 nm. Therefore, the ratio of these two peaks (1932/1936 nm) actually contained information of both surface and bound water. The peak ratios of 1932 versus 1936 nm of the second derivative spectra for samples of varying water content are shown in Figure 5b. It is visually apparent that 8% (w/w) is a ‘‘turning point.’’ Samples with water level over 8% (w/w), likely belong to the same group with surface water, while those with water less than 8% (w/w) fall into another group. Because the trihydrate form contains 6.8% (w/w) water, this analysis suggests the existence of high hydrates that contain bound water between 7 to 8% (w/w). Separate NMR studies10 likewise suggest the existence of high hydrates above 7% (w/w). When the water level is near or above 8% (w/w), samples most likely contain free surface water. The peak ratios reach the lowest range for samples with 4.5–6.5% (w/w) water that corresponds to the di- and trihydrate forms of the drug substance. Compared to PCA analysis, this peak ratio of 1932/1936 nm is more empirical. However, it also provides a simple, visual insight into the drying process, which could supplement the quantitative prediction by NIRS calibration models. In-Line Process Monitoring In-line monitoring experiments were carried out to further demonstrate these findings. To predict water level quantitatively, calibration models were constructed on three runs with KF analysis as reference. The PLS calibration employed four factors in the 1890–1950 nm range of the second derivative spectra. The prediction of a separate dynamic drying run using the optimized calibration model is presented in Figure 6. This particular run started with wet cake of 45% (w/w) water under vacuum. Nineteen hours into the drying a humid nitrogen sweep was applied. After 43 h of drying, the NIRS model predicted the formation of trihydrate, which was subsequently confirmed by off-line KF analysis. Figure 7 depicts the PCA results on the second derivative spectra for this drying run. As the drying proceeds, the product moves from the region of surface and bound water to the region of JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
Figure 6. On-line water prediction profile during the drying process of M1 wet cake by NIRS, line—NIRS; circles—KF.
bound water, and/or trihydrate region as indicated by the ‘‘diamond,’’ which corresponds to 6.5% (w/w) water by the off-line KF analysis. Clearly, this PCA approach graphically illustrates a drying process and could be used to control the drying process and to achieve the desired hydrate form.
CONCLUSION For the drying of drug substance with different hydrate forms, reliable NIRS calibration models can be constructed by PLS regression in the spectral region of combination bands of water. NIRS calibration models provided SEPs 1.81% (w/w) in the range of 1.5–40% (w/w) water and 0.42% (w/w) in the range of 1.5–10% (w/w) water where the end point of drying is located. Similarly valuable was the direct spectral analysis providing
Figure 7. Principle component analysis plot of PC2 versus PC3 during the drying process of M1 wet cake by NIRS in-line monitoring.
SURFACE AND BOUND WATER IN DRUG SUBSTANCES
graphical process information that augmented the tight control of the drying process. The separate and characteristic bands of surface and bound water around 1920 nm were used to distinguish samples with bound water from those with both bound and surface water. The results of this study demonstrated the capability of in-line monitoring of drying process by NIRS to achieve the desired hydrate form of a drug substance.
5.
6.
7.
ACKNOWLEDGMENTS The authors would like to thank Chad Rush, Jeanne Tomaszewski, and John Hiney for participating in the experiments, Rich Varsolona for X-ray analysis and Robert Wenslow for ssNMR analysis.
8.
9. 10.
REFERENCES 1. MacDonald BF, Prebble KA. 1993. Some applications of near-infrared reflectance analysis in the pharmaceutical industry. J Pharm Biomed Anal 11:1077–1085. 2. Lonardi S, Viviani R, Mosconi L, Bernuzzi M, Corti P, Dreassi E, Murratzu C, Corbini G. 1989. Drug analysis by near-infrared reflectance spectroscopy. Determination of active ingredient and water content in antibiotic powders. J Pharm Biomed Anal 7(3):303–308. 3. Windham WR. 1989. Comparison between liquid and paste Karl Fischer extractions for NIRS [nearIR reflectance spectrometry] water calibrations. J Assoc Off Anal Chem 72(2):255–258. 4. Kirsch JD, Drennen JK. 1995. Near-infrared spectroscopy: Application in the analysis of tablets and
11.
12.
13.
14.
1065
solid pharmaceutical dosage forms. Appl Spec Rev 30:139–174. Zhou X, Hines PA, White KC, Borer MW. 1998. Gas chromatography as a reference method for moisture determination by near infrared spectroscopy. Anal Chem 70:390–394. Zhou X, Hines PA, Borer MW. 1998. Moisture determination in hygroscopic drug substances by near infrared spectroscopy. J Pharm Biomed Anal 17:219–225. Beebe KH, Kowalski BR. 1987. An introduction to multivariate calibration and analysis. Anal Chem 59:1007A–1017A. Isaksson T, Naes T. 1988. The effect of multiplicative scatter correction (MSC) and linearity improvement in NIR spectroscopy. Appl Spec 42: 1273–1284. PLSplus/IQ user’s guide. 1991. Chapters 4 and 5. Salem, NH: Galactic Industries Corporation. Wenslow R. 2000. Merck & Co. internal communication. Rantanen J, Antikainen O, Mannermaa J, Yliruusi J. 2000. Use of the near-infrared reflectance method for measurement of moisture content during granulation. Pharm Dev Technol 5(2):209– 217. Hogan SE, Buckton G. 2001. Water sorption/ desorption-near IR and calorimetric study of crystalline and amorphous raffinose. Int J Pharm 227(1–2):57–69. Pyper JW, Buettner HM, Cerjan CJ, Hallam JS, King RJ. 1985. The measurement of bound and free moisture in organic materials by microwave methods. Moisture Humidity Proc Int Symp. pp 909–917. Published by ISA, Research Triangle Park, N. C. Walling PL, Dabney JM. 1989. Moisture in skin by near-infrared reflectance spectroscopy. J Soc Cosmetic Chem (England) 40:151–171.
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
Received 22 October 2002; revised 13 December 2002; accepted 13 December 2002
ABSTRACT: Near infrared spectroscopy (NIRS) was utilized to determine the water content during the drying of a drug substance with Karl Fisher titration as the reference measurement. NIRS calibration models were built with Partial Least Squares (PLS) regression based on spectral region of 1822–1948 nm for samples with 1–40% (w/w) water from five batches. A standard error of prediction (SEP) of 1.85% (w/w) was obtained from validation of the model with additional batches. A second NIRS calibration model using PLS was constructed in the region of 1–10% (w/w) water with samples from the same five calibration batches. This calibration model improved the accuracy of the prediction in the critical region around the end point of drying and provided a standard error of prediction 0.42% (w/w). In addition, direct spectral analyses with Principal Component Analysis (PCA) and peak ratios were applied to distinguish between surface (unbound) water and bound water incorporated into the crystal lattice of the drug substance. Direct spectral analysis indicated the existence of significant numerical and graphical differences between samples containing both surface and bound water, and those containing bound water only. Applying this method to monitor an actual drying process with the graphical assistance of spectral analysis, the drying process can be controlled, and the end point of drying clearly determined to ensure the desired hydrate form of the product. This study has demonstrated the in-line monitoring capability of NIRS to differentiate the surface and bound water simultaneously in addition to the determination of the water level. ß 2003 Wiley-Liss, Inc. and the American Pharmaceutical Association J Pharm Sci 92:1058–1065, 2003
Keywords: near-infrared spectroscopy; surface water; free water; bound water; water determination; on-line monitoring
INTRODUCTION Conventional methods for water determination in pharmaceutical drug products are based on weight loss on drying and Karl Fischer titration.1,2 Other methods such as gas chromatography are alternative choices that provide comparable results.3 However, each method has pros and cons in terms of accuracy, speed, ease of operation,
Correspondence to: George X. Zhou (Telephone: 732-5942277; Fax: 732-594-3887; E-mail: [email protected]) and Bill Izzo (Telephone: 732-594-8431; E-mail: [email protected]) Journal of Pharmaceutical Sciences, Vol. 92, 1058–1065 (2003) ß 2003 Wiley-Liss, Inc. and the American Pharmaceutical Association
1058
and compatibility with certain analytes. For example, Karl Fischer titration is a routine assay in pharmaceutical laboratories, but generally it is labor intensive, time consuming, and often uses toxic reagents. Determining low moisture levels is often difficult requiring control of ambient moisture to achieve reproducible results. In addition, the physical properties of the sample, such as particle size, which affects the dissolution time of solids during Karl Fisher titration, could result in substantial variability in measurement. If large particles are present, it may need a longer time to dissolve entrapped water in the titrator. Above all, all of these techniques are off-line methods that cannot provide real-time water levels during the drying of a drug substance.
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
SURFACE AND BOUND WATER IN DRUG SUBSTANCES
Drying of a drug substance is typically the last step in its production prior to formulation. In some instances, the drying step can become the critical step in the production from either a quality or time cycle perspective. This is certainly true for drug substances exhibiting several hydrate forms. Maintaining an intermediate hydrate form of a drug substance can be challenging. Therefore, inline monitoring of drying processes can assist in both the development and optimization of the drying process as well as the production of quality bulk drug substance. Near infrared spectroscopy (NIRS) has found diverse applications for rapid and nondestructive analysis.4 It is especially suitable for moisture determination because water has strong absorption bands in the near infrared region that provide the sensitivity needed for accurate determinations.1–6 NIRS methods for water determination have been demonstrated to be not only rapid and nondestructive, with no sample pretreatment, but also highly accurate.5,6 Several previous studies have used Karl Fischer (KF) titration as the reference method for NIRS development.2,6 NIRS methods built with GC reference methods have also been reported to provide accurate prediction of water values.5 During a fluidized bed granulation, NIRS was successfully applied for in-process control of a placebo formulation to measure granule moisture content.11 These applications have made NIRS an attractive candidate for the development of an in-line monitoring method for moisture. Water in organic materials can be divided into free (surface) and bound water based on how it exists. In the study of water sorption and desorption of an amorphous raffinose, Hogan and Buchton12 observed and tentatively assigned the penta-, tetra-, tri-, and dihydrate forms in crystalline with a NIR peak around 1440 nm related to OH. The nature of water binding in stratum corneum on skins was studied by Walling and Dabney13 with NIRS and regression equations for free water and bound water were derived. In addition Pyper etc. reported14 a microwave method to measure bound and free water in organic materials. Even though bound and free (surface) water has been studied, none of the work to date has reported the application of NIRS for in-line monitoring of a drying process with concomitant distinction between bound and free water levels. In this study, the drying process of a drug substance M1 has been investigated using NIRS, with KF titration as reference, to in-line monitor the water content. This compound can be isolated
1059
in several hydrated forms with different amounts of bound water, trihydrate 6.8% (w/w), dihydrate 5.6% (w/w), monohydrate 3.7% (w/w), and dehydrate 1.5% (w/w), which were identified by X-ray powder diffractometry (XRPD), NMR, and DSC. The desired hydrate form has been designated as Form I, a trihydrate, and the drying process has been design to preserve the final bulk compound as Form I. During the drying, process control is required to determine not only when the drying step is complete but also when to switch over to the humid nitrogen sweep to preserve the desired form of the compound, and when to stop constant agitation to reduce shear-induced dehydration at low water levels. With KF off-line analysis, one has to interrupt the drying process by releasing the vacuum, pull a sample, and wait for the results of KF analysis. Even though the desired water level has been achieved based on KF analysis, the sample could contain additional surface water and might be a different hydrate because KF analysis cannot differentiate surface water from bound water. Therefore, NIRS has also been evaluated for differentiating between free surface water and bound water, serving as a quality control tool to help achieve the desired hydrate form.
EXPERIMENTAL Apparatus and Materials Diffuse reflectance spectra were obtained with a Foss NIRSystems 6500 spectrophotometer (Foss NIRSystems, Silver Spring, MD) equipped with an in-line probe. Karl Fischer titration was performed with a 701 KF Titrino (L-43406) (Metrohm, Switzerland), and the KF titrant is a single solution Karl Fisher Reagent (stabilized) diluted with KF reagent, both from Fisher Scientific, Pittsburgh, PA. The drying process was monitored in an in-house designed 1-liter filterdryer interfaced with a NIRS probe, and with connections to nitrogen, vacuum, water bath, and mechanical stirrer (Scheme 1). Drug substance M1 under development at Merck and Company was used. Drying Process The drying process involves drying a batch with an approximately 50% (w/w) initial water content, inside an agitated filter dryer, under full vacuum, JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
1060
ZHOU ET AL.
Unscrambler Chemometrics Software (ver. 7.5) from Camo Inc., Norway, was used to perform the Principal Component Analysis for direct spectral analysis in this study. Volumetric Karl-Fischer Titration
Scheme 1. Drying setup for the in-line monitoring.
The volumetric KF titration was performed by charging 30–120 mg of the drug substance solid into the titration vessel. The percent (w/w) moisture level (Cw%) of the sample was calculated by the following equation: Cw% ¼
constant agitation and heat (358C jacket temperature) down to a KF of 25% water. At this point the filter dryer is operated using a humid nitrogen sweep, intermittent agitation, and 358C jacket temperature to dry the wet cake down to the specification of 5–8% (w/w) water and less than 0.5% (w/w) n-propanol and preserve the desired form of the compound. The n-propanol is removed relatively fast, leaving the bulk of the drying time cycle to remove the free surface water.
Spectrum Collection and Data Analysis NIR spectra were collected for each drying sample prior to KF titration. Each spectrum was the average of 32 scans over the range of 1100 to 2500 nm. A reference spectrum of an internal white ceramic was collected automatically for every sample spectrum. The software package Vision (ver. 2.22) accompanying the NIRSystems 6500 analyzer was used to collect the spectra. Calibration models were built with partial least squares (PLS) regression on PLSplus/IQ, an addon program of GRAMS/32 (Galactic Industries Corporation, Salem, NH). During the calibration process, full crossvalidation was applied with a calibration data set, and the standard error of crossvalidation (SECV) was obtained for each calibration model. The number of factors providing the lowest SECV5 was selected as the optimal and the corresponding model was used to predict a separate validation data set. During validation, the standard error of prediction (SEP)5 was calculated by predicting an independent data set. In addition, spectral pretreatment has been reported to improve the quality of the spectra and thus improve the performance of calibration models.8 In this study, the baseline shift was minimized using second derivative pretreatment. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
V T 100 W
where V is the volume of titrant in mL; T is the titer of reagent in mg water per mL; and W is the weight of the sample in mg. The amount of solid powder is adjusted according to its moisture level. In addition, it was found that grinding samples prior to analysis resulted in more representative KF values, especially for samples with large particles at high water levels. For this compound, large particles need longer time to dissolve entrapped water in the titrator and thus introduce additional variation in water measurement due to titration drifting.
RESULTS AND DISCUSSION Spectral Features In the near infrared region, water has large absorption bands with peak maxima around 1420 (overtones) and 1920 nm (combination bands). These absorption bands are very strong, especially the combination bands of OH, and the exact position and width of these bands vary slightly, depending on the chemical and physical environment. Significant changes in spectral features of the sample studied were observed at varying moisture levels. Figure 1 shows the effect of water on the raw reflectance spectra of representative samples, and Figure 2 shows the corresponding second derivative spectra. In Figure 1, the absorbance around 1920 nm increases when water level increases from 1.5 to 33.5% (w/w). These variations in the raw absorbance spectra are retained in the corresponding second derivative spectra (Figure 2). By visual inspection of these representative spectra, it is clear that a quantitative relationship exists between the magnitude of absorption and the moisture level.
SURFACE AND BOUND WATER IN DRUG SUBSTANCES
1061
samples with bound water from those with both bound and surface water. Calibration
Figure 1. NIR spectra of M1 samples with different amount of surface and bound water.
In addition, the NIR peaks at 1904 and 1936 nm vary as water level changes in different samples (Figure 2). Peak 1904 predominately attributes to surface water, while peak 1936 largely corresponds to bound water. In Figure 2, for samples containing only bound water, peak 1936 shrinks as the sample changes from trihydrate to dihydrate, monohydrate, and dehydrate, but there are no changes at 1906 nm. However, the opposite phenomenon is observed for samples at high water level, containing both surface and bound water. As water levels change from 8.7 to 9.3, 16.9, 22.9, and 33.5% (w/w) water the peak at 1904 nm increases correspondingly, while the peak around 1936 nm undergoes almost no changes because similar amount of bound water is present in these samples. These characteristic features clearly separate the
Figure 2. Second derivative spectra of M1 samples with different levels of water with eight data points averaging.
In this work, partial least squares (PLS) regression was selected to construct NIRS calibration models. PLS regression is a factor-based analysis that could enhance S/N ratio. During the process, the variables of the original data sets (spectra and concentrations) are reconstructed with new variables (factors). These factors are believed to better describe the analytical variations in the data sets using few dimensions.7 If experimental conditions are properly chosen, PLS will provide a model that minimizes the effect of minor sources of variance and illustrates the quality of our overall calibration process. The spectral range and the number of PLS factors are the two most crucial parameters in the PLS regression process.7 Based on the absorption features in Figures 1 and 2, several subranges within the 1350–1500 and 1822–1950 nm were investigated. The first region focuses on the 1420 nm absorption bands, while the second one covers the 1920 nm absorption bands. Calibration models were tested by using various combinations of spectral ranges and different number of factors. Initial experiments were carried out on M1 samples with varying levels of water from eight batches to demonstrate the feasibility of NIRS for this application. In total, 94 M1 samples were divided into two data sets, calibration and prediction. There were 62 samples from five batches in the calibration set and 32 samples from the remaining three prediction batches. The former was used to build the calibration models with KF as the reference method while the latter was used to evaluate the performance of these calibration models. For the early stage of drying, an optimum NIRS calibration model for a wide water range of 1–40% (w/w) was built with PLS regression. This model was constructed on the second derivative spectra with four PLS factors in the spectral range of 1822–1948 nm. The corresponding SECV was 1.81% (w/w) water. Regression analysis of the concentration correlation (water by KF versus by NIRS) of this model resulted in a slope of 0.981 and a y-intercept of 0.319% (w/w). Using this calibration model to predict the other three batches, a SEP of 1.85% (w/w) was attained. The large error for this model was attributed primarily to the KF JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
1062
ZHOU ET AL.
analysis, as these samples contained relatively large particles at high water content. To accurately determine the end point of the drying process [6–8% (w/w)], a second NIRS calibration model based on the second derivative spectra with four PLS factors in the range of 1822–1948 nm for samples with 1–10% (w/w) water was developed. This model provided a SECV of 0.43% (w/w). Regression analysis of the concentration correlation of this model resulted in a slope of 0.991 and a y-intercept of 0.062% (w/w). Applying this model to predict the validation batches, the corresponding SEP was 0.42% (w/w) (see Table 1), which was within the desired level of accuracy for control of the drying process. The performance of this model far exceeded that of the model covering the entire (wide) concentration range. This improvement was partially attributed to sample characteristics, namely reduced particle size at low water level, which ultimately improved the precision of the reference KF analysis. Direct Spectral Discriminant Analysis In addition to the quantitative determination of the water level, graphically monitoring the drying process could provide rapid interpretation of the water content without the intensive processing of calibration models. This is of particular importance when targeting an intermediate hydrate form that rapidly transitions to an undesired
Table 1. Prediction Results of External Batches by a Calibration Model of PLS in Concentration Range of 1–10% (w/w)
Sample 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
KF % (w/w)
NIRS % (w/w)
Difference % (w/w)
6.37 6.77 7.00 8.50 6.63 6.82 6.91 6.65 9.63 6.47 7.27 6.91 8.87 9.08 8.86 7.17
6.82 7.08 7.11 8.76 6.95 6.92 7.22 7.20 9.64 6.26 7.16 7.29 9.05 9.06 8.97 6.69
0.45 0.31 0.11 0.26 0.32 0.10 0.31 0.55 0.01 0.21 0.11 0.38 0.18 0.03 0.11 0.48
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
lower hydrate. Such ‘‘overdrying’’ may require extensive rework. To this end, Principal Component Analysis (PCA) was used to examine the spectra of all samples. In PCA, the spectral data set is decomposed into its most common variations (factors or loadings), generating a small subset of well defined numbers (scores) for each sample representing the amount of each variation present in the spectrum.9 The scores can be used for discrimination because they provide an accurate description of the entire set of spectra. In this study, the spectral range of 1890–1950 nm and second derivative pretreatment with 16 data point averaging were used in PCA analysis. Figure 3 shows the first three loadings (factors) of these samples. Compared to Figure 2, factor one (PC1) describes mainly the dried drug substance as observed for dehydrate [1.5% (w/w)], and PC2 mainly characterizes the surface water with a maximum around 1904 nm. PC3 is largely related to bound water with a maximum around 1940 nm. When the scores of PC2 are plotted against PC3 in Figure 4, the samples fall into two distinct groups, one with bound water and the other with both surface and bound water. This figure clearly illustrates that samples with water level less than 8% (w/w) are statistically and graphically different from those over 9% (w/w). At a high water level [above 9% (w/w)], samples contain abundant free surface water whose absorption overwhelms that of bound water, resulting in a low PC2 (high surface water level) and low PC3 in the left corner of Figure 4a. As water level decreases this impact of surface water absorption on that of bound water decreases, and the sample
Figure 3. Loadings of PCA analysis on the spectra of M1 samples in 1890–1950 nm with second derivative pretreatment.
SURFACE AND BOUND WATER IN DRUG SUBSTANCES
1063
water. In this approach, the spectral range of 1890–1950 nm and second derivative pretreatment with 6 to 20 data point averaging were evaluated. The number of consecutive data points used for second derivative calculation affected the performance of this approach with 20 data point averaging yielding significant overlapping of the bound and unbound water bands. Eight-point averaging provided the best separation of NIR bands of surface and bound water (Figure 2); however, 16 data point averaging (Figure 5a) was selected for optimum performance yielding adequate graphical difference between surface and bound water with robust performance. In the region of 1890–1950 nm, the difference can be observed between samples with both bound and surface water, and those with no surface water (Figure 4a). The peak at 1904 nm largely
Figure 4. Score plot of PCA for all M1 samples from eight batches with different water levels in 1890– 1950 nm with second derivative, (a). PC2 versus PC3 and (b) PC3 versus water level.
reaches a high PC3 around 8–9% (w/w) water with minimal amount surface water but substantial amount of bound water. As the water level decreases further, bound water begins to decrease sharply within a narrow range of water content. At a water level below 6.8% (w/w) in Figure 4b where PC3 was plotted against the water level in the samples, samples progressed from trihydrate, to dihydrate, to monohydrate, and eventually became a dehydrated form of the drug substance M1 as drying continued. This corresponds to the changes of peak 1936 nm in Figure 2. Between water levels of 9 and 6.8% (w/w), samples transit from the surface water region to bound water region with trihydrate at the bottom of Figure 4. This relationship could be used to determine if trihydrate has been reached during the drying process. An alternate semiquantitative procedure involving simple peak ratios was investigated to differentiate samples with and without surface
Figure 5. (a) Second derivative spectra of M1 samples with different levels of water with 16 data points averaging; (b) peak ratio versus samples with different amount of water based on all M1 samples from eight batches. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
1064
ZHOU ET AL.
corresponds to surface water while the peak at 1936 nm corresponds primarily to bound water. Because the second derivative pretreatment was based on 16 data point averaging and the spectral resolution was 2 nm, this procedure superimposed some surface water information on the peak at 1932 nm. Therefore, the ratio of these two peaks (1932/1936 nm) actually contained information of both surface and bound water. The peak ratios of 1932 versus 1936 nm of the second derivative spectra for samples of varying water content are shown in Figure 5b. It is visually apparent that 8% (w/w) is a ‘‘turning point.’’ Samples with water level over 8% (w/w), likely belong to the same group with surface water, while those with water less than 8% (w/w) fall into another group. Because the trihydrate form contains 6.8% (w/w) water, this analysis suggests the existence of high hydrates that contain bound water between 7 to 8% (w/w). Separate NMR studies10 likewise suggest the existence of high hydrates above 7% (w/w). When the water level is near or above 8% (w/w), samples most likely contain free surface water. The peak ratios reach the lowest range for samples with 4.5–6.5% (w/w) water that corresponds to the di- and trihydrate forms of the drug substance. Compared to PCA analysis, this peak ratio of 1932/1936 nm is more empirical. However, it also provides a simple, visual insight into the drying process, which could supplement the quantitative prediction by NIRS calibration models. In-Line Process Monitoring In-line monitoring experiments were carried out to further demonstrate these findings. To predict water level quantitatively, calibration models were constructed on three runs with KF analysis as reference. The PLS calibration employed four factors in the 1890–1950 nm range of the second derivative spectra. The prediction of a separate dynamic drying run using the optimized calibration model is presented in Figure 6. This particular run started with wet cake of 45% (w/w) water under vacuum. Nineteen hours into the drying a humid nitrogen sweep was applied. After 43 h of drying, the NIRS model predicted the formation of trihydrate, which was subsequently confirmed by off-line KF analysis. Figure 7 depicts the PCA results on the second derivative spectra for this drying run. As the drying proceeds, the product moves from the region of surface and bound water to the region of JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003
Figure 6. On-line water prediction profile during the drying process of M1 wet cake by NIRS, line—NIRS; circles—KF.
bound water, and/or trihydrate region as indicated by the ‘‘diamond,’’ which corresponds to 6.5% (w/w) water by the off-line KF analysis. Clearly, this PCA approach graphically illustrates a drying process and could be used to control the drying process and to achieve the desired hydrate form.
CONCLUSION For the drying of drug substance with different hydrate forms, reliable NIRS calibration models can be constructed by PLS regression in the spectral region of combination bands of water. NIRS calibration models provided SEPs 1.81% (w/w) in the range of 1.5–40% (w/w) water and 0.42% (w/w) in the range of 1.5–10% (w/w) water where the end point of drying is located. Similarly valuable was the direct spectral analysis providing
Figure 7. Principle component analysis plot of PC2 versus PC3 during the drying process of M1 wet cake by NIRS in-line monitoring.
SURFACE AND BOUND WATER IN DRUG SUBSTANCES
graphical process information that augmented the tight control of the drying process. The separate and characteristic bands of surface and bound water around 1920 nm were used to distinguish samples with bound water from those with both bound and surface water. The results of this study demonstrated the capability of in-line monitoring of drying process by NIRS to achieve the desired hydrate form of a drug substance.
5.
6.
7.
ACKNOWLEDGMENTS The authors would like to thank Chad Rush, Jeanne Tomaszewski, and John Hiney for participating in the experiments, Rich Varsolona for X-ray analysis and Robert Wenslow for ssNMR analysis.
8.
9. 10.
REFERENCES 1. MacDonald BF, Prebble KA. 1993. Some applications of near-infrared reflectance analysis in the pharmaceutical industry. J Pharm Biomed Anal 11:1077–1085. 2. Lonardi S, Viviani R, Mosconi L, Bernuzzi M, Corti P, Dreassi E, Murratzu C, Corbini G. 1989. Drug analysis by near-infrared reflectance spectroscopy. Determination of active ingredient and water content in antibiotic powders. J Pharm Biomed Anal 7(3):303–308. 3. Windham WR. 1989. Comparison between liquid and paste Karl Fischer extractions for NIRS [nearIR reflectance spectrometry] water calibrations. J Assoc Off Anal Chem 72(2):255–258. 4. Kirsch JD, Drennen JK. 1995. Near-infrared spectroscopy: Application in the analysis of tablets and
11.
12.
13.
14.
1065
solid pharmaceutical dosage forms. Appl Spec Rev 30:139–174. Zhou X, Hines PA, White KC, Borer MW. 1998. Gas chromatography as a reference method for moisture determination by near infrared spectroscopy. Anal Chem 70:390–394. Zhou X, Hines PA, Borer MW. 1998. Moisture determination in hygroscopic drug substances by near infrared spectroscopy. J Pharm Biomed Anal 17:219–225. Beebe KH, Kowalski BR. 1987. An introduction to multivariate calibration and analysis. Anal Chem 59:1007A–1017A. Isaksson T, Naes T. 1988. The effect of multiplicative scatter correction (MSC) and linearity improvement in NIR spectroscopy. Appl Spec 42: 1273–1284. PLSplus/IQ user’s guide. 1991. Chapters 4 and 5. Salem, NH: Galactic Industries Corporation. Wenslow R. 2000. Merck & Co. internal communication. Rantanen J, Antikainen O, Mannermaa J, Yliruusi J. 2000. Use of the near-infrared reflectance method for measurement of moisture content during granulation. Pharm Dev Technol 5(2):209– 217. Hogan SE, Buckton G. 2001. Water sorption/ desorption-near IR and calorimetric study of crystalline and amorphous raffinose. Int J Pharm 227(1–2):57–69. Pyper JW, Buettner HM, Cerjan CJ, Hallam JS, King RJ. 1985. The measurement of bound and free moisture in organic materials by microwave methods. Moisture Humidity Proc Int Symp. pp 909–917. Published by ISA, Research Triangle Park, N. C. Walling PL, Dabney JM. 1989. Moisture in skin by near-infrared reflectance spectroscopy. J Soc Cosmetic Chem (England) 40:151–171.
JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 92, NO. 5, MAY 2003