Assessment of Higher Order Structure Comparability in Therapeutic Proteins Using Nuclear Magnetic Resonance Spectroscopy

Assessment of Higher Order Structure Comparability in Therapeutic Proteins Using Nuclear Magnetic Resonance Spectroscopy CARLOS A. AMEZCUA, CHRISTINA M. SZABO Baxter Healthcare Corporation, Round Lake, Illinois 60073 Received 30 November 2012; revised 4 March 2013; accepted 15 March 2013 Published online 9 April 2013 in Wiley Online Library (wileyonlinelibrary.com). DOI 10.1002/jps.23531 ABSTRACT: In this work, we applied nuclear magnetic resonance (NMR) spectroscopy to rapidly assess higher order structure (HOS) comparability in protein samples. Using a variation of the NMR fingerprinting approach described by Panjwani et al. [2010. J Pharm Sci 99(8):3334–3342], three nonglycosylated proteins spanning a molecular weight range of 6.5–67 kDa were analyzed. A simple statistical method termed easy comparability of HOS by NMR (ECHOS-NMR) was developed. In this method, HOS similarity between two samples is measured via the correlation coefficient derived from linear regression analysis of binned NMR spectra. Applications of this method include HOS comparability assessment during new product development, manufacturing process changes, supplier changes, next-generation products, and the development of biosimilars to name just a few. We foresee ECHOS-NMR becoming a routine technique applied to comparability exercises used to complement data from other analytical techniques. © 2013 Wiley Periodicals, Inc. and the American Pharmacists Association J Pharm Sci 102:1724–1733, 2013 Keywords: Higher order structure comparability; biopharmaceuticals characterization; NMR; spectroscopy; ECHOS-NMR; NMR fingerprinting; proteins; protein structure; diffusion

INTRODUCTION An assessment of comparability in biological products after introduction of manufacturing changes1,2 or for licensing purposes of biosimilars3 is essential for the demonstration of quality, safety, and efficacy. Physicochemical characterization is a key component of the comparability exercise and, in some cases, this alone can help avoid costly preclinical or clinical studies. Analytical techniques such as mass spectrometry, capillary electrophoresis, size-exclusion chromatography (SEC), circular dichroism (CD), Fourier transformed infrared spectroscopy (FT-IR), fluorescence, ultraviolet (UV) spectroscopy, and peptide mapping are commonly used.4,5 For protein products, these techniques provide information about their amino acid sequence and modifications, mass, size, charge, aggregation, glycosylation, and secondary structure. Information about the tertiary and quaternary structures can be obtained indirectly through functional Correspondence to: Carlos Amezcua (Telephone: +224-2704484; Fax: +224-270-2269; E-mail: carlos [email protected]) Journal of Pharmaceutical Sciences, Vol. 102, 1724–1733 (2013) © 2013 Wiley Periodicals, Inc. and the American Pharmacists Association

1724

assays or directly by near-UV CD, nuclear magnetic resonance (NMR) spectroscopy, and X-ray crystallography when the molecule is amenable to crystallization. Optical spectroscopic techniques such as CD and FT-IR have been used extensively in the biopharmaceutical industry to characterize protein structure in comparability exercises. However, with the everincreasing need to better understand the product, orthogonal analytical techniques that can probe for higher order structure (HOS) should be implemented. NMR spectroscopy is ideally suited for such analyses. In NMR, the resonance frequency (or chemical shift) of a given atom depends on the surrounding electronic environment. Therefore, the chemical shifts of NMRactive nuclei (such as protons, carbons, and nitrogens) are dependent on the molecule’s three-dimensional structure. NMR spectra are thus structural fingerprints representing a protein’s HOS and can be compared to assess similarity between samples. Protein NMR has traditionally been used for structural and dynamic characterizations of isotopically enriched polypeptides.6 Isotopic labeling is generally not feasible for biopharmaceuticals; however, recent

JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

ASSESSMENT OF HIGHER ORDER STRUCTURE COMPARABILITY BY NMR

advances in instrument sensitivity have made it possible to analyze proteins at natural abundance and low concentrations. Despite the promising potential of NMR as an orthogonal analytical technique for HOS comparability studies of therapeutic proteins, only a few examples can be found in the literature. The use of two-dimensional (2D) 1 H/1 H-nuclear overhauser effect spectroscopy (NOESY) spectra has been proposed for characterization of the structure and consistency of manufacturing7 and to compare the structural similarity of a protein in solution from different preparations.8,9 NMR fingerprinting using 1 H/15 Nheteronuclear single quantum correlation (HSQC) spectra has been used to assess the identity of the bioactive conformation for recombinant human granulocyte macrophage-colony stimulation factor (rhGMCSF).10,11 More recently, solid-state NMR spectra of lyophilized bovine pancreatic trypsin inhibitor (BPTI) and insulin samples have been developed and used for structural comparisons.12 One-dimensional (1D) proton NMR was used as part of the comparability exR (filgrastim) (Sandoz ercise for the biosimilar Zarzio 13 GmbH, Kundl, Austria). Here, we explore the potential of NMR for the HOS comparability in therapeutic proteins. First, we examined the effects of a source change in a nextgeneration product by assessing the chemical shift differences of assigned peaks. Alternatively, we analyzed the NMR data with easy comparability of HOS by NMR (ECHOS-NMR). This method couples the NMR fingerprinting principle11 with a simple statistical analysis to easily quantify the degree of structural similarity. Second, using ECHOS-NMR, we compared the HOS of an active pharmaceutical ingredient (API) from a different supplier to that of a reference-listed drug (RLD). And third, ECHOS-NMR was used to examine the effects of a manufacturing process and a variation of this process on the structural integrity of a protein. In addition to the HOS comparability assessment by ECHOS-NMR, we also propose the use of NMR-derived diffusion coefficients (Ds) to further characterize the structural similarity between samples.

1725

and finally lyophilized. Test solutions were prepared by dissolving 30 mg of the lyophilized protein in approximately 500 :L of D2 O or H2 O/D2 O (90:10). Final concentration was approximately 30 mg/mL (∼5 mM).

API and RLD Test solutions of the API were made by dissolving the lyophilized protein in solution conditions mimicking the RLD matrix. Five hundred microliters of the protein solution was mixed with 50 :L D2 O. RLD solutions were prepared by aliquoting 500 :L from the drug vial and adding 50 :L D2 O. The final protein concentration was approximately 4 mg/mL (∼0.6 mM). Disulfide bond reduction was accomplished by the addition of incremental amounts of a 100 mM 2-mercaptoethanol (BME) in D2 O stock solution to an RLD sample. There were three moles of disulfide bonds per mole of RLD. Solutions with estimated 6, 28, 50, and 100 mol % of reduced disulfides were made.

Albumin Test solutions were made by dissolving 50 mg of lyophilized protein in 1 mL of D2 O (∼0.7 mM). NMR Spectroscopy

Bovine Pancreatic Trypsin Inhibitor One-dimensional 1 H and a series of 2D homonuclear 1 H spectra were acquired including total correlation spectroscopy (TOCSY), NOESY, and double-quantum filtered correlation spectroscopy (DQF-COSY). All data were obtained on a 600 MHz Bruker DRX NMR spectrometer (Bruker Corporation, Billerica, Massachusetts) with a triple resonance inverse probe, and the sample temperature was maintained at 37◦ C. Felix 2000.1 (Felix NMR, Inc., San Diego, California) was used for data processing and chemical shift assignments. Chemical shift assignments were based on TOCSY, NOESY (300 ms mixing time), and DQF-COSY experiments.

API, RLD, and Albumin

EXPERIMENTAL NMR Samples

Bovine Pancreatic Trypsin Inhibitor Synthetic and bovine-derived proteins were equilibrated under conditions that mimicked the finished drug product, the buffer for which included sodium citrate and histidine. The protein solution was further dialyzed against this product buffer using a 3500 molecular weight cut off regenerated cellulose tube (Spectra/Por, Spectrum Laboratories, Inc., Rancho Dominguez, California), then dialyzed against water, DOI 10.1002/jps

Data were acquired on a 600 MHz Bruker AVANCE III NMR spectrometer equipped with a Dual-13 C/1 H cryoprobe. The sample temperature was maintained at 40◦ C during acquisition of the 1D 1 H, 13 C, and 2D 1 H/13 C- HSQC datasets. Diffusion data were obtained at 25◦ C. Spectra of materials from disulfide bond reduction experiments performed on the RLD were acquired at 25◦ C on an 850 MHz Bruker AVANCE III NMR spectrometer equipped with a QCI cryoprobe. TopSpin 3.0 software (Bruker Biospin, Rheinstetten, Germany) was used for data acquisition and processing. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

1726

AMEZCUA AND SZABO

Spectral Bucketing and Linear Regression AMIX 3.9 (Bruker Biospin) was used for spectral binning. The intensity of each bin was measured and divided by the total intensity to normalize the spectra. Normalization eliminates protein concentration effects. Unless otherwise noted, selected spectral regions were divided into bins of width 0.05 ppm (proton), 0.5 ppm (carbon), or 0.05 by 0.5 ppm (1 H/13 C, HSQC). The bin width was selected to be able to detect structurally related chemical shift changes and to average out chemical shift differences arising from nonidentical solution conditions and minor differences observed during data acquisition between samples. Excel 2007 (Microsoft Corporation, Redmond, Washington) was used for linear regression analyses.

Bovine Pancreatic Trypsin Inhibitor The full proton spectrum, excluding regions containing signals from the buffer and excipients, as well as the aliphatic region (between 2.3 and –0.3 ppm) were used for binning. A single sample of each BPTI variant was used for the analysis.

API and RLD Methyl group regions of the proton (1.5 to –0.2 ppm region) and 1 H/13 C-HSQC (1 H: 1.5 to −0.2 ppm, 13 C: 27.5 to 11.0 ppm) spectra were used for binning. Averaged bin intensities were obtained from bins of three different lots. The average values were then used in linear regression analyses between the different samples. A single RLD lot was used for disulfide bond reduction, the pH comparison, and against Lysozyme.

Albumin Regions of the proton (4.5 to −0.5 ppm), carbon (185 to 10 ppm), and 1 H/13 C-HSQC (1 H: 3.3 to 0.5 ppm, 13 C: 41.0 to 10.0 ppm) spectra were used for binning. Linear regression analyses were performed between the different samples.

RESULTS AND DISCUSSION HOS Comparison Because of a Source Change (Bovine Vs. Synthetic Protein) Bovine pancreatic trypsin inhibitor is a small globular protein of 58 amino acids with a molecular weight of approximately 6.5 kDa. A synthetic variant of this protein was produced in an effort to eliminate animalderived materials from an approved product. NMR was used to assess the structural similarity between the bovine and synthetic proteins. Figure 1 shows a stacked plot of their proton spectra. Visual inspection readily indicated a high degree of similarity between the spectra of the two protein samples. To more quantitatively describe the degree of spectral similarity, approximately 80% of the peaks were assigned to JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

specific hydrogen atoms throughout the protein using a combination of 1D- and 2D-1 H NMR experiments. The average chemical shift difference between the assigned peaks in the synthetic versus the bovine proteins was 0.005 ppm. This was an extremely small difference and less than what would typically be expected for the various forms of experimental variations that can occur. Such a close match between the bovine and synthetic BPTI spectra was an extremely strong indicator that the two forms of the protein must possess identical HOS. An alternative approach for the estimation of spectral similarity is the comparison of peak intensities from 2D 1 H/1 H-NOESY spectra.7,8 However, both the peak intensity and chemical shift comparison approaches require spectra with well-resolved peaks. As the molecular weight of a protein increases, the peak overlap and broadness also increase, thus reducing the number of resolved peaks that can be used for such analyses. We, therefore, developed ECHOS-NMR, a spectral comparison method that is relatively independent of the protein’s molecular weight and spectral resolution. In ECHOS-NMR, the normalized bin intensities from two different spectra are then subjected to linear regression analysis. The correlation coefficient (R2 ) obtained from this analysis is used as a measure of spectral similarity. The higher the R2 value is, the more similar the spectra and protein structures are. A similar strategy has been applied to quantitatively compare FT-IR spectra by Prestrelski et al.14,15 ECHOS-NMR can be particularly useful when only a limited number of spectra need to be compared, as opposed to multivariate statistical methods. This was the case for the examples presented in this manuscript. We tested ECHOS-NMR initially with the BPTI proton spectra shown in Figure 1. Linear regression analysis yielded a correlation coefficient of 0.99 when examining either the full spectra (excluding buffer and excipient signals) or only the aliphatic region. The bin size used during the regression analysis only had a small effect on the R2 value. When the bin size was changed to 0.01 ppm, R2 changed to 0.98 (full spectrum) or 0.97 (aliphatic region). A close examination of the spectra indicated that the reduced R2 was a product of unequal spectral line widths caused by shimming differences. The high R2 values were consistent with the high degree of chemical shift similarity obtained above for the two bovine and synthetic samples. These data indicated that the ECHOS-NMR method can be used instead of direct chemical shift comparisons. Data collection and analysis time can then be significantly reduced because ECHOS-NMR does not require peak assignments. In addition, as demonstrated below, our method is applicable to a broad range of NMR spectra and protein sizes. DOI 10.1002/jps

ASSESSMENT OF HIGHER ORDER STRUCTURE COMPARABILITY BY NMR

His

His

His His

1727

Cit

HOD

Bovine

Synthetic

10

8

6

4

2

0

Proton (ppm) Figure 1. Proton NMR spectra of bovine (top) and synthetic (bottom) BPTI. The signals from buffer components, histidine (His) and citrate (Cit), are shown off scale so that the protein signals can be observed.

HOS Comparison Between a RLD and an API Next, we compared the HOS of an approved biopharmaceutical RLD to that of an API from an alternative supplier. The protein subject to this analysis consisted of a molecular species of approximately 35 kDa. As expected, the larger molecular weight resulted in broader line widths and increased signal overlap in the proton spectra (Fig. 2a). The solutions contained large amounts of protonated excipients and preservatives whose signals masked most of the amide region and interfered with the 4.0 to 2.0 ppm area of the aliphatic region (not shown). An attempt to obtain a 1 H/15 N-HSQC spectrum from an unlabeled API sample on a 900 MHz NMR spectrometer equipped with a cryoprobe did not yield useful data even after approximately 62 h of acquisition. Fortunately, the low-frequency region (1.8 to −0.2 ppm) was free of interfering signals and could be used to compare the samples. This spectral region mainly contains signals from methyl groups, which are useful structural probes because of their tendency to reside within a protein’s hydrophobic core. To aid with the peak overlap observed in the proton spectra, 2D 1 H/13 C-HSQC DOI 10.1002/jps

spectra of the methyl region (Fig. 2b) were obtained. Visual inspection of proton and 1 H/13 C-HSQC spectra (Fig. 2) indicated a high degree of similarity between samples, thus suggesting similar HOS. ECHOS-NMR was then used to assess the HOS similarity in a more quantitative manner. Proton and 1 H/13 C-HSQC spectra were divided into bins as described in the Experimental section. Linear regression analyses of the proton and 1 H/13 C-HSQC bins were performed for three different RLD and API lots to determine an average R2 value and to determine whether there was any inherent lot-to-lot variation as reported by the standard deviation (Table 1). R2 values near unity and very small standard deviations were obtained indicating high homogeneity between lots. In general, R2 values from HSQC data were lower than those for the proton data. This result was likely due to the lower number of points defining the 2D dataset and the lower S/N. Regression analysis between two approved RLDs (RLD and RLD-2) of the same protein species from different manufacturers provided an additional control (Table 1). The variation between the two approved RLDs defined an acceptable R2 for comparability. Together, these data JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

1728

AMEZCUA AND SZABO

a

API

RLD 1.5

1.0

0.5

0.0

Proton (ppm)

20 25 30

Carbon (ppm)

15

b

RLD 1.5

1.0

0.5

API

0.0

1.5

1.0

0.5

0.0

Proton (ppm)

Figure 2. (a) Proton and (b) 1 H/13 C-HSQC spectral comparison between a 35-kDa protein species from an RLD and an API. Shown are the spectral regions typical of peaks corresponding to methyl groups that were used for linear regression analyses.

Table 1. List of Correlation Coefficients Obtained from Linear Regression Analyses of Binned Proton and 1 H/13 C-HSQC Spectra for a 35 kDa Protein Species Proton-R2 RLD lot-to-lot 0.99 ± 0.01 API lot-to-lot 0.99 ± 0.00 RLD versus RLD-2 0.97 pH3 versus pH 7 0.42 RLD versus Lysozyme 0.38 Disulfide bond reduction (%, per molar basis) 6% 0.97 28% 0.84 50% 0.74 100% 0.60 RLD versus API 1.00

HSQC-R2 0.97 ± 0.01 0.93 ± 0.04 0.94 0.14 0.00 0.96 0.88 0.80 0.60 0.96

indicated that R2 values of 0.97 or higher (for proton) and 0.93 or higher (for HSQC) were acceptable measures of structural comparability. The sensitivity of ECHOS-NMR to various degrees of structural differences in this system was further investigated. Different conformations of the same protein were examined by comparing spectra of RLD samples at pH 3 (the protein was aggregated) and pH 7 (active conformation). We also compared the JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

RLD against a different protein (lysozyme). As shown in Figure 3, poor spectral overlap resulted in a lack of correlation (R2 ≈ 0; Table 1). Local conformational changes were artificially generated by reduction of disulfide bonds with various amounts of BME. Figure 4 shows spectral overlaps and linear regression graphs for a partially reduced (∼28%) and an untreated RLD as well as for two different RLD lots. The HSQC spectra suggest that the overall protein conformation was not perturbed because only a few peaks, those in proximity to the disulfide bonds, displayed chemical shift changes. ECHOSNMR gave R2 values of 0.98 (6% reduction), 0.88 (28% reduction), 0.80 (50% reduction), and 0.60 (100% reduction). Proton data followed a similar trend (Table 1). The R2 value between spectra of two untreated RLD lots was 0.98 (Fig. 4). Therefore, ECHOS-NMR was unable to detect differences due to the smallest level of disulfide reduction tested (6%). Even when the bin size was reduced to 0.01 ppm in the proton or to 0.01/0.1 ppm in the HSQC spectra, the correlation coefficient only changed slightly to 0.96 (for both spectra). Overall, these data demonstrated that the correlation coefficients derived from DOI 10.1002/jps

1729

ASSESSMENT OF HIGHER ORDER STRUCTURE COMPARABILITY BY NMR

15

RLD: pH 3 versus pH 7 0.008

y = 0.26x + 0.00 R² = 0.14

RLD at pH 3

20 25 30

RLD-pH 3 (green) RLD-pH 7 (black)

Carbon (ppm)

0.007 0.006 0.005 0.004 0.003 0.002 0.001 0 0

0.002

0.004

0.006

0.008

0.01

0.012

0.014

15

RLD at pH 7

Lysozyme versus RLD

0.005

1.0

0.5

0.0

Lysozyme

20 25 1.5

30

Lysozyme (blue) RLD (black)

y = 0.00x + 0.00

0.004

R² = 0.00 0.003

0.002

0.001

0 0

0.002

0.004

0.006

0.008

0.01

0.012

0.014

RLD

Proton (ppm) Figure 3. 1 H/13 C-HSQC spectral overlays (left) and linear regression graphs (right) for the RLD at pH 7 versus pH 3 (top) and Lysozyme versus RLD (bottom). Both the spectral overlays and the regression graphs indicate different higher order structure between the samples being compared. Axes in the regression charts represent normalized intensities of spectral bins.

ECHOS-NMR can be used as a quantitative measure of the structural perturbations visible in NMR spectra. However, the sensitivity of ECHOS-NMR toward small structural changes and/or alternate conformations present at small levels will depend on the spectrometer’s limit of detection and the magnitude of chemical shift and peak intensity changes for each particular system under study. The ECHOS-NMR comparison between RLD and API samples provided strong evidence of highly similar HOS (Table 1). R2 values for both, proton and 1 H/13 C-HSQC spectra, were within the accepted R2 values for protein comparability established above. The HOS comparability assessment derived from ECHOS-NMR can be complemented by diffusion measurements. A molecule’s D is a physicochemical constant that, under given solution conditions, is a function of its hydrodynamic radius, that is, its effective molecular weight, size, and shape.16 Ds thus provide information about the overall protein structure in solution similar to SEC, analytical ultracentrifugation, and dynamic light scattering. Ds were measured for three RLD and API lots. The average Ds and DOI 10.1002/jps

standard deviations were 9.6 ± 1.3 × 10−11 and 9.6 ± 0.3 × 10−11 m2 /s for the RLDs and APIs, respectively. The agreement between Ds provided further evidence of comparable HOS between the RLD and API samples. Although not the focus of this investigation, it is worth noting that ECHOS-NMR can also report the presence of impurities such as residual solvents. For example, one of the API lots examined contained a detectable amount of ethanol. The signal associated with the methyl peak fell within the aliphatic region of the spectra used for the comparison against the RLD. The regression analysis resulted in correlation coefficients of 0.76 and 0.68 for proton and HSQC spectra, respectively. These values were clearly outside the acceptable comparability range determined for this system. HOS Comparison Because of a Manufacturing Process Change For the final example, three albumin samples were examined: albumin before processing and albumin JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

1730

AMEZCUA AND SZABO

RLD: lot 2 versus Lot 1 15

0.035

y = 1.00x + 0.00 R² = 0.98

25

RLD lot 2

20

0.03 0.025 0.02 0.015 0.01

30

0.005 0

15

0

Carbon (ppm)

RLD-Lot 2 (red) RLD-Lot 1 (black)

0.005

0.01

0.015

0.02

0.025

0.03

0.035

RLD Lot 1

~28% Reduced versus untreated RLD

1.5

1.0

0.5

RLD + 0.5 mM BME

RLD + 0.5 mM BME (magenta) RLD (black)

30

25

20

0.05

y = 0.88x + 0.00 R² = 0.88

0.04

0.03

0.02

0.01

0 0

0.01

0.02

0.03

0.04

0.05

RLD

0.0

Proton (ppm)

Figure 4. 1 H/13 C-HSQC spectral overlays (left)and linear regression graphs (right) for two RLD lots (top) and a partially reduced (˜28 %) versus an untreated RLD (bottom). The spectral overlays indicate identical HOS between the two RLD lots. The partially reduced sample seems to preserve the overall protein fold given that only a few peaks, those near the cysteines involved in disulfide bond formation, show chemical shift changes and/or reduced intensities. The correlation coefficients obtained from the regression analyses of these two examples suggests that a structural change has occurred. Axes in the regression charts represent normalized intensities of spectral bins.

after having been subject to two different manufacturing processes. Sample availability was limited in this investigation; therefore, only one sample of each albumin type was analyzed. The goal was to determine whether either manufacturing process affected the protein’s structural integrity. At 67 kDa, albumin is considered a large protein for traditional NMR structural characterization analyses. Peaks in the spectra are rather broad and the NMR signal decays rapidly because of relaxation phenomena typical of large protein, and this is a factor especially in multipulse 2D NMR spectra. To determine the applicability of different NMR spectra for structural comparisons of large proteins, proton, carbon, and 1 H/13 C-HSQC NMR spectra were acquired and compared for each sample. In addition, Ds were also measured. Figure 5 shows the proton (a, aliphatic region), carbon (b), and 1 H/13 C-HSQC (c, methyl region) spectra JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

of the three different albumin samples. Broad peaks were observed in all three spectra; however, the effect was more severe in the HSQC. Nevertheless, no significant differences were observed between the samples in either spectrum. This was also reflected by unity or close to unity R2 values obtained from linear regression analyses (Table 2). Furthermore, the measured Ds 4.95, 4.90, and 4.95 × 10−11 m2 /s for the preprocess, process 1, and process 2, respectively, were within experimental error (± 0.03 × 10−11 m2 /s) of one another. All together, the NMR data suggested a lack of process-related HOS perturbations.

CONCLUSIONS In this paper, we applied NMR to assess the HOS comparability of three nonglycosylated proteins with molecular weights in the range of approximately DOI 10.1002/jps

1731

ASSESSMENT OF HIGHER ORDER STRUCTURE COMPARABILITY BY NMR

c

a

15

Process 2

25

20

Process 2

35

30

Process 1

3

2

Process 1

0

(ppm)

35

30 40

Process 2

(ppm)

b

Carbon

25

20

Proton

1

15

4

40

Preprocess

25

20

15

Preprocess

Process 1

30

Preprocess 100

0

(ppm)

40

Carbon

50

35

150

3.0

2.5

2.0

1.5

Proton

1.0

0.5

(ppm)

Figure 5. Proton (a, aliphatic region), carbon (b), and 1H/13C-HSQC (c, methyl region) spectra of the preprocess, process 1, and process 2 albumin samples.

6.5–67 kDa. A statistical method called ECHOS-NMR was developed to compare the NMR spectra. ECHOSNMR provides a rapid and quantitative way to analyze the structural similarity of proteins via the R2 value derived from linear regression of binned NMR spectra. In addition, there is no need to isotopically enrich the samples or perform peak assignments. Biologics typically possess some degree of inherent structural and size variability, especially compared DOI 10.1002/jps

with discrete chemical entities. Therefore, the baseline value and variability of R2 that corresponds to an acceptable measure of “sameness” for each system should be determined. This could be achieved, for example, by determining the variability between NMR spectra of reference samples. The sensitivity of ECHOS-NMR seems to be highly related to the ability of the spectrometer to detect spectral differences. For example, alternate protein conformations present JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

1732

AMEZCUA AND SZABO

Table 2. List of Correlation Coefficients Obtained from Linear Regression Analyses of Binned Proton, Carbon, and 1 H/13 C-HSQC Spectra for Albumin Proton-R2

Carbon-R2

HSQC-R2

1.00 1.00 1.00

1.00 1.00 0.99

0.96 0.96 0.98

Preprocess versus process 1 Preprocess versus process 2 Process 1 versus process 2

below the detection level can be missed. Therefore, the sensitivity of this method should also be evaluated for each system. Although an entire NMR spectrum may be used for analysis, specific regions, containing relevant structural information, may alternatively be selected. It has been demonstrated that the amide region of the 1 H/15 N-HSQC spectrum is a useful structural fingerprint.10,11 However, in cases wherein 1 H/15 N-HSQC data is inadequate, such as low protein concentration and/or large molecular weight, the aliphatic region can be used as an indicator of overall three-dimensional structure. The applicability of ECHOS-NMR was demonstrated using proton, carbon, and 1 H/13 C-HSQC spectra. However, it could similarly be applied to compare other NMR spectra (e.g., 1 H/1 H-NOESY, 1 H/15 N-HSQC, etc). Spectral regions may be selected based on the absence of interfering signals such as water, buffer, solvent, and/or excipients. We also showed that comparison of Ds between samples could be used independently or in addition to ECHOS-NMR to further analyze the structural similarity between samples. Several articles have been published in the last few years describing the qualification of CD17 and FT-IR18 spectroscopy as well as strategies for qualification of other biophysical techniques in terms of precision and sensitivity to structural perturbations.19 In addition, evaluation of the performance of different spectral comparability algorithms has also been reported.20 Similar efforts should be followed to further qualify NMR regarding its applicability to different protein sizes, effects of glycosylation and other chemical modifications on the spectra, sensitivity to different degrees of structural changes, alternative methods for spectral comparison, and so forth. On-going work by us, as well as by other groups, is being aimed at understanding many of these parameters. Nevertheless, the work presented in this manuscript, together with the references cited within, demonstrate that NMR can be used as an orthogonal technique for HOS comparability exercises of a variety of therapeutic proteins.

ACKNOWLEDGMENTS The authors wish to thank Drs. Sarah Lee, Christine Rebbeck, and Andreas Goessl for providing the protein samples used in this study. We also thank JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013

Drs. Joseph Ray and Edward Chess for critical review of the manuscript.

REFERENCES 1. Center for Drug Evaluation and Research (CDER) and Center for Biologics Evaluation and Research (CBER). Food and Drug Administration. 2005. Q5E comparability of biotechnological/ biological products subject to changes in their manufacturing process. 2. European Medicines Agency. 2005. ICH Topic Q5E, Comparability of Biotechnological/Biological Products. 3. Committee for Medicinal Products for Human Use. European Medicines Agency. 2006. Guideline on similar biological medicinal products containing biotechnology-derived proteins as active substance: Non-clinical and clinical issues. 4. Chirino AJ, Mire-Sluis A. 2004. Characterizing biological products and assessing comparability following manufacturing changes. Nat Biotechnol 22(11):1383–1391. 5. Kranz J, Al-Azzam F, Saluja A, Svitel J, Al-Azzam W. 2013. Techniques for higher order structure determination. In Biophysics for therapeutic protein development, biophysics for the life sciences; Narhi L, Ed. New York: Springer, pp 33– 82. 6. Wuthrich K. 2003. NMR studies of structure and function of biological macromolecules (Nobel Lecture). J Biomol NMR 27(1):13–39. 7. Freedberg DI. 2005. Using nuclear magnetic resonance spectroscopy to characterize biologicals. Dev Biol 122:77–83. 8. Zuperl S, Pristovsek P, Menart V, Gaberc-Porekar V, Novic M. 2007. Chemometric approach in quantification of structural identity/similarity of proteins in biopharmaceuticals. J Chem Inf Model 47(3):737–743. 9. Parsons IC, Lansing J, Bosques CJ, Beccati D. 2010. Comparative analysis of protein conformations by using 2D NOESY NMR spectra. United States Patent Publication US 2010/ 0188084 A1. pp 1–11. 10. Panjwani N, Hodgson DJ, Sauve S, Aubin Y. 2010. Assessment of the effects of pH, formulation and deformulation on the conformation of interferon alpha-2 by NMR. J Pharm Sci 99(8):3334–3342. 11. Aubin Y, Gingras G, Sauve S. 2008. Assessment of the threedimensional structure of recombinant protein therapeutics by NMR fingerprinting: Demonstration on recombinant human granulocyte macrophage-colony stimulation factor. Anal Chem 80(7):2623–2627. 12. Zhou DH, Shah G, Mullen C, Sandoz D, Rienstra CM. 2009. Proton-detected solid-state NMR spectroscopy of naturalabundance peptide and protein pharmaceuticals. Angew Chem 48(7):1253–1256. 13. Sorgel F, Lerch H, Lauber T. 2010. Physicochemical and biologic comparability of a biosimilar granulocyte colony-stimulating factor with its reference product. BioDrugs 24(6):347–357. 14. Prestrelski SJ, Arakawa T, Carpenter JF. 1993. Separation of freezing- and drying-induced denaturation of lyophilized proteins using stress-specific stabilization. II. Structural studies using infrared spectroscopy. Arch Biochem Biophys 303(2):465–473. 15. Prestrelski SJ, Tedeschi N, Arakawa T, Carpenter JF. 1993. Dehydration-induced conformational transitions in proteins and their inhibition by stabilizers. Biophys J 65(2):661–671. 16. Dehner A, Kessler H. 2005. Diffusion NMR spectroscopy: Folding and aggregation of domains in p53. Chembiochem 6(9):1550–1565. 17. Li CH, Nguyen X, Narhi L, Chemmalil L, Towers E, Muzammil S, Gabrielson J, Jiang Y. 2011. Applications DOI 10.1002/jps

ASSESSMENT OF HIGHER ORDER STRUCTURE COMPARABILITY BY NMR

of circular dichroism (CD) for structural analysis of proteins: Qualification of near- and far-UV CD for protein higher order structural analysis. J Pharm Sci 100(11):4642– 4654. 18. Jiang Y, Li C, Nguyen X, Muzammil S, Towers E, Gabrielson J, Narhi L. 2011. Qualification of FTIR spectroscopic method for protein secondary structural analysis. J Pharm Sci 100(11):4631–4641.

DOI 10.1002/jps

1733

19. Jiang Y, Li C, Gabrielson J. 2013. Qualification of biophysical methods for the analysis of protein therapeutics. In Biophysics for therapeutic protein development, Biophysics for the life sciences; Narhi LO, Ed. New York: Springer, pp 99–126. 20. D’Antonio J, Murphy BM, Manning MC, Al-Azzam WA. 2012. Comparability of protein therapeutics: Quantitative comparison of second-derivative amide I infrared spectra. J Pharm Sci 101(6):2025–2033.

JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 102, NO. 6, JUNE 2013