On the Analytical Superiority of 1D NMR for Fingerprinting the Higher

Apr 30, 2015 - The horizontal bars designate 95% confidence limits for the mean similarity values (red circles). Each blend was replicated in five dif...
0 downloads 6 Views 2MB Size
Article pubs.acs.org/ac

On the Analytical Superiority of 1D NMR for Fingerprinting the Higher Order Structure of Protein Therapeutics Compared to Multidimensional NMR Methods Leszek Poppe,*,† John B. Jordan,† Gary Rogers,‡ and Paul D. Schnier†,∥ †

Discovery Attribute Sciences, Amgen Inc., One Amgen Center Drive, Thousand Oaks, California 91320, United States Product Attribute Sciences, Amgen Inc., One Amgen Center Drive, Thousand Oaks, California 91320, United States



S Supporting Information *

ABSTRACT: An important aspect in the analytical characterization of protein therapeutics is the comprehensive characterization of higher order structure (HOS). Nuclear magnetic resonance (NMR) is arguably the most sensitive method for fingerprinting HOS of a protein in solution. Traditionally, 1 H−15N or 1H−13C correlation spectra are used as a “structural fingerprint” of HOS. Here, we demonstrate that protein fingerprint by line shape enhancement (PROFILE), a 1D 1H NMR spectroscopy fingerprinting approach, is superior to traditional two-dimensional methods using monoclonal antibody samples and a heavily glycosylated protein therapeutic (Epoetin Alfa). PROFILE generates a high resolution structural fingerprint of a therapeutic protein in a fraction of the time required for a 2D NMR experiment. The cross-correlation analysis of PROFILE spectra allows one to distinguish contributions from HOS vs protein heterogeneity, which is difficult to accomplish by 2D NMR. We demonstrate that the major analytical limitation of twodimensional methods is poor selectivity, which renders these approaches problematic for the purpose of fingerprinting large biological macromolecules.

A

key challenge in analytical measurement science is the rapid characterization of higher order structure (HOS) of protein therapeutics. The need for rapid, precise, sensitive, and robust HOS assays has increased as regulatory agencies move to create abbreviated pathways for the approval of “biosimilar” biological products.1 It is well-known that the three-dimensional structures of therapeutic proteins correlate closely with biological function.2 Therefore, the accurate assessment of HOS is paramount when characterizing protein therapeutics produced by different processes or sources. A variety of different analytical approaches to accomplishing this comparison have been proposed including circular dichroism (CD), Fourier transform infrared spectroscopy (FT-IR), differential scanning calorimetry (DSC), nuclear magnetic resonance (NMR) spectroscopy, X-ray diffraction, cryo electron microscopy (cryo-EM), and H/D exchange mass spectrometry (HDXMS).3−5 Solution NMR analysis has several analytical advantages compared to the last three techniques, as it does not require sample manipulation (i.e., data are collected on intact and/or formulated proteins). High-resolution NMR spectra are exquisitely sensitive to subtle structural changes. In fact, higher order protein structure in solution can be derived from NMR measurements based on first-principles, which is not the case for other current approaches, like CD or FT-IR. NMR spectroscopists usually associate the three-dimensional “fingerprint” of a protein with the peak pattern observed in a 2D 1H−15N or 1H−13C or even 1H−1H correlation spectra.6 © XXXX American Chemical Society

These two-dimensional NMR experiments, with the exception of the latter, typically require isotopic enrichment of proteins with either 15N or 13C, but it is possible to obtain these types of spectra on natural abundance material with sufficiently high concentrations and/or long experiment times.7,8 Recently, this methodology (using natural isotope abundance) was applied to recombinant human granulocyte macrophage colony stimulating factor (rhGM-CSF), a relatively small (∼14 kDa), nonglycosylated protein.8 In this particular case, the assessment of structural similarity by NMR was based on the visual inspection of the two 1H−15N spectra. Aside from the difficulties implicit with drawing conclusions based on a qualitative and subjective assessment of similarity, recording 2D correlation spectra at natural abundance is substantially more challenging for large proteins such as monoclonal antibodies (mAbs). Broad lines, spectral crowding, and carbohydrate heterogeneity significantly decrease the availability and resolution of signal in these experiments and thus make use of natural abundance 2D NMR as a viable technique for structural comparison highly impractical (vide inf ra). We recently proposed the protein fingerprint by line shape enhancement (PROFILE) methodology as an alternative Received: November 3, 2014 Accepted: April 30, 2015

A

DOI: 10.1021/acs.analchem.5b00950 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry approach to two-dimensional NMR fingerprinting methods.9 This method exploits differences in the diffusion properties of large antibodies and formulation components to generate a highly resolved one-dimensional 1H spectrum of the entire mAb, devoid of solvent and excipients. We demonstrated that, for isotopically labeled protein samples, the PROFILE similarity is commensurable and more sensitive to structural changes than spectral similarity as measured from 2D spectra acquired for the same mAbs. Recently, 2D NMR approaches for antibody fingerprinting have been proposed for mAbs that are enzymatically cleaved into their smaller fragments.10 These fragments, due to their smaller size, can yield improved 2D spectra at natural abundance as compared to an intact antibody, yet may still require very long acquisition times. However, in our experience with this method and as we demonstrate here, the integrity of the protein is compromised upon enzymatic cleavage, and the resulting spectra are not representative of the intact mAb structure and therefore are inadequate for fingerprint determination. The assessment of HOS similarity between different protein samples can be complicated by intrinsic structural heterogeneity, due mostly to differences in glycosylation, which may or may not be associated with the differences in HOS. The process or lot-dependent glycosylation heterogeneity is one of the biggest hurdles in the manufacturing of protein therapeutics. This structural heterogeneity and its inherent variation is often responsible for differences in biological activity or even function of therapeutic proteins.11 With the exception of NMR, differences in glycosylation are difficult to establish by analytical or biophysical characterization of intact samples, usually requiring enzymatic protein decomposition followed by LC-MS analysis.12 Here, we separate contributions of higher order structure from primary structural heterogeneity (including glycosylation) in the similarity analysis of glycoprotein and antibody samples. Also, in this work, a logarithmic scale for the similarity measure of the PROFILE spectra is introduced, which facilitates the assessment of precision and selectivity across different methods (eq 1). This scale is convenient to represent seemingly small (but statistically significant) differences in correlation coefficients; i.e., differences in R values of 0.990 and 0.999 may not “appear” significant but a difference between 30 and 40 dB certainly does. However, when interpreting spectral similarity in terms of known protein differences, we use the original scale in % units. Thus, the dB and % scales serve here for either operative or descriptive purposes, respectively. The cross-correlation analysis of NMR spectra with the outcome in dB (decibels) units was applied to samples of IgG1 and IgG2 molecules in their native, deglycosylated, or unfolded states and to mixtures of both molecules. We applied the same analysis to the samples of a glycoprotein therapeutic obtained from two different manufacturing processes: the licensed manufacturing process (rhEPO-A) and a candidate manufacturing process with potentially higher product yield (rhEPO-B). The common workflow of the different samples comparability adopted in this work is described in Figure 1.

Figure 1. Workflow of the comparative sample analysis by NMR adopted in this work. Stock solutions of samples A and B were used to prepare five NMR samples in 4 mm Shigemi tubes. The NMR measurements were performed on either Bruker 800 or 600 MHz spectrometers. PROFILE spectra data processing was performed with Topspin 3.0 and in-house developed scripts in Matlab, as described in ref 9. The similarity scores were finally compared by the ANOVA statistical analysis.

PROFILE analysis were prepared in 4 mm Shigemi tubes at 300 μM in 180 μL. The samples for the 2D experiments were prepared in 5 mm Shigemi tubes at 2 mM in 300 μL. All NMR experiments were recorded at 305 and 333 K (EPO) or 318 K (mAb) on Bruker Avance III 800 and 600 MHz spectrometers equipped with TCI cryoprobes and automatic sample changers (BACS). The PROFILE experiments were recorded with the PGSTE pulse sequence,13 where the diffusion delay and diffusion gradients lengths were set to 100 and 1 ms, respectively. The Z gradients were applied at maximum strength (57 G/cm). The acquisition time and relaxation delay were both 1 s, with 2560 and 128 scans accumulated for the 300 μM and 2 mM samples, respectively. For statistical analysis, the PROFILE spectra were recorded in triplicate utilizing a Bruker BACS sample changer running in randomized mode. The total acquisition time for a single PROFILE experiment was approximately 7 and 140 min for 2 mM and 300 μM, respectively. The S/N was at least 95:1 in all the spectra. Both sample concentrations generated the same similarity analysis outcome. The samples with the unfolded



EXPERIMENTAL SECTION Sample Preparations and NMR Measurements. Six different lots of rhEPO produced with two different manufacturing processes, A and B (three lots of each process), were chromatographically purified and prepared in 20 mM phosphate buffer, 95% H2O/5% D2O, at pH 7. The samples for B

DOI: 10.1021/acs.analchem.5b00950 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

t2 and t1 dimensions equal to 80 and 22 ms, respectively, and a relaxation delay of 0.5 s. The spectra were recorded with 16, 64, and 256 scans per t1 increment and processed with Topspin 3.0 as described above. All sample information has been compiled in Table S1, Supporting Information. Spectral Similarity Calculations. The similarity of two spectra, either 1D 1H, PROFILE or 2D, was calculated as the maximum value of the normalized cross-correlation function calculated from the corresponding vectors9 (see Figure S1, Supporting Information). Here, we introduce the logarithmic scale of similarity (eq 1), in analogy to the evidence scale from probability theory.17

protein were generated by elevating the temperature of the NMR experiment to 60 °C, where EPO reversibly unfolds.14 The 1H−13C and 1H−15N correlation spectra were recorded with the sensitivity-enhanced HSQC15 and TROSY16 pulse sequences, respectively. Although TROSY works optimally for deuterated proteins, in the case of the intact 15N labeled (nondeuterated) antibody, in our hands, TROSY was found to provide superior sensitivity compared to HSQC/HMQC experiments. For unlabeled samples, the TROSY pulse sequence yielded far better suppression of subtraction artifacts. The HSQC spectra were recorded with 1 and 0.05 s relaxation and acquisition times, respectively, and using 80 scans per t1 increment (resulting in 26 h of total acquisition time). The TROSY spectra were recorded with 0.3 and 0.05 s acquisition times, respectively, using 4000 scans per increment, resulting in 96 h of total acquisition time. The spectra were processed with Bruker Topspin 3.0 software. Prior to Fourier transform, the raw data were multiplied by a Gaussian window function (LB = −1, GB = 0.05 and LB = −10, GB = 0.05 for 1D and 2D spectra, respectively). The native IgG1 and IgG2 antibodies were protein standards generated in house with the same CDR regions and overall 4% differences in amino acid composition. The NMR samples were prepared in 10 mM acetate buffer containing 9% sucrose at pH = 5.2 and at 50 mg/mL concentrations (5% D2O for field lock). The 15N labeled proteins were obtained as described previously and measured at 16 mg/mL concentrations. The 5% and 10% blended samples were prepared from the two proteins at the same stock concentrations. For the statistical analysis, five dif ferent NMR samples of the same solution were prepared in 4 mm Shigemi tubes with 180 μL sample volumes. The deglycosylated IgG1 samples for NMR were obtained by incubating a dilute solution of native protein with PNGase F (New England Biolabs) in 20 mM sodium phosphate, pH 7.0 overnight at 37 °C. Samples were then analyzed for complete (>98%) deglycosylation by LC-MS analysis. NMR samples were then prepared by exchanging the diluted antibody into the aforementioned acetate buffer using a 30 000 MWCO concentrator (Millipore). The unfolded samples of IgG1 and IgG2 for NMR analysis were prepared by adding concentrated solutions of TCEP (tris(2-carboxyethyl)phosphine) and DMU (dimethyl urea) to final concentrations of 1 mM and 4 M, respectively, at pH 2. The PGSTE experiments were acquired as described above with diffusion delays of Δ = 250 ms and Δ = 150 ms at 800 and 600 MHz proton frequencies, respectively (these were adjusted for complete suppression of buffer resonances). The shorter Δ on the 600 MHz system was compensated for by the stronger pulse field gradient at its maximum (67 G/cm). The PROFILE experiments were recorded with 1 and 1.7 s acquisition and relaxation times, respectively, with 256 scans for the unlabeled samples and 512, 1024, and 2560 scans for the 15N labeled samples. The raw data were processed with TOPSPIN 3.0, multiplied by a Gauss-Lorentz window function (LB = −1, GB = 0.005); the residual water signal was removed by using a digital filter (150 Hz window). Data were Fourier transformed and automatically zero-order phase corrected. The final zeroorder phase and baseline corrections and the PROFILE spectral decomposition was performed within the Matlab R2012a (MathWorks Inc.) programming environment. Unless indicated, the S/N in the fingerprint spectra was at least 60:1. The 1H−15N spectra were acquired with the sensitivity optimized TROSY pulse sequence16 with acquisition times in

S(dB) = 10 × log

R 1−R

(1)

R is the cross-correlation coefficient, calculated as described above. On this scale, a correlation coefficient of 0.999 corresponds to a similarity score of 30 dB. For convenience, we show the correlation coefficient in the dB scale in Figure S2 (Supporting Information). Selectivity and Precision of the NMR Experiments. To quantify differences in NMR spectra, we define the selectivity of the experiment (Φexp) by the following relationship: ε Φexp = tε × c M 2 (2) where ε is the similarity difference (precision) in dB units which can be resolved in the experimental time tε (in minutes) and CM is sample concentration in mM. For PROFILE spectral processing, the ε value depends on the smoothing function which generates the contour spectrum. Through this work, we used a Gaussian function with fwhm (full width at halfmaximum) of 330 Hz. Figure S3 (Supporting Information) shows the dependences of ε on the fwhm parameter (graph A) or the S/N (graph B) for the 5% IgG2 in IgG1 sample. Both graphs plateau at fwhm of 300 Hz and S/N of 65:1 and demonstrate that instrumental factors other than thermal noise limit the precision, which would otherwise monotonically increase with tε (or S/N) It is convenient, due to the nature of the fingerprint signal, to define the signal-to-noise ratio (S/N) as the ratio of the RMSD of the signal to the RMSD of the noise. For the PROFILE spectra, we used the fingerprint spectrum for the S/N calculation with 6−11 ppm and −2 to −4 ppm intervals for the signal and for the noise calculations, respectively. In the case of the 1H−15N 2D experiments, the signal contribution was calculated from the 6−11 ppm (1H) and 100−135 ppm (15N) range and the noise was calculated from the 1 ppm × 1 ppm region with no signal.



RESULTS HOS vs Structural Heterogeneity. The PROFILE analysis of rhEPO-A and rhEPO-B samples is shown in Figure 2, where the blue and red colors correspond to the native and thermally unfolded protein samples, respectively. The similarity (S) variations between samples from the different lots of A or B were within the precision of the measurement for the same sample (Figure 2). We introduce the logarithmic scale (in dB units) for the similarity measure (vide infra). Notably, the unfolded samples of A and B are only slightly more dissimilar than the corresponding native proteins. This implies that structural differences not attributable to HOS are responsible for the PROFILE dissimilarity between A and B. A closer look

C

DOI: 10.1021/acs.analchem.5b00950 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

as glycosylated and deglycosylated forms, as depicted in Figure 3. In contrast to the EPO data, the difference between the

Figure 2. Similarity analysis and 1D 1H spectra of the native (blue) and unfolded (red) EPO samples from processes A and B. S(X,X) denotes the autosimilarity and S(X,Y) denotes the cross-similarity measure between samples A and B. The standard deviations correspond to the measurements of the three different lots of A and B.

Figure 3. Similarity analysis and 1D 1H spectra of the native and deglycosylated IgG1 (blue) and the corresponding unfolded forms (red, superscript U). S(X,X) denotes the autosimilarity for the native (N) and deglycosylated (D) samples while S(X,Y) denotes the crosssimilarity measure. Standard errors were obtained from measurements of the same protein solutions in five different NMR tubes.

at the differences of the 1D 1H NMR spectra of A and B clearly reveals differences in the characteristic resonances of the glycan residues. This is in agreement with other observations by more traditional HOS methods (CD, FT-IR, and DSC) where no significant differences in the rhEPO-A and rhEPO-B process samples were detected. Interestingly, differences in the distribution of glycan structures in these particular samples have been observed by specific glyco-profiling methods (unpublished data). Because EPO is an extremely soluble glycoprotein, we could also examine differences between samples A and B by 2D heteronuclear correlation methods using concentrated protein samples. The corresponding 2D 1H−13C and 1H−15N spectra and their similarity values are shown in Figure S5 (Supporting Information). In the case of the 1H−13C correlation spectra, the differences between A and B are less pronounced than those detected by the PROFILE method (SAA,BB = 28 ± 0.2 dB and SAB = 18 ± 0.2 dB (equivalent to a 1.4% spectral difference obtained by inverting eq 1) vs SAA,BB = 33 ± 4 dB and SAB = 16.2 ± 0.4 (equivalent to a 2.3% spectral difference)).The same, and even more pronounced effect, can be observed in the 1 H−15N correlation spectra, with similarities of SAA,BB = 17 ± 0.5 dB and SAB = 13 ± 0.5 dB (equivalent to a ∼2.8% spectral difference). In this case, the low autosimilarity and precision is compromised to some extent by the low signal-to-noise (∼10:1) of the experiment (Figure S4, Supporting Information). This is primarily due to the heterogeneity and dynamics of the attached glycans, resulting in broadening of the 1H−15N NMR signals and a concomitant decrease in sensitivity. Interestingly, the 2.3 ± 0.3% difference in similarity between the PROFILE spectra of samples A and B is in excellent agreement with the difference measured by specific glycan analysis (2.45% differences between samples A and B from the N-glycan composition). Advantages of this methodology in resolving structural contributions to the spectral fingerprint were tested in a similar manner for an IgG1 antibody in its native and unfolded as well

native glycosylated/deglycosylated proteins is 9.2 dB lower as compared to the unfolded, glycosylated/deglysosylated IgG1 samples (9.2 ± 0.1 dB vs 18.4 ± 0.6 dB). This can only be attributed to significant changes (∼10%) in HOS upon deglycosylation and is consistent with changes of similar magnitude observed in the corresponding 2D 1H−15N correlation spectra obtained using 15N labeled IgG1 (∼5.3 dB, Figure S6, Supporting Information). The variations in the canonical Fc glycans have profound effects on the biological function of the antibodies;11 however, the effect on HOS is not well characterized. Since the differences detected here purely represent changes in HOS due to the deglycosylation, it is conceivable to expect differences in HOS associated with the glycan heterogeneity, resulting in an ensemble of structures in the heterogeneous product. Notably, the PROFILE difference of 1.4 ± 0.4% for the unfolded glycosylated vs deglycosylated is close to the 2% of the glycan weight per molecule. Selectivity of PROFILE vs 2D Methods. In order to demonstrate how small changes in HOS of a protein population are reflected in the PROFILE and 2D 1H−15N correlation spectra, we used two highly homologous standard antibodies of types IgG1 and IgG2, which are known to differ in overall structure.18 Initially, we confirmed HOS differences using the PROFILE method (Figure 4) and found that the folded IgG1 and IgG2 differ by ∼68%. In contrast, the spectra of the unfolded proteins (Figure 4) differed by only 4%, which is consistent with 4% difference in the amino acid composition. When measured by 1H−15N correlation spectra, the two molecules differed by ∼42% in their native states. To test the selectivity of PROFILE, we prepared three blended samples of IgG1 with 0%, 5%, and 10% IgG2 and recorded data on two different instruments with proton resonance frequencies of 800 and 600 MHz. The statistical D

DOI: 10.1021/acs.analchem.5b00950 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

calculated selectivity (eq 2) of both PROFILE experiments in Figures 5 and 6 are virtually the same, ΦPROFILE ≅ 2, while the

Figure 4. 1D 1H spectra of the IgG1 and IgG2 proteins in their folded (N, blue) and unfolded (U, red) states. The PROFILE autosimilarity and cross-similarity analysis is shown in the bar graph. Standard deviations were obtained from the measurements of five replicated NMR samples.

Figure 6. ANOVA analysis of similarity for the 15N labeled IgG1 and 95% IgG1/5% IgG2 samples by 2D 1H−15N TROSY (A) and 1D PROFILE (B) recorded at 800 MHz. The horizontal bars designate 95% confidence limits for the mean similarity values (red circles). Each sample was replicated in five different NMR samples yielding 20 independent measurements of the autosimilarity and 25 independent measurements for each cross-similarity. The sample concentrations were 16 mg/mL, and the data was acquired for different texp with the different S/N values listed on the panel.

analysis of the PROFILE similarity scores is shown in Figure 5. All samples can be distinguished on both instruments with a

selectivity of the 2D experiment is significantly lower, Φ2D ≅ 0.5. It should be noted that Φexp was not optimized in this study. However, for the amounts of measured samples, they should be near the optimum values.



DISCUSSION AND CONCLUSIONS The value of 1D 1H NMR and its autocorrelation measure for the quantitative assessment of a protein fold in solution has been recognized before,19 yet its application seems to have been overshadowed by the universal practice of isotope labeling, where the initial assessment of an 1H−15N HSQC spectrum establishes the feasibility of 3D structure determination. Despite the lower selectivity, the 1H−13C and 1H−15N HSQC spectra show only a subset of protein resonances and, in this way, may serve to compromise the obvious advantages of multinuclear spectra for the purpose of structural fingerprinting. In the case of protein therapeutics, however, the PROFILE approach provides as much, or more, information about a protein fingerprint of HOS and microheterogeneity than twodimensional methods, while being acquired in a fraction of the experimental time required for the intact samples at natural isotope abundance and at their original formulation conditions. The structural information encoded in the NMR spectrum can be obscured by (i) insufficient signal-to-noise (S/N) and (ii) instrumental variability/imperfections in the preparation of the NMR sample. To gain insight into how S/N alone impacts similarity magnitude and precision in the case of different experiments, we performed similarity calculations for two pairs of Gaussian signals shown in Figure S4 (Supporting Information). In this simulation, we introduced different levels of random noise and calculated the similarity as a function of

Figure 5. ANOVA analysis the PROFILE similarity for the IgG1/IgG2 blended samples measured on 800 MHz (top) and 600 MHz (bottom) instruments. The horizontal bars designate 95% confidence limits for the mean similarity values (red circles). Each blend was replicated in five different NMR samples yielding 30 independent measurements of the autosimilarity and 25 independent measurements for each cross-similarity. The S/N for the 800 and 600 MHz spectra was 64 ± 1 and 76 ± 2, respectively. Sample concentrations were 50 mg/mL.

high degree of confidence. We performed the same analysis with 15N labeled IgG1 and IgG2 molecules (Figure 6). In this case, both pure and mixed samples can be distinguished, with precisions of ε ≅ 2 dB and ε ≅ 3 dB, nonetheless after 360 min vs 129 min for the 2D and 1D experiments, respectively. The E

DOI: 10.1021/acs.analchem.5b00950 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry

relaxation time in solution. It is conceivable that, through this process, a small population of IgG2 molecules resembles, in terms of the NMR signal, partially unfolded states of IgG1. Since recording 2D spectra of a small protein therapeutic has been demonstrated to be possible,8 Marino and co-workers proposed enzymatic cleavage of an intact antibody into Fc and Fab fragments and demonstrated feasibility of obtaining the 1 H−15N correlation spectra in an experimental time on the order of days or even minutes in case of the methyl 1H−13C spectra.10,20 It is important to note that in our previous work we used the IdeS protease (FabRICATOR, Genovis) for antibody digestion, which has been shown to have a single, specific cleavage site on human IgG1 and IgG2 molecules.9 In contrast, Arbogast et al. in their most recent study20 used papain to generate IgG1 fragments, which is a nonspecific sulfhydryl protease whose activity can be highly dependent on the degree of glycosylation of the mAb.21 In addition, when using papain, the digestion conditions need to be carefully controlled and may require significant method optimization for the individual IgGs.22 Obviously, enzymatic digestion eliminates the ability to measure HOS on an intact drug product. More importantly, as we explored this methodology in house, we found that the superposition of the spectra of the antibody fragments are significantly different from the spectra of the intact antibody. This is demonstrated in Figure S7, which shows the overlays of the full spectra from the intact antibody with the combined spectra from the Fc and F(ab′)2 fragments (yielding a “full” spectrum). The spectral similarity is only 5.2 dB (77%) and 0.52 dB (51%) for the IgG1 and IgG2 molecules, respectively. These dissimilarities appear to be more pronounced than what would be expected solely from T2 differences between the whole antibody and its fragments, especially in the present study, where the spectra were recorded with the TROSY pulse sequence which is not particularly sensitive to T2 changes. The more pronounced differences in the case of the IgG2 molecule are consistent with more complex network of disulfide bonds. Thus, it is legitimate to ask, how can an analytical method which compromises a protein’s HOS be a fingerprint of HOS? The protocol described in Figure 1 was used to compare different protein samples. The average magnitude and standard deviation for the NMR PROFILE measurements of the autosimilarities was 33 ± 3 dB (99.95 ± 0.05%) where the random spectral variation originated from variations in the NMR sample preparation and measurement. The choice of five samples for the statistical analysis was a compromise between experimental time and adequate sampling. Notably, all the significant spectral dissimilarities observed in this study could be explained in terms of the protein structural differences only because we knew the structural differences a priori. For unknown samples, the similarity cutoff must be first established; PROFILE differences suggest differences in HOS may be present, but to elucidate the origin of these differences additional NMR experiments and/or other analytical methodologies would be needed. In this study, PROFILE could clearly distinguish the difference between 5% IgG2 content, while the typical 1D 1H spectra corresponding to 5% and even 10% of the IgG2 content cannot be distinguished from the spectrum of the pure IgG1 by the naked eye. Thus, the robustness, quantitative output, and high selectivity of the PROFILE similarity measure, as demonstrated here, should prove useful in various applications of 1D 1H NMR to process development and quality control of protein therapeutics.

the S/N for a 1D PROFILE and 2D experiment. The impact of random instrumental variations and imperfections in NMR sample preparation on precision were simulated by artificially introducing small random zero-order phase variations into the spectra. This is shown in Figure S3, graph B (Supporting Information). It appears from this figure that the precision of the PROFILE experiment does not significantly improve with S/N greater than 65:1, which corresponds to 13 min experiment time for a 50 mg/mL antibody sample. Indeed, increasing this time to 104 min did not significantly improve the precision which remained close to the 3 dB threshold (Figure 5). Remarkably, the instrumental and sample variability have significantly less impact on similarity of the PROFILE spectra as compared to the similarity of the whole 1D spectra, which is demonstrated in Figure 7. As for the 2D spectra, we

Figure 7. Similarity calculations (A, upper panel) for the blended samples with the PROFILE method using algorithmically phased spectra (B, Δφ0 = 0) and the same spectra with random, zero-order phase variations of 0.4° (B, Δφ0 = 0.4°) and with the normal 1D spectra (A,B bottom panel). The ANOVA analysis is displayed as the horizontal bars designating 95% confidence intervals around the mean values displayed as the red circles. Each protein solution (designated by the different colors) was prepared in five different NMR tubes, yielding 30 autosimilarity and 2 × 25 cross-similarity values. Note that the phase scrambling destroys sample resolution (precision) only in the case of the normal 1D spectra.

anticipate the same limitation on the similarity with the same threshold value for S/N of 65:1. Thus, according to Figure S4 the precision of the 2D experiment can still be improved, however at the cost of significantly longer experiment times (tε), such that the 10-fold increase of experimental time should result in 6 dB precision. To achieve the borderline selectivity of Φ2D ≅ 0.5 for the experiment shown in Figure 6, but for the sample at natural isotope abundance (0.37% for the 15N isotope), one would need to extend tε up to an impractical 50 years! This time could be shortened by increasing the concentration of the sample measured; however, high concentrations of mAbs increase solution viscosity and lead to NMR line broadening and loss of the signal intensity. Here, at the 3 dB precision, the PROFILE method could detect changes in HOS of less than 5%. Obviously, this detection level is by no means general to the method. It is worthwhile to note that IgG2 molecules yield ∼40% weaker PROFILE NMR signal compared to IgG1 molecules, primarily due to the intramolecular dynamics that shorten the transverse F

DOI: 10.1021/acs.analchem.5b00950 Anal. Chem. XXXX, XXX, XXX−XXX

Article

Analytical Chemistry



(14) Arakawa, T.; Philo, J. S.; Kita, Y. Biosci. Biotechnol. Biochem. 2001, 65, 1321−1327. (15) Schleucher, J.; Schwendinger, M.; Sattler, M.; Schmidt, P.; Schedletzky, O.; Glaser, S. J.; Sørensen, O. W.; Griesinger, C. J. Biomol. NMR 1994, 4, 301−306. (16) Lescop, E.; Schanda, P.; Brutscher, B. J. Magn. Reson. 2007, 187, 163−169. (17) Jaynes, E. T. Probability Theory: The Logic of Science; Cambridge University Press: Cambridge, U.K., 2003. (18) Ryazantsev, S.; Tischenko, V.; Nguyen, C.; Abramov, V.; Zav’yalov, V. PLoS One 2013, 8, No. e64076. (19) Hoffmann, B.; Eichmüller, C.; Steinhauser, O.; Konrat, R. Methods Enzymol. 2005, 394, 142−175. (20) Arbogast, L. W.; Brinson, R. G.; Marino, J. P. Anal. Chem. 2015, 87, 3556−3561. (21) Raju, T. S.; Scallon, B. J. Biochem. Biophys. Res. Commun. 2006, 341, 797−803. (22) An, Y.; Zhang, Y.; Mueller, H.-M.; Shameem, M.; Chen, X. mAbs 2014, 6, 879−893.

ASSOCIATED CONTENT

S Supporting Information *

Principles of the correlation coefficient calculations and the relationship between the correlation coefficient and the similarity in dB units. Simulations of similarity differences as a function of experimental and processing parameters. 2D 1 H−13C and 1H−15N correlation spectra of the EPO samples. Overlay of 2D 1H−15N TROSY spectra for the native and deglycosylated IgG1. Overlays of 2D 1H−15N TROSY spectra for intact IgG1 and IgG2 with the corresponding spectra from F(ab)2 and Fc fragments. Table with the list of protein samples and corresponding experimental conditions used in this study. The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.analchem.5b00950.



AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Present Address ∥

P.D.S.: Department of Protein Chemistry, Genentech, Inc., South San Francisco, CA 94080. Notes

The authors declare no competing financial interest.



ACKNOWLEDGMENTS We would like to thank Gino Grampp, John Gabrielson, and Izydor Apostol for valuable comments and suggestions to the manuscript. We would also like to thank Peter Grandsard and Philip Tagari who supported this work.



REFERENCES

(1) International Conference on Harmonisation (IHC) Expert Working Group Comparability of Biotechnological/Biological Products Subject to Changes in Their Manufacturing Process; 2004; http://www. ich.org/fileadmin/Public_Web_Site/ICH_Products/Guidelines/ Quality/Q5E/Step4/Q5E_Guideline.pdf. (2) Finkelstein, A. V.; Ptitsyn, O. Protein physics: A course of lectures; Academic Press: New York, 2002. (3) Berkowitz, S. A.; Engen, J. R.; Mazzeo, J. R.; Jones, G. B. Nat. Rev. Drug Discovery 2012, 11, 527−540. (4) Alsenaidy, M. A.; Jain, N. K.; Kim, J. H.; Middaugh, C. R.; Volkin, D. B. Front. Pharmacol. 2014, 5. (5) Sörgel, F.; Lerch, H.; Lauber, T. BioDrugs 2010, 24, 347−357. (6) Cavanagh, J.; Fairbrother, W. J.; Palmer, A. G., III; Rance, M. Skelton, N. J. Protein NMR Spectroscopy: Principles and Practice, Second ed.; Elsevier, Inc.: Amsterdam, 2007. (7) Chen, K.; Freedberg, D. I.; Keire, D. A. J. Magn. Reson. 2015, 251, 65−70. (8) Aubin, Y.; Gingras, G.; Sauvé, S. Anal. Chem. 2008, 80, 2623− 2627. (9) Poppe, L.; Jordan, J. B.; Lawson, K.; Jerums, M.; Apostol, I.; Schnier, P. D. Anal. Chem. 2013, 85, 9623−9629. (10) Marino, J. P.; Brinson, R. G.; Hudgens, J. W.; Ladner, J. E.; Gallagher, D. T.; Gallagher, E. S.; Arbogast, L. W.; Huang, R.Y-C. In Monoclonal Antibody Therapeutics: Structure, Function, and Regulatory Space; Schiel, J. E., Davis, D. L., Borisov, O. V., Eds.; American Chemical Society: Washington, D.C., 2014. (11) Clynes, R. A.; Towers, T. L.; Presta, L. G.; Ravetch, J. V. Nat. Med. 2000, 6, 443−446. (12) Carr, S. A.; Huddleston, M. J.; Bean, M. F. Protein Sci. 1993, 2, 183−196. (13) Cotts, R. M.; Hoch, M. J. R.; Sun, T.; Markert, J. T. J. Magn. Reson. (1969-1992) 1989, 83, 252−266. G

DOI: 10.1021/acs.analchem.5b00950 Anal. Chem. XXXX, XXX, XXX−XXX