Two-Trace Two-Dimensional Correlation Analysis for the Confirmation of Therapeutic Protein Identification and Evaluation of Extent of Glycosylation: A Rapid Method for Testing Critical Quality Attributes

Abstract

Therapeutic proteins are an effective and highly selective, yet complex class of drugs for the treatment of diseases to improve the quality of life of patients. Their manufacturability is subject to evaluation to ensure product quality, safety, and efficacy. To this end, a need exists within the biopharmaceutical industry for a streamlined platform method that allows end-to-end therapeutic protein comparability assessment of multiple critical quality attributes. The implementation of a single platform method would de-risk development and accelerate speed to market. Herein, a novel method for therapeutic protein identification has been proposed that can be used to monitor regulated drug products throughout their lifecycle. The innovative streamlined method evaluates the intact protein in its formulation conditions. This breakthrough technology comprises a quantum cascade laser microscope with a heated accessory, an innovative slide cell array with a fixed path-length that allows for experimental flexibility and quantitative analysis, and dedicated software. Two-trace two-dimensional correlation spectroscopy (2T2D) was proven useful for the determination of spectral differences between two samples. The weighted absorbance differences, defined using the correlation between the cross-peaks, can be used to determine substantial molecular information, including secondary structure, the extent of glycosylation, amino acid content, and solvent accessibility, which is directly related to the therapeutic protein’s critical quality attributes. Four different comparability assessments are discussed to demonstrate the usefulness of the platform method for the confirmation of protein identification and the comprehensive evaluation of the extent of glycosylation.

Graphical abstract

This is a visual representation of the abstract.

Introduction

Therapeutic proteins have proven to be highly effective for the treatment of multiple diseases such as autoimmune and metabolic disorders, neurodegenerative diseases, and cancer, among others. These therapeutic proteins are complex in design, and yet they are a dominant class of drugs with a pipeline of ∼1400 clinical candidates for regulatory review.^1–5 These proteins also require a complex drug development process, and as such are subject to the evaluation of a series of critical quality attributes to ensure quality, safety and efficacy of the drug product for patients worldwide.^3,6–21

Protein identification is one such critical quality attribute, yet efficient tools to comprehensively address this attribute to enable fast and accurate characterization are lacking. Traditionally, protein identity was considered to involve determining the amino acid sequence and the glycan profile separately. Each component of the analysis is time-consuming, requiring extensive sample preparation protocols for the protein and separately for the glycan, thereby exponentially increasing the complexity and the number of samples needed to be analyzed, and also potentially increasing the variability of the results obtained. The biomolecule fragments are purified by high performance liquid chromatography, at times followed by multiple mass spectrometry analyses. Although well established, these methods cannot address other key attributes that can affect the therapeutic protein and contribute to the comprehensive determination of its identity, such as weak interactions, secondary structure, solvent accessibility, and stability. Furthermore, the glycan profile can be used to predict loss of stability and loss of function that can in turn lead to an unwanted immune response.^3,4,7,16,18 All of these critical quality attributes can adversely impact safety and efficacy.

Here, we demonstrate the use of a platform method that combines quantum cascade laser (QCL) microspectroscopy, an innovative slide cell array, and dedicated software that employs a two-trace two-dimensional (2T2D) correlation algorithm to address the challenge of protein identification using a fast screening comparability assessment for accurate determination of therapeutic protein quality, to ensure safety and efficacy.^21–24 2T2D correlation spectroscopy was first described by Dr. Isao Noda,^25–27 and the algorithm has been instrumental in the comparative analysis of two therapeutic protein spectra, providing enhanced resolution while enabling the correlation between vibrational modes of the same protein to establish the molecular differences between pairs of proteins. For the purposes of this study, four samples of different proteins in the same buffer were evaluated within a single slide cell array. As a result, each sample received a comprehensive evaluation, and each pairwise comparison provided understanding of the molecular differences. The evaluation included the amino acid composition to identify variants (amino acid substitutions for 11 of the 20 amino acid types), differences in solvent accessible surface area (SASA), differences in secondary structure, presence of aggregates, and differences in the glycan profile. The significance of this method is that, to our knowledge, it is the first comprehensive fast screening method for end-to-end comparability assessment of therapeutic proteins with the potential to be implemented throughout the drug’s lifecycle, addressing an unmet need in biopharma.

A wide-field QCL transmission microscope provides significant advantages in speed during hyperspectral image acquisition with enhanced signal-to-noise ratio, making it ideal for the study of biotherapeutics in their formulation conditions. The hyperspectral images are acquired in real time, a capability that has remained elusive to traditional Fourier transform infrared (FT-IR) microscopes. More importantly, the comprehensive assessment of an array of intact therapeutic proteins including, but not limited to, the evaluation of weak interactions, conformational changes, aggregation or self-association, and post-translational modifications, provides key critical quality attribute information for comparative analysis vital for therapeutic protein candidate selection to ensure their quality, safety, and efficacy. The spectral changes are highly sensitive and selective and can be used to provide a comprehensive molecular understanding of the protein even under high-concentration/high-dosing conditions. Other multi-attribute methods use liquid chromatography–mass spectrometry (LC-MS), which requires extensive protein sample preparation including enzymatic digestion, generating hundreds of peptide fragments for subsequent analysis. This increases the workload exponentially, extending the time to results and increasing the costs associated with the analysis of just one protein. More importantly, the fact that hundreds of peptides are derived from each protein sample increases the variability of the data obtained due simply to the extensive sample manipulation required. Furthermore, the multiple LC-MS methods available provide minimum understanding of protein dynamics and stability. In comparison, QCL microspectroscopy (ProteinMentor) provides for the real-time monitoring of an array of clinical antibodies with just 1 μL fixed volume per sample under accurate thermal control (at 24.0 °C). The platform has been quality-by-designed for the biopharma industry to ensure a true multi-attribute method that delivers on the acceleration of speed to market, while ensuring accuracy, reproducibility, and statistical robustness of the results generated for the array of therapeutic proteins analyzed.

Experimental

Materials and Methods

Four different pre-clinical monoclonal antibody (mAb) candidates were generously supplied by Dr. Eduardo Canto. These mAbs contained the same IgG scaffold with varying amino acid sequence within the variable heavy (VH) chain and variable light (VL) chain regions of the antigen binding fragments. These pre-clinical mAb samples were buffer exchanged to PBS at pH 7.4 verified purity by SDS-PAGE and subject to multiple kinetic binding assays by the Charles River Labs. Because the candidates are proprietary, they are referred here generically as the reference mAb, mAb1, mAb2, and mAb3. The amino acid composition for the Fabs has been included in Tables S1–S2 in the Supplemental Material. Pairwise comparison of the amino acid sequences was performed to confirm the results of the different comparability assessments discussed herein, given that amino acid sequencing is the gold standard for protein identification. The reference mAb had 81.5% (VH) and 88.9% (VL) sequence identity with mAb1; mAb1 had 83.3% (VH) and 89.2% (VL) sequence identity with mAb2; and mAb2 had 55.8% (VH) and 96.5% (VL) identity with mAb3. All samples were formulated in phosphate-buffered saline at pH 7.4, and concentration was determined by optical density at λ_max at 280 nm. No further sample preparation was performed.

Quantum Cascade Laser Microspectroscopy

A 1 μL aliquot of each pre-clinical antibody sample was loaded onto a pre-defined well within the innovative CaF₂ slide cell array and placed in the heated slide cell accessory at 24 ± 0.3 °C. The slide cell array is comprised of two CaF₂ components: (i) a bottom slide containing an array of wells and (ii) a top lid with both faces having an optically polished surface providing a fixed path-length of 7.94 ± 0.71 μm . Real-time hyperspectral images of each intact mAb in its formulation or buffer were acquired under thermal equilibrium. The resulting spectral data for the pre-clinical antibodies were within the spectral region of 1775–1015 cm⁻¹ at 4 cm⁻¹ spectral resolution using a QCL infrared microscope. The QCL transmission microscope model ProteinMentor from Protein Dynamic Solutions, Inc. (San Juan, Puerto Rico), was comprised of four broadly tunable external cavity quantum cascade lasers, a 4× low-magnification objective with a numerical aperture of 0.3, a 2 × 2 mm² field of view providing a pixel size of 4.25 × 4.25 μm², a linear response focal plane array (480 × 480 pixels), resulting in a field of view comprised of 223 000 spectra and as such representing the fastest infrared microscope commercially available to our knowledge, a heated slide cell accessory, and dedicated software for hyperspectral image acquisition and spectral data analysis of biotherapeutics. This platform provides significant advantages: (i) true comparability assessment of multiple critical quality attributes of an array of therapeutic proteins in their formulation, (ii) a series of real-time hyperspectral images of an array of samples, (iii) enhanced signal-to-noise ratio within the entire spectral region of interest (1775–1015 cm⁻¹), which includes both the amide I and II bands and the glycosylation region, and (iv) ability to determine differences in the extent of glycosylation, secondary structure, amino acid composition, and SASA by taking advantage of key vibrational modes that serve to inform protein characterization in all of its dimensions. ²⁸

Hyperspectral Image Acquisition and Spectral Data Analysis

The hyperspectral images were acquired using AutoPDS, the subsequent spectral data were generated using DataPDS and analyzed using Correlation Dynamic or IdCQ software suite from Protein Dynamic Solutions, Inc. (San Juan, Puerto Rico). The raw spectral data was subject to linear 2-point baseline correction at 1775 and 1485 cm⁻¹ for the spectral region that was comprised of the amide I and II bands, while the full spectral region was subject to a spline baseline correction at 1775, 1485, 1367, 1136, and 1015 cm⁻¹. No other pre-processing spectral manipulations were employed, thus ensuring the quality of the spectral data for the subsequent 2T2D correlation spectroscopy analysis which included the weighted spectral subtraction between selected samples within the array. Spectral baseline correction, covariance estimation, and weighted spectral calculations were performed using Correlation Dynamics or IdCQ software (Protein Dynamic Solutions; San Juan, Puerto Rico), followed by generation of 2T2D asynchronous plots. Excel in Microsoft Office 365 (Microsoft, USA) was employed to generate spectral overlays and bar graphs. Details of the 2T2D correlation algorithm initially described by Isao Noda, as well as the cross-peak coordinates and intensity differences for each comparative analysis, are summarized in the Supplemental Material.^25–27

Results

This novel comparative approach was applied to four pre-clinical antibodies against a different sample or a reference standard. A subset of the data generated is presented herein to demonstrate the value of the approach. In addition, triplicate results were generated for the mAb comparability assessments that exhibited differences in SASA, secondary structure, amino acid composition and glycan profile have been included as Supplemental Material and are cited throughout.

Spectral Overlay and Weighted Difference Spectra

The hyperspectral images were comprised of 223,000 QCL spectra at 24.0 ± 0.3 °C, enabling fast turnaround time for comparative analysis of multiple mAb attributes. The quality of the baseline-corrected spectral data is clearly demonstrated, yet very little difference can be discerned within the 1775–1015 cm⁻¹ spectral region (Figure 1). Especially for the case of the reference mAb against itself, the overlaid spectra are highly similar within the entire spectral region, suggesting identity (Figure 1a). Slight differences were observed for the mAb2 against mAb1 spectral overlay in the amide II band within the spectral region of 1600–1500 cm⁻¹ and the glycan region (1176–1015 cm⁻¹), suggesting differences in amino acid composition and glycosylation, respectively, as shown in Figure 1b. Greater spectral differences were observed for mAb2 versus mAb3, including differences within the amide I and amide II bands, suggesting both the secondary structure and amino acid composition were different (Figure 1c). This initial overview also considers the comparative evaluation of the reference mAb against mAb1, where differences in absorbance were observed for the amide II band and the glycan spectral region (Figure 1d), suggesting differences in both amino acid composition and the extent of glycosylation. Weighted difference spectra were generated for each of these comparisons within the spectral region of 1775–1485 cm⁻¹ (Figure S1, Supplemental Material). For the reference mAb against itself (our negative control), a single line with a slope equal to zero was observed, confirming identity. For the other samples, the weighted spectra show definite differences, but they are difficult to interpret in the spectral overlay analysis. These spectral differences warranted a more in-depth analysis to provide a comprehensive understanding of the features of the engineered mAbs that impact their stability and function.

Figure 1.

Typical overlaid baseline-corrected QCL infrared microspectroscopy spectra within the spectral region of 1775–1015 cm⁻¹ at 24 °C. (a) Reference mAb (black solid line) against itself (grey, dashed line), (b) mAb2 (olive line) against mAb1 (blue line), (c) mAb2 (olive line) against mAb3 (orange line), and (d) reference mAb (black line) against mAb1 (blue line). The spectral region includes the amide I and II bands (1775–1485 cm⁻¹) along with the glycan region (1176–1015 cm⁻¹) to allow for screening of protein identification. The comparison yielded no spectral difference when comparing the reference with itself, while differences were observed for all subsequent spectral comparisons.

We thus turned to 2T2D correlation analysis which can provide both a quantitative and qualitative assessment of the two weighted difference spectra to ascertain the molecular differences between the mAbs evaluated in each sample. 2T2D analysis provides an in-depth molecular evaluation for each sample within two separate spectral regions: the amide I and II bands (1775–1485 cm⁻¹) and the glycan region (1176–1015 cm⁻¹). Within the amide I and II bands, we explored three distinct features: (i) SASA, for which the aspartate and glutamate residues serve as internal probes (1760–1700 cm⁻¹), (ii) secondary structure (1700–1640 cm⁻¹), and (iii) amino acid content (1660–1490 cm⁻¹).^{21–24,28–32}

Solvent-Accessible Surface Area

Figure 2 shows the amide region for each of the sample sets in the spectral overlay and 2T2D asynchronous plots. Panels d, f, and h illustrate the characteristic cross-peak pattern of the 2T2D asynchronous plots within the spectral region of 1775–1485 cm⁻¹ that serves to establish the molecular differences in solvent accessibility, secondary structure, and amino acid composition. Surface exposed glutamate and aspartate side chains that interact with the immediate shell of water molecules via hydrogen bonding serve as surface probes for the empirical determination of aqueous solvent accessibility for the different mAbs. The weighted absorbance differences and assignments Figure 2 corresponding to solvent accessibility are summarized in Figure S2 and Tables S3–S6 and S7–S10 (Supplemental Material).^29,30 We used bar graphs to quantify the molecular differences determined in the 2T2D correlation analysis (Figure 3), assuming the absorptivities for these Asp/Glu side chain hydrogen bonded vibrational modes (1750–1705 cm⁻¹) are the same. Here and throughout, all of the positive cross-peaks belong to the minuend (first mAb stated in the comparability analysis) and the negative cross-peaks belong to the subtrahend (second mAb stated in the comparability analysis). Also, the color scheme used for all the bar graphs is the following: mAb2 against mAb1 (green bar), mAb2 against mAb3 (orange bar), and reference mAb against mAb1 (blue bar). All cross-peak assignments have been summarized in Table I.

Figure 2.

Typical overlaid QCL infrared microspectroscopy spectra and 2T2D correlation asynchronous plots for the amide I and II bands within the spectral region 1775–1485 cm⁻¹ at 24 °C. (a, c, e, g) Linear baseline corrected spectra (line color scheme is the same as per Figure 1) and (b, d, f, h) asynchronous plots. (a, b) Reference mAb against itself, (c, d) mAb2 against mAb1, (e, f) mAb2 against mAb3 and (g, h) reference mAb against mAb1.

Figure 3.

SASA associated bar graphs summarize the differences in surface exposed Asp/Glu residues within the mAbs β-turns (1692 and 1688 cm⁻¹) and random coil (1643 cm⁻¹) regions. Weighted absorbance differences in the cross-peaks of the 2T2D asynchronous plots were associated with Asp/Glu ν(C=O) in the spectral region 1760–1700cm⁻¹. (a) 1-Hydrogen bonded and (b, c) 2-hydrogen bonded. Color scheme: mAb2 versus mAb1 (green bar), mAb2 versus mAb3 (orange bar), and reference mAb versus mAb1 (blue bar). Assuming the absorptivity is the same within this spectral region (1760–1700 cm⁻¹) and that these surface exposed residues serve as effective probes for the empirical determination of SASA, then absorbance differences suggest that reference mAb has the highest SASA > mAb3 > mAb1 > mAb2. Further information is provided for the average of triplicate results and standard deviation determinations in the Supplemental Material (Figure S2; Tables S3–S6 to S7–10).

Table I.

Summary of cross-peak assignments for all 2T2D asynchronous plots analyzed.

Cross-peak(cm−1)	Assignment
SASA
1760–1730	Glu/Asp ν(C = O) non-hydrogen bonding
1728–1720	Glu/Asp ν(C = O) 1-hydrogen bonding
1718–1704	Glu/Asp ν(C = O) 2-hydrogen bonding
Secondary structure
1695, 1688, 1684, 1680	β-turn
1663	hinge loop
1653	π-helix
1643	random coil (RC)
1636	β-sheet
1620	β-strand
Side chains
1672	Gln ν(C = O)
1612–1611	Asn δ(NH2)
1608–1596	His ν(C = C)
1584–1581	Gln δ(NH2)
1572–1567	Asp ν(COO–)
1550–1534	Glu ν(COO–)
1518	Tyr ν(C = C)
1505	Trp ν(C = C)
Glycan
1176	C–O–C glycosidic linkage
1155–1148	ν(C–O) glycosidic linkage
1110–1100	ν(C–O), ring from C4–O, C6–O
1090	ν(C–O), ν(C–C), δ(COH) from C1–H
1058	ν(C–O), ν(C–C), δ(COH) from C1–H
1047	ν(C–O), ν(C–C), δ(COH) from C1–H
1040	ν(C–O), ν(C–C), δ(COH) from C1–H (galactose)

The cumulative results suggest that the Asp/Glu 1-hydrogen bonding (1724 cm⁻¹) located within random coil (1643 cm⁻¹) and β-turns (1692 cm⁻¹), that is, the SASA for the 1-hydrogen bonded state residues, was greatest in mAb3, followed by mAb1 and finally mAb2 (Figure 3a; Figure S2 and Table S3–S6, Supplemental Material). However, the reference mAb and mAb1 did not exhibit differences in SASA within the random coil (1643 cm⁻¹) or the β-turns (1692 cm⁻¹). Also, amongst the aspartate and glutamate side chain modes that exhibited 2-hydrogen bonding with their aqueous environment, the reference mAb was ranked as having the highest SASA, followed by mAb3 > mAb1 > mAb2 (Figure 3b,c; Figure S2, Tables S7–S10, Supplemental Material). These Asp/Glu side chains 2-hydrogen bonded with their aqueous environment were located within the random coil (1643 cm⁻¹), β-turns (1695 and 1688 cm⁻¹) and β-sheet (1636 cm⁻¹). Also, for mAbs1–3, the β-sheet (1636 cm⁻¹) regions exhibited comparable SASA for the Asp/Glu side chain residues that were 2-hydrogen bonded (1720–1700 cm⁻¹). More importantly, amongst the surface exposed aspartate and glutamate side chains that were 2-hydrogen bonded, it was the glutamates that exhibited greater solvent accessibility. Finally, we need to consider the possibility that some cross-peak coordinates (1724, 1550) and (1716, 1545) may also be due to sialic acid capped N-glycan, which is discussed further below.

Secondary Structure Differences

In general, the amide I band is primarily comprised of the peptide bond carbonyl stretching vibrational modes appearing in the spectral region of 1700–1600 cm⁻¹. These are used to determine secondary structure composition and are sensitive to hydrogen bonding interactions, causing the observed broad band of almost 100 cm⁻¹.^21,31 However, some side chain modes may be assigned within the amide I band as well, and these vibrational modes have been reviewed and validated.

The four comparative evaluations are once again shown in the overlaid baseline corrected spectra (Figures 2a, c, e, and g) and 2T2D asynchronous plots (Figures 2b, d, f, and h) within the spectral region of 1775–1485 cm⁻¹. The cross-peak absorbance differences and their assignments are summarized in Figure S2 and Tables S11–S14, and quantitative information associated with the analysis of the asynchronous plots is presented as a bar graph in Figure 4a. Yet, the method is highly valuable in that it also serves to establish some secondary structure content as well. The evidence for such is observed for the reference mAb against itself, where no spectral or 2T2D asynchronous plot cross-peaks could be discerned (Figures 2a, b). The 2T2D asynchronous plot yielded a featureless plot with an absorbance difference sensitivity of 1 × 10⁻¹⁵, as per the color bar scale, suggesting the secondary structure was the same. Interpretations of these data serve to establish the differences in secondary structure content between the different mAbs.

Figure 4.

Bar graphs summarize the 2T2D asynchronous cross-peak absorbance differences associated with secondary structure and amino acid composition. (a) Secondary structure in relation to side chain interaction and (b) differences in the amino acid composition. Color scheme: mAb2 against mAb1 (green bar), mAb2 against mAb3 (orange bar) and reference mAb against mAb1 (blue bar). Further information is provided for the average of triplicate results and standard deviation determinations in the Supplemental Material (Figure S2; Tables S11–S14 to S15–S18).

The greatest difference was observed for the reference mAb when compared to mAb1, where the reference mAb had greater β-turn content as opposed to random coil (cross-peak coordinates (1695, 1643) and (1684, 1643) with 15.98 ± 6.63 and 14.10 ± 4.93 milli absorbance unit (mAU) differences, respectively) or the anti-parallel β-sheet (1695, 1636) with 15.90 ± 6.12 mAU difference for triplicate data sets. Yet, for the two types of β-turns (1684, 1680 cm⁻¹) the content was about the same for both mAbs.

For the mAb2 comparability assessments versus mAb1 and mAb3, mAb2 had lower β-turn content as opposed to random coil (cross-peak coordinates (1695, 1643) with –8.05 ± 2.09 and –13.20 ± 1.05 mAU differences, respectively; and for (1684, 1643) –5.39 ± 2.36 mAU and –10.6 ± 2.23 mAU, respectively; the negative sign denotes the subtrahend) and greater anti-parallel β-sheet, crosspeak coordinate (1695, 1636) with differences of –5.79 ± 2.06 mAU and –12.20 ± 1.11 mAU, respectively. However, no differences were observed between mAb2 and mAb3 for the β-turn (1684, 1680 cm⁻¹) vibrational modes. Furthermore, the loop to π-helix (cross-peak coordinate (1660, 1653)) content was greater for mAb1 and mAb3 when compared to mAb2. Meanwhile, for the π-helix to random coil relationship (cross-peak coordinate (1653, 1643)), mAb1 and mAb3 compared to mAb2 for this specific relationship exhibited highly dispersed data. However, mAb2 had slightly higher π-helix to β-sheet (cross-peak coordinate (1653, 1636)) content when compared to mAb1.

For the comparison of reference mAb against mAb1, with regard to π-helix against the other structural components (cross-peak coordinates (1660, 1653), (1653, 1636), and (1653, 1643)) it was the reference mAb that had the higher content. However, for the random coil and β-sheet relation (cross-peak coordinate (1643, 1636)), the reference mAb and mAb1 had similar content (Figure 4a; Figure S2 and Table S11–S14, Supplemental Material). When considering the β structures overall, the highest secondary structure content can be ranked as: reference mAb > mAb3 > mAb1 > mAb2. These results demonstrate that secondary structure comparability assessment is an important element for establishing protein identification and that clear secondary structure differences can be established with this screening approach.

Amino Acid Composition Differences

The amide I and amide II bands within the spectral region of 1700–1500 cm⁻¹ are comprised of backbone associated vibrational modes and multiple side chain vibrational modes that serve to establish differences in amino acid composition.^21,31,32 In general, the greatest difference in amino acid composition was observed for the reference mAb when compared to mAb1 (Figure 4b; Figure S2 and Tables S15–S18), specifically for the glutamates located within the loops (cross-peak coordinate (1660, 1545)) and random coil regions (cross-peak coordinate (1643, 1545)). We must also consider the possibility of these cross-peaks also representing the N-acetyl group (1550 cm⁻¹) of sialic acid (see Extent of Glycosylation section below) due to the magnitude of the absorbance difference observed when compared to the other amino acid components. Meanwhile, mAb2 had less glutamine (cross-peak coordinate (1670, 1643)) and asparagine (cross-peak coordinate (1678, 1643)) content within the random coils compared to mAb1 and mAb3. For the case of the reference mAb against mAb1, the reference mAb was observed to have higher glutamine and asparagine content within the random coil regions. Finally, the aspartate-to-glutamate variant (cross-peak coordinate (1572, 1545)) was observed for all the cases studied. However, the reference mAb against mAb1 case had a greater absorbance difference, which may arise from differences in stability due to the location within the sequence, and in this case the number of substitutions would need to be evaluated further. Similarly, the absorbance difference for the aspartate-to-tyrosine variant (cross-peak coordinate (1572, 1518)) was greater for the reference mAb against mAb1 case when compared to the other samples studied. Finally, the sensitivity of the evaluation has been demonstrated in this study to be able to detect the substitution of a single amino acid based on the pairwise sequence alignment (data not shown). This result also demonstrates the selectivity of the method that allows for the direct monitoring of 11 of the 20 types of amino acids with known absorptivity.^31,32

Extent of Glycosylation

The post-translational modification of glycosylation may account for differences in stability, an unwanted immune response in patients, and/or loss of efficacy. Cell culture and media conditions can lead to variations in the extent of mAb glycosylation. To our knowledge this is the first fast screening comparability assessment for the evaluation of the extent of glycosylation for an array of therapeutic proteins (mAbs) against a reference standard. The method provides the absorbance differences in the glycan region of the spectra (1176–1015 cm⁻¹) as shown in Figure 5 for the 2T2D asynchronous plots for the full spectral region of 1775–1015 cm⁻¹ (panels a, c, e, and g) and for the spectral overlay within the spectral region of 1136–1015 cm⁻¹ (panels b, d, f, and h). The 2T2D asynchronous plot cross-peak information is summarized in Figure S3 and Tables S19–S22 to S31–34 (Supplemental Material), and differences are quantified in Figure 6. Assignments of the infrared carbohydrate vibrational modes were based on the published work of Kačuráková and Mathlouthi. ³³ More importantly, the comparison of the extent of glycosylation is direct, providing both qualitative and quantitative results that allow for the assessment of the glycan differences, including sialic acid-capped N-glycan, for the mAbs studied.

Figure 5.

Extent of glycosylation examined using representative 2T2D correlation and zoomed in overlaid QCL infrared microspectroscopy spectra of the glycosylation region at 24 °C. (a, c, e, g) 2T2D asynchronous plots within the full spectral region 1775–1015 cm⁻¹ and (b, d, f, h) linear baseline corrected spectra within the spectral region of 1136–1015 cm⁻¹ (line color scheme is the same as per Figure 1). (a, b) Reference mAb against itself, (c, d) mAb2 against mAb1, (e, f) mAb2 against mAb3 and (g, h) reference mAb against mAb1. The absence of cross-peaks in the asynchronous plot in (a) suggests identical extent of glycosylation for the reference mAb against itself.

Figure 6.

Bar graph summarizing the 2T2D asynchronous cross-peak absorbance differences associated with differences in the extent of glycosylation. Color scheme: mAb2 against mAb1 (green bar) and reference mAb against mAb1 (blue bar). More importantly, mAb2 and mAb3 were determined to have similar glycan profiles (orange bars). Further information is provided for the average of triplicate results and standard deviation determinations in the Supplemental Material (Figure S3, and associated Tables S19–S22 to S31–S34).

The 2T2D asynchronous plots for the full spectral region show the correlations between the amide I and II bands and the glycan spectral regions for the different samples (Figures 5a, c, e, and g). The reference mAb has a confirmed identity based on the highly sensitive and selective 2T2D asynchronous plot (Figure 5a) in which no cross-peaks are observed for the full spectral region (1775–1015 cm⁻¹). Also, the overlaid spectra for the glycan region (1136–1015 cm⁻¹) show complete overlap (Figure 5b), suggesting the same glycan profile.

For the comparison of mAb2 against mAb1, strong correlations of the glycan modes with the structural and side chain (amide I and II bands) associated modes are evident (Figure 5c). The overlaid spectra for the glycan region show a greater extent of glycosylation observed for mAb1 when compared to mAb2 (Figure 5d). Further evaluation of these absorbance differences was done by using the 2T2D asynchronous correlation analysis, and the triplicate results of the weighted absorbance differences along with the cross-peak coordinates obtained from the asynchronous plot are presented in Figure 6, and in Figure S3 and Tables S19–S22 of the Supplemental Material. The C–O–C glycosidic linkages (1176 cm⁻¹) are correlated with the β-turn (1680 cm⁻¹), β-sheet (1636 cm⁻¹), π-helix (1653 cm⁻¹) and asparagine (1612 cm⁻¹), all assigned to mAb2 when compared against mAb1. Also, the β-sheet (1636 cm⁻¹) has the potential to be an antiparallel sheet due to the presence of the β-turn (1680 cm⁻¹), thus suggesting the C–O–C glycosidic linkage is observed in an antiparallel β-sheet and a π-helix (1653 cm⁻¹) within mAb2, but when one considers the standard deviation for this specific comparison the C–O–C glycosidic linkage contributions (1176 cm⁻¹) they are highly dispersed. Different interactions between the backbone vibrational modes and the glycan ν(C=O) glycosidic linkage (1149 cm⁻¹), potentially via hydrogen bonding with β-turn (1680 cm⁻¹), and random coil (1643 cm⁻¹), were observed for mAb1 when compared with mAb2 (Tables S19–S22, Supplemental Material). These cross-peaks suggest a correlation exists between the different secondary structures and the ν(C–O) glycosidic linkage for glucose or galactose (1149 cm⁻¹) and fructose (1154 cm⁻¹). The cross-peak coordinates involving ν(C–O) and the ring from C₄–O, C₆–O for glucose (1100 or 1108 cm⁻¹) indicate correlations with the peptide backbone, for which greater content was observed for mAb1 when compared to mAb2 (Supplemental Tables S23–S26). Also, a correlation between surface exposed Asp/Glu ν(C = O) 2-hydrogen bonded with ν(C–O), ring from C₄–O, C₆–O for glucose (cross-peak coordinate (1710, 1100)) was observed only for mAb1. In addition, a significant difference was observed for ν(C–O), ν(C–C), and δ(COH) from C₁–H for galactose with lower hydration level for the sialic acid end cap (1047 cm⁻¹) in greater content for mAb1 when compared to mAb2. Also, once again solvent accessible Asp/Glu residues 2- and 1-hydrogen bonded potentially with the δ(COH) of the galactose (Figure 6; Figure S3 and Tables S27–S30 to S31–S34, Supplemental Material).

For the comparison of mAb2 against mAb3, the 2T2D asynchronous plot (Figure 5e) does not have any cross-peaks with absorbance difference above the cut-off range of detection (Figure S3 and Tables S19–S22 to S31–S34, Supplemental Material), suggesting the glycosylation profile is very similar. This is also evident in the spectral overlay within the glycan spectral region (Figure 5f). The cumulative analysis suggests the glycan profiles for both mAb2 and mAb3 are the same.

For the comparison of the reference mAb against mAb1, the 2T2D asynchronous plot does have cross-peaks within the glycan spectral region that are correlated to the amide I band (Figure 5g). The spectral overlay within the glycan spectral region (1136–1015 cm⁻¹) also shows differences in both spectral contribution and features (Figure 5h), suggesting differences in the glycan profile of these two mAbs. A significant difference for ν(C–O), ν(C–C), and δ(COH) from C₁–H was observed for galactose (1058 and 1040 cm⁻¹) in greater content for mAb1 when compared to the reference mAb. Further analysis of the cross-peak absorbance differences is shown in Figure S3, and in Tables S19–S22 to S31–S34 (Supplemental Material) and summarized in Figure 6. Interestingly, the absorbance difference for mAb1 within the β-turn (1680 cm⁻¹) and random coil region (1643 cm⁻¹) was ν(C–O), ring from C₄–O, C₆–O (1100 cm⁻¹); the asparagine 1612 cm⁻¹) was a n(C–O), ring from C₄–O, C₆–O (1108 cm⁻¹) and also within the random coil region (1643 cm⁻¹) was the n(C–O), ν(C–C), δ(COH) from C₁–H (1090 cm⁻¹) (Figure S3 and Tables S23–S26, Supplemental Material; Figure 6); no other mAb within the samples contained this type of glycan profile.

All of the pre-clinical mAb samples studied were in buffered phosphate saline solution at pH 7.4 and as such the buffer associated phosphate stretching modes ν _as (PO₂⁻), ν _as (PO₃) and ν _s (PO₂⁻) at 1158, 1080, and 1077 cm⁻¹, respectively and deformation mode δ(PO–H) at 1250 cm⁻¹;^34,35 these buffer associated modes do not overlap with the glycan vibrational modes identified above and are therefore valid for the comparability assessment of the glycan profiles Figure S4.

Discussion

Therapeutic proteins are complex due to their architecture, post-translational modifications, and the weak interactions that govern structure, stability, and function. Their evaluation needs to consider product-related impurities such as aggregates, differences in the extent of glycosylation, and the presence of degradation products such as Asn/Gln deamidation and oxidation. A need exists for a method capable of directly addressing this level of complexity in all of its dimensions. Although amino acid sequencing is the gold standard, that method cannot address all the attributes mentioned above.

Herein, we have proven through a comprehensive comparability assessment the capability to determine key attributes of amino acid composition, SASA, secondary structure, and post-translational modification such as the extent of glycosylation using a powerful method for the fast screening of therapeutic protein samples. We determined that for the aspartate-to-tyrosine variant, only one substitution was observed within the full-length mAb, demonstrating the sensitivity and selectivity of the approach. The use of the 2T2D correlation algorithm was instrumental in the spectral analysis to determine the extent of glycosylation using this novel approach whereby, mAb2 and mAb3 were observed to be similar, while the distinct carbohydrate content of mAb1 compared to either mAb2 or the reference mAb suggested differences in the cell culture conditions that led to a different glycan profile. Even for the case involving mAb2 and mAb1, which had highly similar secondary structures and amino acid sequences, their glycan profile was distinctly different, demonstrating that all attributes presented herein are vital to the determination of the therapeutic protein’s identity. More importantly, these differences in glycan profile can lead to differences in function, stability, and an unwanted immune response.^3,4,18,36

This fast-screening method also served to establish differences in SASA, amino acid composition for the residues monitored, and slight differences in β-sheet, β-turn and random coil content observed for the array of mAb samples studied. This was demonstrated with the case of mAb2 and mAb3, in which the glycan profile was highly similar, yet the secondary structure and amino acid composition were different, as observed by evaluating the amide I and II bands using the 2T2D correlation algorithm. Finally, the comparability assessment of the reference mAb against mAb1, in which differences were observed for all attributes of the evaluation, demonstrated the need for such a comprehensive evaluation to determine protein identification in all of its dimensions.

Collectively, greater molecular level understanding is obtained from this fast-screening method when compared to the traditional amino acid sequencing methods or the separate glycan profiling methods, thus demonstrating the value of the approach for protein identification to de-risk pre-clinical candidate selection. The method involves a streamlined evaluation of an array of pre-clinical antibodies by directly assessing the infrared spectral differences within 1775–1015 cm⁻¹. The full-length mAbs are in the same buffer conditions, requiring only 1 μL of sample per well. The value of this approach consists of: (i) a platform using QCL microspectroscopy and the innovative slide cell that provides the flexibility for comparability assessment of an array of full-length mAbs in solution under controlled conditions; (ii) no additional sample preparation, thus effectively streamlining the analysis; and (iii) the highly sensitive and selective 2T2D asynchronous plots that allow for the determination of molecular differences. This approach is not limited to the study of mAbs nor to evaluations at room temperature. The current spectral data analysis demonstrated its usefulness in providing an in-depth understanding of the molecular factors that determine protein identity, with the added potential of being a highly effective fast screening method.

Conclusion

To our knowledge, we are the first to provide a method for fast screening of protein identification and extent of glycosylation for an array of therapeutic proteins in a single test. This multi-attribute comparability assessment study determined identity for the reference mAb, while also establishing differences in secondary structure, amino acid composition, solvent accessibility, and the extent of glycosylation and sialic acid capped N-glycan for multiple mAb samples. This was all done in a single experiment and for the intact proteins in the same formulation conditions, proving the value of a comprehensive evaluation that allows for the molecular understanding of the differences amongst an array of pre-clinical mAb candidates that share high sequence identity. Unlike other techniques that require extensive sample preparation steps that exponentially increase the time to result and add variability to the data; while only examining the amino acid sequence or the extent of glycosylation, our method addresses multiple attributes to evaluate identity in a comprehensive manner. This method has the capability of also evaluating aggregation and degradation, such as that arising from Asn/Gln deamidation and oxidation, within the same spectral regions studied, thus ensuring the value of using a highly sensitive and selective analytical tool such as the platform presented in this study. The significance is that one rapid screening method can be implemented early in developability to de-risk pre-clinical candidate selection or during release testing to provide accurate and comprehensive understanding of a drug throughout its lifecycle. The platform method presented herein explores many of the attributes needed for decision making in a single rapid screening test ideal for the biopharma industry.

Supplemental Material

sj-docx-1-asp-10.1177_00037028261451288 - Supplemental material for Two-Trace Two-Dimensional Correlation Analysis for the Confirmation of Therapeutic Protein Identification and Evaluation of Extent of Glycosylation: A Rapid Method for Testing Critical Quality Attributes

Supplemental material, sj-docx-1-asp-10.1177_00037028261451288 for Two-Trace Two-Dimensional Correlation Analysis for the Confirmation of Therapeutic Protein Identification and Evaluation of Extent of Glycosylation: A Rapid Method for Testing Critical Quality Attributes by Belinda Pastrana, Sherly Nieves and Eduardo I. Canto in Applied Spectroscopy