Native mass spectrometry combined with enzymatic dissection unravels glycoform heterogeneity of biopharmaceuticals

Robust manufacturing processes resulting in consistent glycosylation are critical for the efficacy and safety of biopharmaceuticals. Information on glycosylation can be obtained by conventional bottom–up methods but is often limited to the glycan or glycopeptide level. Here, we apply high-resolution native mass spectrometry (MS) for the characterization of the therapeutic fusion protein Etanercept to unravel glycoform heterogeneity in conditions of hitherto unmatched mass spectral complexity. Higher spatial resolution at lower charge states, an inherent characteristic of native MS, represents a key component for the successful revelation of glycan heterogeneity. Combined with enzymatic dissection using a set of proteases and glycosidases, assignment of specific glycoforms is achieved by transferring information from subunit to whole protein level. The application of native mass spectrometric analysis of intact Etanercept as a fingerprinting tool for the assessment of batch-to-batch variability is exemplified and may be extended to demonstrate comparability after changes in the biologic manufacturing process.

Page 8: "We determined the most abundant mass in the deconvoluted spectrum at 118,941. Taking into account…." Mass of the most abundant species in the intact sample minus desialylated sample is 9548. (128,489 -118,941 = 9548). 9548/291(mass of sialic acid) = 32.8 sialic acids. It seems they are missing somewhere 58 Da. Can they explain that? This also raises the question about mass accuracy and correctness of the spectra annotation in this manuscript in general. There is not accurate mass mentioned in the whole manuscript and I am forced to believe that everything is correct without chance to evaluate the data. Fig. 2 The annotation of the native spectra is confusing and probably also not completely correct. Every peak assigned with different symbol corresponds to sialic acid variants of each glycoform. Peaks, which are in some certain ion series are certainly those glycoproteoforms with different number of sialic acids. However, these ion series are not in one "compact" cluster of ions consecutively differing by one sialic acid, but many of them are interrupted and continue from other masses. For example, a range of yellow squares (with annotated peak 18 O-cores + 22 Neu5Ac) ranging from 114000 Da to 116600 Da somehow continues also approx. 2kDa later. The same problem can be seen also for another peak series. I also would like to know, how the authors assumed that for example black half-moon series ends with 20 Neu5Ac. Next, the legend of the Fig. 2b is confusing. Every symbol corresponds only to a number of O-glycan cores, however it looks like it is the whole composition together with the number of sialic acids. Although it is mentioned in the I suggest to re-annotate the spectra based on the peak series and their abundances revealed in Fig. 2a. It shows that the dominant O-cores range from 16-21 and all other contribute only in minority. This would be actually also be consistent with the logic of the used approach. Fig. 3 Here authors assigned with asterisks peaks as Na+ adducts. However, these assignments seem to be quite random and hardly justified. Fig. 4 Is there no chance that HexNAcHex is not O-glycan but a polyLacNAc elongation on one of the N-glycan branches? Fig. 5 In the panel "a" authors observe differences between Pre-change and Post-change batches. Authors should provide information about HCD energies since the main difference between the two batches seems to be a shift towards lower sialylation states. Higher HCD energies certainly result in loss of Neu5Ac on natively measured proteins. Supplementary Fig. 9 There is unidentified peak series between 23+ and 22+ charge state 22+ and 21+, and 21+ and 20+. Can authors comment about it? Supplementary table 2 A2Ga1G2F shows α-Gal structure, which is potentially immunogenic and undesirable on a biopharmaceutical product. This should be mentioned in the main text. Convincing MS/MS spectra corroborating structure must be included to support this specific structure.
Online Methods: Why did authors use only HCD fragmentation technique? Would not EthcD as this is known to be better for glycopeptide analysis? Did the peptide data correlate well with the native MS measurements? Please discuss that in the text.
Discussion: Reference 22. Should be discussed in the 1st paragraph as it is also focused on combining bottom up and native level data. Second paragraph: Limitations should be discussed more in detail including the issues with data analysis, mainly deconvolution: The ReSpect algorithm does not produce a deconvoluted mass spectrum, it is rather deconvoluted mass value plotted as a function of ppm accuracy ("peak" width in deconvoluted spectra is determined by ppm accuracy rather than the real width of peak in original spectrum). This results in a weird observation wherein each peak in deconvoluted "spectrum" has different peak width. Based on my experience, "intensity value" obtained from deconvoluted spectra is highly irreproducible. This remains a major bottleneck in data analysis of complicated nonisotopically resolved spectra. The authors should discuss this

Reviewer #2 (Remarks to the Author):
The authors combine native mass spectrometry of an intact protein, native mass spectrometry of large fragments of the protein, and peptide analysis, to identify glycoforms on different sites in the protein, relative stoichiometries of different glycoforms, and glycoform composition of the major forms of the intact protein. The protein they studied is heavily glycosylated (much more complicated than antibodies, which are the main proteins that have been studied by similar techniques), so presents a real challenge to characterize, but they have demonstrated what is achievable with a reasonable amount of work. While they did not achieve site-specific information about the O-glycosylation, they were able to deconvolute a fair amount of detail about the N-glycosylation stoichiometry on different sites. They emphasize the advantage of native MS for spacial resolution of isotope clusters. Visually this is obviously significant, but even if isotope clusters overlap deconvolution software may be able to separate them; it would be good for the authors to comment on this, particularly as an Orbitrap will have higher resolution at the lower m/z observed in denatured analysis. Would this be easier / more effective if the data was acquired at higher resolution? Spectra in Figure 1 were acquired at resolutions ranging from 17.5K to 70K. If Fig 1b/1c were acquired at 70K then baseline separation of all of these peaks may have been obtained. I am not suggesting they need to re-acquire data at higher resolution, but mentioning why it was acquired at lower resolution would be appropriate.

Specific Comments:
I think the spectrum of the deconvoluted intact protein ( Supplementary Fig. 3b) needs to be in the main part of the manuscript: this is the ultimate focus of the whole paper.
There should be a sentence that explains why versions with different numbers of lysines were expected; this is annotated without any explanation.
In the sialidase-treated data, how confident can the authors be that the removal was complete; i.e. they can assign two fucose over a residual NeuAc?

Reviewer #1 (Remarks to the Author):
Wohlschlager et al. here report on the analysis of Etanercept using high-resolution native mass spectrometry combined with enzymatic treatment. As this is a highly relevant biopharmaceutical and exhibits a very complex glycosylation profile their work is a tour de force, brought about by the high-resolution mass spectra obtained with an Orbitrap EMR. The method is complemented also by bottom up mass spectrometry measurements. The authors highlight that the higher mass resolution achieved by native MS is a key parameter to facilitate the separation of different protein proteoforms differing from each other by only a small mass. Next, they applied enzymatic treatment of the sample in order to reduce sample complexity and molecular mass, which partially overcomes limitations resulting from the extremely high number of possible PTMs combinations and therefore difficulties with confident peak assignment. In general, this reviewer highly appreciates the presented work which demonstrates an unique ability of native MS measurements, which in combination with enzymatic processing of the sample can describe extreme glycoprotein micro heterogeneity. However, in its current form, the manuscript needs severe further improvements and detail especially due to incomplete data representation and relatively poor discussion.

Reviewer comment 1:
Major critique: although the language in the manuscript is correct, I find it very difficult to read. The text is written very technically with many details, but the structure seems to be a bit chaotic. The Results part contains paragraphs and sentences, which should rather be in the discussion part.
Our reply 1: The manuscript has been fully revised with regard to structure and phrasing. Several sentences of the results section were moved to the discussion part as suggested by the reviewer. In order to provide an overview on the experimental workflow and the transfer of information for peak assignment, Supplementary Fig. 4 has been revised and now illustrates the different structural levels described in this study. Phrasing was improved upon linguistic review by a native English speaker.

Reviewer comment 2: Major problem is the absence of clear tables (in supplementary material) containing vital information about detected and described glycoproteoforms of the analyzed sample and a list of detected glycopeptides. The table should contain a list of the all masses of "
presented" glycoproteoforms with mass accuracy achieved, retention time (in case of glycopeptides), some score metrics etc.
Our reply 2: Files containing detected masses, relative abundances and assigned glycoforms have been added (Supplementary Data 1). The files also provide information on search parameters, mass deviation of assigned glycoforms with respect to the experimental mass, and hit scores where available. In case of glycopeptides, precursor m/z, product charge, mass error, total area, retention time and a confidence score provided by the software BioPharma Finder are listed.

Reviewer comment 3:
Result part could be probably structured better. For example, native spectra annotation could start from the enzymatically processed samples, which are "easier" to annotate. From that information, one could consecutively interpret more complicated spectra. Next, the discussion part needs to be improved.
Our reply 3: In our intention, the order in spectra annotation does reflect an increase in sample complexity for O-glycans, N-glycans, and the combination of both. As the assignment of O-and N-glycans requires different enzymatic and computational approaches, an order purely based on spectral complexity is not coherent in our opinion. The logics we tried to elaborate in our argumentation was as follows: (1) the raw spectrum of intact Etanercept provides insight into the complexity of the molecule, illustrating the need for simplification at the experimental level (2) assignment of O-glycoforms and their sialylated variants is possible upon enzymatic deglycosylation at the whole protein level (3) assignment of Nglycans requires separate analysis of TNFR and Fc subunits and integration of glycopeptide data (4) information on both O-and N-glycans can be transferred to the whole protein level. We have outlined this strategy at the end of the section "Native mass spectrometry of Etanercept at the intact level" in order to clarify the sequence of experiments and data interpretation. Following the reviewer´s suggestion, we have revised and simplified the annotation in Figure 2b, which now better illustrates how glycan complexity increases with each level of enzymatic trimming. Finally, the discussion section has been completely rewritten to emphasize the essential aspects of our workflow and to outline the limitations of the approach.
general, detail discussions on limitations of the approach and other difficulties should be removed from the results section, and mentioned in the discussion section.

Our reply 5:
We have added a paragraph to the first section of the results part to clarify the strategy behind the analysis at different structural levels. We have rearranged the text accordingly and shifted the sentences suggested by the reviewer to the discussion section. In addition, we have revised the titles of the subheadings to more clearly reflect the workflow employed for deciphering the multitude of glycoforms.
Reviewer comment 6: Page 8: "We determined the most abundant mass in the deconvoluted spectrum at 118,941. Taking into account…." Mass of the most abundant species in the intact sample minus desialylated sample is 9548. (128,489 -118,941 = 9548). 9548/291(mass of sialic acid) = 32.8 sialic acids. It seems they are missing somewhere 58 Da. Can they explain that? This also raises the question about mass accuracy and correctness of the spectra annotation in this manuscript in general. There is not accurate mass mentioned in the whole manuscript and I am forced to believe that everything is correct without chance to evaluate the data.
Our reply 6: As outlined on page 7 as well as in Supplementary Figure 6 of our original manuscript, mass accuracy and mass resolution achieved for large biomolecules at high m/z are not sufficient to resolve the different glycoforms present under the peaks obtained for intact Etanercept. Therefore, we cannot simply derive the number of sialic acids from the mass difference observed in the spectra of the intact Etanercept and the desialylated molecule. Our viewpoint is corroborated by a recent publication on the limitations of mass resolution for distinguishing species of slightly differing mass, such as water-or sodium adducts (see Philip Lössl, Joost Snijder, Albert J. R. Heck, Boundaries of Mass Resolution in Native Mass Spectrometry, J. Am. Soc. Mass Spectrom., 2014, 25:906Y917), which is also discussed in our manuscript (reference 29 of the revised manuscript). Following the reviewer´s critique on the transparency of spectra annotation, files containing experimental and theoretical average masses have been added to the revised manuscript in order to document peak assignment and mass accuracy (Supplementary Data 1). Fig. 2. The annotation of the native spectra is confusing and probably also not completely correct. Every peak assigned with different symbol corresponds to sialic acid variants of each glycoform. Peaks, which are in some certain ion series are certainly those glycoproteoforms with different number of sialic acids. However, these ion series are not in one "compact" cluster of ions consecutively differing by one sialic acid, but many of them are interrupted and continue from other masses. For example, a range of yellow squares (with annotated peak 18 O-cores + 22 Neu5Ac) ranging from 114000 Da to 116600 Da somehow continues also approx. 2kDa later. The same problem can be seen also for another peak series.

Our reply 7:
The annotation of Fig. 2b has been revised with regard to clarity: taking into account relative abundances of the respective O-glycoforms observed in Fig. 2a (PNGase F+sialidase digest), only one, i.e. the predominant O-glycoform was assigned to each annotated peak. A file listing all theoretically possible glycoforms and their estimated abundances has been added (Supplementary Data 1, fig_2b/2b_7_figure_data.csv). The absence of certain peaks within sialic acid series is due to larger mass deviations than the accepted mass tolerance of +/-5 Da. This may be due to overlapping proteoforms, which were not resolved. In the new version of Fig. 2b this is only the case for one glycoform, which is now also discussed in the text.

Reviewer comment 8: I also would like to know, how the authors assumed that for example black half-moon series ends with 20 Neu5Ac.
Our reply 8: As discussed in the previous paragraph, the absence of certain peaks within sialic acid series is due to larger mass deviations than the accepted mass tolerance of +/-5 Da. This may be due to overlapping proteoforms, which were not resolved. Considering the low relative abundance of the black half-moon O-glycoform corresponding to 22 O-glycan cores in Fig. 2a, this symbol has been completely removed from the annotation in Fig. 2b. Based on the relative abundance of each glycoform, a hit score was introduced in order to transfer this information between different digestion levels. For clarity, the revised version of Fig. 2b only shows highly abundant O-glycoforms, i.e. the most probable variants. A file listing all theoretically possible glycoforms and their estimated abundances has been added (Supplementary Data 1, fig_2b/2b_7_figure_data.csv).

Reviewer comment 9:
Next, the legend of the Fig. 2b  Our reply 9: Figure 2b has been revised as described in replies 7 and 8. Additionally, the number of incorporated sialic acids is now indicated above each annotated peak.
Reviewer comment 10: I suggest to re-annotate the spectra based on the peak series and their abundances revealed in Fig. 2a. It shows that the dominant O-cores range from 16-21 and all other contribute only in minority. This would be actually also be consistent with the logic of the used approach.
Our reply 10: This is a very helpful suggestion and Figure 2b was changed accordingly. As mentioned in reply 8, a hit score based on the relative abundance of each glycoform was introduced in order to transfer this information between different digestion levels. A peak list containing all possible glycoform assignments with their hit scores has been added as supplementary data. Fig. 3. Here authors assigned with asterisks peaks as Na+ adducts. However, these assignments seem to be quite random and hardly justified.

Reviewer comment 11:
Our reply 11: The mass of 22.99 Da corresponding to one Na + ion was included in the list of possible modifications for the annotation of glycoforms. The Na + adducts indicated are within the accepted mass accuracy and occur in tandem with an M+H + peak at a relative abundance of around 10%-13%. This pattern of a main M+H + peak and a Na + peak of lower intensity is characteristic in case of adduct formation and supports the annotation of the respective peaks. Fig. 4

. Is there no chance that HexNAcHex is not O-glycan but a polyLacNAc elongation on one of the N-glycan branches?
Our reply 12: PolyLacNAc was neither detected in glycopeptide analysis (this study) nor in the analysis of released N-glycans (Houel S. et al., Anal. Chem., 2014), the latter representing the gold standard in glycan analysis. This structure was therefore not included in the applied N-glycan library (compare Supplementary Table 1). Fig. 5. In the panel "a" authors observe differences between Prechange and Post-change batches. Authors should provide information about HCD energies since the main difference between the two batches seems to be a shift towards lower sialylation states. Higher HCD energies certainly result in loss of Neu5Ac on natively measured proteins.

Reviewer comment 13:
Our reply 13: As outlined in the methods section of the manuscript, comparison of different Enbrel ® batches was performed at identical instrument settings. Mass spectra shown in Fig.   5 were all acquired with fragmentation energy settings of SID 100 eV and CE 25 eV, ruling out the possibility that the observed difference in intact masses is due to an experimental artifact. Information on fragmentation energy settings has been added to the corresponding figure legend. In addition, raw files containing information on instrumental settings will be made publicly available. With regard to fragmentation energies, settings were initially optimized by systematically increasing fragmentation energies while following the formation of characteristic glycan fragment ions at low m/z. Fragmentation energies were chosen in order to achieve a compromise between maximum desolvation and minimum glycan fragmentation. The observed shift towards lower masses for the two post-change batches compared to the pre-change batch is in-line with a previous report describing a decrease in variants containing the N-glycan G2F and an increase in G0F glycoforms, corresponding to lighter variants due to a lower number of incorporated galactoses (Schiestl M. et al., Nat. Biotechnol., 2011). Supplementary Fig. 9. There is unidentified peak series between 23+ and 22+ charge state 22+ and 21+, and 21+ and 20+. Can authors comment about it?

Reviewer comment 14:
Our reply 14: Charge states of this species were labelled in blue in Supplementary Fig. 8b of the revised manuscript ( Supplementary Fig. 9 of the submitted manuscript) and can be assigned to a truncated Etanercept dimer lacking amino acids 1 to 186 of one TNFR chain within a mass accuracy of ± 30 ppm.

Reviewer comment 15:
Supplementary table 2. A2Ga1G2F shows α-Gal structure, which is potentially immunogenic and undesirable on a biopharmaceutical product. This should be mentioned in the main text. Convincing MS/MS spectra corroborating structure must be included to support this specific structure.

Our reply 15:
The reviewer is correct that identification of the immunogenic A2Ga1G2F structure requires further consideration. As the identification was solely based on intact glycopeptide mass (upon identification by MS/MS in a different Enbrel ® batch), the respective structure has been eliminated from Supplementary Tables 1 and 2. The revised tables now contain only structures for which reliable MS/MS spectra were acquired. Because of the low abundance of the eliminated glycan structures (1.7%), the resulting effect on glycoform annotation at the subunit and whole protein level is minimal.

Reviewer comment 16:
Online Methods: Why did authors use only HCD fragmentation technique? Would not EthcD as this is known to be better for glycopeptide analysis? Did the peptide data correlate well with the native MS measurements? Please discuss that in the text.

Our reply 16:
The reviewer is correct that for in-depth analysis aiming at maximum sequence coverage and identification of minor glycoforms, EThcD is the method of choice. In this study, the purpose of glycopeptide analysis was the generation of a database containing the main N-glycan structures. Therefore, HCD was applied as a generic method fit for purpose. The identified N-glycopeptides are well in accordance with previously published structures identified by released N-glycan analysis, which is the gold standard for glycan analysis (Houel S. et al., Anal. Chem., 2014). A comparison of N-glycan structures and their relative abundances as identified by glycopeptide and released glycan analysis is shown below (released glycans up to a fractional abundance of 1.0% were derived from Houel, S. et al. 2014; peptide mapping data originate from the current study).
The main advantage of glycopeptide data is the information on glycosylation site occupancy, which is considered in the annotation of subunit and whole protein spectra. In order to assess correlation of glycopeptide data with native MS measurements, we compared a theoretical spectrum simulated based on our desialylated glycan library (Supplementary Table 1) with the experimental spectrum obtained for Etanercept lacking sialic acid and Oglycans (Fig. 4a). The mirror plot shown below demonstrates coherence of the theoretical spectrum based on glycopeptide data and the experimentally derived spectrum: almost all experimentally detected masses correspond to a theoretically calculated mass, providing evidence that the applied N-glycan library is adequate for annotation of native mass spectra. Nevertheless, it should be noted that sialylated N-glycans were not evaluated in the same way, as annotation of native mass spectra of sialylated Etanercept was not possible due to the limits in mass resolution discussed in the manuscript.
Reviewer comment 17: Discussion: Reference 22. Should be discussed in the 1st paragraph as it is also focused on combining bottom up and native level data.
Our reply 17: As outlined in reply 4, we have added more detailed information on the results presented in reference 22 as well as a comparison with the findings described in this study (introduction and discussion sections).