Native mass spectrometry and ion mobility characterization of trastuzumab emtansine, a lysine-linked antibody drug conjugate

Native versus denaturing MS analysis of intact trastuzumab emtansine

As T-DM1 is a lysine-linked ADC, only covalent species are expected and subsequent LC-MS34 or size-exclusion-chromatography–MS35 analysis in classical denaturing conditions can be used to determine both the average drug load and the drug load profile. To reduce sample heterogeneity, ADCs are usually deglycosylated prior to denaturing MS analysis, leading to the detection of a distribution of both odd and even drug loads. However, since pretreatments should be kept to a minimum in studies of ADCs or mAbs, we first directly analyzed untreated (glycosylated) T-DM1 by denaturing MS. As expected, the resulting raw electrospray ionization (ESI) mass spectrum reveals many overlapping charge states distributions [Fig. 1(A)], reflecting the combination of both glycosylation and drug loading heterogeneities. Further deconvolution is not straightforward, the inability of the algorithms to cope with such heterogeneity leaving many low-abundance drug-load species (e.g., D0, D7, and D8) undetected, while only the major glycoforms of the most abundant conjugates (i.e., G0F/G0F, G0F/G1F, G1F/G1F, and G1F/G2F of D1–D6) are observed. After deglycosylation (see the “Materials and Methods”), the raw mass spectrum is easier to interpret [Fig. 1(B)]. Better deconvoluted spectra are obtained, comprising D0–D7, leading to an average DAR of 3.5 in agreement with Lazar et al.35 However the D8 species are still missing from these spectra, mostly because of overlapping charge state distributions. As described previously,34 the adducts at ∼220 Da visible in Figure 1(B) for each deglycosylated drug-load species correspond to linkers that have modified the antibody but do not contain conjugated DM1 due to a side-reaction.

In consideration of the many benefits described in the literature that native MS offers for the analytical characterization of noncovalent cysteine-linked ADCs,24, 36, 37 we used this technique on the covalent ADC T-DM1. As expected, the same analysis performed under nondenaturing conditions (150 mM ammonium acetate at pH 7.4, see the “Materials and Methods”) strongly reduces the charge state distribution, such that the different drug load species are more clearly resolved with fewer overlapping ion distributions. After fine tuning of the deconvolution parameters, D1–D6 and D0–D7 are resolved for the nondeglycosylated [Fig. 2(A)] and deglycosylated [Fig. 2(B)] T-DM1 samples, respectively. The average DARs calculated from the individual charge states (see “Materials and Methods”) were 3.1 ± 0.1 and 3.3 ± 0.3 for untreated and deglycosylated T-DM1, respectively, which is in good agreement with the value of 3.5 measured by LC-MS.34 Nonetheless, even for the deglycosylated sample, D8 is not detected, while D0 and D7 are also missing in the spectrum for untreated T-DM1. This is due to the relatively low resolution of the classical quadrupole Q-TOF instruments (R ∼1500 at m/z = 6200) used here. However, since the missing payload contributions compensate each other, the average DARs obtained are fortuitously accurate. Interestingly, D0 species are visible for the highest and lowest charge states (28+, 23+, 22+) in the spectrum of deglycosylated T-DM1 [as indicated by black dots in Fig. 2(B)] but not for those in between, reflecting the partial overlap of the different species and the importance of the choice of the charge states used for deconvolution. Altogether, our results illustrate—as already reported for cysteine-ADC conjugates²⁴—that native MS is a powerful technique for covalent lysine-ADC analysis, even if a certain number of issues remain to be solved. However, due to deconvolution bias, the average DARs deduced from native MS should be determined using deglycosylated (rather than glycosylated) ADCs and from raw (rather than deconvoluted) spectra.

The benefits of high resolution native MS for lysine-linked ADC analysis

To avoid problems arising from overlapping ion signals in the native MS spectra, measurements were performed on a higher resolution mass analyzer. Both untreated and deglycosylated T-DM1 samples were injected into a high resolution orbitrap instrument optimized for native MS analysis.38 As has already been reported for cysteine-linked ADCs,24 at a nominal resolution of 17,500, an effective resolution of ∼2600–3000 at m/z = 6500 was obtained with all glycoforms resolved to the baseline. These measurements [Fig. 3(A)] allowed average DARs to be obtained for each species detected, that is, for the most intense glycoform of each individual charge state. For the most intense charge states, up to seven glycoforms were detected [Fig. 3(A)]. However, the highest Dn species of a given charge state (D7/D8) still overlap with the lowest (D0/D1) from the neighboring charge state. Consequently, the DARs obtained from this spectrum only account for species D0–D6, explaining the rather low average DAR (3.0 ± 0.2, see Table 1) obtained. Even with a high-resolution instrument, the mass distribution of large (150 kDa) molecules, such as T-DM1 intrinsically limits the resolution that can be achieved. Indeed, the full width at half maximum (FWHM) of 24+ or 23+ charge state peaks can reach 1 Th for a 150 kDa molecule. Thus, even at infinite instrument resolution (see Fig. S1 in Supporting Information), the D8 G0F/G1F 24+ peak (m/z = 6495.93) will always overlap with the D1 G1F/G1F 23+ peak (m/z = 6494.10). This means that increasing the effective resolution up to ∼6100 at m/z = 6500 will narrow the peaks, but that a mass resolution of ∼150,000 would be needed to start distinguishing the different isotopes (still with the same isotopic distribution width). As for the untreated sample, the high-resolution native MS instrument provides more clearly resolved spectra for deglycosylated T-DM1 than the Q-TOF instrument. Nevertheless, D8 species are still absent from the deconvoluted spectrum [Fig. 3(B)] measured at a resolution of 17,000, leading to an underestimated DAR of 3.0 ± 0.2.

Combining charge state reduction with high-resolution native MS for the characterization of lysine-linked ADCs

Charge state reduction has been proposed as a means to separate overlapping peaks, which would further improve average DAR calculations from native MS measurements.39, 40 This was achieved here by supplementing the T-DM1 sample with increasing imidazole concentrations. Theoretically, shifting the charge states lower than 20+ should avoid the problem of overlapping peaks described in the previous section. Indeed, for glycosylated T-DM1, the theoretical m/z values for charge states lower than 20+ mean that neighboring D8 and D0 peaks should not overlap (the D8 and D0 peaks, respectively, of the G1F/G1F 19+ G0F/G0F 18+ states appear at 8213.64 and 8226.71 Th, respectively). The same applies for deglycosylated T-DM1 (D8–19+ at 8081.57 and D0–18+ at 8104.46 Th). The charge state reduction method described above leads to a shift of up to 10 charges for both the glycosylated and deglycosylated T-DM1 samples on the Q-TOF instrument (Fig. 4, in comparison with Fig. 2). Imidazole improves both the separation of the different Dn species (Fig. 4 insets) and the desolvation of the samples, leading to sharper peaks within each charge state. However, longer acquisition times are required to compensate for the increase in the number of ionized species and a lower transmission of the charged reduced species at higher m/z values.40 Furthermore, this charge reduction strategy is limited by the fact that lower charge states (at high m/z values) are desolvated less efficiently, as seen for the 14+ to 16+ states in Figure 4(A). Nevertheless, a good compromise can be achieved between the ion transmission, charge reduction, and separation of the different Dn species, allowing unambiguous observation of all of the latter (D0–D8) for deglycosylated T-DM1 [Fig. 4(B)].

We next investigated whether the benefits of high-resolution MS and charge state reduction could be synergic for the characterization of T-DM1. In the presence of 20 mM imidazole, the charged species that lead to the overlap of the D0/D1 and D7 peaks of neighboring groups were effectively reduced. At a nominal resolution of 17,500 [Fig. 5(B)], this lead to an average DAR of 3.3 ± 0.2, closer to the expected value (viz. 3.5). Further charge state reduction is observed with 40 mM imidazole [Fig. 5(C)], but the associated decrease in intensity of the peaks, already reported by Mehmood et al.,40 prevents the detection of low abundance D7/D8 species, leading to an underestimation of the average DAR (see Table 1). Because of the lower number of species in deglycosylated T-DM1, this sample could be analyzed at a higher nominal resolution of 37,000. Under these experimental conditions, an effective resolution of 4500 at m/z = 6200 was obtained, allowing the detection of all Dn species. In the absence of imidazole, the deconvoluted mass spectrum obtained from deglycosylated T-DM1 shows the presence of all species from D0 to D7, leading to an average DAR of 3.4 ± 0.2 [Fig. 5(D)]. Adding 20 or 40 mM imidazole enables the unambiguous detection of the D8 species in the deconvoluted mass spectra [Fig. 5(E,F)], with a corresponding average DAR of 3.5 ± 0.1, in agreement with the expected value (Table 1).

Native ion mobility MS analysis for average DAR determination of trastuzumab emtansine

Following the determination of the average DAR of the cysteine-linked reference ADC, BV, by IM-MS,24 we set out to analyze the more heterogeneous lysine conjugate T-DM1 with the same approach. A comparison of the IM-MS driftscope plots obtained for deglycosylated trastuzumab and T-DM1 shows that IM-MS provides a direct picture of lysine-conjugate heterogeneity (Supporting Information Fig. S2). The intensity distribution of the arrival time distribution (ATD) peaks corresponding to the Dn species of the most intense 24+ charge state of deglycosylated T-DM1 is similar to that of the corresponding m/z peaks [Fig. 6(A)]. The intensities of the drift peaks extracted for each charge state were plotted across the series for each drug binding stoichiometry and a Gaussian curve was fitted to the resulting distribution [Fig. 6(C)]. The relative intensity of these ATDs indicates the relative abundance of each species. Averaging the semiquantitative IM-MS data among the three most intense charge states (23+ to 25+) gives an average DAR of 3.4 ± 0.2 [Fig. 6(B) and Table 1], in good agreement with the expected value of 3.5, this despite the fact that the D8 species were not detected under these conditions.

Molecular modeling of the trastuzumab emtansine D4 positional isomers

On the basis of these results, IM-MS appears to be a promising analytical technique to characterize positional isomers in both cysteine and lysine conjugates. We have previously reported that positional isomers were not detected for BV, the reference full length cysteine ADC,24 suggesting that current resolutions of IM cells are insufficient for the detection of positional isomers. For deglycosylated T-DM1, the ATDs are resolved at 92.5% of the valley with the IM cell resolution used here [∼20, Fig. 6(C)]. Theoretical simulations of the ATDs indicate that a separation at 50% of the valley would require a resolution of ∼70 (and a higher number of bins per mobility cycle) [Fig. 6(D)].

The current absence in the Protein Data Bank (PDB) of any 3D ADC structural model makes our assumption (that current IM resolution is insufficient) rather difficult to confirm. Here we used a molecular model of the D4 positional isomers of T-DM1 to see how challenging it would be to separate DAR isoforms in a ion mobility cell. This model was generated based on the structure of human IgG1 and on the Fab structure of trastuzumab (PDB IDs 1HZH and 1N8Z, respectively). The theoretical collision cross-sections (CCSs) were calculated with Mobcal (see “Materials and Methods”) for D0 and each isoform from D1 to D4 [Fig. 7(A, B) and Table 2]. The CCS values measured by IM-MS (Table 2) are slightly higher than those calculated from the masses41 but considerably lower than those expected from molecular modeling [Fig. 7(A) and Table 2]. We have already pointed out the latter of these two discrepancies,24 as have others,42 which is rationalized by collapse in the gas-phase, in good agreement with in vacuo molecular dynamics simulations.42 In contrast with the results obtained for BV,24 the values calculated assuming spherical proteins41 do not agree with those measured by IM-MS. Each drug-load binding event induced a CCS increment of ∼25 Ų, which matches the contribution on binding expected from the mass of a single DM1 molecule (Table 2). This suggests that the conformational changes that occur in trastuzumab upon drug binding are very slight. However, the increases in CCS upon drug binding predicted from modeling (∼100 Ų) and from IM-MS (∼25 Ų) do not agree. The surface-exposure of lysine-linked payloads may explain this result. This indeed makes them more susceptible to gas-phase collapse than equivalent payloads in cysteine-linked conjugates.

Subsequent simulations of the ATDs of the different D1 isoforms indicate that an IM resolution of ∼330 would be necessary to separate them in an ion mobility cell [Fig. 7(C)], indicating that further instrumental developments are required before the quantitative characterization of positional isoforms by native IM-MS becomes a possibility. With the resolution currently provided by commercial instruments, middle-down approaches following partial enzymatic digestion are probably better suited to study subtle structural differences using IM-MS.

Sample preparation

Trastuzumab emtansine (Kadcyla^®, Genentech/Roche) was deglycosylated by incubation for 30 min at 37°C with IgGZero (1 unit/µg, Genovis, Lund, Sweden). Glycosylated and deglycosylated T-DM1 samples were desalted against 150 mM ammonium acetate (pH 7.4) during five concentration/dilution cycles on a 30 kDa Vivaspin ultracentrifugation device (Sartorius, Göttingen, Germany). For charge state reduction, a 200 mM stock solution of imidazole (Sigma Aldrich, Steinheim, Germany) in 150 mM ammonium acetate was directly added to the buffer up to the desired concentration.

Denaturing and noncovalent mass spectrometry experiments

Mass spectrometry experiments were performed in positive mode, on a Synapt-G2 HDMS Traveling-Wave-Ion Mobility (TWIMS) instrument (Waters, Manchester, UK) and on an Orbitrap Exactive Plus Extended Mass Range mass spectrometer (Thermo Fisher Scientific, Bremen, Germany), both coupled with a Nanomate chip-based autosampler (Advion, Ithaca, USA). The parameters of the instruments were optimized to allow the stabilization and transmission of high molecular weight species. Electrospray ionization was conducted at a capillary voltage of 1.75 kV and a nitrogen nanoflow of 0.55 psi.

Denaturing experiments were acquired on the Synapt G2 instrument only, in the m/z 500–5000 mass range with a sample cone (Vc) of 50 V and a backing pressure of 2.0 mbar. Noncovalent experiments on the Synapt G2 instrument were performed in the m/z 1000–10,000 mass range, with a Vc of 150 V and a backing pressure of 6.0 mbar.

External calibration was performed with 2 µM horse heart myoglobin (Sigma Aldrich, Steinheim, Germany) in H₂O/ACN/FA (50/50/1 v/v) for the experiments performed under denaturing conditions, and 2 mg/mL cesium iodide (Sigma Aldrich, Steinheim, Germany) in isopropanol/H₂O (1/1) for those performed under nondenaturing conditions.

The measurements on the Exactive Plus EMR instrument were performed as follows. The data were acquired in positive mode for 2 min with the Nanomate, after external calibration in EMR mode with cesium iodide at 0.1 mg/mL (using the heated electrospray ionization source). The in-source collision-induced dissociation was set to 75 eV and the higher-energy collisional dissociation cell was turned off. The automatic gain control target was set to 5 × 10⁵ with nominal resolutions of 17,000 or 35,000, as described in Table 1. The gas pressure in the trap was set to 7.0 mbar and the high-mass species were transmitted with voltages on the inject, inter, and bent flatapoles of 8, 7, and 6 V, respectively.

Ion mobility experiments

Ion mobility MS experiments were performed on a TWIM-MS Synapt G2 instrument (Waters, Manchester, UK). The parameters were optimized with respect to the ion transmission and resolution in the drift cell, with as little gas-phase unfolding as possible: the sample cone, trap collision energy, transfer collision energy, and trap DC bias were set to 100, 2, 15, and 37 V, respectively. The backing pressure was 6.0 mbar and gas flows of 1.5, 130, and 45 mL/min were set in the collision (Ar), helium (He) and mobility cells (N2), respectively. The IM wave height and velocity were 40 V and 950 m/s, respectively. The CCS calculations were calibrated using concanavalin A, pyruvate kinase, and alcohol dehydrogenase (Sigma Aldrich, Steinheim, Germany) desalted against 150 mM ammonium acetate on 0.5 mL Zeba columns (Thermo Scientific, Rockford, USA). The CCS measurements were run in triplicate and the data were analyzed with MassLynx 4.1 and Driftscope 2.2 (Waters, Manchester, UK).

Homology modeling of trastuzumab and of the four payloads

The full trastuzumab structural model was assembled using Discovery Studio 3.0 modeling tools with the CHARMm forcefield (Biovia, France). The full human IgG1 structure (PDB ID 1HZH) and the Fab structure of trastuzumab (PDB ID 1NZ8) serve as template structures. The structure of IgG1 was completed by modeling the unresolved Thr-His-Thr sequence in the hinge region. The hinge region structure was then modified to obtain a more symmetrical structure of the two Fab domains compared with the Fc domain. The two Fab domains were removed and replaced by that of trastuzumab.

A structural model of was prepared for DM1 and four DM1 molecules were arbitrarily grafted on the four more exposed lysine side chains (L1-169, H1-325, L2-190, and H2-305).

Theoretical rotationally averaged collision cross sections

Collision cross-sections were calculated using the Mobcal exact-hard-sphere-scattering (EHSS) and projection-approximation models (PA).44, 45 The projection approximations were scaled by a factor of 1.14 as reported elsewhere.39

Deconvolution and average DAR calculation

The data acquired on the Q-TOF instrument was deconvoluted using the maximum entropy MaxEnt^™ algorithm using MassLynx (Waters, Manchester, UK), after background subtraction (polynomial order 25, 2% area beneath the curve), from 145 to 157 kDa and with a minimum intensity ratio of 25% (left and right). The resolution (Da/channel) and uniform Gaussian FWHM were set to the same value, which was however adjusted for each spectrum in order to maximize the number of DARs in the deconvoluted spectra. The data acquired on the orbitrap EMR were deconvoluted using the Manual Respect^™ algorithm in the Protein Deconvolution software (version 3.0, Thermo Scientific) with an output mass range of 145–157 kDa, 4 iterations, 1–3 minimum adjacent charges, and both the left and right peak shapes set to 2. The m/z range, the mass tolerance, the charge state range, and the target mass were optimized for each spectrum, in order to maximize the number of DARs in the deconvoluted spectra. In both cases, the charge state ranges used are indicated on the corresponding figures with a double arrow.

The average DARs were calculated, from both the denaturing and nondenaturing MS data, as a weighted average of each drug-load, based on the relative intensities of each peak in the deconvoluted spectra or of each charge state in the raw spectra, with an automated Excel macro developed inhouse and applying Eq. 1. The standard deviations in Table 1 represent the errors that result from averaging the values from each charge state.

For the glycosylated samples, the DARs were averaged over the main glycoforms (G1F/G1F and G1F/G2F), when the resolution was sufficient.

(1)

Similarly, the average DARs from ion mobility MS were calculated as a weighted average of each drug-load, based on the relative intensities of each DAR drift trace, using the same equation but with the intensities of the ATDs used instead of those of the charge states.