Overcoming Lot-to-Lot Variability in Protein Activity Using Epitope-Specific Calibration-Free Concentration Analysis

Concentration determination is a fundamental hallmark of protein reagent characterization, providing a means to ensure reproducibility and unify measurements from various assays. However, lot-to-lot differences in protein activity often still occur, leading to uncertainty in the accuracy of downstream measurements. Here, we postulate that those differences are caused by a misrepresentation of the protein concentration as measured by traditional total protein techniques, which can include multiple types of inactive protein species. To overcome this, we developed a standardized method to quantify a protein’s active concentration via calibration-free concentration analysis (CFCA). As a pilot study, we compare the biophysical and immunoassay responses from three batches of recombinant soluble lymphocyte-activation gene 3 (sLAG3), as defined by either their total or active concentrations. Defining the sLAG3 reagents by their assay-specific concentration improved consistency in reported kinetic binding parameters and decreased immunoassay lot-to-lot coefficients of variation (CVs) by over 600% compared to the total protein concentration. These findings suggest that the total concentration of a protein reagent may not be the ideal metric to correlate in-assay signals between lots, and by instead quantifying the concentrations of a reagent’s assay-specific epitopes, CFCA may prove a useful tool in overcoming lot-to-lot variability.


Antibodies and Proteins
NISTmAb, Humanized IgG1k Monoclonal antibody (NIST 8671, lot 14HB-D-002) was used as a standard for the CFCA measurements.Biotinylated recombinant human anti-sLAG3 monoclonal capture antibody (BMS) as well as the unlabeled and sulfo-tagged recombinant mouse anti-sLAG3 monoclonal detection antibody (BMS) were used for the sandwich immunoassay in concert with each of the recombinant human sLAG3 calibrator lots.The sLAG3 calibrator construct (M1-H449) is a full-length ectodomain expressed from stable-expressing CHO-S cells and purified without the signal peptide (L23-H449) through affinity chromatography of a C-terminal 6-His tag followed by size exclusion chromatography.Lot 1 was purified in December of 2019, Lot 2 was purified in February of 2020, and Lot 3 was purified in July of 2022.Calibrators were stored in buffer at -80°C for long-term storage, sub aliquoted and refrozen once prior to analysis, and each subaliquot was then used within 2 weeks of thawing, stored at 4°C.All CFCA and MSD analyses for this study were conducted from Q1 to Q3 of 2023.

Biolayer Interferometry (BLI) kinetics
All BLI experiments were conducted on a Sartorius Octet HTX system in 1X HBS-EP + 1% BSA.Octet AHC or AMC tips (for capture mAb or detection mAb, respectively) were first equilibrated offline with 1X HBS-EP + 1% BSA for 10 minutes or more prior to usage.sLAG3 reagents were then serial diluted into 1X HBS-EP + 1% BSA based on their total protein concentrations.Following a 2 minute on system equilibration, toolkit mAbs were loaded onto the tips at 10ug/mL for 10 minutes, followed by a 2-minute baseline.The tips were then assayed with a 10-minute association and dissociation phase in the sLAG3 serial dilutions.Curves were globally fit using a 1:1 binding model in ForteBio Data Analysis Version 12.0.2.3.Lot-to-lot differences in binding parameters were assessed using the 2way ANOVA multiple comparisons test in GraphPad Prism.

sLAG3 Ligand Binding Assay
Free sLAG3 was measured using the MSD platform as previously described. 1Briefly, biotinylated capture mAb was loaded on streptavidin-coated MSD plates.Following incubation at room temperature (RT) for 1 to 2 hours, the plates were washed with PBST and blocked with Blocker Casein for 2 to 2.5 hours at RT.Following plate wash with PBST, the standards and QCs were added to each plate, diluted in MSD Diluent 3/Reagent Additive 1 and incubated at RT for 1 hour.The plates were then washed with PBST and the sulfo-tagged detection antibody was added to the plate and incubated for 1 hour at RT.Following the final plate wash with PBST, MSD Read Buffer T was added to each well and the plates were read on an MSD SECTOR ® Imager 6000.Data analysis was performed using GraphPad Prism.A major assumption in CFCA is that analyte binding is at least partially mass transport limited (MTL).This is achieved by saturating the sensor surface with ligand (Figure S1A).However, if there is a marked loss in conjugated ligand activity between cycles (e.g.nonideal regeneration conditions), later cycles may not have the same degree of MTL as earlier cycles, skewing the resulting active concentration measurement.Previous work has suggested that comparable CFCA results can be attained by capturing the ligand on the chip each cycle prior to analyte injection. 2To increase robustness and avoid regeneration condition scouting, this study sought to load the antibodies of interest onto ProteinA and/or ProteinG sensor chips.However, the default CFCA form factor is stated to only be accurate for CM5-based chips, not including other sensors.

Flow cell form factors can be calibrated using NISTmAb
To first assess the degree of inconsistency across sensor types, the reported default active concentrations of NISTmAb were compared across both flow cell pairs on a ProteinG sensor, ProteinA sensor, PrismA sensor and CM5 sensor with ProteinG pre-conjugated.Each CFCA run was performed with NISTmAb at 4 dilutions (50nM, 25nM, 5nM, and 2nM) and at two flow rates (5µL/min and 100µL/min), injecting only on the second (sample) flow cell.NISTmAb is a stable and highly characterized monoclonal antibody reference standard, which made it an ideal choice to compare sensors. 3,4 he generated data was processed with Biacore T200 Evaluation Software to perform default calibration-free concentration analysis fitting with reference subtraction.Using MALDI-TOF, the molecular weight of glycosylated NISTmAb was found to be approximately 149kDa.Since the NISTmAb hydrodynamic radius is known to be 4.98nm by the certificate of analysis DLS measurement, 4 the Stokes-Einstein equation was employed to define the diffusion coefficient in water (similar viscosity to PBST) at 20°C: where Dt is the diffusion coefficient, k is the Boltzmann constant, T is the temperature in Kelvin (293.15K),η is the viscosity (water: 1.024 mPa*s at 20°C), 5 and Rh is the hydrodynamic radius, so the diffusion coefficient of NISTmAb at 20°C is 4.211E-11 m 2 s -1 .
As Cytiva mentions, different sensor chips reported different active concentrations for NISTmAb.However, even the CM5 chip with ProteinG conjugated to it, which should in theory be in agreement with the hardcoded form factor for the CM5 chip in the Evaluation Software, 6 did not report an active concentration close to the expected 67.15µM (10.003mg/mL).Interestingly, flow cell pairs on the same sensor chip on the same Biacore T200 system reported different active concentrations for NISTmAb (Figure S1B), using CM5 sensor chips or other sensor chips.Adding to these inconsistencies, the mass transport rate constant kt (which is calculated from the form factor 6 ) has been predicted to fluctuate by up to 15% between Biacore instruments but the influence of other components (e.g.sensor chip type & lot) was not as clear. 7ile seemingly antithetical to a technique characterized as "calibration-free", a method was therefore developed to re-calibrate the form factors of each Series S sensor chip used in this study.This was deemed necessary, as variability in the experimental form factor would affect the kt parameter and, by extension, the reported active concentration. 6Given that NISTmAb is a highly characterized standard with a well-defined concentration, it was treated as a gold standard to adjust the form factor of each flow cell pair per chip, with an assumed 100% activity in its Fc domain.While the immunoreactive fraction of monoclonal antibodies is often considerably lower, IRF assays are designed to report specifically on the paratope activity. 8,9 ince NISTmAb was likely affinity purified against its Fc domain during manufacturing (e.g.Protein A chromatography), molecules with an inactive Fc domain were likely discarded. 10onsidering the CFCA binding rate equation under mass transport that was previously defined 11,12 and showcased below, chip-specific calibrations could be calculated by assuming that the default form factor (G) was the sole source of error skewing the reported NISTmAb active concentration.The dR/dt, Lm, MW, Lr, kd, and R values should remain consistent whether the algorithm was fit with the default (incorrect) form factor or adjusted (corrected) form factor, since the intrinsic properties of the molecular interaction and the raw NISTmAb CFCA data would not change based on how the data was analyzed.With all variables other than the form factor (G) and active concentration (Abulk) being static, setting a "default" and "adjusted" version of the right-hand side of the above equation condenses to:

𝑑𝑅
where the default form factor,   , was 0.81 (times 10 9 but this power-scaling was not accessible in the software) 6 and the molar concentration of NISTmAb, [   ], was determined to be 67.15µMgiven that the CoA-reported concentration of 10.003 mg/mL took glycan mass into account. 3NISTmAb CFCA runs were then re-evaluated with the adjusted form factors, demonstrating concentration agreement across flow cells and across sensor types (Figure S1B).However, it is important to note that the accuracy of a CFCA measurement is reliant on the accuracy of the measured molecular weight and diffusion coefficient.Reprocessing the NISTmAb ProteinG FC2-1 data with theoretical errors in the MW or D value markedly changed the reported active concentration of NISTmAb (Figure S1C).Therefore, the most precise methods available to measure MW and D of the analyte should be utilized to avoid propagating errors to the active concentration measurements.Given that SPR is linearly related to mass bound to the chip, relevant PTMs such as glycosylation should be included in these measurements.The CFCA and purity data in Figure S4A were compared using a two-tailed Pearson correlation, additionally fitting the data with a simple linear regression with 95% confidence intervals (dotted lines).

Supporting data for
Table S1: Quantification of the variance and deviation of sLAG3 QCs of each lot vs the Lot3 standard curve.Related to Figure 3C and Figure S7, this data demonstrates the nominal concentration of each QC that is prepped, the back-calculated concentration based on the Lot3 standard curve interpolation, the mean %CV of each QC in each set, and the mean %deviation of each back-calculated concentration from the nominal concentration.These values were determined for both the Bradford-based total concentration of each lot and the intersection active concentration of each lot.All QC pairs that passed the acceptance criteria (<25% %CV between pairs) were included in this analysis.Coefficients of variation did not change between the Bradford and active concentrations because both datasets are using the same MSD dataset.So, the differences in %Deviation are likely due to increased relative accuracy of the calibrator concentration.Table S2: P(D|C) comparison of the Rmax-based method to multiple variations of a bivalent analyte nonlinear least squares curve fit to SCK data.The Rmax-based P(D|C) values were calculated in the main text of the paper, with one concern being that the detection mAb binding injection may not have reached a steady state to be considered an Rmax value.Fitting curves would allow us to instead determine the P(D|C) value using the full kinetic trace.Locally fitting the rate constants and fitting the conversion factor for each lot did yield P(D|C) values higher than the Rmax-based technique, which agrees with the hypothesis that the detection mAb association had not reached steady state.Likewise, the globally fit rate on and off-rate constants with a conversion factor equal to the form factor times 1e3 calculated P(D|C) values higher than the Rmax-based technique.However, when the globally fit model was employed with a conversion factor equal to the form factor * 1e9, 6 the P(D|C) values were curiously low.Overfitting may be an issue with how many variables are manipulatable by the NLLS engine in this model.However, in both globally fit models, the P(D|C) values were Lot3 < Lot2 < Lot1, like the Rmax-based technique.3 SUPPLEMENTAL DISCUSSION

Alternative methods to determine the intersection active concentration of a calibrator
Protein A/Protein G chips were attractive for this initial study since their robust regeneration conditions and reusability allow for direct calibration of each individual chip flow cell and assessment of the active concentration of multiple antibody epitopes per biomarker.However, initial assessment of CM5 chips indicates that the form factor may not vary considerably between chips in the same lot, so that the NISTmAb-calibrated form factor of one CM5 chip may be applied to another chip conjugated with an antibody of interest.While the sLAG3 detection mAb does not bind Protein G well, direct conjugation of the capture mAb or Biotin CAPture kits may be preferrable for other sandwich immunoassay pairs where both mAbs bind Protein A/G to better isolate biomarker-specific signal during the secondary mAb injection.One requirement being that the regeneration conditions would have to be near complete to avoid detection mAb binding biomarker from a previous CFCA cycle.However, more work is required to quantify the form factor variation between chips to ensure form factor consistency from chip to chip.
If each chip must be individually calibrated, there may be cases where the capture mAb cannot saturate a ProteinG/ProteinA surface fully to prevent detection mAb:ProteinG/A binding, potentially obscuring the P(D|C) calculation.Therefore, three additional strategies are outlined for determining the intersection active concentration.(1) Use a ProteinG, ProteinA, or similar antibody-binding standard at a well-defined concentration to calibrate the CM5 chip with the capture mAb already conjugated to the second flow cell (reverse of the NISTmAb calibration) to directly calibrate a CM5 chip.(2) A detection scFv/Fab could be employed to avoid any ProteinA/G binding using the same protocol as described above.(3) A CFCA (without a second mAb injection) can be performed loading capture mAb alone, detection mAb alone, or a 1:1 mixture of capture and detection mAb onto the ProteinA/G chip.This mixture may report on the union of the capture and detection epitopes using the default CFCA algorithm if the MTL was near saturation so that the rate of binding of a biomarker that has both epitopes would be equivalent to the rate of binding of a biomarker with only one of the two.Accurate quantitation of the individual epitope active concentrations and the union active concentration could then derive the intersection active concentration.
( ∩ ) = () + () − ( ∪ ) (S4) However, CFCA does not necessarily reach a saturated MTL in practice. 12It is not yet clear if a more advanced algorithm could account for the differences in MTL in a mixed antibody loaded on a chip to accurately quantitate the union/intersection of the two active epitopes.Additionally, this method assumes that binding the capture mAb does not alter the detection epitope accessibility, and so may be less applicable to harmonizing calibrator lots for sandwich immunoassays than the former method.

Areas for method improvement to enhance measurement accuracy
(1) Given the reliance on molecular weight in CFCA calculations, the most accurate technology available should be employed to determine the MW of the recombinant biomarker and secondary mAb.However, highly accurate MW measurements (such as ESI-MS) highlight a potentially flawed assumption in the CFCA algorithm: glycoproteins often have heterogeneous molecular weights, but the CFCA algorithm assumes a single species of known diffusion coefficient and molecular weight.Here, using the average MW of a glycoprotein calibrator appears to produce reliable CFCA results that far outperform total protein concentration methods to unify calibrator lots for a sandwich immunoassay.However, accounting for heterogeneous species may help improve accuracy further.
(2) In this work, the diffusion coefficients were theoretically calculated from the molecular weight instead of experimentally determining both.However, this work demonstrates that errors in the diffusion coefficient value can propagate to errors in the active concentration calculated from a dataset.Since the theoretical methods make assumptions about protein shape and specifically use water instead of the assay conditions, it may be preferrable to use an experimentally determined diffusion coefficient for CFCA if absolute accuracy is desired.But if the goal is instead just to harmonize calibrator lots, a theoretical diffusion coefficient may suffice, as the errors in the assumptions would likely be equivalent across all calibrator lots.Additionally, in considering the heterogeneity of glycoprotein molecular weights, it is difficult to envision an experimental method that would be able to accurately measure the diffusion coefficients of each protein species.Whether experimental or theoretical, it is important to note that including the weight of PTMs such as glycosylation in the molecular weight should be strongly considered when defining the MW and D values for the CFCA run, so that the CFCA run calculates a relatively accurate molar concentration for each calibrator lot.
(3) Instead of fitting the NISTmAb data by the default CFCA method to then estimate a new form factor manually, it may be advantageous to directly fit the form factor of the chip flow cell with a static reference standard active concentration.However, the way in which the present study calibrates the system by modifying a single form factor per flow cell may be an oversimplification for two reasons: In this example, a single cycle kinetics experiment followed each cycle of the CFCA to provide multiple detection mAb concentrations to fit (Figure S8A).The model was first fit locally by each on-and off-rate, diffusion rate constant, and an unknown (to be fit) conversion factor (Figure S8C, Table S2).The model was then adjusted to globally fit the on-and off-rates, locally fit the diffusion rate constants per lot, and include a constant conversion factor that had been calibrated for each sensor chip using NISTmAb.Multiplying the form factor by the base multiplier used for CFCA analysis (1e9) to convert molar concentrations to RU yielded P(D|C) values lower than the Rmax-derived P(D|C) values, which was not expected.But when this multiplier was lowered to be more in line with the locally fit conversion factors (1e3), the P(D|C) values were again above the Rmax-derived values and Lot3 < Lot2 < Lot1, like the Rmax-derived values.
However, one benefit of CFCA over conventional bridging studies is that the method does not require direct comparison between lots to derive an active concentration and unify lots.Fitting the detection mAb binding curves to a bivalent model locally (i.e. for each lot) risks unjustified lot-specific adjustment of the on-and off-rates, potentially reporting incorrect fraction activities.On the other hand, globally fitting the kinetic parameters to define the on-and off-rates of the mAb:antigen interaction is no longer measuring each lot's active concentration independently.It is therefore unclear how to apply bivalent analyte model curve fitting to potentially better harmonize calibrator lots while retaining the benefits of the above method (e.g. each measurement is lot-intrinsic).Given this, and the general concerns that the NLLS engine may have been overfitting the data given the large number of adjustable variables, the basic Rmax method was chosen to estimate the P(D|C) values for this initial study.

Figure S1 :
Figure S1: Calibration of Biacore sensor flow-cell pairs using the NISTmAb ProteinG/A interaction unifies the reported active concentrations.(A)A representative CFCA assay where an Fc-binding protein such as ProteinG is conjugated at high density to the surface and an antibody of interest is flowed over the surface, creating a concentration gradient based on its diffusion rate, used to calculate the active concentration.(B) The reported active concentrations of NISTmAb using different sensor types and flow cell pairs on the same Biacore T200 instrument using the default 0.81 form factor, two independent replicates (left panel).Dotted line represents the CoA-reported total concentration for NISTmAb.After applying adjusted form factors to respective chip flow cell pairs, the new reported active concentrations for NISTmAb are in strong agreement (right panel).(C) To see how strongly errors propagate in CFCA, the measured diffusion coefficient or molecular weight was modified (0.5X to 2X artificial error) and the corrected ProteinG 2-1 NISTmAb datasets were reanalyzed.Errors in the molecular weight appeared to have a slightly greater impact on the reported active concentration compared to diffusion coefficient.

Figure S2 :
Figure S2: Capture mAb bind ProteinG well, but detection mAb does not.(A) The biotinylated capture mAb loads well (>7000RU) onto a ProteinG sensor chip and remains tightly bound at the end of the loading step.The MTL therefore does not change considerably with little mAb dissociating during the CFCA step.Regeneration of the ProteinG chip with pH 1.5 glycine removes the mAb:biomarker complex from the surface so the next CFCA cycle can be reloaded with fresh mAb.(B)The detection mAb for the sLAG3 LBA does not load on ProteinG well and dissociates relatively quickly from the surface.An in-line CFCA run does report an active concentration for the detection mAb epitope, but (C) the detection mAb active concentration measured by direct amine coupling the sulfo-tagged detection mAb to the chip did not completely agree with this ProteinG loaded CFCA method (three independent replicates of the Detection-ProteinG, two independent replicates of the Detection-Conjugated (CM5)).(D) ProteinA and PrismA chips were also tested to see if the detection mAb would bind with higher affinity to another antibody binding protein, but similar fast off-rates were seen with all three sensor types.

Figure S3 :
Figure S3: The purity of sLAG3 lots used in this study.(A) SDS-PAGE of 2ug, 4ug, and/or 6ug of the recombinant sLAG3 lot.The main species runs between 54 and 71kDa band in the denatured form.(B) SEC-HPLC of each lot of sLAG3.(C) The purity of the main species was calculated by integrating the main peak vs all other protein staining 13 or A280 signal for SDS-PAGE (4ug lane) and SEC-HPLC respectively.

Figure S4 :
Figure S4: Reagent purity moderately correlates with epitope-specific activity, but CFCA activity still varies between reagents with high purity.(A) sLAG3 reagents produced at BMS or purchased from vendor suppliers were assessed for purity, capture epitope active concentration, detection epitope active concentration, and total concentration.The total concentration of each reagent was assessed by BCA following the manufacturer protocol for comparability.Purity was assessed by SEC-HPLC, quantifying the area under the curve for the expected main protein peak (not in the void volume or near the salt peak) compared to the total area under the curve.Capture mAb CFCA was performed by loading on a ProteinA chip while detection mAb CFCA was performed by directly conjugating detection mAb to a CM5 chip.Vendor A -Reagent 1 was expressed in E. coli while all other reagents were mammalian-expressed. (B) The CFCA and purity data in FigureS4Awere compared using a two-tailed Pearson correlation, additionally fitting the data with a simple linear regression with 95% confidence intervals (dotted lines).