Analyzing protein conjugation reactions for antibody‐drug conjugate synthesis using polarized excitation emission matrix spectroscopy

Abstract Antibody‐drug conjugates (ADCs) are promising anticancer therapeutics, which offer important advantages compared to more classical therapies. There are a variety of ADC critical quality attributes (CQAs) such as the protein structure, aggregation, and drug‐to‐antibody ratio (DAR), which all impact on potency, stability, and toxicity. Production processes can destabilize antibodies via a variety of physical and chemical stresses, and or by increased aggregation after conjugation of hydrophobic drugs. Thus, a proper control strategy for handling, production, and storage is necessary to maintain CQA levels, which requires the use of in‐process quality measurements to first identify, then understand, and control the variables which adversely affect ADC CQAs during manufacturing. Here, we show how polarized excitation emission matrix (pEEM) spectroscopy, a sensitive, nondestructive, and potentially fast technique, can be used for rapidly assessing aggregation and DAR in a single measurement. pEEM provides several sources of information for protein analysis: Rayleigh scatter for identifying aggregate/particle formation and fluorescence emission to assess chemical and structural changes induced by attachment of a linker and/or a small molecule drug payload. Here, we used a nontoxic ADC mimic (monoclonal antibody with linker molecule) to demonstrate efficacy of the measurement method. Emission changes caused via light absorption by the attached linker, allowed us to predict DAR with good accuracy using fluorescence signal from the final purified products (6% relative error of prediction [REP]) and also from unpurified alkylation intermediates (11% REP). pEEM changes could also be correlated with size (hydrodynamic radius, R h) and aggregate content parameters obtained from dynamic light scattering and size exclusion chromatography (SEC). For the starting material and purified product samples, pEEM correlated better with R h (R 2 = 0.99, 6% REP) than SEC determined aggregate content (18% REP). Combining both fluorescence and light scatter signals also enabled in‐process size quantification (6% REP). Overall, combining polarized measurements with EEM and Rayleigh scatter provides a single measurement, multi‐attribute test method for ADC manufacturing.

This is done to increase specificity toward tumor cells as well as improving pharmacokinetic profiles, and providing a wider therapeutic window (Wu & Senter, 2005). There are currently nine FDA approved ADCs with more than 80 molecules in clinical studies (Joubert et al., 2020). IgG1 is the most widely used antibody type for therapeutic purposes and also the most common protein found in ADCs in the market or in late-stage clinical trials (Joubert et al., 2020).
Chemical conjugation via lysine and cysteine are the most common ADC synthetic strategies (Joubert et al., 2020). Conjugation to the lysine amine group is widely used because it is relatively simple, often a single-step reaction. However, it generates heterogeneous products because IgG has approximately 80 lysine-derived amine groups (Mueller et al., 1988) of which approximately 10 are readily accessible for chemical modification. Cysteine conjugation is the major alternative, but because of the lack of free cysteine thiol groups in most proteins, the process usually first involves disulfide bond reduction under carefully controlled conditions to create sulfhydryl groups.
These free thiols are then available for conjugation to reagents containing groups like maleimide. This route usually restricts the attachment sites to eight, leading to more homogeneous product mixtures compared to lysine conjugation (Jain et al., 2015), although site-specific conjugation methods are now becoming available (Coumans et al., 2020).
Mishandling of proteins can lead to protein unfolding and aggregation, which can cause a loss in function and potentially cause immunogenicity issues for the patient (Sharma, 2007). In general, the most important critical quality attributes (CQAs) to be considered during protein modification are: homogeneity, purity, degree of conjugation, total protein concentration, and lot-to-lot variability of starting materials, intermediates, and final conjugated products. The CQAs ("Guidance for industry: Q8 [R2] pharmaceutical development, Guideline ICH Harmonized Tripartite," 2009) are the physical, chemical, or biological attributes of the drug substance/product known to impact product quality in terms of potency, pharmacokinetics, and toxicity (Alt et al., 2016;Raynal et al., 2014;Wagh et al., 2018). Protein and ADC analysis is technically demanding because of increased structural complexity compared to small molecules and the need to monitor both tertiary and quaternary structures. For ADC's, the problems are compounded by the fact that the small molecule payload also has to be characterized and that the structural changes caused by the payload (e.g., increased hydrophobicity) lead to products which are significantly more sensitive to aggregation.
Here, we investigated the use of polarized EEM (pEEM) to monitor an ADC reaction process with the key objectives being to: (1) measure variance of the mAb starting reaction materials, (2) monitor the course of the reaction and assess variance in the reaction intermediates, (3) predict final product drug to antibody ratio (DAR) during the alkylation reaction, and (4) measure variance and DAR of the final purified products. Most ADC analysis studies focus on characterizing final conjugate stability and DAR (Wakankar et al., 2011) with fewer looking at conjugation reaction monitoring.
UV-vis absorbance spectroscopy (Andris et al., 2018) was used to monitor DAR during conjugation of two different drug mimics to an engineered mAb. A RP-HPLC and TOF mass spectrometry-based method (Tang et al., 2017)  HPLC grade water (Fisher chemicals) was used for all solutions which were membrane filtered (0.10 μm) using Captiva filters before use.
A single 11.0 ml mAb aliquot (see Supplemental Information, for more details) was buffer exchanged to remove the formulation buffer to facilitate synthesis. The mAb in reaction buffer was then aliquoted into smaller vials (5.0 ml LoBind tubes) suitable for single experiments, refrozen, and stored at -70°C until required. This reduces and controls the number of freeze-thaw cycles ensuring that all samples in an experimental campaign have the same number of cycles. Absorbance spectroscopy was used to check the final concentration of mAb in PBS/EDTA and fluorescence measurements and dynamic light scattering (DLS) were also carried out (data not shown). Ellman's test (Supporting Information: Figure S1, Table S1) was used to verify the number of free thiol sites under the different conditions employed.

| Linker and conjugation
A big challenge when studying ADCs is the often very high potency of the payload drug requiring the use of strict safety protocols and controlled environments. Thus, the safest alternative for preliminary studies particularly for analytical method development is to use a nontoxic model with payload molecules that mimic the structure/ behavior of the real drug-linker which facilitates the safe study of all the key steps from starting material preparation to final purification of conjugated products. The drug mimic used here was selected to be similar (e.g., solubility, absorptivity) to common drug-linkers used in marketed ADCs but avoided toxicity issues. We used a molecule composed of valine, citruline, PAB-OH, with an active maleimide terminal group, which is a common commercial ADC linker (Joubert et al., 2020). We selected a nonfluorescent small molecule because we wanted to minimize interference, and only measure the changes in protein fluorescence (Supporting Information: Figures S2 and S3).
Conjugates were prepared by partial reduction of IgG disulfide bonds with TCEP. HCl for 2 h followed by alkylation with an excess of the "drug" linker for 2 h, at 20°C. The two-step reactions were conducted in 1 × 1 cm path length quartz cuvettes with slow stirring (using a flea magnetic follower in the cuvette with the sample holder stirrer) in the spectrometer (Supporting Information: Figure S3). A total of 24 reactions were performed, using 8 reducing agent TCEP concentrations from 0 to 50 molar excess (0, 1.25, 2.50, 5.0, 7.5, 10, 25, and 50) in triplicate. For all reactions, the same amount of "drug" linker (20-fold excess compared to mAb concentration) and NAC quencher (12-fold excess with respect to linker concentration) was used. The reaction mixtures were measured by absorbance and pEEM spectroscopy at 45-minute intervals during the reaction (designated Red1-/2-/3-IgG, Alk1-/2-/3-/4-IgG). The starting material (IgG-SM), reduced intermediate (Red3-IgG), unpurified final reaction mixture (Alk4-IgG), and partially purified final products (Pur-ADC) were also analyzed by DLS to look for aggregate formation (see Supporting Information: Tables S2 and S3). Because DLS, absorbance, and fluorescence measurements were made on the same cuvette sample we have confidence that all data relates to the exact same sample in terms of chemical and physical composition.
Unpurified reaction mixtures were transferred to Eppendorf tubes and then stored at -70°C until purification was undertaken, several days post reaction. A simple filtration-based purification (see Supporting Information) was implemented to remove unreacted small molecules (<10 kDa) and these partially purified samples (Pur-ADC) were characterized by absorbance and fluorescence spectroscopy, DLS, SEC, and sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE).

| Instrumentation and data collection
Absorbance spectra were collected using a Cary 60 spectrometer (Agilent) from the same cuvettes used for fluorescence measurements. Polarized EEM spectra were collected from 1 × 1 cm quartz cuvettes (Lightpath Optical) using a Cary Eclipse fluorescence spectrophotometer (Agilent) fitted with wire grid polarizers (Casamayou-Boucau & Ryder, 2017) and a temperature controlled multicell holder. pEEM spectra were collected over an excitation range of λ ex = 250-320 nm and λ em = 290-450 nm emission range (2 nm increments in each case) with 10 nm excitation/emission slit widths and a scan rate of 1200 nm/min. Samples were measured under the four different polarization configurations, as previously described (de Faria e Silva et al., 2020a, 2020b, but only the parallel polarized EEM (EEM || ) was used here because it is more sensitive to changes induced by conjugation and aggregation (de Faria e Silva et al., 2020a, 2020b), and does not need G factor correction, simplifying measurement. Data were blank subtracted before any data analysis using the corresponding buffer for each step (see Supporting Information). pEEM data were corrected for differences in instrument response using a correction factor calculated using a Spectral Fluorescence Standard Kit (Sigma, product No. 69336) (Resch-Genger et al., 2005), for a restricted λ em = 302-450 spectral range (this was because the standards only covered the 300-700 nm emission range). Following this, the Rayleigh scatter (RS) area was replaced with missing data, and the fluorescence signal smoothed using Savitzky-Golay filter. For most analyses, the smoothed spectra were normalized to the point of the maximum intensity. The extracted RS band (RS || ) is the first order Rayleigh scattering spectrum, and from this, the area under the curve (RS volume) was calculated. Both of these parameters were used for qualitatively assessing protein aggregation. Descriptions of sample types, data collected, and datasets used for modeling are provided in Supporting Information: Table S2.
Chemometric analysis was performed using PLS_Toolbox8.2.1 ® , MATLAB (ver. 9.1.0), and in-house written codes. Exploratory data analysis was carried out using ROBust principal component analysis (ROBPCA), which minimizes the effect of outliers (Hubert et al., 2005) compared to classical PCA. It was implemented using the Venetian blind method (four splits) for cross-validation and the root mean square error (RMSE) values to select the optimum number of PCs. Quantitative modeling for DAR and percent of aggregates was implemented using unfolded PLS, u-PLS, (Haaland & Thomas, 1988). Model performance was assessed by coefficient of determinations (R 2 ), RMSE, and relative error of prediction (REP). The elliptical joint confidence region (EJCR) test was used to compare the accuracy and precision of different models at a 95% confidence interval (Mandel & Linnig, 1957). iPLS was used for variable selection and it works by comparing u-PLS performance with and without each variable (here each excitation wavelength), selecting variables that return lower cross-validation errors (Nørgaard et al., 2000). The pooled relative standard deviation (RSD P ) of SEC and DLS parameters was calculated using the RSD of values obtained from replicate measurements (see SI).
SEC was performed using a 300 × 7.8 mm mAb PAC-SEC1 column (ThermoFisher) with a 5 µm particle size with an Agilent 1260 HPLC system equipped with a DAD detector. Solutions were filtered using a 0.20 μm Captiva PES filter and 10 µl of sample were injected in triplicate at 30°C with 50 mM Sodium Phosphate pH 6.8 + 300 mM NaCl buffer as the mobile phase, and a 0.8 ml/min flow rate. The important output parameter used for characterization and modeling was the %Agg. value which for this study was defined as: the ratio in % terms of the sum of the area under the curve (AUC) of the aggregate peaks (all those at Rt < 10 min) divided by the total peak area in the SEC chromatograms.
Because no fragment peaks were observed and the buffer components peak (~14 min) was excluded, the percent of aggregates corresponded to 100-(% Monomers). Here, the aggregate peaks occur between Rt =~7 and 10 min. However, we do have to note that the %Agg. values here represent the soluble species and do not take into account large aggregates which might precipitate out or otherwise be lost. DLS data were collected at 20°C, after filtration (0.20 μm PES filter), using a Malvern Zetasizer Nano ZS (173°detection angle). Each sample was measured five times (each measurement was an average of 10 runs of 10 s duration) in disposable plastic cuvettes. Z-average size (radius) and polydispersity index (PdI) were obtained by cumulants analysis while the hydrodynamic radius (R h ) was extracted from the distribution fit (both from the Intensity PSD), using the Zetasizer software, ver. 7.13 (Malvern Panalytical). SDS-PAGE was performed according to the BioRad TGX Precast Gels ® specifications (see Supporting Information for details).

| RESULTS AND DISCUSSION
Starting materials (IgG-SM) and partially purified products (Pur-ADC) were first characterized using conventional methods to measure aggregation by SEC, size by DLS, and DAR via UV-visible absorbance spectroscopy. These data were then used to explain the observed spectral changes in the pEEM spectra before using multivariate data analysis to build predictive models for DAR and aggregation content.

| DAR quantification
As all reactions were undertaken in cuvettes, we expected high recovery, however, absorbance spectroscopy suggested a decrease in total protein concentration over the course of the reaction (Figure 1a).
Average absorbances of 1.08 ± 0.02 (IgG-SM), 1.04 ± 0.01 (Red3-IgG), and 0.98 ± 0.02 g/L (Alk4-ADC) were measured, with a decrease due mostly to dilution via reagent addition. Recovery after purification was 75% which was a significant protein loss experienced during handling and reconstitution of purified material which is to be expected with the small reaction volumes used. The final purified solutions used for measurements had a concentration (measured by absorbance) of 1.05 ± 0.02 g/L. DAR can be considered one of the most important CQAs because it determines the final product potency and stability.
Absorbance spectroscopy is often used to determine DAR if drug and mAb have different absorption maxima (Chen, 2013;Wakankar et al., 2011). However, free drug is a problem if its absorbance spectrum overlaps that of the conjugated drug, which could lead to DAR overestimation. Here, we collected absorbance spectra from the purification washings to ensure complete free "drug" removal. DAR was calculated from the ratio of the concentrations of "drug" and mAb (see Supporting Information) which were determined using the extinction coefficients at the two wavelengths of maximum absorbance (Hamblett et al., 2004). It confirmed that varying degrees of conjugation, 1.0 ± 0.0 to 8.2 ± 0.1 DAR (Table 1) (Table 1 and Supporting Information: Table S3) showed that all IgG-SM had very similar aggregation profiles (99.2 ± 0.1% monomer) and sizes. These measurements also showed that aggregation increased after purification (87.9 ± 13.3% monomer).

| pEEM measurements
3.2.1 | Fluorescence spectroscopy Normalized EEM ǁ difference spectra showed significant intensity changes (at λ ex < 260 nm) of up to 20% for the highest DAR samples ( Figure 2a) which was probably related to secondary changes in fluorescence induced by varying Inner filter Effect (IFE) from light absorption of the attached linker, rather than large changes in intrinsic protein emission. IFE is usually considered a problem because it causes a nonlinear dependence between intensity and concentration. However, it is a valuable source of information about protein-based samples because of the high sensitivity to changes in sample composition (Panigrahi & Mishra, 2019;Ryder et al., 2017) and can be incorporated into variance assays once identified and taken into account.
To better interpret the spectral changes and their significance we used ROBPCA, which generates loadings plots to provide information about the types of spectroscopic changes occurring and scores plots that provide data on the magnitude/significance of these changes.
ROBPCA of IgG-SM + Pur-ADC samples (Table 2, Figure 2b,c) easily discriminated conjugated from nonconjugated mAb with most separation along PC1 (78% explained variance) which is caused by F I G U R E 1 (a) Normalized absorbance spectra of purified products; (b) sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) gels, coomassie blue stained, of Pur-ADC samples (first reaction replicate only, see Supporting Information: Figure S4 for replicate measurements); (c) normalized size exclusion chromatography (SEC) chromatograms; (d) normalized DLS intensity obtained from analysis of starting materials (IgG-SM) and purified products (Pur-ADC); the data shown is the mean of the triplicate reactions carried out for each condition.
T A B L E 1 Average sample data: nominal and PLS predicted DAR, percent of aggregates (from SEC), hydrodynamic radius, R h (from DLS), and Rayleigh scatter (RS) volumes, see also Supporting Information: Figure S5 (

| DAR quantification
Based on these observations, we built a u-PLS model (Table 3)

| Aggregation analysis and prediction
Aggregation is a critical CQA to be monitored during ADC production and it is thus important to monitor changes in both tertiary and quaternary structure. An increase in insoluble aggregates was

T A B L E 3 Summary of u-PLS DAR prediction results using normalized spectra of alkylation intermediates and Pur-ADC with and without iPLS variable selection
Absorbance spectra pEEM Note: In all cases, the total sample number was 24, which were split into calibration (n = 18) and external validation (n = 6) sets (the same samples in all cases). The best model are shown in bold.
indicated by a general rise in the ultraviolet aggregation index Aggregate content (%Agg. from SEC measurements) was better correlated to fluorescence than to scatter signals (but best when both signals were used), which was probably due to the weaker scatter contribution of the nm-sized soluble aggregates (most samples were T A B L E 4 Summary of u-PLS modeling results obtained for percent of aggregate and R h quantification using the fluorescence signal from EEM ǁ , the RS ǁ (not normalized), and the combined fluorescence and scatter signal. Note: For u-PLS modeling EEM spectra were normalized to maximum intensity, EEM with RS were normalized to λex/em 294/336 nm, which is a data point with high intensity value and low StDev. Abbreviations: ADCs, antibody-drug conjugates; EEM, excitation emission matrix; RMSE, root mean square error. a The same samples were used for calibration and prediction in each sample group to enable comparison between various measurements/parameters. composed of monomers, dimers, and trimers according to SEC), and also because of the noisy/variable absolute scatter signal measured. This is clearly seen in EJCR plots (Figure 4c) which confirmed that the model using both EEM ǁ + RS ǁ data with variable selection was best.
Here, with a relatively large sample set and using a mAb which better represents therapeutically relevant molecules, pEEM can be used for the quantification of soluble aggregates (as determined by SEC measurements) down to~1% ( pEEM spectral variance correlated better with R h , with the RS || based models being slightly better than those using EEM ǁ spectra (Table 4). The major issue with using non-normalized RS data for modeling aggregation was noise and poor reproducibility because the model relied on absolute intensity changes. We attempted to quantify aggregation using normalized RS || (by either band maximum or area), but this produced worse results (data not shown). Despite this, the error was relatively low and could be improved by spectral averaging. Another solution is to simply use normalized full pEEM spectra to minimize unwanted signal fluctuations (fluorescence and RS), which should be a more reproducible source of size-related changes. Here, a specified spectral data point (λ ex/em 294/336 nm, point of maximum fluorescence intensity) was used for normalization.
Using this data resulted in better R h models with good correlation coefficients (R 2 > 0.95) and a 2%-3% decrease in REP for EEM ǁ + RS ǁ compared to RS ǁ . Larger improvements were obtained for aggregate content (Table 4) using EEM ǁ + RS ǁ , with R 2 > 0.98 and a REP decrease of 10%-13%. iPLS selected variables and loadings (Supporting Information: Figure S6) showed that areas of the fluorescence signal and scatter band were important for size prediction of ( Figure S6d,g,h) and aggregate content (Supporting Information: Figure S6o,l,p), with the RS signal always being the strongest contribution. EJCR plots (Figure 4d) confirmed that the best model used both EEM ǁ + RS ǁ data and variable selection.

| Spectral changes
To investigate reaction-dependent spectral changes, we collected pEEM spectra every 45 min. during the reaction (nine measurements in total): one IgG starting material (IgG-SM), three during reduction (Red1/2/3-IgG), four during alkylation (Alk1/2/3/4-IgG) and one of the partially purified product (Pur-ADC). The mean (Figure 5a with a maximum of 2% for a single data point, and a significantly higher variance among alkylation intermediates and purified product (EEM RSD = 8.8 and 4.4%, n = 96 and 24, respectively). These were significant emission changes compared to 1% changes previously obtained for control measurements (de Faria e Silva et al., 2020b). It was not possible to collect more spectra during the reaction because the scanning-based spectrometer took~7 min. to collect one full spectrum, which limited the number of sample points available for analysis and analysis of reaction rates.
We assessed the changes at each reaction step using StDev calculated after successively adding data collected at each data point for example, for Step 3 of reduction StDev was calculated using IgG-SM + Red1-IgG + Red2-IgG + Red3-IgG samples ( Figure 5). This again shows that alkylation was responsible for most of the reaction spectral variance, which increased from 8.2% to 11% (EEM RSD ) from the first to last sampling point. The equivalent variance (i.e., reproducibility), between replicates of same reaction at a specific timepoint, was much lower (<2%). Overall, we can say that the spectral changes induced by alkylation were large, significant, and thus suitable for ROBPCA and quantitative modeling (Table 2).
ROBPCA ( Figure 6) was then used to better understand the source of these spectral changes and three ROBPCs were required to explain the spectral variance when models were built using all samples. This ROBPCA model contains more complex samples than that depicted in Figure 2  We attempted DAR prediction using pEEM spectra collected during Alkylation, before purification (Table 2). Here, absorbance spectroscopy was ineffective because of spectral overlap between free and conjugated linker, which resulted in very small spectral differences between in-reaction samples with different DAR (the same amount of linker was added to all reactions). u-PLS results suggested however, that there were small variances in pEEM spectra of the alkylation intermediates (Alk1-4), which correlated with DAR.
Alk3-IgG had a better correlation with DAR (R 2 > 0.92, REP = 11%) compared to Alk1/Alk2 because it was later in the reaction whereas the poorer correlation obtained with Alk4-IgG seems to be caused by interference from addition of the reaction quencher, NAC. When Alk1-4 absorbance spectra were used for DAR quantification, fairly good calibration results were obtained (relative error of calibration [REC] = 11%), but prediction errors were significantly worse (REP > 28%) compared to pEEM. Quantification here seems to be based on small absorbance changes at~310 nm (and also IFE) due to  Table S2, for details of the sample sets.
loss of conjugation in the maleimide linker (-C=C-C=O → -CH-CR-C=O) after alkylation (Liu et al., 2013), with the high quantification errors caused by increased scattered light at~310 nm.
Overall, these results suggested that EEM ǁ was the better reaction monitoring option because of significant alkylation induced spectral changes. However, when we looked at the normalized score changes (Supporting Information: Figure S8) there was very little change from Alk1 to Alk3 which suggests that alkylation was faster than anticipated and was nearly complete before the first pEEM measurement was completed.

| Physical stability (in reaction)
Previously we assessed polyclonal IgG solution quality using pEEM which indicated that most of the variance originated from longer wavelength emission (de Faria e Silva et al., 2020b). Here, long wavelength emission did not play a role in discriminating solutions according to aggregation which suggested that the previous observations were related to lower purity and higher variability of polyclonal IgG. The emission ratio between solvent exposed and buried Trp residues (I 350 /I 330nm using 296 nm excitation) is commonly used to assess protein stability (Beckley et al., 2013). Here, the ratio did not correlate with R h , which is clear evidence that there were no major structural changes like unfolding. As expected, however, the reaction/purification did induce some physical sample changes as observed by DLS in the reported R h (Table 1) and Z-average size values (Supporting Information: Table S3).
Although disulfide bond reduction could increase flexibility, and linker addition causes a small increase in product mass (~3% for a DAR of 8) these did not cause large changes in the R h values: R h (IgG-SM) = 6.4 ± 0.3 nm, R h (Red3) = 6.1 ± 0.1 nm, and R h (Alk4) = 6.7 ± 0.1 nm (Supporting Information: probably due to a larger sample set size and greater protein size variability when these samples were included. Overall, use of full (EEM + RS) ǁ spectra enabled size prediction with low errors (REP < 8%) and good correlation with nominal values (R 2 Pred = 0.97), implying a more robust size change assessment for complex in-reaction samples.

| Reaction end point determination
One goal of reaction monitoring is to accurately determine reaction end points, and this usually involves collecting multiple spectra throughout the reaction and then extracting kinetic and end-point data. Here, the long spectral collections (because of th scanning spectrometer design) times coupled with the fast reaction kinetics, prevented this and thus we investigated a different approach. We used nonlinear, support vector machine (SVM) classification (Supporting Information: Table S4) to quantitatively assess if these low numbers of EEM ǁ measurements could classify samples according to reaction stage and the ultimate product DAR (i.e., predict endpoint for different performing reactions). Samples were split into calibration and validation sets and 10 different classes were used for classification: IgG-SM and reaction intermediates/ Pur-ADCs from reactions producing low (1.0-2.0), medium (3.2-4.3), and high (5.9-8.2) DAR. The classification errors indicated a good performance for successfully classifying both reaction stages and DAR (errors lower than 10% for all the classes) for the medium and high DAR ranges. The low DAR-related classes (and the IgG-SM) showed lower specificity, which was probably due to the very small spectral changes induced by the lower number of linker molecule attachments.

| CONCLUSIONS
Using non-destructive pEEM measurements for reaction monitoring of the important mAb linker conjugation process and product variance analysis has been demonstrated. Here, although the model reaction was limited to looking just at the linker addition, in small scale, and with insufficient time resolution, it does show the very significant spectral differences between each stage in the reaction process. Although spectral changes were relatively small (EEM RSD = 10 for all samples, n = 216), they were significant and reproducible. By using the full pEEM spectral information, that is both the scatter and fluorescence signals, one is able to build quantitative models for predicting DAR using the alkylation intermediates (R 2 > 0.90 and REP 11%) and that correlate with aggregation and size (R h , R 2 PRED = 0.97 and REP = 6%) parameters extracted from SEC and DLS measurements, respectively. We also showed that simple UV-AI measurements were poorly correlated with both %Agg. from SEC, and DLS-derived size parameters, confirming its unsuitability for ADC reaction monitoring with these levels of soluble aggregates.
For reaction monitoring, the spectral profile changes observed here although small, were still significant considering that we only used normalized data. The u-PLS modeling showed that it was possible to generate in-process accurate correlations for both size and aggregation parameters when using the combined scatter and fluorescence signals. For wider application using payload molecules with different absorption spectra that overlap less with the protein absorption spectra, and potentially more with protein emission spectra, we suggest that fluorescence spectral changes will be larger and thus easier modeled using these techniques. Larger signal changes should make DAR quantification and reaction monitoring easier to implement and potentially more accurate and this is the focus of future studies.
Although fluorescence and pEEM are not currently, widely used analytical techniques for monitoring ADC synthesis and manufacturing processes (it is mostly used to assess the effect of conjugation on higher order structure and stability [Turecek et al., 2016;Wakankar et al., 2011]), these results are very promising. However, this alkylation reaction was too fast for this type of spectrometer and was substantially complete by 10 min when the first in-process pEEM data was collected. For more accurate, continuous in-process monitoring we require either faster data acquisition using spectrometers with multichannel detectors or slower reactions, both of which are being investigated, along with the use of more therapeutically drug linker moieties. Overall, this use of the full pEEM measurement shows considerable promise as a robust PAT tool for ADC manufacturing.