Transforming Downstream Bioprocesses With Real-Time Monitoring

Enhancing Downstream Bioprocesses with Raman Spectroscopy Compendium of application notes Compendium The accurate measurement of protein and excipient concentration in biomanufacturing processes is essential to production of therapeutics like monoclonal antibodies (mAbs), nucleic acids and other important drug products. Raman spectroscopy, a Process Analytical Technology (PAT) based on specific molecular signatures, is particularly suitable for monitoring complex biological processes characterized by multi-component interactions and dynamic changes. In this compendium, we present a collection of studies demonstrating the application of Raman spectroscopy towards monitoring downstream bioproduction processes. When deployed as a PAT, Raman spectroscopy provides both qualitative and quantitative feedback, in real time, of proteins, amino acids, and even buffer quality. As the application notes contained herein show, Raman spectroscopy’s intrinsic capabilities allow it to be highly useful in bioprocess monitoring even when the presence of complex biomolecules or interferents like clarified harvest might render other analytical methods ineffective. Using a Process Raman Analyzer as an In-line Tool for Accurate Protein Quantification in Downstream Processes 4 A Classical Least Squares (CLS) Approach for Protein Quantification in Downstream Processing Using Raman Spectroscopy 10 Direct in-line quantification of titer in clarified harvest using Raman spectroscopy 16 Raman-based Accurate Protein Quantification in a Matrix that Interferes with UV-Vis Measurement 21 Process Raman as a comprehensive solution for downstream buffer workflow 27 Protein secondary structure analysis using in-line process Raman and offline SERS surface 33 Table of Contents Using a Process Raman Analyzer as an In-line Tool for Accurate Protein Quantification in Downstream Processes Application note Authors Michelle Nolasco¹, Kristina Pleitt¹, Nimesh Khadka² ¹ BioProduction Group, Thermo Fisher Scientific, St. Louis, Missouri USA ² Analytical Instrument Group, Thermo Fisher Scientific, Tewksbury, Massachusetts, USA Significance In-line and real-time measurement of accurate protein (monoclonal antibody, or mAb) concentration has been demonstrated over a wide dynamic range (0-135 g/L) using the Raman process analyzer for the ultrafiltration/ diafiltration (UF/DF) as shown in Figure 1. The methodology described in the paper offers actionable results for downstream processing, providing valuable insights for monitoring and controlling protein concentration. Additionally, we have demonstrated that the Raman models developed for measuring mAb concentration are transferrable to a similar class of monoclonal antibodies. Introduction Therapeutic proteins (e.g., monoclonal antibodies, insulin, Fc-fusion protein, antibody-drugs conjugates, hormones) are used to treat a variety of conditions and diseases such as cancer, immunologic diseases, communicable diseases like COVID-19, and many others.¹ The efficacy and functionality of a therapeutic protein are dependent on dosage and its physical state and structure. Abrupt changes in the environment of the protein, such as shifts in pH, temperature, shear force, or chemical modification, can induce conformational changes within the protein. The structural changes may lead to protein denaturation, aggregation, and even degradation—all of which can lower the efficacy of the therapeutic protein or trigger negative health consequences for the patient. Thus, ensuring the quantity and quality of therapeutic proteins throughout the manufacturing process—from production, through the fill and finish stage, and during storage— is of paramount importance.² Figure 1. Diagram showing integration of process Raman in a UF/DF process using a flowcell probe for in-line estimation of accurate protein concentration. Buffer Pump Pump Membrane Retentate Permeate Outlet FlowCell Process Raman Ultrafiltration/ Diafiltration Accurate Protein Concentration Therapeutic proteins like monoclonal antibodies are produced in a bioreactor where cells are maintained at optimum conditions by controlling nutrition and physical factors like pH, temperature, and dissolved oxygen. The bioreactor’s controlled environment leads to improved monoclonal antibody production. Following the upstream process, operations like centrifugation, filtration, chromatography, viral inactivation, concentration, buffer exchange, formulation, and fill and finish are performed to purify the monoclonal antibodies and formulate them into drug products. These operations are collectively categorized as downstream processes (Figure 2). In this study, we placed a Thermo Scientific™ MarqMetrix™ AllIn-One Process Raman Analyzer in a UF/DF run, as described in Figure 1, for use as an in-line process analytical technology (PAT) solution for measuring mAb concentration. In-line PAT enables real-time monitoring and control of critical process parameters that are key to reducing batch-to-batch process variability and ensuring uniformity of products. A well-controlled process also implies high efficiency, product quality, and minimized manufacturing costs.³ Currently, ultraviolet- visible (UV-Vis) spectrometry is the standard technology for in-line quantification of protein concentration during downstream purification.⁴ However, in recent years, process Raman has gained popularity as a complementary technology for in-line monitoring of protein concentration with its additional benefits of being able to measure excipient concentrations, buffer components, and critical quality attributes (CQAs) of products.⁵ Here, we demonstrate the use of process Raman as a viable PAT solution for accurate and real-time quantification of mAb concentration during downstream processing. In addition, we also show the model’s transferability to a different monoclonal antibody. Experimental design and data analysis Calibration model development Partial least square (PLS) calibration models for mAb quantification were developed using the calibration samples of product A ranging in concentration from 0 to 135 mg/mL. The samples were passed through the FlowCell probe integrated with process Raman at a 100 mL/min flow rate. The Raman spectra were acquired using a 785 nm laser with the following acquisition parameters: laser power 450 mW; integration time 3000 ms; average 3 (i.e., a single spectrum per 18 s); and ten replicates per concentration. PLS models were developed for mAb quantification using two spectral regions. The “Amide I PLS model” utilized the spectral region of approximately 1550 to 1850 cm-1, while the “Extended Region PLS model” utilized a broader spectral region from approximately 850 to 1850 cm-1. The terminal region of the spectra (~ 3098 to 3230 cm-1) that corresponds to water band vibrations was also included in the normalization model. The models were built after following the data preprocessing steps: 1) normalization to the water band in every spectrum using infinity norm; 2) SavGol filter 1st derivative (smoothing window 13 and polynomial order 2); and 3) mean centering. Both models were internally validated using a leave-one-out cross-validation (LOOCV) strategy (contiguous block of 10). Initially, both models were built using the latent variables ranging from 1 through 20. The root mean square error of calibration (RMSEC), and the root mean square error of crossvalidation (RMSECV) was calculated for each model. Finally, the optimum latent variable PLS model was selected using the following criteria: 1) adding more latent variables did not significantly improve the RMSECV; and 2) the values of RMSEC and RMSECV were similar. To evaluate model specificity, the variable importance in the projection (VIP) scores was calculated for the above two models. All chemometric works were performed using Solo 9.3 (2024) Eigenvector Research, Inc, Manson, WA USA 98831 software. Figure 2. Upstream and downstream workflow for production of monoclonal antibodies. Monoclonal Antibody Production Centrifugation → depth filtration Upstream Process Downstream Processes Cells Monoclonal Antibody Bioreactor Capture chromatography Viral inactivation Viral filtration Intermediate and polishing chromatography Monoclonal Antibody Purification and Formulation Ultrafiltration/ Diafiltration Bulk fill Fill and Finish Validation of model performance The Raman spectra for the test samples were collected for two different monoclonal antibodies (products A and B) using the procedure described for the training data. The formulation buffers for products A and B were different during the diafiltration steps. These acquired spectra were fed into the models to get the real-time prediction of protein concentration. The predicted concentrations were then compared with the reference in-line and offline UV-Vis values to estimate the prediction error. Results The correlation plot between measured and predicted (during cross-validation) protein concentrations for the Amide I PLS models is shown in Figure 3A. The Amide I PLS model was developed using four latent variables as the RMSECV did not improve by adding more latent variables (data not shown). The RMSEC and RMSECV for the Amide I PLS model were 0.526 mg/mL and 0.607 mg/mL respectively. The ratio of RMSEC and RMSECV being close to 1 suggests that the model is not overly fitted. Similarly, the R² of CV is close to 1, and negligible CV bias indicates that the model fits well with the training data. The model statistics are summarized in Table 1. Figure 3. The correlation plot for measured and cross validation predicted protein concentration for mAb using Amide I and Extended Region PLS model. Table 1. PLS Model Latent Variables RMSEC (mg/mL) RMSECV (mg/mL) CV Bias R² CV Amide I PLS model 4 0.526 0.607 -0.036 ~1 Extended Region PLS model 4 0.237 0.319 0.014 ~1 0 20 40 60 80 140 Y Measured 1 Protein Conc (mg/mL) Y Measured 1 Protein Conc (mg/mL) Amide I PLS Model Extended Region PLS Model Y CV Predicted 1 Protein Conc (mg/mL) Y CV Predicted 1 Protein Conc (mg/mL) 0 20 40 60 80 100 120 140 6 100 120 0 20 40 60 80 140 100 120 0 20 40 60 80 100 120 140 4 Latent Variables RMSEC = 0.52639 RMSECV = 0.60764 Calibration Bias = 0 CV Bias = -0.03626 R² (Cal, CV) = 1.000, 1.000 4 Latent Variables RMSEC = 0.2379 RMSECV = 0.31917 Calibration Bias = -3.5527e-15 CV Bias = -0.014102 R² (Cal, CV) = 1.000, 1.000 A B The Amide I model is primarily based on the Raman signature of the carbonyl group (–C=O) of the peptide bond (–CO–NH–) of mAb. The carbonyl group on different secondary structures has different Raman shifts ranging from 1600 to 1750 cm-1. The mAb secondary structure is primarily a β-sheet structure. The carbonyl in the β-sheet secondary structure has a Raman peak at ~ 1670 cm-1. Thus, it can be hypothesized that the Raman shift at ~1670 cm-1 should influence the Amide I PLS model. The VIP scores were calculated to assess the influence of each Raman shift on the model. The Raman shifts with VIP scores over one are considered significant for the model. Figure 4A shows the VIP scores plot for the Amide I PLS model. The red dotted line represents the threshold of a VIP score equal to 1. The region from 1640 to 1700 cm-1 have VIP scores of more than one and are influential to the model. The region around 1670 cm-1 has the highest VIP score in the Amide I PLS model. This indicates the carbonyl Raman signature in the β-sheet secondary structure highly dominates the Amide I PLS model. In other words, the VIP scores plot for the Amide I PLS model indicates that the model has high specificity for mAb. Based on the published work, the Amide I region of mAb is unique. It has a distinct Raman signature compared to other molecules (excipients and buffers) commonly used in the downstream processes. No spectral interference in the Amide I region suggests that the Amide I PLS model is across different matrixes. In addition, most of the mAb have similar mass (a similar number of carbonyl residues) and secondary structure. Thus, the Amide I PLS model is transferable across multiple mAb within the same classes. This was the basis of our intent to develop the Amide I PLS model. The correlation plot for the four latent variables of the Extended Region PLS model is shown in Figure 3, and the model statistics are summarized in Table 1. Like the Amide I PLS model, the Extended Region PLS model’s excellent statistics suggest that it adequately captures the spectral information and correlates it with the measured concentration. The VIP scores plot for the Extended Region PLS model is shown in Figure 4B. Besides the Amide I region, the Raman shift that corresponds to the CH deformation (~1450 cm-1), the breading mode of the phenylalanine ring (~1005 cm-1), and the tyrosine vibration mode (~850 cm-1) are influential in this model. Thus, the VIP scores plot for the Extended Region PLS model indicates the specificity of the model for mAb. The RMSEC and RMSECV for the Amide I PLS model are higher than the Extended Region PLS model, as shown in Table 1, which suggests that the former model is less accurate than the latter. The Extended Region PLS model leverages additional variables (Raman shifts) in explaining the information of the training dataset. Although the Extended Region PLS model provides more accuracy, in some cases, the transferability of this model across different matrices may result in higher prediction error due to spectral overlap. Thus, both models have pros and cons, and depending on the need, one model might be a better choice. Figure 4. The VIP scores plot for Amide I PLS model (plot A) and Extended Region PLS model (plot B). 0 0 0.5 0.5 1.0 1.0 1.5 1.5 2.0 2.0 2.5 2.5 Raman Shift (cm-1) Raman Shift (cm-1) VIP Scores for Amide I Model VIP Scores for Extended Region Model 1,600 1,800 2,000 2,200 2,400 2,600 2,800 3,000 1,000 1,500 2,000 2,500 3,000 A B 3.0 3.5 4.0 3,200 3.0 3.5 4.0 4.5 Performance of models on the UF/DF run using product A (mAb) Transferability of models on product B (mAb) UF/DF run The Amide I and Extended Region PLS models were developed using the Product A data set. A UF/DF run was carried out with different protein (product B) in a different formulation matrix to validate the model performance with other monoclonal antibodies. Figure 6A shows the real-time monitoring and prediction of Product B concentration during the UF/DF process using Amide I and Extended Region PLS models. The prediction from the Amide I model (blue trace) and Extended Region Model (orange trace) excellently overlay each other. To the right in Figure 6B, the absolute prediction error between the predicted (Raman) and reference (offline UV-Vis) is shown as a function of protein concentration. The blue and orange bars represent the prediction errors from the Amide I PLS and Extended Region PLS models, respectively. The overall prediction errors were below 3% across the concentration range tested. The real-time prediction of protein concentration during the process of UF/DF using data from in-line UV-Vis (grey), Amide I PLS model (blue trace), and Extended Region PLS model (orange) is demonstrated in Figure 5. As shown in the figure, all three predictions overlayed each other, demonstrating excellent agreement. Minute data discrepancies between in-line UV-Vis and Raman were observed. The Raman data was acquired every 18 sec, while the in-line UV-Vis data was acquired every 12 sec. Since the UF/DF process is highly dynamic, the difference in the acquisition times for the in-line UV-Vis and Raman instruments explains the discrepancies. Such discrepancies can be easily overcome with proper control of the process dynamics or acquisition settings. Finally, when the Raman prediction from both models was compared with the offline UV-Vis reference values, the absolute prediction errors were below 5% throughout the process. Figure 5. Showing excellent correlation of the real-time monitoring of UF/DF of product A. Figure 6. Data demonstrate excellent model transferability between proteins with different formulation matrices. The calibration models for predicting protein concentration were built with Product A and applied to Product B that has a different formulation buffer. The absolute prediction error was < 3 %. 0 80 100 120 140 160 Time (min) mAb concentration (mg/mL) 0 60 120 180 240 20 40 UV-Vis Amide I Extended Region 60 0 0% 40 5% 60 10% 80 100 120 Spectra Protein Concentration (g/L) Protein Concentration (g/L) % Error 200 600 1,000 42.1 0 400 800 35.0 94.5 20 Amide I Extended Region Amide I Extended Region A B 2.2% 2.5% 0.5% 1.0% 2.2% 0.5% Learn more at thermofisher.com/marqmetrixAIO For research use only. Not for use in diagnostic procedures. For current certifications, visit thermofisher.com/certifications © 2024 Thermo Fisher Scientific Inc. All rights reserved. All trademarks are the property of Thermo Fisher Scientific and its subsidiaries unless otherwise specified. MCS-AN1145-EN 9/25 Conclusions 9 We demonstrated process Raman to be a rapid, reliable, and easy-to-use PAT solution for in-line monitoring of protein concentration during downstream processes, with an accuracy comparable to the reference values from in-line and offline UV-Vis instrumentation. 9 In this study, we outlined two strategies for developing the chemometric models for the quantification of mAbbased on specific Raman signatures. Because of the high specificity of both models for mAb, we also demonstrated their excellent transferability to a different mAb in a UF/DF run, regardless of differences in buffer formulations. 9 Based on works published in the literature, process Raman is a broader PAT solution for downstream applications because of the additional benefits it brings regarding measurement of buffer components, critical quality attributes of proteins (aggregation, secondary structure, disulfide region), protein modification (antibody-drugs conjugates), formulation components, and more.6–8 References 1. Leader, B.; Baca, Q. J.; Golan, D. E. Protein Therapeutics: A Summary and Pharmacological Classification. Nat Rev Drug Discov 2008, 7 (1), 21–39. https://doi. org/10.1038/nrd2399. 2. Alt, N.; Zhang, T. Y.; Motchnik, P.; Taticek, R.; Quarmby, V.; Schlothauer, T.; Beck, H.; Emrich, T.; Harris, R. J. Determination of Critical Quality Attributes for Monoclonal Antibodies Using Quality by Design Principles. Biologicals 2016, 44 (5), 291–305. https://doi.org/10.1016/j.biologicals.2016.06.005. 3. Research, C. for D. E. and. PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance. U.S. Food and Drug Administration. https://www.fda.gov/regulatory-information/ search-fda-guidance-documents/pat-framework-innovative-pharmaceuticaldevelopment-manufacturing-and-quality-assurance (accessed 2023-03-21). 4. Rolinger, L.; Rüdt, M.; Hubbuch, J. Comparison of U.V.- and Raman-Based Monitoring of the Protein A Load Phase and Evaluation of Data Fusion by PLS Models and CNNs. Biotechnol Bioeng 2021, 118 (11), 4255–4268. https://doi.org/10.1002/ bit.27894. 5. Rolinger, L.; Rüdt, M.; Hubbuch, J. A Critical Review of Recent Trends, and a Future Perspective of Optical Spectroscopy as PAT in Biopharmaceutical Downstream Processing. Anal Bioanal Chem 2020, 412 (9), 2047–2064. https://doi.org/10.1007/ s00216-020-02407-z. 6. Sukumaran, S. Protein Secondary Structure Elucidation Using FTIR Spectroscopy. 4. 7. Dolui, S.; Mondal, A.; Roy, A.; Pal, U.; Das, S.; Saha, A.; Maiti, N. C. Order, Disorder, and Reorder State of Lysozyme: Aggregation Mechanism by Raman Spectroscopy. J. Phys. Chem. B 2020, 124 (1), 50–60. https://doi.org/10.1021/acs.jpcb.9b09139. 8. Hernández, B.; Pflüger, F.; López-Tobar, E.; Kruglik, S. G.; Garcia-Ramos, J. V.; Sanchez-Cortes, S.; Ghomi, M. Disulfide Linkage Raman Markers: A Reconsideration Attempt: Disulfide Raman Markers. J. Raman Spectrosc. 2014, 45 (8), 657–664. https://doi.org/10.1002/jrs.4521. A Classical Least Squares (CLS) Approach for Protein Quantification in Downstream Processing Using Raman Spectroscopy Technical note Authors Nimesh Khadka¹, Ph.D. Kristina Pleitt², Ph.D. Michelle Nolasco² ¹Analytical Instrument Group, Thermo Fisher Scientific, Tewksbury, Massachusetts USA ²Bioproduction Group, Thermo Fisher Scientific, St. Louis, Missouri USA Industry/Application Biopharma PAT / Downstream processing Products used Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer, Thermo Scientific™ MarqMetrix™ FlowCell™ Sampling Optic Goals Demonstrate quick and easy chemometric strategies to develop an accurate protein quantification model for downstream applications using only a single spectrum of known concentration. Highlight the developed strategy’s performance and transferability across quantification of other types of monoclonal antibodies (mAbs), different matrices, and processes. Key analytes Protein (mAb) quantification Key benefits • Cost and time savings from eliminating the need for extensive training data collection and laboratory analytics • Enable immediate deployment into process monitoring and control, allowing customers to leverage the numerous benefits of process Raman spectroscopy The successful deployment of Raman spectroscopy to monitor complex processes relies on the robustness of chemometric models, which typically require large datasets and sophisticated algorithms. The time and cost associated with generating extensive and reliable training datasets often hinder the adoption and integration of Raman technology in biopharmaceutical applications. In this work, we introduce a chemical information-based approach to develop a robust and accurate chemometric model with a minimal training dataset. We employed the classical least squares (CLS) algorithm using a selected region of a single spectrum of a protein with a known concentration to develop a protein quantification model for monitoring ultrafiltration/diafiltration (UF/DF) in downstream processing. The performance of the CLS model and its transferability across different monoclonal antibodies are discussed. This approach can also be leveraged to build quick and accurate chemometric models for various other applications, making Raman technology more accessible for adoption and utilization. Data acquisition and chemometric modelling A monoclonal antibody (mAb; protein) at a concentration of 93.56 mg/mL in a formulation aqueous buffer containing histidine, arginine, and sucrose was passed through a Thermo Scientific™ MarqMetrix™ FlowCell Sampling Optic at a flow rate of 100 mL/min. Raman data were acquired during the dynamic flow using a Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer with acquisition parameters of 450 mW power, 3000 ms integration time, and 3 averages. Ten spectra were acquired and preprocessed to remove any cosmic ray interference. The ten spectra were then averaged into a single spectrum that was used to build the CLS model. Two spectral regions were selected before developing the CLS model: 3100 to 3230 cm⁻¹ (water band) and 1570 to 1750 cm⁻¹ (protein amide I region). The infinity norm for the 3100 to 3230 cm⁻¹ spectral region was calculated and used as a weight to normalize the entire spectrum, correcting spectral path differences. Baseline features were removed by applying the Savitsky-Golay filter (2nd derivative, polynomial order = 2, window width = 13). The derivatized spectrum was then used to develop the CLS model, which in essence is a ratio-metric model that translates the ratio of Raman intensity of the amide I band to the water band into predicted concentrations. Demonstrating a simpler chemometric approach for quantifying protein concentration in a downstream ultrafiltration/diafiltration (UF/DF) processing using process Raman. Buffer Pump Pump Membrane Retentate Permeate Outlet FlowCell Process Raman Test Data Ultrafiltration/ Diafiltration CLS Model -14 -12 -10 -8 -6 -4 Raman shift (cm-1) Raman intensity (a.u.) 1,600 1,650 1,700 1,750 -2 0 2 4 x 10- ³ 105 100 95 90 85 80 Amide I region 0 20 40 60 80 100 Measured concentration (mg/mL) Predicted concentration (mg/mL) 0 20 40 60 80 100 120 140 160 120 140 160 RMSEC = 0.74464 RMSEP = 2.2493 Prediction Bias = 0.55971 R² (Pred) = 0.997 Single Spectrum Predictions Following the CLS model development, validation data were acquired using the same parameters over the mAb concentration range of 0 to 155 mg/mL in the same formulation buffer. The model’s performance was further validated by applying it to the UF/DF process with different mAbs and by including tryptophan in the formulation buffer alongside histidine, arginine, and sucrose. A brief comparison of CLS and partial least square (PLS) models was also performed. All data management, cosmic ray removal, averaging, and timestamp alignment were performed in Python™ programming language. The data were then processed in a commercially available chemometric package. All chemometric works were performed using software package SOLO 9.3.1 (2024), Eigenvector Research. Inc., Manson, WA USA 98831. Results and discussion Standard normal variate (SNV) is a widely used normalization technique for spectroscopic data to correct for path length differences.¹ While effective when the primary contribution to variables is noise and they share the same overall signal, SNV may lead to non-linear responses if the overall signal changes significantly between samples.² This may especially be true in downstream processes where concentrations are high and dynamic changes result in rapid changes in spectral features and intensities. In this study, we opted to use the water band as an internal standard to normalize the spectra as the water concentration remains relatively constant throughout the bioprocesses.³– ⁵ The region from 3100 to 3230 cm⁻¹ includes the Raman band attributed to the symmetric stretching of the O-H vibrational bond in water molecules. The O-H stretching vibrational band is susceptible to changes in pH, ionic strength, and temperature; nevertheless these parameters are well-defined and controlled during process development. This ensures the reliability of this region for spectral normalization. Additionally, unlike the symmetric bending vibration of water molecules at 1640 cm⁻¹, the 3100 to 3230 cm⁻¹ spectral region has minimal spectral overlap from the constituents commonly used in the downstream processes. This makes it a viable option for spectral normalization.⁶ One concern is the low quantum efficiency of silicon optical sensors in this high Raman shift region. We, and others, have previously used this region for modelling, and the region’s performance has proven to be acceptable, likely because the high concentration of water (~55.5 M) compensates for the limitation in efficiency.³– ⁵ Figure 1a shows the average spectrum used for building the training model, while Figure 1b presents the 2nd derivative plot of the Raman spectra in the amide I region of mAb. The amide I region spans the Raman shift from approximately 1630 to 1700 cm⁻¹, primarily influenced by the variations in the energy of C=O symmetric stretching vibrations in different secondary structures of mAb⁷. Our previous work has demonstrated that the amide I region is free of spectral interferences and can be effectively utilized to develop models with high specificity for mAb for downstream processes.³, ⁵ Consequently, the amide I region was selected for the CLS model. The Savitsky-Golay filter (2nd derivative) was used in the preprocessing step to remove the baseline shifts as well as the broad water band, ensuring that the Raman information from the mAb is used to train the model. Figure 1. (a) Selected region used for chemometric model development. The amide I region provides specificity to the model for mAb while the water band is used for correction of spectral path length differences. (b) Showing the 2nd derivative preprocessed spectrum of amide I region used for model development. x 10⁴ -14 -12 -10 -8 -6 -4 Raman shift (cm-1) Raman intensity (a.u.) 1,600 1,650 1,700 1,750 0.5 1 1.5 2 2.5 3 Raman shift (cm-1) Raman intensity (a.u.) 2,000 2,500 3,000 105 -2 0 2 4 x 10- ³ 100 95 90 85 80 105 100 95 90 85 80 a b Amide I region Amide I region Water band The CLS model is a quantitative analytical method that explains the observed spectrum of a given sample by using the linear combination of the spectra of the pure components present in the sample.⁸ For the CLS model to perform accurately, it is essential to acquire the pure spectrum for each component of the mixture, which is practically challenging or impossible for a complex bioprocess. This problem is addressed in this study by using the mAb-specific amide I region that is free of spectral overlap. If a larger region is used to create the model, the CLS model showed a decrease in performance (data not shown). The loading for the CLS model is shown in Figure 2. The model has high influence from the Raman shift at approximately 1670 cm-1 in the amide I region, which is assigned to the symmetric C=O stretching of the β-sheet secondary structure of the mAb, thus providing specificity to the model. As expected for the single-component CLS model, the loading and the preprocessed spectrum are similar (compare Figures 1b and 2). The CLS model was then applied to the validation dataset (shown in Figure 3a) across a wide concentration range in different buffer matrices: 0 mg/mL in water; 1 to 33 mg/mL in tris buffer; and 33 to 155 mg/mL in histidine, arginine, and sucrose buffer. Across the concentration range in diverse buffer components, the root mean square error of prediction (RMSEP) was approximately 2.25 mg/mL as shown in the correlation plot of Figure 3b. The low RMSEP for the concentration range of 0 to 155 mg/mL demonstrates excellent model performance and transferability across buffer matrices. Note that the spectra shown in Figure 3a exhibit differences in water band intensities at approximately 3240 cm-1. The water concentration is relatively constant across the samples. These differences in intensities within the water band indicate variations in optical path length during data acquisition, caused by turbidity and occasionally small air bubbles trapped inside the MarqMetrix FlowCell probe. The inclusion of water band normalization in the CLS model appropriately corrects for the path length differences and improves prediction accuracy. Additionally, no baseline was removed before water band normalization. Although the data is not shown, the CLS model demonstrated similar performance with or without baseline removal (using automatic Whittaker filter and automatic weighted least squares) before water band normalization. Figure 2. The regression vector for the CLS model demonstrating the influence of the approximately 1670 cm-1 Raman shift in the amide I region, which is assigned to the symmetric C=O stretching of the β-sheet secondary structure of the mAb. -14 -12 -10 -8 -6 -4 Raman shift (cm-1) Variable/Loadings 1,580 1,600 1,620 x 10⁵ 1,640 1,660 1,680 1,700 1,720 1,740 -2 0 2 4 6 β-sheet Figure 3. (a) Spectral data used for validation of the CLS model. Also, highlighted are the differences in the intensities of water band across spectra which substantiate the need of inclusion of the water band normalization step in the model. (b) Correlation plot for measured vs. predicted concentration along with performance statistics shown as inset. 0 20 40 60 80 100 Measured concentration (mg/mL) Predicted concentration (mg/mL) 0 20 40 60 80 100 120 140 x 10⁴ 0 1 2 3 4 Raman shift (cm-1) Raman intensity (a.u.) 1,000 1,500 3,000 140 120 100 80 40 20 a 60 500 2,000 2,500 5 6 160 120 140 160 RMSEC = 0.74464 RMSEP = 2.2493 Prediction Bias = 0.55971 R² (Pred) = 0.997 b Water band To validate the performance of the CLS model, the training and validation data used to develop the CLS model were combined into a single dataset. The combined data was fed into the PLS algorithm using the same spectral region and preprocessing as the CLS model. A one-latent-variable PLS model was selected based on the leave-one-out crossvalidation (LOOCV) strategy, where each class was left out once. The root mean square error of cross-validation (RMSECV) was calculated, as shown in inset of Figure 4. The RMSECV for the PLS model was 2.88 mg/mL, while the RMSEP for the CLS model was 2.24 mg/mL. The RMSEP of CLS model and RMSECV of the PLS model is not a direct comparison, nonetheless, with some approximation, these results do indicate that the CLS model performed comparable to the PLS model. To further test the model’s performance, scalability, and transferability, we applied the lab-based PLS and CLS models to Raman data collected in-line during a UF/DF pilot run. This run used a different type of mAb and a formulation buffer. The buffer included tryptophan, histidine, arginine, and sucrose that were added during diafiltration. We have described the relevance of this experiment before.³ The predicted protein concentrations from the CLS and PLS models showed a high correlation. This is shown in Figure 5 with the orange and blue traces. The pooled samples (marked by red stars) were measured using HPLC and UV-Vis spectroscopy. The absolute prediction errors for both models are shown in Table 1. The predictions from the CLS model exhibited lower errors compared to the PLS model, however it does not mean CLS is superior to PLS based on a statistically insignificant sample size of n=1 dataset. In this study, the CLS model was built by selecting the amide I region that has high specificity for mAb and minimum spectral interferences If the entire spectral region was used with complex spectral overlap or in cases with ill-conditioning matrix and multicollinearity, PLS or other regression models are better choices with a multitude of other advantages. Additionally, using different regions of spectra and other combinations of preprocessing, the performance of the PLS model may be further improved in the above case. Here, PLS is used only as reference but not for comparison. Figure 4. PLS model shown with model statistics in inset. The training and test dataset used for the CLS model were combined to develop the PLS model. The RMSECV of the PLS model calculated using leave-one-out cross validation is close to the RMSEP for the CLS model, indicating similar model performance. -15 5 25 45 65 165 Linear index Predicted concentrations (mg/mL) 100 200 300 400 500 600 700 800 Figure 5. This plot shows the agreement in prediction of protein concentration from the PLS (orange) and CLS (blue) models for the UF/DF run. Table 1. Performance of CLS model. Reference concentration (mg/mL) PLS model predictions (mg/mL) CLS model prediction (mg/mL) PLS model average abs. % error CLS model average abs. % error 30.72 30.6 ± 0.4 30.8 ± 0.1 1.12 0.36 155.9 146 ± 1 154 ± 1 6.15 0.67 50 100 150 Measured concentration (mg/mL) Predicted concentration (mg/mL) 50 100 150 150 100 50 0 0 0 1 Latent Variable RMSEC = 2.2415 RMSECV = 2.8879 Protein concentration predictions in UF/DF 0 85 105 125 145 CLS Model Predictions PLS Model Predictions Protein concentration Protein concentration Buffer exchange Sampling Conclusion This study demonstrated an alternative approach of a single-spectrum-based CLS model for protein quantification in downstream bioprocesses. The CLS model exhibited lower prediction errors than the broadly acceptable tolerance of < 5-10 % for process monitoring. It also showed that the single-spectrum-based CLS model has scalability from lab to pilot scale and transferability across different monoclonal antibody and buffer matrices. A key factor in the successful implementation of the CLS model was appropriate region selection. We found that the unique Raman signature of the amide I region of monoclonal antibodies has minimal spectral interference from other constituents commonly used in downstream processes.³, ⁵ This allowed us to develop a mAb-specific CLS model using the Raman intensity of the amide I region that linearly scales with the concentration. Identifying similar unique regions in other applications can provide a rapid and straightforward method to build robust and accurate chemometric models. In addition, augmenting more data to the CLS model especially at the upper and lower concentration range will further improve the model. Another important aspect involves normalizing the spectra using the O-H symmetric stretching Raman band of water molecules. Recent literature has also utilized a similar normalization strategy for developing chemometric models for upstream bioreactor monitoring.⁴ This strategy appears to work across all modalities. Finally, the ability to build the chemometric model using a minimal dataset not only facilitates the adoption of Raman technology but also broadens its applicability. References: 1. Barnes, R. J.; Dhanoa, M. S.; Lister, S. J. Standard Normal Variate Transformation and De-Trending of Near-Infrared Diffuse Reflectance Spectra. Appl Spectrosc 1989, 43 (5), 772–777. https://doi.org/10.1366/0003702894202201. 2. Advanced Preprocessing: Sample Normalization - Eigenvector Research Documentation Wiki. https://www.wiki.eigenvector.com/index. php?title=Advanced_Preprocessing:_Sample_Normalization. 3. Nolasco, M.; Pleitt, K.; Khadka, N. Using a Process Raman Analyzer as an In-Line Tool for Accurate Protein Quantification in Downstream Processes. 4. Pétillot, L.; Pewny, F.; Wolf, M.; Sanchez, C.; Thomas, F.; Sarrazin, J.; Fauland, K.; Katinger, H.; Javalet, C.; Bonneville, C. Calibration Transfer for Bioprocess Raman Monitoring Using Kennard Stone Piecewise Direct Standardization and Multivariate Algorithms. Engineering Reports 2020, 2 (11), e12230. https://doi.org/10.1002/ eng2.12230. 5. Nolasco, M.; Pleitt, K.; Khadka, N. Raman-Based Accurate Protein Quantification in a Matrix That Interferes with UV-Vis Measurement. 6. Palacký, J.; Mojzeš, P.; Bok, J. SVD-Based Method for Intensity Normalization, Background Correction and Solvent Subtraction in Raman Spectroscopy Exploiting the Properties of Water Stretching Vibrations. Journal of Raman Spectroscopy 2011, 42 (7), 1528–1539. https://doi.org/10.1002/jrs.2896. 7. Peters, J.; Park, E.; Kalyanaraman, R.; Luczak, A.; Ganesh, V. Protein Secondary Structure Determination Using Drop Coat Deposition Confocal Raman Spectroscopy. 2016, 31, 31–39. 8. Lackey, H. E.; Sell, R. L.; Nelson, G. L.; Bryan, T. A.; Lines, A. M.; Bryan, S. A. Practical Guide to Chemometric Analysis of Optical Spectroscopic Data. J. Chem. Educ. 2023, 100 (7), 2608–2626. https://doi.org/10.1021/acs.jchemed.2c01112. Learn more at thermofisher.com/marqmetrix For Research Use Only. Not for use in diagnostic procedures. © 2025 Thermo Fisher Scientific Inc. All rights reserved. Python is a trademark of Python Software Foundation. SOLO is a trademark of Eigenvector Research. Inc. Manson, WA USA. All other trademarks are the property of Thermo Fisher Scientific and its subsidiaries unless otherwise specified. MCS-TN1384-EN 4/25 Direct in-line quantification of titer in clarified harvest using Raman spectroscopy Application note Authors Nimesh Khadka¹, Ph.D. Lin Zhang¹, Ph.D. Michelle Nolasco² ¹Analytical Instrument Group, ThermoFisher Scientific, Tewksbury, Massachusetts USA ²BioProduction Group, ThermoFisher Scientific, St. Louis, Missouri USA Industry/Application: Biopharma PAT / Downstream Products used: Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer, Thermo Scientific™ MarqMetrix™ Performance BallProbe™ Sampling Optic Goals: Eliminating offline analytics with real time titer quantification using process Raman in a clarified harvest Key Analytes: Titer; monoclonal antibody; clarified harvest Key Benefits: • Raman analysis streamlines downstream biopharma processes with cost and time benefits by eliminating the need for offline chromatographic analysis when calculating loading volume. • This methodology paves a path toward automation and continuous manufacturing by coupling upstream and downstream processes Thermo Scientific MarqMetrix All-In-One Process Raman Analyzer, Thermo Scientific MarqMetrix Performance BallProbe Sampling Optic. Introduction Monoclonal antibodies (mAbs) are manufactured in two stages: upstream and downstream processes, as shown in Figure 1. In the upstream process, cells are cultured to produce mAbs. These mAbs are then purified in the downstream process through a series of chromatography and filtration operations. Before chromatographic purification, cells and debris are removed by centrifugation and/or depth filtration in a process called clarification. The resulting supernatant after clarification, known as clarified harvest, contains the desired product (mAb), along with soluble metabolites and waste products. The clarified harvest is then subjected to further downstream processing to isolate and purify the target mAb. Knowing the concentration of the mAb (titer) in the clarified harvest before loading onto an affinity (capture) chromatography column (e.g., protein A) is essential for several reasons. It allows for the determination of loading volumes, ensuring the column is loaded with the right amount of clarified harvest to maximize the resin binding capacity. Overloading the column can lead to incomplete binding of the mAb, product loss, and, thereby, reduced purification efficiency. Since Raman spectroscopy provides real-time measurements, it facilitates continuous manufacturing and automation from the upstream bioreactor run to the clarification step to downstream purification, eliminating the need for conventional laboratory analytics and significantly reducing the time otherwise required for HPLC to measure titer concentration. In this study, we demonstrate the capability of process Raman to directly quantify the titer in the clarified harvest without any need of sample preparation. Experimental details Data collection Calibration samples with known mAb concentration were obtained from multiple bioreactor conditions. The titer in clarified samples ranged between 0 and 9 g/L. The samples were prepared by mixing titer-free clarified harvest (filtrate collected during protein A affinity column) with purified mAb to create different concentration of training samples using the design of experiment (DoE) approach based on the algorithm of Uniform Design (UD).¹ Each sample was scanned with a Thermo Scientific™ MarqMetrix™ Performance BallProbe™ Sampling Optic integrated with the Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer. The acquisition parameters were set to a power of 450 mW, an integration time of 5000 ms, and an average of 10 spectra, resulting in a spectrum every 2 minute. Chemometric model development The spectral regions 775 to 1920 cm-1 were selected to develop the Partial Least Square (PLS) model for titer quantification (Figure 2A). Baseline was removed from each spectrum using automatic Whittaker filter with the parameters for asymmetry and lambda set to 0.001 and 1000 respectively. Each spectrum was then normalized using the L1 norm calculated for the region 1590 to 1655 cm-1. To improve the model performance, all spectra were derivatized using a Savitzky-Golay filter (1st derivative; order = 2; window width =9) followed by mean centering. Figure 1. The workflow for the manufacturing of monoclonal antibodies. Also highlighted is the clarification step by the blue dotted box where in-line process Raman was used for real time titer concentration measurement. Monoclonal Antibody Production Centrifugation → depth filtration Upstream Process Downstream Processes Cells Monoclonal Antibody Bioreactor Capture chromatography Viral inactivation Viral filtration Intermediate and polishing chromatography Monoclonal Antibody Purification and Formulation Ultrafiltration/ Diafiltration Bulk fill Fill and Finish Protein Concentration Raman 0 2 4 6 8 10 Measured titer concentration (mg/mL) Predicted titer concentratin (mg/mL) 0 2 4 6 8 10 5 Latent Variables RMSEC = 0.11456 RMSECV = 0.23534 RMSEP = 0.3575 Calibration Bias = -4.4409e-16 CV Bias = -0.048638 Prediction Bias = -0.12852 R² (Cal, CV) = 0.998, 0.994 R² (Pred) = 0.978 Fit (slope = 0.9985) 1:1 Calibration Test 0.0 0.5 1.0 1.5 2.0 2.5 Raman Shift (cm-1) Raman intensity (a.u.) (x10⁴) 800 1,000 1,200 1,400 1,600 1,800 3.0 2 3 4 5 6 7 A Figure 2. The spectral region used to develop the titer PLS model is shown in plot A. The selectivity ratio (SR) plot is shown in B and indicates the Amide I spectral region (1650 to 1700 cm-1) has strong influence in the model performance and specificity. Plot C and D show the predictive performance of the model for the DoE samples and different batches of clarified harvest, respectively. The gray filled circle are training data and red diamond are test data. 0 2 4 6 8 10 Measured titer concentration (mg/mL) Predicted titer concentratin (mg/mL) 0 2 4 6 8 10 0 5 10 15 20 Raman Shift (cm-1) Selectivity Ratio for titer prediction 800 1,000 1,200 1,400 1,600 1,800 The number of latent variables (LVs) for the PLS model was selected using leave-oneout cross-validation (LOOCV). During this process, each unique concentration block was left out from model training and used in prediction exactly once, and all replicates for a given concentration were treated as a single block to prevent data leakage. The optimal number of LVs was determined by minimizing the root mean square error of calibration and cross-validation while maintaining their ratios close to 1. The developed PLS model for the clarified harvest titer was evaluated using a validation set prepared with the DoE approach and by applying it to Raman data acquired from samples collected from different clarified harvests. All data management, cosmic ray removal, averaging, and timestamp alignment were performed in Python. All chemometric works were performed using software package SOLO 9.3.1 (2024). Eigenvector Research. Inc. Manson, WA USA 98831. B C D 1 8 9 3 5 7 9 11 1 5 Latent Variables RMSEC = 0.11456 RMSECV = 0.23534 RMSEP = 0.19075 Calibration Bias = -4.4409e-16 CV Bias = -0.048638 Prediction Bias = 0.057436 R² (Cal, CV) = 0.998, 0.994 R² (Pred) = 0.996 Fit (slope = 1.0070) 1:1 Calibration Test Results and discussion The PLS model was developed using five latent variables (LVs). The root mean square error of calibration (RMSEC) and cross-validation (RMSECV) were 0.114 mg/mL and 0.235 mg/ mL, respectively. The specificity of the model was evaluated by calculating the selectivity ratio (SR), as shown in Figure 2B. The SR is the ratio of the explained variance to residual variance for each Raman shift.² A higher SR for a given Raman shift indicates its greater importance for the model, forming the basis for model specificity. The Raman shifts between 1650 to 1700 cm-1 for the developed PLS model had high SR. This region, known as the Amide I region, primarily includes the Raman signature associated with the symmetrical stretching of the carbonyl group of the amide (peptide) linkage.³ Depending on the location of the carbonyl group in different secondary structures, they experience different electronic environments and thus have different energies associated with the symmetrical stretching of the carbonyl group. The Amide I region provides molecular information on the secondary structure of the protein, where its total area is proportional to the total amount of carbonyl functional groups present in the protein, and its features or peak positions depend on the presence of different secondary structures. The titer (mAbs) is a globular protein, and its secondary structure is dominated by β-sheet structure. As shown in Figure 2B, the SR ratio at ~ 1670 cm-1 is most prominent; this can be assigned to the symmetrical stretching of carbonyl group in the β-sheet structure. Thus, Amide I region contributes to the specificity for titer quantification in the clarified harvest. Similarly, the CH deformation (~1440 cm-1), Amide III region (~ 1230 cm-1; symmetric stretching C-N (ν(C-N)), N-H bending (δ(N-H)), symmetric C-C stretch (~1130 cm-1), phenylalanine ring breathing mode (~1005 cm-1), and tyrosine doublet (~830 and 850 cm-1) due to Fermi resonance between the in-plane breathing mode of the phenol ring and an overtone of outof-plane deformation mode are other Raman features of titer that are influential in the model.⁴ All these features collectively provide specificity for the model to quantify titer against the matrix of the host cell proteins (HCPs), metabolites, and other waste products. The model performance when applied to the independent validation set is shown in Figures 2C and 2D. Initially, the model was applied to the validation set samples prepared using the Uniform Design. The root means square error of prediction (RMSEP) was 0.19 mg/mL across the concentration range of 0 to 9 mg/mL. Similarly, when the model was applied to different batches of clarified harvest samples, the average RMSEP was 0.36 mg/mL. The offline analysis on the clarified harvest samples revealed that different batches of clarified harvest had varying matrices, including differences in the concentration and composition of HCPs, metabolites, and other molecules. After preprocessing, overlaying, and color-coding the training and clarified harvest datasets with titer concentration, a clear correlation was observed between Raman intensity and concentration in the spectral regions around 1670 cm-1, 1440 cm-1, 1005 cm-1, 830 cm-1, and 850 cm-1 as shown in Figure 3. However, other spectral regions exhibited strong interference from the Raman signatures of the matrices, likely from Raman signals from the HCPs, resulting in a lack of correlation between Raman intensity and titer concentration in these spectral regions. These findings validated the model’s specificity, as shown in the specificity plot (Figure 2B), and insight into the value of multivariate chemometrics in extracting useful information from complex spectra. Since HCPs vary between batches, it is recommended to augment the titer model with multiple batches of clarified harvest process data. This approach could further optimize the model by capturing process variations, thereby improving performance and lowering the RMSEP. Figure 3. Spectral overlay of the preprocessed training and test datasets (different batches of clarified harvest). The spectra are color-coded by titer concentrations, as indicated by the vertical bar. The red arrow points to the spectral regions that visually correlate with titer concentrations. Uncorrelated spectral regions are mainly due to Raman signals from host cell proteins. 0.00 0.01 0.02 0.03 Raman Shift (cm-1) Raman intensity (a.u.) (x10⁴) 800 1,000 1,200 1,400 1,600 0.04 2 3 4 5 6 7 1 8 9 Conclusion In this study, we demonstrated real-time quantification of titer in the complex matrix of clarified harvest without any sample preparation. The results indicate that by implementing in-line process Raman in the workflow, users can eliminate the need for conventional offline HPLC analysis to quantify titer in clarified harvest. Thus, users can directly proceed to the downstream purification step by reliably calculating the loading volumes for purification columns. This result not only demonstrates the capability of Raman spectroscopy for complex mixture analysis by leveraging unique molecular Raman signatures, but also provides a practical solution with time and cost benefits to the user. We and others have previously demonstrated Raman as a reliable tool for monitoring and feedback control of upstream and downstream processes.⁵– ⁹ The results shown here directly bridge our previous works by coupling upstream with downstream processes and establish process Raman as a single sensor with wide applications for biomanufacturing, as well as achieving automated continuous manufacturing. References 1. Zhang, L.; Liang, Y.-Z.; Jiang, J.-H.; Yu, R.-Q.; Fang, K.-T. Uniform Design Applied to Nonlinear Multivariate Calibration by ANN. Anal. Chim. Acta 1998, 370 (1), 65–77. https://doi.org/10.1016/S0003-2670(98)00256-6. 2. Kvalheim, O. M. Variable Importance: Comparison of Selectivity Ratio and Significance Multivariate Correlation for Interpretation of Latent-Variable Regression Models. J. Chemom. 2020, 34 (4), e3211. https://doi.org/10.1002/ cem.3211. 3. Peters, J.; Park, E.; Kalyanaraman, R.; Luczak, A.; Ganesh, V. Protein Secondary Structure Determination Using Drop Coat Deposition Confocal Raman Spectroscopy. 2016, 31, 31–39. 4. Dolui, S.; Mondal, A.; Roy, A.; Pal, U.; Das, S.; Saha, A.; Maiti, N. C. Order, Disorder, and Reorder State of Lysozyme: Aggregation Mechanism by Raman Spectroscopy. J. Phys. Chem. B 2020, 124 (1), 50–60. https:// doi.org/10.1021/acs.jpcb.9b09139. 5. Abu-Absi, N. R.; Kenty, B. M.; Cuellar, M. E.; Borys, M. C.; Sakhamuri, S.; Strachan, D. J.; Hausladen, M. C.; Li, Z. J. Real Time Monitoring of Multiple Parameters in Mammalian Cell Culture Bioreactors Using an In-Line Raman Spectroscopy Probe. Biotechnol. Bioeng. 2011, 108 (5), 1215–1221. https://doi.org/10.1002/bit.23023. 6. Villa, J.; Zustiak, M.; Kuntz, D.; Zhang, L.; Woods, S.; Scientific, F. Real Time Metabolite Monitoring Using the MarqMetrix All-In-One Process Raman Analyzer and the 500L HyPerforma Dynadrive Single-Use Bioreactor (S.U.B.). 7. Villa, J.; Zustiak, M.; Kuntz, D.; Zhang, L.; Khadka, N.; Broadbelt, K.; Woods, S. Use of Lykos and TruBio Software Programs for Automated Feedback Control to Monitor and Maintain Glucose Concentrations in Real Time. 8. Nolasco, M.; Pleitt, K.; Khadka, N. Using a Process Raman Analyzer as an In-Line Tool for Accurate Protein Quantification in Downstream Processes. 9. Nolasco, M.; Pleitt, K.; Khadka, N. Raman-Based Accurate Protein Quantification in a Matrix That Interferes with UV-Vis Measurement. Learn more at thermofisher.com For research use only. Not for use in diagnostic procedures. For current certifications, visit thermofisher.com/certifications © 2025 Thermo Fisher Scientific Inc. All rights reserved. All trademarks are the property of Thermo Fisher Scientific and its subsidiaries unless otherwise specified. MCS-AN1439-EN 06/25 Raman-based Accurate Protein Quantification in a Matrix that Interferes with UV-Vis Measurement Application note Authors Michelle Nolasco¹, Kristina Pleitt¹, Nimesh Khadka² ¹BioProduction Group, Thermo Fisher Scientific, St. Louis, Missouri USA ²Analytical Instrument Group, Thermo Fisher Scientific, Tewksbury, Massachusetts, USA Significance Ultraviolet-Visible spectroscopy (UV-Vis) is widely used for accurate quantification of concentrations of purified proteins. In cases where the matrix strongly absorbs photons of the key wavelength (280 nm), UV-Vis leads to the inaccurate quantification of protein concentration. The issue is further complicated if the matrix is dynamic, as it adds potential error to mathematical operations applied to remove the background information. This work presents a case study of an ultrafiltration/ diafiltration (UF/DF) experiment with a matrix containing tryptophan. As shown in Figure 1, both in-line UV-Vis and in-line Raman were used to monitor the process. Our findings demonstrate that Raman-based estimation of protein concentration proves to be more accurate and reliable when matrix interference is present. These results serve as a paradigm, reinforcing that Raman spectroscopy, based on specific molecular signatures, is particularly suitable for monitoring complex biological processes characterized by multi-component interactions and dynamic changes. Buffer Pump Pump Membrane Retentate Permeate Outlet Raman Ultrafiltration/ Diafiltration Accurate Protein Concentration Figure 1. Advantage of Raman based estimation of protein concentration in a UF/DF run in presence of interfering background due to matrix which varies with progress of the process. UV-Vis Inaccurate Protein Concentration Matrix (absorbs 280 nm light) Introduction The industrial biomanufacturing of therapeutic monoclonal antibodies (mAb) involves upstream and downstream processes,¹, ² as depicted in Figure 2. In the upstream process, cells are cultivated in bioreactors to produce the mAb. These mAb then undergo a series of operations in the downstream processing, including clarification, chromatography, filtration, viral inactivation, viral filtration, ultrafiltration/diafiltration (UF/DF), final filtration, and fill and finish steps. These operations are collectively categorized as downstream processing and are crucial for ensuring the purity, stability, and efficacy of the mAb. Accurate quantification of the protein at different stages is important for downstream processes.³, ⁴ It enables real-time decision-making. For example, accurate quantification of the target protein after the clarification step provides actionable results such as yield estimation, GO/NO GO decisions (e.g., discontinuing a batch if the product yield falls below the threshold to save costs, time, and resources), and selection of the appropriate number of chromatography cycles required to maintain the process within expected resin loadings. Accurate protein quantification is essential during UF/DF to monitor and control the protein concentration to a predefined target. Similarly, precise protein quantification in fill and finish products ensures correct dosages are delivered to patients. Thus, accurate protein quantification is a critical process parameter for downstream processing. The widely accepted standard method for estimating total protein concentration is at-line or in-line UV-Vis absorption spectroscopy using a light source at approximately 280 nm.⁵, ⁶ This method relies on the absorption of aromatic amino acid residues (tyrosine and tryptophan) and disulfide bonds, which act as chromophores contributing to the UV-Vis absorption.⁷ In UV-Vis spectroscopy, the absorbance is linearly correlated with concentration as defined by the Beer-Lambert Law: In cases where the matrix absorbs photons at 280 nm, the total absorbance (matrix + mAb) will be higher than the absorbance of the pure mAb for a given concentration, resulting in overprediction. If the absorbance of the matrix is low and constant, a simple correction factor can be calculated and subtracted from the total absorbance to obtain an accurate mAb concentration. However, estimating the correction factor can be challenging when the matrix is dominant and dynamic. One example of complex matrix interference in downstream processing is the clarified harvest. The clarified harvest pool contains interfering components of various concentrations from the upstream process, such as host cell proteins, amino acids, and other biomolecules. This prevents the use of UV-Vis to estimate mAb concentration. Another example of matrix interference is diafiltration buffer excipients that absorb UV-Vis photons, leading to overestimating protein concentration during UF/DF. In both these examples, mAb is first purified using affinity chromatography (Protein A). Then, the concentration is determined by reading through the at-line UV-Vis instrument (i.e., an HPLC titer assay). This additional sample preparation and analytical steps increase the cost and process time from a couple of hours to potentially days, depending on the workload. An alternative technology that can be used instead of UV-Vis for downstream monitoring, particularly in situations where there is a matrix effect, is process Raman. Raman spectroscopy is an optical technique that involves measuring the vibrational modes of molecules by observing the inelastic scattering of photons (known as Raman scattering) after interacting with a light wave.⁸ It is highly specific for the target molecules and, thus, widely used for identification and quantification. In this study, we present a case study where we used process Raman and in-line UV-Vis to monitor a UF/DF run with a UV-Vis interfering matrix and demonstrate the advantage of Raman for such applications. Monoclonal Antibody Production Centrifugation → depth filtration Upstream Process Downstream Processes Cells Monoclonal Antibody Bioreactor Capture chromatography Viral inactivation Sterile filtration Intermediate and polishing chromatography Monoclonal Antibody Purification and Formulation Ultrafiltration/ Diafiltration Bulk fill Fill and Finish Figure 2. Upstream and downstream processes for monoclonal antibody production. Concentration = Absorbance path length * molar extinction coefficient Experimental details Calibration model development The protein quantification partial least square (PLS) models for Raman were developed using the spectra collected on a single monoclonal antibody (concentration range 0 to 157 mg/mL measured at-line using UV-Vis instrument) in various buffer and excipient-containing solutions. The samples were passed through the flowcell probe integrated with Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer at a 100 mL/min flow rate. The Raman spectra were acquired using a 785 nm laser with the following acquisition parameters: laser power 450 mW, integration time 3,000 ms, an average of 3 (i.e., a single spectrum per 18 s), and ten replicates per concentration. As preprocessing steps, each spectrum was normalized by dividing it by the infinity norm, which was calculated using the water band of each spectrum using region 3,098 to 3,230 cm-1. The normalized spectra were further processed with the SavGol filter (1st derivative, order = 2, window width = 13) and mean centering. The PLS model was developed using the spectral region 1,600 to 1,750 cm-1 and water band region 3,098 to 3,230 cm-1. The Raman spectral region 1,600 to 1,750 cm-1 of mAb is assigned to the vibration of the carbonyl group in the amide bond (-CO-NH-) located at different secondary structures like α-helix, β-sheet, turns, and random coils.⁹ Thus, the model is termed the “Amide I model.” Another PLS model was developed using the same preprocessing but utilizing the spectral region 800 to 1,750 cm-1 and water band region 3,098 to 3,230 cm-1, which we termed the “Extended Region model.” The models were internally validated using a leave-one-out cross validation strategy. The variable importance in projection (VIP) scores was calculated for each model to calculate the importance of each Raman shift for the development of the model.¹⁰ It also allows us to validate the statistical model with chemical information. The details on strategies for model development, validation, and VIP analysis were discussed previously. All chemometric works were performed using SOLO 9.3.1 (2024). Eigenvector Research. Inc. Manson, WA USA 98831. Tryptophan as an excipient in the ultrafiltration/ diafiltration (UF/DF) Ultrafiltration diafiltration (UF/DF) is a common unit operation in downstream bioprocessing involving a known pore-size tangential flow filtration membrane. It is used to buffer exchange and concentrate the desired biomolecules to prepare the product for final formulation. In this study, we started the experiment with ~10 mg/mL mAb in tris buffer and concentrated it to ~23 mg/mL in tris buffer. At this point, we initiated diafiltration to exchange the tris buffer with the excipient buffer. One of the components of the excipient buffer was tryptophan (20mM). The mAb concentration throughout the UF/DF run was monitored using an in-line process Raman and in-line UV-Vis spectrometer, as shown in Figure 1. The reference mAb concentration at the different stages of the UF/DF run was quantified by analytical protein A titer assay. The in-line Raman predicted mAb concentration and at-line HPLC-titer measured mAb concentration were used to calculate the root mean square error of prediction (RMSEP). Test of model transferability The Amide I and Extended Region calibration models were developed without incorporating any training dataset on tryptophan. Instead, the training dataset for both models consisted of mAb (IgG 4) samples in various buffers, including tris, histidine, arginine, sucrose, and polysorbate. As a result, the predictive performance of these models also assesses their transferability across different processes. Results The results of the PLS calibration models for predicting protein concentration based on Raman data are shown in Figure 3. The details of the strategy of model development, validation, and the VIP plot have been discussed in previous work. Two latent variables were selected to build the Amide I and Extended Region PLS model. The root means square error of cross-validation (RMSECV) of approximately 0.70 mg/mL for both models (Figure 3A and 3C inset) across the training concentration of 0 to 157 mg/mL mAb indicates that both models have high accuracy. The R2 for cross-validation being close to 1 for both models indicates that the models explain the variance in the training data related to mAb very well. The VIP score plot shows the importance of the Raman shift for the model. Here, the VIP score one is defined as the threshold (red dotted line in Figure 3B and 3D). Any Raman shifts with a VIP score of more than one are considered important for the model. Similarly, the higher the scores, the more important the Raman shift is for the model. For the Amide I model, the Raman shift at ~1,670 cm-1 is influential in the model (Figure 3B). On the other hand, the amide region (~1,670 cm-1), CH deformation (~1,450 cm-1), breathing mode of phenylalanine (~1,005 cm-1), and fermi doublet of tyrosine (~830-850 cm-1) are dominant in the Extended Region model (Figure 3D). Figure 3. Plots A and B show the Amide I calibration model and its VIP score plot, respectively. Similarly, plots C and D show the Extended Region calibration model and its VIP score plot. The model statistics are shown in the inset of plots A and C. 0 20 40 60 80 100 Measured Titer concentration (mg/mL) Extended Region Model CV Predicted Titer concentration (mg/mL) 0 0.0 1.0 2.0 3.0 4.0 Raman Shift (cm-1) VIP Scores for Extended Region Model 1,000 1,500 2,000 2,500 3,000 Measured Titer concentration (mg/mL) Amide I Model CV Predicted Titer concentration (mg/mL) 0 20 40 60 80 Raman Shift (cm-1) VIP Scores for Amide I Model 100 120 140 160 1,600 2,000 2,400 2,800 3,200 20 40 60 80 100 120 140 160 120 140 160 0.5 1.5 2.5 3.5 4.5 0 20 40 60 80 100 120 140 160 0.0 1.0 2.0 3.0 4.0 0.5 1.5 2.5 3.5 1,800 2,200 2,600 3,000 2 Latent Variables RMSEC = 0.49881 RMSECV = 0.6318 Calibration Bias = 0 CV Bias = NaN R² (Cal, CV) = 1.000, 1.000 2 Latent Variables RMSEC = 0.75816 RMSECV = 0.72856 Calibration Bias = 3.5527e-15 CV Bias = NaN R² (Cal, CV) = 0.999, 1.000 A B C D The results for the UF/DF run are shown in Figure 4. The process was initiated with 10 mg/mL mAb in tris buffer and concentrated to 23 mg/mL in tris buffer. In this concentration step, the prediction of mAb concentration from in-line UV-Vis and in-line Raman (for both models) showed excellent agreement with an overall absolute prediction error of less than 5%. In the diafiltration step, the mAb in the tris buffer was exchanged with the buffer containing tryptophan. As the tryptophan was introduced into the system, the prediction of mAb concentration from in-line UV-Vis had an absolute prediction error of 20 to 70% compared to the HPLC-titer (Figure 4). In contrast, the real-time predictions from in-line process Raman using Amide I and Extended Region model were more accurate. The absolute prediction error for the Amide I model was below 5%, while for the Extended Region model, it was below 10% (Figure 5). Why are Raman predictions accurate? The accuracy of Raman predictions can be explained based on molecular specificity, which forms the fundamental basis of all Raman applications. Raman is based on the unique signature exhibited by molecules with different atomic compositions or molecular bonds. In this study, the molecular specificity for mAb is provided by the Raman signature of the Amide I region that extends from ~1,640 – 1,700 cm-1. Amide bonds (-CO-NH-) are absent in this study’s tryptophan and other buffer components. This is evident in the second derivative plot of mAb concentration (~ 31 mg/mL) in tris buffer (red) and excipient buffer with tryptophan (green), as shown in Figure 6A. The spectral overlay of the Amide I region with and without tryptophan is approximately identical in intensity (at ~1,670 cm-1), and peak features indicate no spectral interference. As shown in the VIP plot for the Amide I model (Figure 3B), the 1,670 cm-1 region is dominant in the calibration model and crucial in making predictions. This unique Raman signature of the Amide I region of mAb explains the accurate transferability of the Amide I model with an absolute prediction error of less than 5% (Figure 5). Tryptophan, an amino acid part of mAb, has a strong Raman signal that overlaps the protein spectra beyond the Amide I region, as shown in Figure 6B. The Raman peaks from tryptophan at ~1,623 cm-1 and ~1,555 cm-1 are assigned to the stretching vibration of the benzene ring; the peak at ~1,436 cm-1 to the stretching vibration of the pyrrole ring; and at ~882 cm-1 to the skeletal vibration, with significant contribution from the pyrrole NH in-plane deformation.¹¹ However, the prediction errors from the Extended Region model were less than 10% (Figure 5) due to the molecular specificity provided by including the Amide I region in the calibration model. The slightly higher error in the prediction from the Extended Region Model compared to the Amide I Model is mainly due to spectra overlap with tryptophan. By augmenting the Extended Region model with an appropriate tryptophan dataset using the design of the experiment approach, the prediction error could be improved further. Figure 4. Overlay of mAb predictions from in-line UV-Vis (grey) and in-line process Raman (blue and orange) in the UF/DF run. The blue and orange traces represents the Raman predictions of the mAb concentration using Amide I and Extended Region models respectively. The excipient buffer containing tryptophan was introduced in the system at time ~120 min. The mAb prediction using in-line UV-Vis and in-line Raman were accurate and similar (<5% differences) before tryptophan was introduced (0 to 120 min). After introducing tryptophan, the predicted error form in-line UV-Vis was between 20 to 70%. Figure 5. Plot showing absolute error for mAb predictions from Amide I and Extended Region models at the reference concentration of mAb of 30.72 and 155.9 mg/mL in presence of 20mM tryptophan. Note, none of the tryptophan dataset were included in the training set for either model. 0 50 100 150 200 Time (min) Concentration (g/L) 0 100 200 300 400 Amide I Extended Region UV-Vis 0 2 4 6 8 Reference concentration (mg/mL) Absolute Prediction Error % 30.72 155.9 Amide I Prediction Error Extended Region Prediction Error Learn more at thermofisher.com/marqmetrixAIO For research use only. Not for use in diagnostic procedures. For current certifications, visit thermofisher.com/certifications © 2024 Thermo Fisher Scientific Inc. All rights reserved. All trademarks are the property of Thermo Fisher Scientific and its subsidiaries unless otherwise specified. MCS-AN1163-EN 8/24 Conclusion • We presented a case study demonstrating the advantage of in-line process Raman for accurately estimating mAb concentration in downstream processes, specifically UF/DF, where tryptophan interfered with direct in-line UV-Vis analysis. Raman spectroscopy overcame this interference and provided reliable and real-time mAb concentration predictions. • Our study showcased the transferability of Raman models across processes with high accuracy. This transferability is attributed to the fact that Raman spectroscopy relies on specific molecular signatures, making it highly specific for the target analytes. This characteristic of Raman spectroscopy makes it particularly suitable for monitoring and controlling complex and dynamically changing biological processes. • In downstream processing for mAb production, the analytes are typically present at high concentrations and purity levels. This makes Raman spectroscopy an ideal tool for monitoring and controlling these processes. Not only can Raman quantify different analytes accurately, it also provides valuable insights into critical quality attributes. The unique advantage of process Raman is its ability to provide real-time and actionable results by measuring multiple analytes in a single scan. This capability aligns with the objectives outlined in the FDA guidance for Process Analytical Technology (PAT). It accelerates progress towards reducing batch-to-batch variability and ensuring uniformity in the quality of products. References 1. Bergemann, K.; Eckermann, C.; Garidel, P.; Grammatikos, S.; Jacobi, A.; Kaufmann, H.; Kempken, R.; Pisch-Heberle, S. Production and Downstream Processing. In Handbook of Therapeutic Antibodies; John Wiley & Sons, Ltd, 2007; pp 199–237. https://doi.org/10.1002/9783527619740.ch9. 2. Gronemeyer, P.; Ditz, R.; Strube, J. Trends in Upstream and Downstream Process Development for Antibody Manufacturing. Bioengineering 2014, 1 (4), 188–212. https://doi.org/10.3390/bioengineering1040188. 3. Rolinger, L.; Rüdt, M.; Diehm, J.; Chow-Hubbertz, J.; Heitmann, M.; Schleper, S.; Hubbuch, J. Multi-Attribute PAT for UF/DF of Proteins—Monitoring Concentration, Particle Sizes, and Buffer Exchange. Anal. Bioanal. Chem. 2020, 412 (9), 2123–2136. https://doi.org/10.1007/s00216-019-02318-8. 4. Rathore, A. S.; Bhambure, R.; Ghare, V. Process Analytical Technology (PAT) for Biopharmaceutical Products. Anal. Bioanal. Chem. 2010, 398 (1), 137–154. https://doi.org/10.1007/s00216-010-3781-x. 5. Brestrich, N.; Rüdt, M.; Büchler, D.; Hubbuch, J. Selective Protein Quantification for Preparative Chromatography Using Variable Pathlength UV/Vis Spectroscopy and Partial Least Squares Regression. Chem. Eng. Sci. 2018, 176, 157–164. https://doi. org/10.1016/j.ces.2017.10.030. 6. McKechnie, W. S.; Tugcu, N.; Kandula, S. Accurate and Rapid Protein Concentration Measurement of In-Process, High Concentration Protein Pools. Biotechnol. Prog. 2018, 34 (5), 1234–1241. https://doi.org/10.1002/btpr.2695. 7. Biter, A. B.; Pollet, J.; Chen, W.-H.; Strych, U.; Hotez, P. J.; Bottazzi, M. E. A Method to Probe Protein Structure from UV Absorbance Spectra. Anal. Biochem. 2019, 587, 113450. https://doi.org/10.1016/j.ab.2019.113450. 8. Mulvaney, S. P.; Keating, C. D. Raman Spectroscopy. Anal. Chem. 2000, 72 (12), 145–158. https://doi.org/10.1021/a10000155. 9. Dolui, S.; Mondal, A.; Roy, A.; Pal, U.; Das, S.; Saha, A.; Maiti, N. C. Order, Disorder, and Reorder State of Lysozyme: Aggregation Mechanism by Raman Spectroscopy. J. Phys. Chem. B 2020, 124 (1), 50–60. https://doi.org/10.1021/acs.jpcb.9b09139. 10.Chong, I.-G.; Jun, C.-H. Performance of Some Variable Selection Methods When Multicollinearity Is Present. Chemom. Intell. Lab. Syst. 2005, 78 (1), 103–112. https://doi.org/10.1016/j.chemolab.2004.12.011. 11. Characterization of a few Raman lines of tryptophan - Hirakawa - 1978 - Journal of Raman Spectroscopy - Wiley Online Library. https://analyticalsciencejournals. onlinelibrary.wiley.com/doi/abs/10.1002/jrs.1250070511 (accessed 2024-06-30). Figure 6. Spectral overlap of mAb in tris (red) and tryptophan containing buffer (green) after applying SavGol filter (2nd derivative). The Amide I region is free of spectral overlap, thus providing specificity to the models. -40 -30 -20 -10 0 10 Raman Shift (cm-1) Amide I Region Preprocessed Raman Spectra 1,600 1,650 1,700 1,750 mAb in tris buffer mAb in buffer with tryptophan -150 -100 -50 0 50 Raman Shift (cm-1) Extended Region Preprocessed Raman Spectra 800 1,000 1,200 1,400 1,600 1,800 mAb in tris buffer mAb in buffer with tryptophan 20 A B Process Raman as a comprehensive solution for downstream buffer workflow Application note Authors Michelle Nolasco¹ Andrew Siemers¹ Kristina Pleitt¹, Ph.D. Nimesh Khadka², Ph.D. ¹BioProduction Group, ThermoFisher Scientific, St. Louis, Missouri USA ²Analytical Instrument Group, ThermoFisher Scientific, Tewksbury, Massachusetts USA Industry/Application: Biopharma PAT / Downstream Products used: Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer, Thermo Scientific™ MarqMetrix™ FlowCell Sampling Optic, Thermo Scientific™ MarqMetrix™ BallProbe™ Sampling Optic Goals: Enabling real-time excipient quantification and quality assessment using process Raman Key Analytes: Arginine, Histidine, Sucrose Key Benefits: • Enables real-time excipient quantification with cost and time benefits from eliminating the need for laboratory analytics • Provides a platform to take actionable decision using real time data rather than theoretical estimation. • Demonstrates potential of process Raman as a PAT tool for automation of Ultrafiltration/Diafiltration (UF/DF) and other downstream processes through simultaneous and real-time monitoring, quality assessment, and allowing multimodal feedback controls. Introduction Raman technology is rapidly gaining interest as a promising Process Analytical Technology (PAT) solution for real-time, non-invasive monitoring and control of downstream biopharma processes, especially for therapeutics like monoclonal antibodies (mAbs) and nucleic acids. Raman measurement, based on the vibration of molecular bonds, is highly specific for identification and quantification, even in complex or interfering matrices. As an in-line PAT tool, Raman spectroscopy offers direct and rapid measurement in aqueous phases without sample preparation. These features make it ideal for monitoring and controlling dynamic processes such as downstream processing. This study demonstrates a real-time methodology for accurately quantifying formulation excipients in the dynamic ultrafiltration/diafiltration (UF/DF) process using the Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer (Figure 1). In addition, this study also illustrates a case where process Raman was able to provide real-time information on buffer quality. To allow the generalization of these excipient models to their use in ultrafiltration/diafiltration (UF/DF) process for IgG1 mAb, Raman spectra for IgG1 mAb (5 to 150 mg/mL in various matrices) were added to the training dataset, and new models were developed. The addition of these protein spectra allowed the resulting PLS model to better distinguish Raman signals among L-arginine, L-histidine, and protein. Ultrafiltration/Diafiltration process The Raman Process Analyzer with FlowCell Probe was integrated in-line to monitor an UF/DF process (Figure 2). A PES membrane was equilibrated with tris buffer pH 7.0 prior to feeding a purified IgG1 mAb at 10 g/L to a target loading of 500 g/m². In the first ultrafiltration (UF) step, the mAb was concentrated at feed rate of 300 L/m²/hr, and TMP was maintained between 10-15 psi via manual flow restrictor. The mAb was then buffer exchanged into final formulation matrix containing L-histidine, L-arginine, and sucrose by manually feeding in the diafiltration (DF) buffer to the recirculation tank to maintain constant volume. After buffer exchange, the mAb was further concentrated to the desired final concentration in the second UF step. Figure 2. Ultrafiltration/Diafiltration (UF/DF) Process Diagram. Recirculation Tank TMP Control Valve Ramina Process Analyzer Feed Pressure Permeate Pressure Permeate Waste Retentate Pressure Diafiltration Buffer Product Feed Experimental details Excipient quantification models Calibration samples with defined concentrations of L-histidine, L-arginine, and sucrose were prepared using a design of experiments (DoE) approach called Uniform Design (UD) derived from number theory.¹ UD significantly reduces the total number of experiments while optimally spans the whole process space for model building and validation. These excipients were chosen due to their relevance in high-concentration monoclonal antibody (mAb) formulations. The analyte concentrations in the mixtures were designed with UD, with ranges of L-histidine (0-15 mg/mL), L-arginine (0-40 mg/mL), and sucrose (0-200 mg/mL) to develop calibration models. Each sample was passed in randomized order through a FlowCell Probe integrated with the MarqMetrix All-In-One Process Raman Analyzer at a flow rate of 100 mL/min. The acquisition parameters were set to a laser power of 450 mW, an integration time of 3000 ms, and an average of 3 spectra, resulting in an 18 second total collection time per spectrum. A Partial Least Squares (PLS) chemometric model was developed using the spectral range of 800 to 3235 cm-1 Raman shift. The spectra were normalized using infinity norm calculated in the spectral region of 2900 to 3230 cm-1 and preprocessed with a Savitzky-Golay (SavGol) filter (1st derivative, polynomial order = 2, window width = 13) and mean-centered. Overfitting was minimized by selecting appropriate latent variables using a leave-one-out cross-validation (LOOCV) strategy. To initially validate the model performance, seven different validation samples were collected in three different instruments using the same acquisition parameters as used in training data acquisition. Figure 1. Thermo Scientific MarqMetrix All-In-One Process Raman Analyzer, Thermo Scientific MarqMetrix FlowCell Sampling Optic. B -1 0 1 2 3 4 5 L 0 60 120 180 240 - h i s t i d i n e c o n c . ( m g / m L ) Time (min) UF1 DF A -1 1 3 5 7 9 0 60 120 180 240 L - a r g i n i n e c o n c . ( m g / m L ) Time (min) UF1 DF C UF1 DF UF2 -10 10 30 50 70 90 0 60 120 180 240 300 S u c r o s e c o n c . ( m g / m L ) Time (min) B -1 0 1 2 3 4 5 L 0 60 120 180 240 - h i s t i d i n e c o n c . ( m g / m L ) Time (min) UF1 DF A -1 1 3 5 7 9 0 60 120 180 240 L - a r g i n i n e c o n c . ( m g / m L ) Time (min) UF1 DF C UF1 DF UF2 -10 10 30 50 70 90 0 60 120 180 240 300 S u c r o s e c o n c . ( m g / m L ) Time (min) B -1 0 1 2 3 4 5 L 0 60 120 180 240 - h i s t i d i n e c o n c . ( m g / m L ) Time (min) UF1 DF A -1 1 3 5 7 9 0 60 120 180 240 L - a r g i n i n e c o n c . ( m g / m L ) Time (min) UF1 DF C UF1 DF UF2 -10 10 30 50 70 90 0 60 120 180 240 300 S u c r o s e c o n c . ( m g / m L ) Time (min) Buffer stability Buffers in downstream processes are often prepared in advance and typically used based on prior knowledge of their stability, rather than confirming stability through analytical analyses before each UF/DF run. In two of our downstream runs, we noticed that the process Raman spectroscopy predicted a lower-than-expected sucrose concentration in the excipient buffer. To mitigate potential risks, we discarded the previously prepared buffer and made a fresh batch. The predictions from process Raman analyzer on the new buffer were much closer to the reference values obtained from HPLC (high performance liquid chromatography). To further validate this capability in a controlled experiment, we measured the Raman spectra of the excipient buffer, which contains L-histidine, L-arginine, and sucrose, at room temperature for 15 days. The buffer was monitored through a Thermo Scientific MarqMetrix BallProbe Sampling Optic integrated with the MarqMetrix All-In-One Process Raman Analyzer. The acquisition parameters were set to a laser power of 450 mW, an integration time of 3000 ms, and an average of 3 spectra, resulting in an 18 second total collection time per spectrum. Additionally, we collected data on pH, osmolarity, and performed HPLC analysis at various time intervals. Results The Partial Least Squares (PLS) models for L-histidine, L-arginine, and sucrose were initially tested using seven independent samples collected on three different process Raman analyzers. Data were mathematically processed to standardize spectra across instruments before applying the models. All the spectra were interpolated to have a common x-axis by equally spacing the 2048 pixels across 60 to 3250 cm-1 Raman shift, followed by relative y-axis standardization using the SRM fluorescence data as described in the NIST standardization protocol.² Figure 3 shows the correlation plot of predicted versus reference values for L-histidine, L-arginine, and sucrose for the validation samples. A correlation coefficient of over 95% and a root means square error (RMSE) of less than 5% of the reference value for calibration, cross-validation, and prediction across three instruments demonstrate the reliability of process Raman to accurately predict the concentrations of these excipients, as well as easy model transferability. Figure 3. Correlation plot of predicted vs reference values for L-histidine, L-arginine, and sucrose across three different instruments. Figure 4. Raman concentration predictions during bench-scale UF/DF of IgG1 mAb for L-arginine (A), L-histidine (B), and sucrose (C); target end conc. shown with light blue dashed line. 0 0 10 3 20 5 30 7 Measured L-Arginine mg/mL Time (min) Predicted L-Arginine mg/mL L-arginine conc. (mg/mL) 30 180 10 60 0 0 20 120 40 240 0 40 80 120 Measured Sucrose mg/mL Predicted Sucrose mg/mL 0 50 100 150 200 0 5 10 15 Measured L-Histidine HCI mg/mL Predicted L-Histidine HCI mg/mL 0 1 2 3 Time (min) L-histidine conc. (mg/mL) 60 120 180 240 -10 50 70 90 Time (min) Sucrose conc. (mg/mL) 60 120 180 240 300 UF1 DF -1 9 A B C 0 0 4 5 UF1 DF 10 30 UF1 DF UF2 RMSEC = 0.44192 RMSECV = 0.44823 RMSEP = 1.1258 Calibration Bias = 3.5527e-15 CV Bias = -0.0005461 Prediction Bias = 0.17372 R² (Cal, CV) = 0.99, 0.999 R² (Pred) = 0.995 RMSEC = 0.90539 RMSECV = 2.3907 RMSEP = 3.5746 Calibration Bias = -7.1054e-15 CV Bias = NaN Prediction Bias = 1.0537 R² (Cal, CV) = 1.000, 0.998 R² (Pred) = 0.997 RMSEC = 0.19718 RMSECV = 0.99342 RMSEP = 0.66593 Calibration Bias = 0 CV Bias = NaN Prediction Bias = 0.015517 R² (Cal, CV) = 0.998, 0.931 R² (Pred) = 0.981 Fit 1:1 BOX 1 BOX 2 BOX 3 Fit 1:1 BOX 1 BOX 2 BOX 3 Fit 1:1 BOX 1 BOX 2 BOX 3 40 50 160 200 0 2 4 6 8 10 12 14 Concentration g/L Glucose 0 20 40 60 80 Fructose Sucrose Before After The L-histidine, L-arginine, and sucrose models were then applied to the data acquired during the ultrafiltration/diafiltration (UF/DF) process. The predicted values for L-histidine, L-arginine, and sucrose are shown in Figures 4A, 4B, and 4C, respectively. The average predicted concentrations of L-histidine, L-arginine, and sucrose at the end of buffer exchange were compared to the reference values, resulting in a prediction error of less than 5% of the reference values (Table 3). This clearly illustrates the capability of process Raman analyzers to monitor and quantify excipients in real-time. The real-time prediction of sucrose concentrations over 15 days at room temperature in a briefly air-exposed formulation buffer, containing L-histidine, L-arginine, and sucrose, is shown in Figure 5A. This scenario mimics what may occur during the storage of the formulation buffer. Initially, the predicted sucrose concentration was 86 mg/mL and remained stable for 5 days, but then steadily decreased to 57 mg/mL by day 15. Predictions from the arginine and histidine PLS models behaved similarly to those from the sucrose model, showing accurate and stable values up to day 5, but then steadily increasing until day 15 (data not shown). Raman spectral analysis revealed that the decrease in the sucrose peak was accompanied by the appearance of glucose and fructose Raman peaks (Figures 5B and 5C). HPLC analysis confirmed the intactness of arginine and histidine for all 15 days, and the hydrolysis of sucrose into glucose and fructose (Figures 5D and 5E). Figure 5. Showing the sucrose prediction over 14 days (A). Initially, the sucrose concentration was 86 mg/mL that lowered to about 57 mg/mL over the span of 14 days. The decrease in sucrose prediction were evident by decrease in sucrose specific intensity of 835 cm-1 that is assigned to the twisting (τ (CH₂)) with some contribution form symmetric stretching (ν(CC)) vibrational mode (B). Showing in figure C, the decrease in sucrose specific band (red arrow; ~550 cm-1 assigned as in-plane bending (β(OCO)) is followed by increase in glucose (blue; ~525 cm-1) and fructose (green; ~ 640 cm-1) specific Raman band that are assigned mainly to the deformation of CCC, CCO, and OCO bands. In figure D and E the result of day 1 (Before; blue) and day 15 (After; cyan) are compared where arginine and histidine remained unchanged while sucrose hydrolyzed to glucose and fructose. Table 3. Prediction Error Calculation. Excipient Reference concentration (g/L) Predicted concentration (g/L) % absolute error L-histidine 4.2 4.1 1.0 L-arginine 7.0 7.1 1.4 Sucrose 92.4 95.6 3.4 55 70 75 80 Linear index (x10⁴) Predicted sucrose mg/mL 0.5 1.0 1.5 2.0 2.5 85 90 60 65 A Raman shift (cm-1) Raman shift (cm-1) Raman intensity (a.u.) Raman intensity (a.u.) 600 1,000 1,400 -0.005 B C -0.010 0.005 0.000 0.010 500 550 650 -8 600 -6 -4 -2 0 2 4 6 Sucrose Glucose Fructose HPLC analysis Concentration g/L Histidine D E 0 2 4 6 8 Arginine Before After 100 Sucrose Glucose Fructose Hydrolysis Acidic hydrolysis of sucrose is well documented in the literature.³ To investigate if the root cause had pH-association, we examined the osmolarity and pH profiles during the experiment (Figures 6A and 6B). The onset of sucrose hydrolysis coincided with an increase in osmolarity and a decrease in pH, suggesting that the hydrolysis of sucrose into glucose and fructose is most likely driven by the lower pH. Since no acid was added to the system, the pH decrease was likely due to external factors. Although we did not identify the exact cause of the pH decrease, potential factors in practice could include bacterial growth, improper pH adjustment during buffer preparation, or dissolution of carbon dioxide or other acid-producing gases, among others. Note that the sucrose PLS model was developed using a mixture of L-histidine, L-arginine, sucrose, and protein, and lacks any spectral information from glucose and fructose. Since glucose and fructose have significant overlaps in a wide spectral region, this explains why the Raman predictions were higher compared to the reference HPLC values.⁴ The same reasoning applies to the discrepancies observed in Raman predictions for L-arginine and L-histidine (data not shown) when compared to the HPLC values (Figure 5D). In both cases, predictions can be improved by augmenting the model with additional training data that includes glucose and fructose spectral information. However, this was beyond the scope of the current work. Not including glucose and fructose spectral information in the models is advantageous for monitoring buffer quality. As glucose and fructose are produced by sucrose hydrolysis, new spectral features appear that were not present in the training dataset. The Q residual is one of the model statistics calculated using the residual spectra remaining after projecting the original spectra into the model space.⁵ As the spectral information of glucose and fructose increases with the progress of sucrose hydrolysis, the magnitude of the Q residual increases over time, as shown in Figure 7. Users can leverage this information to design quality control measures based on the reduced Q vs. T² plot to assess buffer quality.⁵ For instance, in this study, a mean value of 0.15 for the reduced Hotelling T² and 0.5 for the reduced Q residual, with 95% confidence intervals for upper and lower limits (red dotted oval in Figure 7), can be used as quality control thresholds. Any buffer with reduced Q and T² values beyond these limits is deemed to fail quality control. All spectra after day 5 had low reduced T² and high reduced Q values, thus failing the quality control. Figure 7. Showing increase in Q residual with sucrose hydrolysis, calculated by projecting the Raman data into the PLS sucrose model. The red dotted oval showing one of the possible control limits for quality assessment. 0 0.5 1.0 2.0 2.5 3.0 Hotelling T² Reduced (p=0.950) (96.94%) Q Residuals Reduced (p=0.950) (3.06%) 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 Figure 6. Showing increase in osmolarity by 40% (A) and decrease in pH by 1 unit during the hold period (B). 400 4.0 450 4.5 500 550 5.0 5.5 Hold Time (days) Hold Time (days) Osmolality (mOsm/kg) pH 0 2 4 6 8 10 12 14 0.15 1.5 3.5 4.0 A B 650 650 16 0 2 4 6 8 10 12 14 16 Learn more at thermofisher.com For research use only. Not for use in diagnostic procedures. For current certifications, visit thermofisher.com/certifications © 2025 Thermo Fisher Scientific Inc. All rights reserved. All trademarks are the property of Thermo Fisher Scientific and its subsidiaries unless otherwise specified. MCS-AN1440-EN 06/25 Conclusion 1. Real-time quantification of L-arginine, L-histidine, and sucrose were demonstrated with absolute error of < 5% in a UF/DF process using process Raman analyzer configured with FlowCell probe. In absence of in-line PAT tools to quantify these excipients, the volume needed for diafiltration (Vdf ) is carried out based on following mathematical expression: To balance the osmotic pressure difference, sucrose is excluded along with water.⁶ This effect was accurately captured in the real-time predicted data from the process Raman analyzer. Thus, process Raman provides unique capabilities to ensure product quality by offering real-time data, rather than relying on empirical hypotheses. 2. This work, combined with our previous demonstrations of accurate in-line protein quantification during UF/DF processes,⁷, ⁸ clearly highlights the value of process Raman for downstream process monitoring. Raman spectroscopy allows for the simultaneous measurement of multiple critical process parameters (CPPs) with a single scan. These findings establish process Raman as a PAT tool with unparalleled benefits compared to other analytical methods. 3. Simultaneous measurement of protein and excipient concentrations not only allows tighter process control but also opens opportunities for automating downstream processing. 4. The ability of process Raman to provide real-time insights into buffer quality before its use in UF/DF runs offers substantial value by preventing batch failures. This capability enhances quality control, making Raman spectroscopy an essential tool for integration as an in-line sensor to improve downstream process monitoring, control, and automation. References 1. Zhang, L.; Liang, Y.-Z.; Jiang, J.-H.; Yu, R.-Q.; Fang, K.-T. Uniform Design Applied to Nonlinear Multivariate Calibration by ANN. Analytica Chimica Acta 1998, 370 (1), 65–77. https://doi.org/10.1016/S0003-2670(98)00256-6. 2. Choquette, S. J.; Etz, E. S.; Hurst, W. S.; Blackburn, D. H.; Leigh, S. D. Relative Intensity Correction of Raman Spectrometers: NIST SRMs 2241 through 2243 for 785 Nm, 532 Nm, and 488 Nm/514.5 Nm Excitation. Appl Spectrosc 2007, 61 (2), 117–129. https://doi.org/10.1366/000370207779947585. 3. Torres, A. P.; Oliveira, F. a. r.; Silva, C. l. m.; Fortuna, S. p. THE INFLUENCE of pH ON the KINETICS of ACID HYDROLYSIS of SUCROSE. Journal of Food Process Engineering 1994, 17 (2), 191–208. https://doi.org/10.1111/j.1745-4530.1994. tb00335.x. 4. Wiercigroch, E.; Szafraniec, E.; Czamara, K.; Pacia, M. Z.; Majzner, K.; Kochan, K.; Kaczor, A.; Baranska, M.; Malek, K. Raman and Infrared Spectroscopy of Carbohydrates: A Review. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 2017, 185, 317–335. https://doi.org/10.1016/j.saa.2017.05.045. 5. Kumar, S.; Martin, E. B.; Morris, J. DETECTION OF PROCESS MODEL CHANGE IN PLS BASED PERFORMANCE MONITORING. IFAC Proceedings Volumes 2002, 35 (1), 125–130. https://doi.org/10.3182/20020721-6-ES-1901.00752. 6. Agrawal, P.; Wilkstein, K.; Guinn, E.; Mason, M.; Serrano Martinez, C. I.; Saylae, J. A Review of Tangential Flow Filtration: Process Development and Applications in the Pharmaceutical Industry. Org. Process Res. Dev. 2023, 27 (4), 571–591. https://doi. org/10.1021/acs.oprd.2c00291. 7. Nolasco, M.; Pleitt, K.; Khadka, N. Using a Process Raman Analyzer as an In-Line Tool for Accurate Protein Quantification in Downstream Processes. 8. Nolasco, M.; Pleitt, K.; Khadka, N. Raman-Based Accurate Protein Quantification in a Matrix That Interferes with UV-Vis Measurement. where: • V0 is the initial volume of the solution. • D is the number of diavolumes required. • R is retention factor • C0 is initial concentration of solute • Cf is the final concentration of solute In practice, the retention factor for excipients is typically assumed to be 0, as the pore size of the diafiltration membrane is significantly larger than the hydrodynamic size of excipients. However, charge buildup across the membrane results in electrochemical potential which in turn prevents the free mobility of charged excipients, thereby increasing their retention factor above 0. This effect is known as the GibbsDonnan effect.⁶ In such scenarios, the empirically calculated volume needed for diafiltration (Vdf ) can result in incomplete buffer exchange, which may affect the functionality and stability of monoclonal antibodies (mAbs). An in-line process Raman analyzer provides a reliable solution to this issue by offering real-time monitoring of excipient concentrations. This enables tighter process control and ensures product quality by allowing for immediate adjustments to the diafiltration process, thereby preventing incomplete buffer exchange and maintaining the stability and functionality of the therapeutic product. Similarly, in Figure 4C, the sucrose concentration in the retentate decreased during UF2, as confirmed by offline HPLC analysis. Given the hydrodynamic size of sucrose relative to the pore size of the membrane, sucrose should theoretically exchange freely between the retentate and filtrate, resulting in equal concentrations in both. However, as the protein concentration increases, the osmotic pressure also rises, making the exclusion of water thermodynamically unfavorable. Equation 1. Vdf = V0 * D = V0 * ln(C0 /Cf ) / ln(1−R) Protein secondary structure analysis using in-line process Raman and offline SERS surface Application note Authors Nimesh Khadka, Ph.D. Bin Han, Ph.D. Analytical Instrument Group, ThermoFisher Scientific, Tewksbury, Massachusetts USA Industry/Application: Biopharma PAT / R&D laboratories / Downstream Products used: Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer, Thermo Scientific™ MarqMetrix™ FlowCell Sampling Optic, Thermo Scientific™ MarqMetrix™ BallProbe™ Sampling Optic, Thermo Scientific™ OMNIC™ Software Suite Goals: • Demonstrate the capability of the MarqMetrix All-In-One Process Raman analyzer to acquire high-quality data for in-line protein secondary structure determination. • Highlight the easy probe swap feature, facilitated by an optical fiber cable, which allows for seamless deployment for offline uses. In this case, SERS measurements to determine protein structure in samples in low concentration samples are featured. Key Analytes: Lysozyme, protein secondary structure Key Benefits: • Process Raman provides real-time insight into protein secondary structure during downstream purification that ensures quality of products. • Process Raman’s ability to provide simultaneous information on protein concentration and its quality makes it a valuable process analytical technology (PAT) tool. • Use of process Raman for offline SERS measurements enable users to leverage MarqMetrix All-In-One Process Raman analyzer for all mode of applications: in-line, at-line, online, or offline. A. Thermo Scientific MarqMetrix All-In-One Process Raman Analyzer with Thermo Scientific MarqMetrix FlowCell Sampling Optic. B. Thermo Scientific MarqMetrixBallProbe Sampling Optic C. Easy swapable fiber head and other sampling probes A B C Introduction Proteins are made up of amino acids linked together by peptide bonds (amide bonds; -CO-(NH)-), which are formed when the α-carboxyl group of one amino acid reacts with the α-amino group of another, releasing a water molecule in the process.¹ This sequence of amino acids is called the primary structure. The primary structure can fold spontaneously, or with the help of molecular chaperones, into various secondary structures such as α-helices, β-sheets, β-turns, and 3₁₀ helices. These secondary structures then interact in three-dimensional space to form the protein’s tertiary structure. For proteins with multiple subunits, these individual units come together to form a multimeric three-dimensional complex through various interactions known as the quaternary structure. Protein secondary structures are crucial for determining the functions of proteins because they maintain the structural integrity of the molecules. Research has shown that tracking changes in these structures can reveal functional losses due to conformational changes, degradation, denaturation, and aggregation of proteins. Although X-ray crystallography and cryo-electron microscopy (cryo-EM) provide accurate information on protein secondary structures, they are impractical for routine use in protein biomanufacturing due to their high costs, lengthy processing times, and resource demands.² A more practical and efficient solution is optical analysis. Techniques like Fourier transform infrared spectroscopy (FTIR) and Raman spectroscopy have proven effective for quickly analyzing protein secondary structures in offline operation mode.³ For routine assessments, the spectroscopic data are processed using multivariate analyses to create chemometric models. These models are validated by comparing their results with established secondary structure information from X-ray crystallography or cryo-EM. Once confirmed for accuracy, the models can be used for rapid analysis of future batches of protein samples, enabling users to evaluate protein secondary structure in a laboratory setting.⁴ This study presents the feasibility of using the in-line Thermo Scientific™ MarqMetrix™ All-In-One Process Raman Analyzer for real-time protein secondary structure analysis during downstream processing. Our previous works have already demonstrated accurate protein quantification in clarified harvest and ultrafiltration/diafiltration (UF/DF), as well as excipients quantification in downstream workflow.⁵– ⁷ Thus, this work complements our previous findings and establishes process Raman as a reliable PAT tool for multicomponent monitoring and multi-modal feedback to enable tighter process control. Additionally, this report discusses the integration of process Raman analyzer with an internally developed SurfaceEnhanced Raman Spectroscopy (SERS) solid substrate as an offline alternative for determining secondary structures in samples with low protein concentrations. Experimental Details Data collection for initial proof of concept (PoC) A 17 mg/mL lysozyme solution was prepared in water. Approximately 1 mL of this solution was flushed through the Thermo Scientific™ MarqMetrix™ FlowCell Sampling Optic, which has a sampling volume of approximately 180 μL. The residual lysozyme solution inside the FlowCell after flushing was used to collect the Raman spectrum. The FlowCell was connected to the MarqMetrix All-In-One Process Raman analyzer via an optical fiber. Raman data were acquired using a laser power of 450 mW, an integration time of 5000 ms, and an average of 10 scans. Three acquired spectra were further averaged to improve the signal-to-noise ratio (SNR). The same strategy and acquisition parameters were used to acquire Raman spectra of water. Cation exchange purification A buffered solution of 3.95 mg/mL lysozyme in 20mM MES buffer pH 6.2 was loaded at 0.83 ml/mL into a column packed with POROS™ 50 HS Strong Cation Exchange Resin. After loading, the column was washed with 5 column-volumes of 20mM MES buffer pH 6.2. Finally, the bound lysozyme was eluted at a flow rate of 3mL/min using elution buffer (20mM MES buffer 1M NaCl pH 6.2). This entire chromatographic run was monitored with the UV-Vis detector of the AKTA Pure (300 nm wavelength) and MarqMetrix All-In-One Process Raman analyzer integrated with a FlowCell probe. The Raman spectra were acquired using the acquisition settings of laser power 450 mW, integration time 3000 ms, and average of 3. During the elution step, the lysozymes are concentrated and eluted from the column. When the concentrated lysozyme reached the FlowCell cavity, the flow was paused and the Raman spectrum was acquired with the acquisition setting of power 450 mW, integration time 5000 ms, and average of 10. Three of these spectra were averaged for further analysis. SERS data collection A 5 μL aliquot of 0.1 mg/mL lysozyme in water was added on the SERS surface. The sample was allowed to concentrate through evaporation for approximately 10 min. This step provides enhancement due to both the increased concentration and true SERS effect. Raman spectra were collected using a MarqMetrix All-In-One Process Raman analyzer integrated with the Thermo Scientific™ MarqMetrix™ Proximal BallProbe™ Sampling Optic via optical fiber. The Raman data were acquired using a laser power of 200 mW, an integration time of 5000 ms, and an average of 10 scans after optimizing the focal distance using XYZ three axis micrometer stage to obtain the maximum signal. Data analysis and peak deconvolution For the initial proof-of-concept (PoC) study, the Raman spectra of water and a 17 mg/mL lysozyme solution in water were smoothed using a Savitzky-Golay filter (window width = 7, order = 2) and then each spectrum was then normalized using the weight vector that was obtained by calculating the infinity norm in the region of 3000 to 3240 cm-1, which corresponds to the symmetric stretching of O-H bonds in water molecules. This region was chosen for normalization as it has minimal spectral interferences from the Raman signatures of lysozyme or buffer analytes, and thus can correct for any path length differences. After normalization, the spectral region from 1400 to 1800 cm-1 was selected for each spectrum. The baseline was removed from each spectrum using an Automatic Whittaker filter (lambda = 5000, asymmetry (p) = 0.001). The baseline-removed water Raman spectrum was then subtracted from the baseline-removed lysozyme solution Raman spectrum to isolate the pure lysozyme spectrum. This pure lysozyme Raman spectrum was preprocessed using a Savitzky-Golay filter (window width = 9, order = 2, second derivative) to identify the number and positions of peaks. After identifying the peaks, the pure lysozyme Raman spectrum was exported as a .spc file. All data analysis explained above were performed using the SOLO 9.3.1 software package (2024, Eigenvector Research, Inc., Manson, WA, USA 98831). The exported .spc file was loaded into Thermo Scientific™ OMNIC™ Software for peak deconvolution in the spectral region of approximately 1500 to 1800 cm-1. The Voigt function was selected for peak fitting. The initial guesses for the number of peaks and their positions were defined based on previously calculated second derivative data. The initial guess for the full width at half maximum (FWHM) for the Voigt peaks was set at 8 cm-1, considering the resolution of instrument was about 6 cm-1. The noise level was calculated using the spectral range of 1750 to 1780 cm-1. With these initial guesses, the global optimization was performed in the OMNIC software to optimize for the number of peaks, peak positions, and peak widths. Convergence was attained by minimizing the residual between the observed spectrum and the fitted spectrum. Finally, the positions and percentage contributions of the fitted peaks within the spectral region of 1630 to 1700 cm-1 were obtained to assign the types of secondary structures and their respective amounts. The Raman spectra collected on lysozyme during cation exchange purification were analyzed similarly , except that instead of pure water, the Raman spectrum of 20 mM MES buffer with 1 M NaCl at pH 6.2 was used to subtract the background information. For SERS data analysis, the background spectrum was collected using water. The rest of the data analysis was the same as explained above. Figure 1. Spectral overlay of 17.5 mg/mL lysozyme in water (red) and water (green) is shown in plot A. The fingerprint region between 800 to 1800 cm-1 Raman shift is shown in plot B. 0.0 0.5 1.0 1.5 2.0 Raman shift (cm-1) Raman intensity (a.u.) 500 1,000 1,500 2,000 2,500 3,000 Lysozyme in water Water 0.1 0.2 0.3 0.4 0.5 0.6 Raman shift (cm-1) Raman intensity (a.u.) 800 1,000 1,200 1,400 1,600 1,800 0.7 0.8 0.9 A B Lysozyme in water Water Results and Discussions Initial PoC study The spectrum of 17.5 mg/mL lysozyme is shown in Figure 1. In Figure 1A, the spectra are normalized using infinity norm calculated in the region 3000 to 3240 cm-1. Distinct Raman features of lysozymes in the fingerprint region are shown in Figure 1B; these match with reference values reported in the literature.⁸ The spectral overlay after removal of baseline for the Raman shift region 1500 to 1800 cm-1 is shown in Figure 2. The pure lysozyme spectrum obtained after subtracting water background is shown in Figure 3A. Likewise, the second derivative plot of Figure 3A is shown in Figure 3B which illustrates the presence of four peaks (blue arrow) within the region of 1630 to 1700 cm-1. The result of peak deconvolution of Figure 3A using Voigt function is shown in Figure 4. The Voigt function was selected based on the recommendation from the literature for aqueous biological samples.⁹ The spectral region 1620 to 1750 cm-1 of protein is called Amide I region. The peaks in the Amide I region are attributed to the symmetric stretching of the carbonyl functional group in the peptide bond. The carbonyl groups in different secondary structures experience different electronic environments, leading to different vibrational energies that appear as peaks at different positions. Thus, the Amide I region provides information on the secondary structure of proteins.9 The symmetric stretching of the carbonyl group in the peptide bond is both Raman and infrared (IR) active because this vibrational mode is associated with change in polarization and dipole moment during transition, making both Raman and IR spectroscopic techniques suitable for secondary structure analysis. In the literature, offline FTIR, FT-Raman and drop coat deposition Raman (DCDR) have been reported for the study of native protein secondary structures as well as monitoring structural changes during protein degradation, denaturation, aggregation, and chemical modification. However, the possibility of elucidating the protein secondary structure in aqueous phase using in-line Raman has not been reported, which sets the foundation for this study. Figure 3. The water background-subtracted lysozyme spectra is shown in plot A and its 2nd derivative is shown in plot B. The blue arrows in the plot B shows four peaks in the spectral region of 1630 to 1700 cm-1. Figure 2. Spectral over of selected region after baseline removal. The lysozyme peak is distinct at approximately 1660 cm-1. 0.00 0.05 0.10 0.15 0.20 Raman shift (cm-1) Raman intensity (a.u.) 1,500 1,550 1,600 1,650 1,700 1,750 Lysozyme in water Water 0.25 0.30 0.35 -0.01 0.00 0.01 0.02 0.03 Raman shift (cm-1) Raman intensity (a.u.) 1,580 1,600 1,620 1,640 1,660 1,680 -1.5 -1.0 -0.5 0.0 0.5 Raman shift (cm-1) Raman intensity (a.u.) x10-3 1,600 1,620 1,640 1,660 1,680 1,700 A B Lysozyme 1,700 1,720 0.04 0.05 0.06 0.07 0.08 0.09 The peak positions and peak contributions in the Amide I region of lysozyme from Figure 4 are summarized in Table 1. The results from process Raman showed that there were four peaks centered at Raman shift of 1640, 1657, 1672 and 1688 cm-1 that were assigned to the extended conformation (random coils), α-helices, β-sheet, and extended/PPII (β-turns) respectively. The assignments were adopted based on published work.8 The peak positions and their relative contribution to the Amide I region were in close agreement with the published work with FTIR, FT-Raman, X-ray crystallography, and DCDR. These results clearly demonstrated that the data quality from the MarqMetrix™ All-In-One analyzer provides a platform for the direct analysis of the secondary structure of proteins in the aqueous phase without any sample preparation. Cation exchange purification After initial validation of the possibility of in-line protein secondary structure analysis, the strategy was applied to the in-line process data collected during elution step of a lysozyme form column packed with POROS™ 50 HS Strong Cation Exchange Resin. The eluted lysozyme concentration was approximately 25 mg/mL as confirmed by offline analysis. The result for peak deconvolution of Amide I region of lysozyme (1630 to 1700 cm-1) from the in-line process data is shown in Figure 5. A small peak at approximately 1700 cm-1 was also observed that was initially assumed to be associated with aggregated lysozyme. However, literature review on Raman based studies on lysozyme aggregation revealed that the peak at 1700 cm-1 was not associated with lysozyme. Thus, four peaks centered at 1640, 1656, 1670, and 1684 cm-1 were only considered Amide I peaks. The percentage area contributions for these selected peaks, shown in Figure 5, were in close agreement with values shown in Table 1. Note, even if the uncertain 1700 cm-1 were taken into consideration, it would have affected the results insignificantly. These results clearly demonstrated that process Raman is a valuable in-line process analytical technology (PAT) that can provide real time insights into protein secondary structure as the downstream process is happening. Secondary Structure Process Raman FT-Raman FT-IR X-Ray Peak Center % Area Peak Center % Area Peak Center % Area Peak Center % Area Solvent Exposed Extended Conformation 1640 24 1637 27 1645 19 NA 19 α-Helix 1657 42 1655 42 1654 40 NA 45 β-Sheet 1672 23 1673 27 1670 20 NA 23 Extended and PPII 1688 11 1685 4 1683 21 NA 13 Secondary Structure Raman Peak Center %Area Solvent Exposed Extended Conformation 1640.803 22.59 α-Helix 1656.157 41.93 β-Sheet 1670.16 24.47 Extended and PPII 1684.597 11.01 Figure 5. Showing result of peak deconvolution performed on inline process data using process Raman. Table 1. Results from peak convolution from process Raman and its comparison with published result using FT-Raman, FTIR, and X-ray crystallography. 0 200 400 600 800 1,000 Raman shift (cm-1) Raman Intensity 1,600 1,620 1,640 1,660 1,700 Figure 4. The result of peak deconvolution of lysozyme spectra of Figure 3A is shown. The four peaks in the spectral region between 1620 to 1750 cm-1 corresponds to four different secondary structure of lysozyme. 0.00 0.02 0.04 0.06 0.08 0.10 Raman shift (cm-1) Raman Intensity 1,550 1,600 1,650 1,700 0.01 0.03 0.05 0.07 0.09 1,680 1,200 1,400 1,600 1,800 2,000 2,200 2,400 SERS SERS is often used as direct or indirect sensor to identify and quantify biological or non-biological analytes due to its high sensitivity. In some cases, the detection limit is close to a single molecule.¹⁰ Several types of SERS exist including colloidal solution, suspended nanoparticle in solution, immobilized nanoparticles on solid surface, nanostructure fabricated on solid surface, or metallic nanowires/foils. Because of high stability, sensitivity, reproducibility of results, low background signal, and homogeneous signal enhancement across the surface, the internally developed SERS substrate was evaluated for this study to analyze the protein secondary structure. The experimental set up and the results for lysozyme secondary structure analysis using SERS substrate is shown in Figure 6. The laser spot size of ~ 500 μm allows sampling from larger area of SERS substrate while the optical fiber provides easy connectivity and sample handling during SERS measurement using the MarqMetrix All-In-One analyzer coupled with the Thermo Scientific™ MarqMetrix™ Proximal BallProbe™ Sampling Optic. A non-contact probe was used to avoid any physical disturbance to the dried protein sample on the SERS substrate and to prevent contamination or degradation that could interfere with the signal. This setup also enabled easier handling of delicate low-concentration samples. The peak deconvoluted results for Amide I of lysozyme from SERS were like that shown in Table 1. Thus, SERS is a viable option to study protein secondary structure when the sample is at low concentration. Figure 6. Demonstrating the workflow for SERS measurement using the MarqMetrix All-In-One Process Raman analyzer and the output of peak deconvolution for Amide I region of lysozyme. Secondary Structure SERS Peak Center % Area Solvent Exposed Extended Conformation 1644 23 α-Helix 1659 41 β-Sheet 1675 24 Extended and PPII 1691 12 0 1,000 2,000 3,000 Raman shift (cm-1) Raman Intensity 1,600 1,700 10,000 4,000 5,000 6,000 7,000 8,000 9,000 Thermo Scientific™ Marqmetrix™ All-In-One Process Raman Analyzer Thermo Scientific™ Marqmetrix™ Proximal BallProbe™ Sampling Optic Fiber optics ~8mm Optimize Distance SERS Chip XYZ micrometer stage As shown in Figure 7, the Raman spectrum of 20 mg/mL lysozyme (red) and 0.1 mg/mL lysozyme on SERS surface (green) have overall similar peaks profile but minute differences in the spectral regions at 1358 cm-1 (CH deformation), 1210 cm-1 (aromatic residue), 935 cm-1(C-C stretch), ratio of 508 cm-1, 523 cm-1, 540 cm-1 (disulfide region) were also observed. This is expected for larger molecules as not all sections of lysozyme experience homogeneous electro-chemical SERS enhancement effect.¹¹ Nonetheless, the result shown in Figure 7 demonstrates these SERS substrates are homogeneous and can be used for protein analysis. To test enhancement efficiency of SERS substrate, using the same methodology the data was collected on the non-SERS substrate for 0.1 mg/mL lysozyme but only a weak spectrum was observed (data not shown) under the experimental conditions. Thus, a significant portion of enhancement was attributed to SERS effect, not increase in concentration due to drying as is the case in DCDR. Another key aspect of SERS measurement is the enhancement of Raman signal relative to fluorescence. The SERS spectra of lysozyme (green) in Figure 7 is shown without baseline removal while aqueous 20 mg/mL lysozyme (red) is plotted after removal of baseline. Thus, SERS can be advantageous for samples that have high fluorescence. Although the details are beyond the scope of this work, a plot is shown in supplementary information that illustrates the enhancement of Raman signals even with high fluorescent background (Figure S1). Conclusion In this study, the feasibility of using the Thermo Scientific MarqMetrix All-In-One Process Raman Analyzer for analyzing the secondary structure of lysozyme is demonstrated. The lysozyme secondary structures that were determined were comparable to those obtained using FTIR, FT-Raman, DCDR, and X-ray crystallography. This method’s accuracy and the lack of need for sample preparation offer significant advantages for probing protein secondary structure in its native aqueous state. This capability can be leveraged to determine protein secondary structure during real-time downstream processing (e.g., UF/ DF), where protein concentrations are relatively high, aiding in the acquisition of high SNR spectra. This ensures the quality of proteins during purification, as published work has shown changes in secondary structure with protein denaturation, aggregation, or degradation. Thus, this capability helps users make actionable decisions to proceed with further downstream processes confidently or, in some cases, scrap the run when quality is compromised. In conjunction with previous work, this study establishes process Raman as a PAT tool for both realtime quantification and quality assessment. SERS-based protein secondary structure analysis is also demonstrated in a sample with low concentration lysozyme. The protein secondary structures determined from SERS and aqueous solutions were similar, and the accuracy was comparable to FTIR, FT-Raman, DCDR, and X-ray crystallography. The SERS measurement was performed using the MarqMetrix All-In-One Process Raman analyzer, leveraging its features to easily swap probe types and use optical fibers for ease of use in laboratory settings. References 1. Branden, C. I.; Tooze, J. Introduction to Protein Structure, 2nd ed.; Garland Science: New York, 2012. https://doi.org/10.1201/9781136969898. 2. Wang, H.-W.; Wang, J.-W. How Cryo-Electron Microscopy and X-Ray Crystallography Complement Each Other. Protein Sci. 2017, 26 (1), 32–39. https://doi.org/10.1002/ pro.3022. 3. Pelton, J. T.; McLean, L. R. Spectroscopic Methods for Analysis of Protein Secondary Structure. Anal. Biochem. 2000, 277 (2), 167–176. https://doi.org/10.1006/abio.1999.4320. 4. Peters, J.; Jin, C.; Luczak, A.; Lyons, B.; Kalyanaraman, R. Machine Learning Enabled Protein Secondary Structure Characterization Using Drop-Coating Deposition Raman Spectroscopy. J. Pharm. Biomed. Anal. 2025, 259, 116762. https://doi.org/10.1016/j.jpba.2025.116762. 5. Khadka, N.; Pleitt, K.; Nolasco, M. A Classical Least Squares (CLS) Approach for Protein Quantification in Downstream Processing Using Raman Spectroscopy. 6. Nolasco, M.; Pleitt, K.; Khadka, N. Using a Process Raman Analyzer as an In-Line Tool for Accurate Protein Quantification in Downstream Processes. 7. Nolasco, M.; Pleitt, K.; Khadka, N. Raman-Based Accurate Protein Quantification in a Matrix That Interferes with UV-Vis Measurement. 8. Dolui, S.; Mondal, A.; Roy, A.; Pal, U.; Das, S.; Saha, A.; Maiti, N. C. Order, Disorder, and Reorder State of Lysozyme: Aggregation Mechanism by Raman Spectroscopy. J. Phys. Chem. B 2020, 124 (1), 50–60. https://doi.org/10.1021/acs.jpcb.9b09139. 9. Peters, J.; Park, E.; Kalyanaraman, R.; Luczak, A.; Ganesh, V. Protein Secondary Structure Determination Using Drop Coat Deposition Confocal Raman Spectroscopy. 2016, 31, 31–39. 10. Schlücker, S. Surface-Enhanced Raman Spectroscopy: Concepts and Chemical Applications. Angew. Chem. Int. Ed. 2014, 53 (19), 4756–4795. https://doi.org/10.1002/ anie.201205748. 11. Pérez-Jiménez, A. I.; Lyu, D.; Lu, Z.; Liu, G.; Ren, B. Surface-Enhanced Raman Spectroscopy: Benefits, Trade-Offs and Future Developments. Chem. Sci. 2020, 11 (18), 4563–4577. https://doi.org/10.1039/D0SC00809E. Figure 7. Demonstrating the similarity of Raman spectrum of aqueous 20 mg/mL lysozyme (baseline corrected and water background subtracted) and 0.1 mg/mL lysozyme (without baseline correction) on SERS substrate. Raman shift (cm-1) Raman intensity (a.u.) 600 800 1,000 1,200 1,400 1,600 0.1 mg/mL Lysozyme SERS 20 mg/mL Lysozyme in water Learn more at thermofisher.com For research use only. Not for use in diagnostic procedures. For current certifications, visit thermofisher.com/certifications © 2025 Thermo Fisher Scientific Inc. All rights reserved. All trademarks are the property of Thermo Fisher Scientific and its subsidiaries unless otherwise specified. MCS-AN1443-EN 06/25 Supplementary information Figure S1. Demonstrating of SERS substrate enhancement of Raman signal even in presence high fluorescence. Both spectra were collected using liquid samples using same acquisition parameters, demonstrating SERS substrate for both liquid and solid samples. 0 0.5 1.0 1.5 2.0 2.5 Raman shift (cm-1) Data (x10-4 ) 500 1,000 1,500 2,000 2,500 3,000 3.0 3.5 Rhodhamine Droplet SERS Rhodhamine Droplet Hydrophobic Surface Notes Learn more at thermofisher.com For research use only. Not for use in diagnostic procedures. For current certifications, visit thermofisher.com/certifications © 2025 Thermo Fisher Scientific Inc. All rights reserved. All trademarks are the property of Thermo Fisher Scientific and its subsidiaries unless otherwise specified. MCS-CM1559-EN 9/25