Custom Spectral Preprocessing Improves Raman Bioreactor Model Predictions Across Scales

Chemometric models calibrated on small-scale high-throughput bioreactors often suffer severe predictive performance loss when deployed at larger manufacturing scales, but tailoring spectral preprocessing pipelines to individual cell culture targets can significantly mitigate this transfer gap. In biopharmaceutical development, monitoring cell culture dynamics in real time is essential for ensuring product quality and operational efficiency. High-throughput mini-bioreactor systems allow researchers to gather large, diverse datasets across varied experimental conditions, yet models derived from these mini-scale cultures frequently falter when scaled up. Investigators evaluated whether variable-specific preprocessing pipelines could address out-of-distribution challenges when transferring Raman spectroscopic models from high-throughput systems to five-liter benchtop bioreactors without relying on multi-scale calibrations.

Biopharmaceutical manufacturing demands precise real-time monitoring and control to ensure product safety, efficacy, and consistent quality, particularly because biologics like recombinant proteins and monoclonal antibodies are structurally complex molecules produced in tightly regulated cellular environments. Even slight fluctuations in critical process parameters can substantially alter critical quality attributes, making robust monitoring essential during development and process transitions. Within the framework of Process Analytical Technology, Raman spectroscopy has emerged as a valuable tool capable of generating molecular fingerprints through inelastic light scattering. The method operates in a label-free, non-destructive, and real-time manner directly within aqueous cell culture environments, allowing the simultaneous tracking of multiple analytes and physical process attributes.

Constructing functional Raman chemometric models requires systematic data collection, spectral preprocessing, and multivariate regression model building. Signals originating from key process analytes are frequently weak, obscured by background fluorescence, or buried beneath overlapping vibrational bands from complex media components and host cell proteins. While high-throughput systems such as the Ambr 250 mini-bioreactor accelerate calibration data generation through design-of-experiments approaches, cross-condition variability often degrades predictive performance during scale-up. In practical settings, transitioning between scales involves distinct instrumentation, optical probe designs, and spectral acquisition parameters, creating systematic variations in baseline shape and scattering intensity that do not reflect true chemical concentrations.

To overcome these cross-scale discrepancies, previous modeling efforts frequently turned to combined-scale calibration models that incorporate spectral measurements collected directly from larger bioreactors, or explored synthetic data augmentation, domain adaptation, and advanced machine learning algorithms. While combined-scale models are effective, acquiring extensive large-scale calibration datasets undermines the economic efficiency and speed of high-throughput development pipelines. Rather than altering regression algorithms or collecting multi-scale data, the researchers investigated whether systematic, variable-specific spectral preprocessing could clean and structure the single-scale spectral data matrix prior to regression, thereby insulating the resulting models from scale-dependent optical discrepancies.

The study evaluated culture runs of a stable Chinese hamster ovary CHO-DG44 cell line producing a recombinant enzyme across thirty total batches. The calibration foundation comprised twenty-three batches cultivated in an Ambr 250 high-throughput mini-bioreactor platform using commercial chemically defined media and a design-of-experiments matrix with varied operating conditions. Cross-scale model testing was conducted on five independent five-liter bench-scale bioreactor batches. Two additional bench-scale batches were allocated as supplemental calibration data to benchmark against combined-scale models, ensuring that bench-scale prediction batches remained strictly separate from all parameter estimation, latent variable selection, and calibration routines.

Raman spectra were collected throughout cell cultivation using distinct optical hardware across the two bioreactor platforms. The Ambr 250 system utilized an at-line BioPAT spectro analyzer collecting ten scans with a thirty-second exposure time per scan, whereas the five-liter bioreactors utilized an in-line Raman Rxn2 analyzer acquiring seventy-five scans with a ten-second exposure time per scan. Both systems operated at a seven hundred eighty-five nanometer laser excitation wavelength with four hundred milliwatts of laser power across a spectral range of one hundred to three thousand four hundred twenty-five reciprocal centimeters, which was subsequently restricted to four hundred fifty to one thousand eight hundred reciprocal centimeters for modeling. Reference analytical measurements, including viable cell density, cell viability, glucose, lactate, glutamine, glutamate, ammonia, and lactate dehydrogenase, were collected daily using automated cell counting and biochemical profiling systems.

Spectral preprocessing pipelines were assembled from three established chemometric methods: a first derivative calculated with a Savitzky-Golay algorithm using a second-order polynomial and a fifteen-point window, Standard Normal Variate transformation to correct multiplicative scattering, and Asymmetric Least Squares baseline correction to remove broad, nonlinear fluorescence baselines. Chemometric models were constructed using Orthogonal Partial Least Squares regression in SIMCA software, with latent variable numbers selected via seven-fold cross-validation on the calibration dataset. An additional leave-one-group-out cross-validation across culture batches verified that internal performance was not an artifact of time-course correlations within individual batches.

Standard single-scale models calibrated using a conventional pipeline combining first derivative preprocessing followed by Standard Normal Variate transformation achieved strong internal cross-validation performance, with low relative cross-validation errors between 3.4 percent and 8.1 percent and high coefficients of determination. However, when applied to the five independent bench-scale batches, the standard single-scale models exhibited severe cross-scale prediction errors and substantial systematic biases. While robust outputs such as glucose, lactate, and viable cell density transferred reasonably well, predictions for viability, glutamine, glutamate, ammonia, and lactate dehydrogenase deteriorated significantly. Combined-scale models demonstrated that incorporating benchtop data reduced these prediction errors, confirming that scale-dependent artifacts were skewing single-scale predictions.

To test how preprocessing order influences spectral information, the researchers intentionally reversed the sequence by applying Standard Normal Variate transformation prior to the first derivative. Normalizing spectra prior to baseline removal caused large fluorescence drifts to dominate the standard deviation calculation, which compressed the dynamic range of true analyte signals in principal component analysis and Orthogonal Partial Least Squares score spaces. This structural distortion resulted in catastrophic cross-scale failure even for robust analytes like glucose and viable cell density, confirming that improper preprocessing can inadvertently destroy subtle chemical variance while preserving deceptive internal calibration metrics.

Systematic screening of variable-specific preprocessing pipelines revealed that matching preprocessing techniques to the unique spectral properties of each analyte dramatically enhanced single-scale transferability. For outputs characterized by weak Raman scattering or severe spectral overlap, incorporating Asymmetric Least Squares baseline correction effectively decoupled chemical signals from fluctuating fluorescence baselines and batch-correlated media shifts. Compared to the standard pipeline, optimized variable-specific pipelines reduced cross-scale root mean square errors of prediction by 36.7 percent for cell viability, 35.0 percent for glutamine, 45.9 percent for glutamate, 14.0 percent for ammonia, and 56.1 percent for lactate dehydrogenase.

Analysis of variable importance in projection scores and model regression coefficients confirmed the spectroscopic basis of these improvements. While variable importance scores broadly reflected general process-wide culture progression, regression coefficient profiles demonstrated that optimized preprocessing shifted the most influential spectral regions toward known literature-reported Raman bands for specific metabolites. For glutamate and cellular biomass attributes, tailored baseline suppression removed broad background shifts and allowed the regression model to focus on true molecular vibrations. Permutation testing with one hundred iterations confirmed that the optimized models reflected genuine chemical correlations rather than statistical artifacts, yielding negative cross-validation values under randomized response permutations.

The findings demonstrate that variable-specific preprocessing enables single-scale high-throughput Raman models to approach the cross-scale predictive accuracy of labor-intensive combined-scale calibrations without requiring large-scale experimental runs. Nonetheless, the authors note that the experimental validation was conducted on a single CHO-DG44 fed-batch culture system, and extending the methodology to different cell lines, media formulations, perfusion operational modes, and alternative spectroscopic instruments will require further study. Combining targeted preprocessing screening with formal mathematical model transfer algorithms represents a promising avenue to secure scalable, cost-efficient Process Analytical Technology monitoring across biopharmaceutical manufacturing platforms.

Subject of Research: Raman spectroscopy cross-scale model transfer in bioprocess monitoring

Article Title: Variable-specific preprocessing enables cross-scale transferable Raman models from high-throughput bioreactor data

Article References: Han, S.-H., Park, J., Lee, K.-B., Park, C.-H., & Lee, D.-Y. (2026). Variable-specific preprocessing enables cross-scale transferable Raman models from high-throughput bioreactor data. Bioresources and Bioprocessing, 13(1), Article 122. https://doi.org/10.1186/s40643-026-01105-5

Image Credits: AI Generated

DOI: 10.1186/s40643-026-01105-5

Keywords: Raman spectroscopy, chemometrics, process analytical technology, bioreactors, bioprocessing, cell culture monitoring, CHO cells, orthogonal partial least squares, spectral preprocessing, asymmetric least squares, standard normal variate, scale-up, high-throughput bioreactors, metabolite monitoring, viable cell density, model transferability, fluorescence interference