Investigating Transferability of Multivariate Calibration Models for Bioprocessing Applications

Abstract

Raman spectroscopy has emerged as a prominent Process AnalyticalTechnology (PAT) in bioprocessing due to its ability to deliver near realtime,non-destructive, multi-attribute measurements that support Quality byDesign (QbD) concepts, such as advanced control strategies, continuedprocess verification (CPV) and real-time release testing (RTRT). However,extracting quantitative information from spectra typically requiresmultivariate calibration models (most commonly partial least squares, PLS),and these data-driven models frequently exhibit limited transferability acrosschanges in instruments, acquisition settings, and process or productconditions. In industrial practice, this lack of transferability drives a relianceon process- or product- specific calibration models, resulting in repeatedrecalibration and revalidation whenever a new domain arises. This increasesthe resources required for PAT model calibration and hinders widespreadadoption.This industrial dissertation, conducted in close collaboration withPAT practitioners and motivated by real-world constraints, investigates thedrivers of calibration model transferability for Raman-based soft sensors,with a focus on upstream mammalian cell culture applications andindustrially relevant transfer scenarios (e.g. change of spectroscopichardware, process parameters and product attributes). The thesis proposes aholistic framework for model transferability by improving transferability atmultiple phases of the calibration lifecycle: (i) high-throughput datageneration, including targeted spiking, (ii) analyte-informed model trainingand interpretability-driven diagnostics, (iii) transfer task characterization viadataset shift concepts (covariate, conditional, and prior shift), and (iv) theselection and development of calibration transfer and maintenance (CTM)strategies that address common practical bottlenecks (e.g. limited labeleddata availability in new processes).Three first-author contributions form the core of the thesis. First, anunsupervised pre-processing optimization strategy is developed to enablecalibration transfer using unlabeled target spectra by selecting preprocessingvia maximum mean discrepancy (MMD) minimization betweensource and target prediction distributions. The method is demonstrated onindustrial case studies from a gene therapy process and systematicallybenchmarked in simulations, showing that pre-processing-based transfer canbe highly effective under covariate shift, while offering limited ability toresolve conditional and prior shift without additional information. Second,the thesis demonstrates that for “difficult-to-transfer” analytes such asmonoclonal antibody (mAb) titer, limited transferability is primarilya selectivity issue, where models can perform strongly in the trainingdomain yet degrade in a target domain (e.g., a different cell line) becausethey rely on indirect, correlation-driven spectral features arising from covaryingcell culture components rather than on analyte-specific signals.Using a high-throughput cell culture platform with automated spectralacquisition and analyte spiking, a transferable Raman titer model is achievedacross two cell lines with distinct metabolic profiles, supported byregression-coefficient interrogation that confirms increased selectivity.Third, analyte-informed hybrid calibration methods (hybrid linear analysis,HLA, and constrained regularization, CR) are evaluated as an algorithmicalternative to extensive spiking, demonstrating that incorporating analytereference information can improve robustness to changing correlationstructures while reducing experimental burden; HLA is found to beparticularly robust across reference choices and practical pre-processingconstraints.Additionally, three supporting contributions expand the scope of thepresented findings, focusing on (i) high-throughput Raman data generationusing a Raman microscope coupled to a multi-well plate including an indepthinvestigation of model selectivity, (ii) benchmarking severalunsupervised transfer learning approaches against practical pre-processingbaselines in simulated and real-world transfer tasks, and (iii) the introductionof a pragmatic framing for transfer-task characterization by jointlyinterpreting observed prediction errors and dataset-shift patterns, therebysupporting more systematic calibration transfer and maintenance methodselection in industrial practice.Overall, this work provides an actionable, industry-orientedframework for moving from local toward transferable Raman calibrationmodels by linking shift characterization, selectivity diagnostics, and transferstrategy selection. Future work should prioritize spectroscopy-specific,operational shift diagnostics (especially for unlabeled targets), systematicvalidation across in-line (probe-based) and at-line (high-throughput) dataacquisition strategies, and extension of transferability concepts to morechallenging targets, such as product quality attributes, where robustnessrequirements are especially critical.