Automated chromatogram evaluation for digital twin based monitoring of protein refolding reactions

Abstract

Inclusion body refolding processes, controlled using model predictive controllers based ondigital twins, require the quantitative determination of protein folding states via chromatography.The current methods for this are slow, inconsistent, and predominantlymanual due to significant peak overlaps in chromatograms, making different folding stateschallenging to distinguish. This master thesis aims to improve and automate the evaluationprocess of highly overlapped chromatograms to reduce the delay between sampledrawing and arrival of the evaluated result, and thus leading to real-time monitoringand control of fed-bach inclusion body refolding processes. The research identifies theessential steps for automatic chromatogram evaluation: pre-processing (signal smoothingand baseline detection), peak detection, peak separation and post-processing. Subsequently,different methods for performing these steps were evaluated and implemented ina Python application as well as partially adapted. Based on this, a procedure for assemblingand parameterizing the application for the automatic evaluation of chromatogramsduring fed-batch refolding was developed and tested in the course of an experiment andcompared to a manual evaluation approach. The developed application encompasses Signalsmoothing methods based on a moving average, a Gaussian, and a Savitzky Golaykernel. In the context of smoothing, the automatic parametrization of smoothing algorithmsthrough autocorrelation coefficients was investigated. Therefore an extractionmethod for the noise of the entire chromatogram was developed which is based on theuse of the Fourier transform as a low-pass filter. In the context of baseline detection andbaseline calculation, methods were implemented using the Python package pybaselines,with particular emphasis on the fast-chrome method, which detects baselines based onthe standard deviation. Regarding peak detection, methods based on the Python scipypackage, the first, second, and multiple derivatives were integrated and compared. Forpeak separation, the methods of perpendicular dropline and model-based deconvolutionwere implemented. Furthermore, a model based halve peak deconvolution method and aGaussian separation method were developed and tested. Assessing the accuracy of thesemethods for inclusion body refolding chromatograms remains complex due to multivariateparameterization challenges and non-comparability with benchmark values. However,with the deconvolution method and the Gaussian separation method, it can be seen thatthe results change in the reverse bias direction compared to the perpendicular precipitation method. For all the tested separation algorithms for the evaluation of the conductedexperiment, it was shown that the method of model-based deconvolution (with the Gaussianmodel) worked best. This work establishes a foundational Python-based platformfor automated chromatogram evaluation, facilitating an easy way for the implementationand testing of additional chromatogram evaluation methods.