% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@PHDTHESIS{Helleckes:999462,
author = {Helleckes, Laura Marie},
othercontributors = {Oldiges, Marco and Matuszynska, Anna Barbara and Wiechert,
Wolfgang},
title = {{A}utomated experimentation, {B}ayesian statistics and
machine learning for high-throughput bioprocess development},
school = {RWTH Aachen University},
type = {Dissertation},
address = {Aachen},
publisher = {RWTH Aachen University},
reportid = {RWTH-2024-12118},
pages = {1 Online-Ressource : Illustrationen},
year = {2024},
note = {Veröffentlicht auf dem Publikationsserver der RWTH Aachen
University 2025; Dissertation, RWTH Aachen University, 2024},
abstract = {The transition to a sustainable, circular bioeconomy is
essential to tackle the socioecological crises of the 21st
century. Industrial biotechnology, a cornerstone of this
bioeconomy, leverages modern biofoundries that integrate
automation and high-throughput experimentation with the
Design-Build-Test-Learn (DBTL) cycle to streamline
bioprocess development. While advances in automated cloning
and genome editing have increased the availability of large
strain libraries for early-stage screening, several
limitations remain in the Test and Learn phases of DBTL.
This work combines automated experimentation, Bayesian
statistical modelling and machine learning to bridge the
remaining gaps towards autonomous bioprocess development.
This requires an $\textit{experiment-in-the-loop}$-approach,
where simulations are closely coupled with experiments on
automated microbioreactor platforms. Consequently, the
toolboxes for experimental workflow development and decision
making based on process models were extended in this thesis.
These improved tools were then applied to biotechnological
case studies, focusing on model-driven experimental design
and iterative screening. First, manual steps in microbial
screening, such as precultures in shake flasks, were
replaced by automated solutions. Existing automated
microbioreactor platforms were thus extended to enable
consecutive screening experiments without human
intervention. For example, an automated deep freezer was
seamlessly integrated, including the connection to the
existing process control infrastructure. Furthermore,
automated precultures and microtiter plate recycling were
achieved for the microbioreactor, leading to the
demonstration of a fully automated, iterative screening with
cutinase-secreting $\textit{Corynebacterium glutamicum}$
strains. With the gaps in automated experimentation closed,
the focus was shifted to high-throughput data analysis and
process modelling. A need was identified for the evaluation
of analytical calibration data, for example from
high-throughput enzymatic assays. This led to the
development of Bayesian calibration models for
biotechnological applications, which describe the
relationship between tested quantities and measured values,
including uncertainty. The open-source Python package
calibr8 was developed to help practitioners with little
programming experience to easily implement complex,
non-linear calibration models. It serves as a toolbox for
high-throughput analytical calibration, as well as a
starting point for advanced process models that account for
bias in measurement systems. Using calibration models as
likelihoods, Bayesian statistical models were developed to
represent the technical and biological parameters of a
screening process. For example, batch effects between
screening experiments were modelled to avoid a bias in the
final ranking of strains and conditions. The process models
were also used to derive key performance indicators with
uncertainties for decision making. In two application
studies, Bayesian hierarchical process models were combined
with Bayesian optimisation to efficiently design iterative
screening experiments. For example, the number of
experiments required to screen a strain library of
catalytically active $\textit{inclusion bodies}$ (CatIBs)
could be reduced by 25\%. At the same time, the
probabilistic approach to calibration and process modelling
allows to identify major sources of uncertainty. This was
exploited to guide workflow development, e.g. leading to a
reduction of the relative standard deviation in the
automated CatIB purification and assay procedures from
11.4\% to only 1.9\% over 42 replicates. Finally, modern
machine learning tools were used to develop process models
and experimental designs for applications with limited
process understanding. The potential of horizontal knowledge
transfer for process models was explored, using data from
historical processes to improve predictions for new
processes. For example, Gaussian processes, popular machine
learning models for small data sets, were combined with
$\textit{meta learning}$ and benchmarked using in silico
cell culture data. In a final step, the established
knowledge transfer models identified optimal experimental
designs to characterise the behaviour of an unseen process,
a procedure called $\textit{calibration design}$. In
conclusion, this work intensifies bioprocess screening by
improving autonomous workflows on automated microbioreactor
systems. The close interaction between experiment and model
is crucial to achieve this goal, as is harnessing the power
of laboratory automation, computational tools and
interdisciplinary research. Overall, this thesis paves the
way for autonomous DBTL cycles, which are essential for a
sustainable bioeconomy in the future.},
cin = {162610 / 160000 / 165230 / 420410 / 057700},
ddc = {570},
cid = {$I:(DE-82)162610_20140620$ / $I:(DE-82)160000_20140620$ /
$I:(DE-82)165230_20220204$ / $I:(DE-82)420410_20140620$ /
$I:(DE-82)057700_20231115$},
typ = {PUB:(DE-HGF)11},
doi = {10.18154/RWTH-2024-12118},
url = {https://publications.rwth-aachen.de/record/999462},
}
guest ::