Methodology for fast development of digital solutions in integrated
continuous downstream processing
Niklas Andersson, Joaquín Gomis Fons, Madelène Isaksson, Simon Tallvod,
Daniel Espinoza, Linnea Sjökvist, Gusten Zandler Andersson and Bernt
Nilsson#
Dept. of Chemical Engineering, Lund University, Lund, Sweden
# Corresponding author, email address:
bernt.nilsson@chemeng.lth.se
Methodology for fast development of digital solutions in integrated
continuous downstream processing
Abstract
The methodology for production of biologics is going through a paradigm
shift from batch-wise operation to continuous production. Lot of efforts
are focused on integration, intensification and continuous operation for
decreased foot-print, material, equipment and increased productivity and
product quality. These integrated continuous processes with on-line
analytics becomes complex processes, which requires automation,
monitoring and control of the operation, even unmanned or remote, which
means bioprocesses with high level of automation or even autonomous
capabilities. The development of these digital solutions becomes an
important part of the process development and needs to be assessed early
in the development chain. This work discusses a platform that allow fast
development, advanced studies and validation of digital solutions for
integrated continuous downstream processes. It uses an open, flexible
and extendable real-time supervisory controller, called Orbit, developed
in Python. Orbit makes it possible to communicate with a set of
different physical setups and on the same time perform real-time
execution. Integrated continuous processing often imply parallel
operation of several setups and network of Orbit controllers makes it
possible to synchronize complex process system. Data handling, storage
and analysis are important properties for handling heterogeneous and
asynchronous data generated in complex downstream systems. Digital twin
applications, such as advanced model-based and plant-wide monitoring and
control, are exemplified using computational extensions in Orbit,
exploiting data and models. Examples of novel digital solutions in
integrated downstream processes are automatic operation parameter
optimization, Kalman filter monitoring and model-based batch-to-batch
control.
Introduction
The methodology for production of biologics is going through a paradigm
shift from batch-wise operation to continuous production. A lot of
efforts are focused on integration, intensification and continuous
operation for decreased foot-print, material, equipment and increased
productivity and product quality (Ng et al 2012, Godawat et al 2012,
Baur et al 2016a, 2016b, 2018, Sellberg et al 2017, Gomis-Fons et al
2020B and 2021, Moreno-González et al 2021). On-line analytics are
needed in these continuous processes to reach their potentials with the
final goal of real-time release testing, RTRT (Farid 2007, Shukla and
Thömmes 2010, Godawat et al 2015, Walthers et al 2015, Konstantinov and
Cooney 2015, Zydney 2015, Shukla et al 2017, Kesik-Brodacka 2018, Vogg
et al 2018). These integrated continuous processes with on-line
analytics become complex processes, which requires automation,
monitoring and control of the operation, even unmanned or remote, which
means bioprocesses with high level of automation or even autonomous
capabilities. The development of these digital solutions becomes an
important part of the process development and needs to be accounted for
early in the development chain. Lab-scale automated continuous
bioprocesses have been reported (Steinebach et al 2017, Feidl et al
2020, Gomis-Fons et al 2020B, Rathore et al 2021, Schwartz et al 2022,
Tiwari et al 2022). The software used in these experimental platforms
were very different from locally developed PLC to commercial process
automation SCADA systems. This work discusses a platform that allows
development, advanced studies and validation of digital solutions for
integrated continuous downstream processes. It uses an open flexible and
extendable real-time supervisory controller developed in Python. Python
is a well-known programming language with a large user community that
supports the language with a large set of external packages for
different kind of applications, for instance in data handling, machine
learning, simulation and optimization.
The paper presents the properties of the supervisory controller, called
Orbit, and networks of controllers implemented in Python, and how it
impacts the development of downstream processes (Andersson et al 2017,
2022). Orbit makes it possible to communicate with a set of different
physical setups and on the same time perform real-time execution.
Integrated continuous processing often imply parallel operation of
several setups and network of Orbit controllers makes it possible to
synchronize complex process system. Data handling and data analysis is
an important property for handling heterogeneous and asynchronous data
generated in complex downstream systems. Advanced model-based and
plant-wide monitoring and control are used as examples of novel digital
solutions in integrated downstream processes.
Materials and methods
Integrated continuous downstream system
A process for continuous production of an antibody was developed and
studied in lab-scale, Scheffel et al 2022, and scaled-up and operated in
pilot-scale, Schwartz et al 2022. The upstream process consisted of a
high cell-density perfusion bioreactor with a membrane system for
production of clear cell-free supernatant. Between the upstream process
and the downstream process there was a harvest tank making it possible
to disconnect the downstream process for up to 20 hours. Under normal
conditions, the feed flow rate to the downstream process was set for the
control of the liquid amount in the harvest tank, meaning that the
production rate in the complete process was controlled by the perfusion
rate selected in the upstream process.
The downstream process consisted by four units; continuous periodic
counter-current chromatography system (PCC) for antibody capture, viral
inactivation unit (VI) based on two parallel batch reactors, one cation
exchange chromatography step (CEX) in bind-and-elute mode mainly for
aggregate removal and finally one anion-exchange chromatography step
(AEX) in flow-through mode for final polishing. The PCC and the loading
of VI was configured on one ÄKTA setup and the VI control together with
the IEX polishing train on a second ÄKTA setup. Each setup was
controlled by an installation of the research software Orbit, through
the Unicorn control system for the ÄKTA system. The lab-scale process
was operated continuously for 9 days and the pilot-scale process for 19
days, for details see Scheffel et al 2022 and Schwartz et al 2022.
Support system automation
An ÄKTA explorer 100 was modified into a buffer preparation unit, BPU,
by introducing additional valves and adapting the standard flow path.
Two pumps were needed to perform the buffer preparation and delivery.
The configuration enabled connection of 9 different stock solutions. In
addition, 14 buffer mixing flasks were installed using the column
valves, the outlet valve and the sample valve. In each of the buffer
mixing flasks, a specific buffer could be prepared. Since the same
buffer was prepared in the same mixing flask the whole time, no cleaning
of the flasks was needed. At each of the client systems, the tube from
the BPU was connected to a client inlet valve, which was used to
distribute the buffers to the corresponding buffer flasks in the client.
This setup allowed for the BPU to deliver buffer directly to the right
buffer flask in the client system. The client inlet valve worked
independently of the other tasks that the client system performed,
meaning that the downstream process did not need to stop every time a
buffer in the client was re-filled. For more detail see Isaksson et al
2023.
Quality analysis systems consists of a sample preparation system, a
modified ÄKTA explorer 100, and a HPLC analysis system, Agilent 1260.
The prep and analysis systems were connected via a tube between the prep
system’s outlet valve and the analysis system’s injection valve. The
client systems were connected to the prep system at several points
allowing for sampling the flow path of for sampling a container with a
pool, before further processing. The retrieved sample is prepared for
analysis, which can mean dilution, buffer exchange, chemical
modification etc., before sending to analysis. The analysis system
consisted of an injection valve with an injection loop, a column valve
with several analysis columns, a diode array detector and a refractive
index detector. When the prep system was sending a sample for analysis,
the flow path on the analysis system went through the injection loop and
directly to waste. Based the kind of sample loaded the corresponding
analysis protocol were performed. For more detail see Tallvod et al
2023.
Automation and operation of continuous downstream system
Orbit was originally created in 2015 and the current version is 3. Orbit
was created to make complex, custom control of processes and to
integrate several instruments together. The fundamental part of the
orbit design is the orbit kernel that consist of one part where the
system configuration and communication is defined and a second part
where the real-time control is defined. A configuration of a system is
designed with unit objects which contains methods to control the object.
There are unit libraries where units for different type of system is
defined. Every unit that communicates with an instrument has a defined
protocol for the available methods. Tubes can also be defined to keep
track of flow paths, dead volumes. The real-time kernel handles the
time, current phases, sampling and communication to the instrument and
between Orbit controllers. The user writes a script to define the phases
of the process that could be run sequential or in parallel. The phases
are defined by the methods in a method library to control the process.
The methods can be commands executed at defined times or more complex
event-based tasks like control. There is an existing library from
different application, such as sequential polishing steps, PCC, MCSGP,
conditioning, etc. A new phase starts when all the instructions in the
previous phase are executed. For automation of integrated continuous
downstream process discussed above two setups were automated using two
Orbit supervisory controllers; one for the PCC operation on an ÄKTA pure
and one for the sequential control of the polishing ion exchange train
also on an ÄKTA pure.
Functionality of the orbit kernel can easily be extended. There are
built-in extensions like monitoring, simulation and plotting and a user
can easily create their own extensions to add for example control,
optimization or predictions based on Kalman filters, see below.
In a process with several orbits there need to be communication between
Orbit controllers. Every Orbit instance has a server enabled where it
can be connected to by other orbit instances or for manual control. For
larger processes this creates a network of Orbit controllers. The
protocol for the communication is to send dictionaries with specified
tasks and arguments to complete the task. One type of Orbit instances is
called orbitX, where a service is offered for other Orbit controllers
that can connect as clients. Examples of this is the quality analysis
system and buffer management system discussed above. When for example
ordering a buffer a task is sent with information about the recipe and
the volume of the ordered buffer. In orbitX instances that handles
requests from several clients there need to be a queue controller
defined. The simplest version of a queue is to have the
first-in-first-out principle. This can also be extended with a priority.
For time critical requests where a buffer or analysis needs to be
performed within a deadline, this can also be added and considered that
the priority is high enough it is allowed to jump the queue.
When working with real-time processes a lot of data is created. A
database has been implemented to handles Orbit data. The data to be
stored is both the system configuration, but also run logs, signal
information and data. This is automatically uploaded directly to the
database during a run. In a process there typically exist sampling
points from where a sample is taken for further analysis. In our QAS
system this analysis is ordered and information about where the sample
is taken is stored together with the result of the analysis.
Modelling and model-based monitoring and control
The Orbit supervisory controller requires a digital representation of
the physical setup, with it is communicating. The representation is
composed of a list of objects and a list of tubing, connecting the
object ports. All objects and tubing have a mathematical representation
in Orbit, allowing a simulation extension to perform a simulation of the
complete setup models, i.e. automatic generation of a digital twin of
the chromatographic system. A set of experiment can be executed to
generate data for parameter estimation and model calibration, see
Tallvod et al 2022A, based on methodologies presented in Borg et al 2014
and Saleh et al 2020. One example of real-time usage of a model is in
on-line monitoring. A dynamic extended Kalman filter is implemented as a
computational extension in Orbit for monitoring of the composition
inside the column. A size exclusion application for the separation of
two components with large size difference is presented for illustration.
Both components can be measured using UV. A sample is introduced at 70s
and passes through the column after 300s, with peak max at about 140s
and 210s. The size exclusion model is calibrated using residence time
experiments and the model were discretized using 55 mesh points. This
means that the Kalman filter estimates 110 states, inside the column.
A batch-to-batch controller was implemented for the retention peak
control of two components using the elution gradient, see for more
details in Espinoza et al 2022B. Two proteins are separated with an
ion-exchange chromatography column by a salt gradient, with a start
value and an end value. These two values are the output from the
controller. After the batch-wise separation, the retention time for the
peak maximum is detected, which is the input to the controller. Based on
the difference between the measured and desired retention time the
controller computes the gradient profile. The controller parameters are
found by using a process gain matrix, derived from the system model.
Results and discussion
Operation of integrated downstream processes
In a previous work a continuous antibody production process was operated
continuously for 9 days in lab-scale and in pilot-scale for 19 days, for
details see Scheffel et al 2022 and Schwartz et al 2022. The perfusion
rate targeted 1.5 reactor volume per day which means that the upstream
process together with the harvest tank had an approximated residence
time of about 24 hours. The downstream process had a residence time of
about 5 hours, with a column switching in the PCC every 40 min, a
product pool passed the polishing ion exchanger train every second
hours. The experimental data from the two different runs in lab-scale
and pilot-scale, clearly indicated that the process dynamics in the
process was slow, dominated by the time-scale in the upstream process.
The time scale of changes in upstream outflow was about 24 hours. The
downstream process consisting of four steps was configured on two ÄKTA
setups for parallel operation of PCC, viral inactivation and polishing,
see Figure 1.