Template for Large Format Poster in PowerPoint Linda Cunningham The Audio Visual Centre, University

1
Application of Bayesian Neural Networks for Early
Indication of Problematic Bioprocess Scale-up
A Collaborative Research Project By Nicholas
Veuger, Gary Montague, Elaine Martin
Ronan OKennedy School of Chemical
Engineering Advanced Materials, Newcastle
University GlaxoSmithKline, Beckenham

1. Introduction
From a theoretical perspective the control of
batch bioprocesses is well documented but their
inherent complexity makes exact characterisation
of operation a non-trivial task. Deviations from
theoretical behaviour may stem from a physical,
chemical or biochemical uncertainty, or usually,
from a combination of all three. It is therefore
the case that the efficient and safe running of
such processes in an industrial environment
requires a certain amount of operator know-how as
well as a sound understanding of process
specifics. Whilst experience in operation exists
for production plant, this is not the case in
process development. Problems can be particularly
pertinent when attempting to change the operating
scale of processes. Moreover, bioprocess scale-up
problems can be difficult, time consuming and
costly to overcome given the impact of a varying
environment on biological behaviour. With this in
mind, in order to produce a robust model to be
used for bioprocess scale, it is wise to fully
utilise heuristic knowledge along with physical
understanding.   Bayesian inference attempts to
do precisely this by starting from a purely
knowledge-based perspective and adding supporting
mathematical evidence in the form of process
data, a combined process model is sought. In
fact, there are many ways in which Bayesian
techniques can be built into a given model
framework, however, in this instance the Bayesian
Neural Network (BNN) will be considered. The
presentation will describe the concept of BNNs
and will explain why BNNs can help overcome the
problems faced during the scale-up of mammalian
cell culture batch processes, focusing
particularly on the production of monoclonal
antibodies.
Changing Scales
Laboratory
Pilot Plant
Production
3. Considering The Effects of Scale
2. Cell Cultivation Batch Reactor

• The effects of scale upon the growth and general
  behaviour of a culture may manifest themselves
  via a number of a physical causes
• Mixing It is common to find that there are
  differences in batch evolution with respect to
  mixing. This may be due to changes in localised
  pressure, flow, bulk viscosity, turbulence or
  other fluid dynamical properties. In order to
  resolve the mechanics behind these, complex
  analysis via computational fluid dynamics (CFD)
  is usually required
• Mass Transfer The transfer of respiratory gases
  and feed to and from a culture may well be
  affected by the scale of the set-up. Localised
  concentration differences, gaseous bubble surface
  area and thermodynamic heterogeneity all have an
  effect on the diffusional dynamics of the
  process. Consideration of a diffusion model (e.g.
  Ficks Law) together with thermodynamics is
  necessary here.
• Biochemical Possibly the most complex effect of
  all to describe is the likelihood that
  biochemical pathways may shift when the process
  scale conditions change. Spectroscopic techniques
  such as NIR and chemometrics may be useful in
  gaining knowledge in this area.
• Measurement The likelihood of a change in
  measurement apparatus when moving across scales
  is very high. Thus, accurate calibration and
  careful placement of measurement instuments is
  imperative when attempting to analyse multi-scale
  data

• The Cell Culture Reactor Consists
• of a number of simple components
• and control loops
• Stainless or Plastic Vessel in various volumes
  from 5ml to 1000L
• Variable Speed Stirrer
• Glucose feed injector
• Heater Coil with Temperature Control
  loop
• Acid/Alkali Injectors with pH Control Loop
• NIR Spectroscopy Probe
• Oxygen disperser ring

4. Bayesian Neural Networks
Data Used for testing and Validation of Network
Pilot Scale
Production Scale
Lab Scale
Bayesian statistics provides a different
perspective of what it means to learn from data,
especially those in which probability is used to
represent a level of uncertainty about the
relationship being investigated. However, before
the analysis of such data can begin Bayesian
models require a series of inputs of operator
opinions, known as prior distributions (as
initial conditions). Hence, Bayesian models are
initially wholly heuristic in nature and are
based upon a human estimation of parameter
characterisation. It is this subtle but important
difference that moves Bayesian techniques away
from other more commonly used statistical
methods. Then, with the addition of data,
revised predictions (posterior distributions)
can be calculated which express a level of
probability that each possibility is an outcome
which can therefore be used to classify the
likelihood of an event. Neural Networks provide a
convenient basis upon which to build such a
Bayesian predictor, as they are essentially
simple and have good representational
capabilities. It is also found that they are
especially suited to bioprocess modelling as they
have a number of favourable properties, such as
their ability to cope with highly non-linear
behaviour and correlated variables.  
Historical Data
Fully Developed Bioprocess 1
Lab Scale
Pilot Scale
Production Scale
Fully Developed Bioprocess 2
Bayesian Neural Network
Data Used to Estimate Likely Success of Scale-up
New Data
Successful Production Scale ?
Lab Scale
Successful Pilot Scale ?
Bioprocess in Development
It has been expressed within industry that it
would be beneficial to obtain a robust model
which has as its output - a level of belief that
a process will scale to a given size without
complications. In this scenario the BNN can be
thought of as a classifier where the output
represents a parameter of likely scale-up
success. In terms of the training and validation
of such a network for the modelling of a
monoclonal antibody reactor, it is proposed to
utilise GSKs specific industrial knowledge as a
prior basis and to add historical data at varying
scales to obtain posterior beliefs. Once this has
been achieved, it is anticipated that this will
be extended other similar cell-lines in order to
build a number of validation models, from which
comparisons and conclusions can hopefully be
drawn.

If you are interested in this research, please
feel free to contact us n.j.veuger_at_ncl.ac.uk
gary.montague_at_ncl.ac.uk e.b.martin_at_ncl.ac.uk