Bioprocess and Biosystems Engineering

, Volume 26, Issue 6, pp 353–359

Intelligent modelling of bioprocesses: a comparison of structured and unstructured approaches

  • Benjamin J. Hodgson
  • Christopher N. Taylor
  • Misti Ushio
  • J. R. Leigh
  • Tatiana Kalganova
  • Frank Baganz
Original paper

DOI: 10.1007/s00449-004-0382-0

Cite this article as:
Hodgson, B.J., Taylor, C.N., Ushio, M. et al. Bioprocess Biosyst Eng (2004) 26: 353. doi:10.1007/s00449-004-0382-0

Abstract

This contribution moves in the direction of answering some general questions about the most effective and useful ways of modelling bioprocesses. We investigate the characteristics of models that are good at extrapolating. We trained three fully predictive models with different representational structures (differential equations, differential equations with inheritance of rates and a network of reactions) on Saccharopolyspora erythraea shake flask fermentation data using genetic programming. The models were then tested on unseen data outside the range of the training data and the resulting performances were compared. It was found that constrained models with mathematical forms analogous to internal mass balancing and stoichiometric relations were superior to flexible unconstrained models, even though no a priori knowledge of this fermentation was used.

Keywords

Genetic programmingPredictiveModellingFermentationDevelopment

List of symbols

Nvars

Number of variables

Nbatches

Number of batches

Nnodes

Number of nodes making up an individual

tF

Time of last sample

Ci, Cj, Ck

Hidden variables used internally by models

CGluc, CNitrate, CBiomass, CRed

Predicted values of measured variables, glucose, nitrate, biomass, red pigment, respectively (g/l)

MGluc, MNitrate, MBiomass, MRed

Measured values of glucose, nitrate, biomass, red pigment, respectively (g/l)

rbatch

Pearson correlation coefficient between measured and predicted values at a given time point varying with respect to initial conditions

rtime

Pearson correlation coefficient between measured and predicted values with respect to time

R2

Root mean squared error

a, b, c, d

Floating point weights

ɛs

Scaled error of model on batch

ɛav

Error between the average profiles of training data and the actual value on that batch

Copyright information

© Springer-Verlag 2004

Authors and Affiliations

  • Benjamin J. Hodgson
    • 1
  • Christopher N. Taylor
    • 2
  • Misti Ushio
    • 1
  • J. R. Leigh
    • 3
  • Tatiana Kalganova
    • 4
  • Frank Baganz
    • 1
  1. 1.The Advanced Centre for Biochemical EngineeringUniversity College LondonLondonUK
  2. 2.Lilly Systems Biology Pte LtdSingapore
  3. 3.Control Engineering Centre, Department of Electrical Engineering and ElectronicsBrunel UniversityUxbridgeUK
  4. 4.Electrical and Computing Engineering DepartmentBrunel UniversityUxbridgeUK