17
Comparison of PLS regression and Artificial Neural Network for the processing of the Electronic Tongue data from fermentation growth media monitoring Alisa Rudnitskaya 1 , Andrey Legin 1 , Dmitri Kirsanov 1 , Boris Seleznev 1 , Kim Esbensen 2 , John Mortensen 3 , Lars Houmøller 2 , Yuri Vlasov 1 1 Laboratory of Chemical Sensors, Chemistry Department, St. Petersburg University, Russia; www.electronictongue.com 2 , Aalborg University Esbjerg, Denmark; 3 Department of Life Science and Chemistry, Roskilde University Centre, Denmark.

Industrial use of filamentous fungi batch fermentation

Embed Size (px)

DESCRIPTION

Comparison of PLS regression and Artificial Neural Network for the processing of the Electronic Tongue data from fermentation growth media monitoring. - PowerPoint PPT Presentation

Citation preview

Page 1: Industrial use of filamentous fungi batch fermentation

Comparison of PLS regression and Artificial Neural Network for the processing of the Electronic Tongue data from fermentation

growth media monitoring

Alisa Rudnitskaya1, Andrey Legin1, Dmitri Kirsanov1, Boris Seleznev1, Kim Esbensen2, John Mortensen3, Lars Houmøller2, Yuri Vlasov1

1 Laboratory of Chemical Sensors, Chemistry Department, St. Petersburg University, Russia; www.electronictongue.com

2, Aalborg University Esbjerg, Denmark;3 Department of Life Science and Chemistry, Roskilde University Centre, Denmark.

Page 2: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 2

Industrial use of filamentous fungi batch fermentation

Fungi: Aspergillus, Penicillium etc

Citric acid

Food stuffs

Enzymes

Pharmaceuticals

Food additives

Page 3: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 3

Purpose of the study

• Development of rapid analytical methodology to follow-up batch

fermentation processes and for quantitative analysis of broths

– Evaluation of Electronic Tongue (ET) for following-up of the batch fermentation

processes and quantitative analysis of broths on the example of Aspergillus

Niger batch culture medium

– Application and comparison of different chemometric techniques for multivariate

calibration of ET

Page 4: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 4

Experimental set-upSamples

Background: 0.5 gL-1 KCl, 1.5 gL-1 KH2PO4, 0.5 gL-1 MgSO4, 1 mlL-1 of Vishniac trace element solution, pH 6

Sample Time, h Citrate Pyruvate Oxalate Glucose Glycerol, Mannitol Erythritol NH4Cl

1 0 0 0 0 45.3 0.14 0.05 0 14.0

2 1.8 0 0 0 45.3 0.17 0.07 0 14.0

3 5.3 0 0 0 44.0 0.24 0.05 0 13.6

4 7.9 0.5 0 0 42.7 0.35 0.05 0.02 13.2

5 10.5 1.4 0 2.6 41.4 0.41 0.05 0.03 12.8

6 11.6 1.7 0 7.8 38.9 0.52 0.03 0.05 12.0

7 12.6 2.2 0 10.4 36.3 0.59 0.03 0.07 11.2

8 13.7 2.6 0 13.0 33.7 0.69 0.03 0.10 10.4

9 14.7 3.3 0 18.1 29.8 0.76 0.05 0.14 9.2

10 15.3 3.6 0 20.7 27.2 0.86 0.05 0.16 8.4

11 15.8 3.8 0 23.3 25.9 0.93 0.07 0.19 8.0

12 16.3 4.0 0 25.9 23.3 1.04 0.09 0.24 7.2

13 16.8 3.8 0 28.5 20.7 1.10 0.10 0.26 6.4

14 17.1 3.8 0 28.5 19.4 1.17 0.10 0.28 6.0

15 17.4 3.8 0 31.1 18.1 1.28 0.12 0.29 5.6

16 17.9 3.8 0 33.7 15.5 1.38 0.13 0.31 4.8

17 18.4 4.0 0 38.9 13.0 1.48 0.17 0.40 4.0

18 18.9 4.3 0 44.0 9.1 1.59 0.21 0.45 2.8

19 19.5 4.7 0.3 46.6 6.5 1.66 0.22 0.47 2.0

20 20 5.0 1.6 49.2 5.2 1.73 0.28 0.52 1.6

21 20.5 5.3 2.6 59.6 1.3 1.90 0.40 0.60 0.4

22 21.1 5.5 2.4 62.2 0 1.90 0.47 0.62 0

1. Solutions simulating growth media of real fermentation processes involving Aspergillus niger

2. Same solutions with 10mM of sodium azide added.

Page 5: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 5

• Measurements • ET comprising 10 potentiometric chemical sensors with polymeric

membranes• Direct and fast (few minutes) measurements• No sample preparation

Experimental set-up

•Data processing•Data splitting into calibration, monitoring and test sets (D-optimal design)•Multivariate calibration•PLS-regression•Feed-forward neural network

Software used: Unscrambler v. 7.8 by CAMO AS, Norway;NeuroSolutions by NeuroDimensions Inc, USA

Page 6: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 6

Determination of ammonium, oxalate, citrate content and time elapsed from the media inoculation in the model growth media using ET

Average relative error of

prediction, %

Ammonium Oxalate Citrate Time

12 6 11 7 without sodium azide

10 6 10 7 with sodium azide

Calibration of ET by PLS regression for each component separately

Results for the test set

Page 7: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 7

Non-linearity of the sensors’ responsesCalibration of ET w.r.t. ammonium concentration using PLS-regression

-2,8 -2,6 -2,4 -2,2 -2,0 -1,8

-2,8

-2,6

-2,4

-2,2

-2,0

-1,81234 5678

9101112

13141516

17

18

19

20

lgC

am (

me

asu

red

)

lgCam

(predicted)

-2,8 -2,6 -2,4 -2,2 -2,0 -1,8-0,12

-0,10

-0,08

-0,06

-0,04

-0,02

0,00

0,02

0,04

0,06

0,08

12

3

4

567

8

9

10

11

12

131415

16

17

18

19

20

lgC

am r

esid

uals

lgCam

(predicted)

-30 -20 -10 0 10 20 30 40-0,8

-0,6

-0,4

-0,2

0,0

0,2

0,4

123456789

101112

13141516

17

18

19

20

U s

core

s (8

5%)

T scores (75%)

Page 8: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 8

Response of the NH4-sensitive electrode to NH4+ on the

growth medium

Detection limits to NH4+:

Discrete electrode - 3.07 pNH4

Sensor array - 3.7 pNH4-5 -4 -3 -2 -1

40

60

80

100

120

140 C(K+) = 0.018M

E,

mV

logC(NH4

+)

jiji akanF

RTEE ln0

Nikolski equation:

Page 9: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 9

Non-linear calibration methods

•Non-linear regression•Artificial neural networks

•Advantages

-Flexibility

-Noise tolerance

•Drawbacks

-Prone to overfitting

Page 10: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 10

Feed-forward neural network

Learning

Local error function: ej = -E/ Ij

for output layer: eo = f’(Io) •(y-ŷ)

for hidden layers: esj = f’(Is

j) • (es+1s• ws+1

kj)

Weight update: wsij = - • ( E/ ws

ij) = • esj • xs-1

i

x1

x3

x3

x2

I, f(I)

I, f(I)

I, f(I) I, f(I)

wsij

wsij

Input layerHidden layer

Output layer

ŷ

Weight - wsij

Input function: Isj = xs-1

i*wsij

Transfer function: f(I)

Forward pass

E =ly-ŷl

Error back-propagation

-8 -6 -4 -2 0 2 4 6 8

-2,0

-1,5

-1,0

-0,5

0,0

0,5

1,0

1,5

2,0

Hyperbolic

tangent: xx

xx

ee

eexf

)(

Page 11: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 11

Neural network validation

Evolution of training and monitoring errors during ANN training. Calibration of ET w.r.t. oxalate concentration

0 200 400 600 800 10000,00

0,02

0,04

0,06

0,08

0,10

0,12

0,14

0,16

0,18

0,20

0,22

Modeling Overfitting

Stopping point

Err

or

Iterations

Training Monitoring

Page 12: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 12

Data splitting into calibration, monitoring and test sets using D-optimal design

Basic idea of D-optimal design: finding a design matrix that maximizes the determinant D of the initial data matrix, i.e. finding a set of samples that are maximally independent of each other.

Ideal distribution: if calibration set contains n samples, monitoring and test sets should contain between n/2 and n samples each.

In this case: calibration set – 22 samples, monitoring set – 11 samples, test set – 21 samples.

-40 -20 0 20 40 60

-10

-5

0

5

10

15

PC

2

PC1

Calibration Monitoring Test

Page 13: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 13

Optimization of the neural network architecture

Aim: minimization of prediction error AND number of network parameters (weights), i.e. hidden and input neurons.

1 2 3 40,000

0,001

0,002

0,003

0,004

0,005

0,006

MS

E o

f C(o

xala

te)

pre

dic

tion

in te

st s

et

Number of hidden neurons

Number of inputs 9 7 5

0

5

10

15

20

25

30

35

40

Nu

mb

er

of w

eig

hts

Number of inputs 9 7 5

Optimized ANN for calibration w.r.t. content of :

Ammonium: 5 2 1

Oxalate: 5 3 1

Citrate: 7 2 1

Page 14: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 14

Determination of ammonium, oxalate and citrate content and time elapsed from the media inoculation in the model growth media using ET

Average relative error of prediction,

%

Calibration method Ammonium Oxalate Citrate Time Samples

ANN

6 6 8 2 without sodium azide

7 7 7 2 with sodium azide

11 6 12 2 all data

PLS

12 6 11 7 without sodium azide

10 6 10 7 with sodium azide

Page 15: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 15

PCA score plot of ET measurements in growth media with and without sodium azide added

-40 -20 0 20 40 60 80-20

-15

-10

-5

0

5

10

15

20

25

1234567

910

11

1213

14

15

161718

1920

21 22

123

456

78

910

111213

14151617

18

19

20

2122

12

345

678

9101112

1314151617

18

2021

22

123

45

678

91011

1213

141516

1718

19202122

1

23

45

6

78

9101112

13

14

15

16

17

1819 2022

PC

2 (1

1%)

PC1 (78%)

without azide with azide

Page 16: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 16

Non-linearity of the sensors’ responses

-1,5 -1,0 -0,5 0,0 0,5 1,0 1,5 2,0 2,5-1,2

-1,0

-0,8

-0,6

-0,4

-0,2

0,0

0,2

0,4

0,6

0,8

1,0A

ctiv

ity

Input

Hidden neuron 1 Hidden neuron 2

Calibration of ET w.r.t. to ammonium concentration using ANN

Page 17: Industrial use of filamentous fungi batch fermentation

WCS-4, February 15—18 , 2005, Moscow (Chernogolovka), Russia

A. Rudnitskaya et alSt. Petersburg University 17

Conclusions

• An ET system comprising a sensor array based on ten PVC-plasticized cross-sensitive potentiometric chemical sensors was successfully applied to simultaneous determination of ammonium, oxalate and citrate content in simulated fermentation media closely resembling real-world samples typical of a process involving Aspergillus niger.

• Feed-forward neural network was found to be superior to PLS regression for the ET data fitting due to better consideration of non-linearity of the sensor potentials/concentration relationship particularly at low concentration levels. The average prediction errors for key metabolites’ concentrations in the given ranges was about 6-8% when using a feed-forward artificial neural network for ET calibration.

• Content of three key components of the growth media can be measured by ET in the presence of 10 mM sodium azide, which is commonly used to suppress microbial activity after sampling.

• ET was demonstrated to be promising for monitoring fermentation processes.