14
Cllent. Anal. (Warsaw),44, 865 (1999) Flow-Injection.Determination of Phenols with Tyrosinase Amperometric Biosen.sor and Data Processing by. Neural Network by Marek TroJanowicz 1*, .AnnaJagielska 2 ,. Piotr Rotkiewicz 2 and AndrzeJKierzek 3 Ilnstiiute plNuclear Chemistry and Technology, Analytical Chemistry Division, 16 Dorodna, Str., 03-195 Warsaw, Poland 2DepartnlentplChemistry,·University of Warsaw, ·l··Pasteura Str., 02-093 Warsaw,· Poland 3lnstitute 0.1 Bioche111istr,V and Biophysics pl Polish Acade/11Y o.lSciences, 5A PawiJlskiego, Sfr., 02-106 Poland Key words: flow-injection analysis, phenol biosensor, neural network, tyrosinase, environmental·.analysis Amulti-melnbrane amperometric biosensor prepared with immobilized tyrosinase on a platinum disk electrode in a large-volume wall-jet flow-through cell was applied for the determination of phenolic compounds via flow-injection measurelnents. For data pro- cessing of measurements carried out simultaneously with several biosensors of different selectivity using different membranes in three-component mixtures of phenol, catechol and 111-cresol,a three layer artificial neural network \\rith feedforward connections, sign10idal transfer function and back propagation learning algorithm waselnployed. The best functional parameters of the network were found to be 5 inputs, 3 neurons in the hid- den layer and 10000 learning cycles. For 36 samples analyzed the best correlation coeffi- cient values were obtained for catechol (0.96) and phenol (0.88). Results for in-cresol, which produced the slnallest an1p.eron1etric signal with all biosensors tested were only semi-quantitative (correlation coefficient 0.67). *. Corresponding author.

Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

  • Upload
    others

  • View
    9

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

Cllent. Anal. (Warsaw),44, 865 (1999)

Flow-Injection.Determination of Phenols withTyrosinase Amperometric Biosen.sor and Data

Processing by. Neural Network

by Marek TroJanowicz 1*, .AnnaJagielska2,. Piotr Rotkiewicz2

and AndrzeJKierzek3

Ilnstiiute plNuclear Chemistry and Technology, Analytical Chemistry Division,16 Dorodna, Str., 03-195 Warsaw, Poland

2DepartnlentplChemistry,·University ofWarsaw, ·l··Pasteura Str., 02-093 Warsaw,· Poland3lnstitute 0.1 Bioche111istr,V and Biophysics pl Polish Acade/11Y o.lSciences,

5A PawiJlskiego, Sfr., 02-106 ~Varsaw, Poland

Key words: flow-injection analysis, phenol biosensor, neural network, tyrosinase,

environmental·.analysis

Amulti-melnbrane amperometric biosensor prepared with immobilized tyrosinase on a

platinum disk electrode in a large-volume wall-jet flow-through cell was applied for the

determination of phenolic compounds via flow-injection measurelnents. For data pro­

cessing ofmeasurements carried out simultaneously with several biosensors ofdifferentselectivity using different membranes in three-component mixtures ofphenol , catechol

and 111-cresol,a three layer artificial neural network \\rith feedforward connections,

sign10idal transfer function and back propagation learning algorithm waselnployed. The

best functional parameters of the network were found to be 5 inputs, 3 neurons in the hid­

den layer and 10000 learning cycles. For 36 samples analyzed the best correlation coeffi­

cient values were obtained for catechol (0.96) and phenol (0.88). Results for in-cresol,which produced the slnallest an1p.eron1etric signal with all biosensors tested were onlysemi-quantitative (correlation coefficient 0.67).

* .Corresponding author.

Page 2: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

866 M. Trqjan0 wic::., A. Jagielska, P Rotkielvicz and A. Kierzek

Do oznaczania wybranych zwi qzk6w fenoIowych w ukladzie przeplywowo-wstrzy­

kO\vYln zastosowano nlembrano\ve bioczujniki enzytnatyczne z tyrozynazq, platynowq

elektrodq dyskowq i naczynkienl przeplywowytn typu wall-jet. P0111iary prowadzono

przy uzyciu kilku bioczujnik6w, w kt6rych r6znq seIektywnosc na fenol, pirokatechin~ i

111-kresol osiqgano przez uzycie r6znych tnctnbran. Do przetwarzania danych ponliaro­

wych stosowano tr6jwarstwowq'sztucznq siec neuronOWq z nletodq wstecznej propaga­

cj i. NajIepsze wyniki osiqgano przy 5 wejsciach, 3 nauronach w warstwie ukrytej i 10000cyklach uczqcych. Dia 36 anaIizowanych tnieszanin najIepsze wyniki osiagni~to dia

pirokatechiny (wsp61czynnik korelacji ze znanq zawartosciq 0.96) i fenolu (0.88).Wyniki dla m-krezoIu, dla kt6rego uzyskiwano najmniejsze sygnaly, byly jedynie

p6lilosciowe (0.67).

The idea of using an artificial neural network (ANN) as a signal processing tool,which lnilnics the processing and translnissionof signals in hlunan neural systeln isknown and realized in various fields of science and technology since forties [1,2]. Itsfast technological developlnent and an increase ofthe nlllnber ofvarious applicationsis observed since beginning ofeighties. The lnost comlnon application ofa neural net­work is its use as a cOlnputer software for the processing ofnUlneric data. Silnilarly tothe hlunanbrain the artificial neural network is cOlnposed ofneurons, which are indi­vidual units of the processing of inforlnation. The biophysical and chelnical pro­cesses that occur in natural neuron during processing of inforlnation, in the artificialneurons are reflected by weights i.e. nlllnerical coefficients which are used to lnulti­ply each signal prior to its translnission to a next neuron and by certain functions thatlnodify the signal. The weights deterlnine the extent of the stilnulation of particularneurons in a higher layer of the network and kind of inforlnation which will be ob­tained in the output layer. The lnost essential feature ofneural networks is their learn­ing ability. Being able to lnilnic the hlllnan cognitive processes the neural networklnay be successfully applied to noisy, incolnplete and apparently inconsistent data.

Alnong a large variety of applications of artificial neural networks in variousfields ofscience and technology an increasing nUlnber ofapplications appears in ana­lytical chelnistry. Fundalnentals and discussion ofearlier chelnical alld analytical ap­plications of neural networks were already reviewed by several authors [3-5]. Theanalytical applications reported so far deal with the spectroscopic calibration andquantitation in near-infrared and UV-Vis spectroscopy [6], X-ray fluorescence spec­troscopy [7] and Inass spectrolnetry [8], evaluation of analyte concentration frolnflow-injection analysis signals of a pH ISFET [9], the optilnization of lnobile phaseparalneters [10] and retention lnodeling in liquid chrolnatography [11]. Several re­ported applications ofneural networks were focused on processing ofsignals froln ar­rays of chelnical sensors, of which each produced a different response to analytes tobe deterlnined. The satisfactory resul ts in such a data processing was reported in si­Inultaneous Ineasurelnents with several ion-selective electrodes [12] and calibrationof an array ofvoltalllinetric Inicro-electrodes [13]. Several applications can be foundfor the arrays of gas sensors [14-17]. The use of a neural network allows as well thequalitative recognition of a given natural or cOllllnerciallnaterial [14,15], as the de-

Page 3: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

Detern1inatiol1 o.lphenols with tyrosinase aTnperoTnetric biosensor 867

terlnination ofpartiGular analytes in gaseous Inixtures [15-21] . Recently, an applica­tionofthisdataprocessing has been reported for simultaneousdeterlnination of theconcentration of sulfur dioxide and relative hlunidity with a single coated piezoelec-\tric crystal [22] and signal diagnosis [23] and optilnization of response [24] in se­quential flow analysis systelns.

The ailn of this work was to exalninewhether a siInple neural network softwaredeveloped· in our group can be used for the processing. of experilnental data inflow-injectionbiosensing systeln •for thedeterlnination of several phenolic COln­pounds in tnixtures. So far, in the field ofbiosensors,ANN were elnployed forevalua­tion·of flow injection signals fro In salnples with. changing pH in Ineasurelnents ofglucose and urea [25].

EXPERIMENTAL

Apparatus

The 111easurenlents were carried out in a two line flow-injectionsystenl (Fig. 1), vvhere the 11lain C0111­

ponents .were a peristaltic. pU111p Isn1atec •• MP 13GJ4 (Zurich, Switzerland), a rotary injection valve

Rheodyne l11ode15020 (Cotati, CA, USA), avoltanl111ograph BAS 1110del CV-37 (W.Lafayette, IN, USA)

and a strip chart recorder Labographn1odelE 586 fro111 Metrohl11 (Herisau, Switzerland). Measurel11e11ts

were carried our with the three electrode syste111 and a large volul11ewall-jet detector with exchangeable

cap for Pt disk electrodes of3.0 nl111 dian1eter. A graphiterod was used as the auxiliary electrode and sil­

ver/silver chloride electrode with saturated potassiul11 chloride solution was used as the reference.

c

R

p s

w

Figure 1. Schenlatic diagranl offlow-injection systenl used for amperol11etric deternlinatiol1 ofphenols.P - peristaltic ptnllp; S - sanlple injection valve; D - large-volul11e wall-jet detector withtyrosinase biosel1sor; W- waste;C carrier strea111 ofdistilled vvater (1.01111 111in- I

); R - streanlof50 11111101 r l phosphate buffer pH 7.0 containing 50 1111nol r l potassiUlll hexacyanoferrate(II)(1.01111 nli11- I

)

Reagents

Forthe preparation ofbiosen SOl'S the tyrosinase (EC 1.14.18.1) ofactivity4200 unitshng frol11 Siglna

was used. Other reagents used were fron1 POCh (Gliwice, Poland). All solutions were prepared using tri­

ply distilled water deionized \vith Waters Milli-Q systen1.

Page 4: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

868 M. TrojanoYvicz, A. Jagielska, P Rotkie}vicz and A. Kierzek

For the immobilization of the enzyme a polyester Nucleopore melnbranes of a pore size 0.4 Jlm

(Pleasanton, PA, USA) ,vas used. A solution containing 2.5 mg lyophilized enzyme powder in 200 III O. I

mol I-I phosphate buffer pH 7.1 and 10 III ofa 250/0 glutaraldehyde solution was used to immerse the poly­

ester membranes of4 mIn diameter for 24 hr at -5°C. The membranes were washed with phosphate buffer

and applied to the platinum disk electrode.

Flow-injection measurements

100 JlI samples"of solutions containing phenols were injected into the stream ofdeionized water and

lnerged with the stream of 50 n111101 I-I phosphate buffer pI-I 7.1 containing 50 mmol I-I hexa­

cyanoferrate(II). Flow-rates in both lines were 1.0 ml min-I. The working electrode (biosensor) was

maintained at 0 V vs Ag/AgCI. An exan1ple of the recorded FIA signals are shown in Figure 2.

3

4

,

9

10

11

J5min

Figure 2. The example ofrecorded flow-injection signals for double injections ofphenol solutions in thesysteln with Nucleopore polyester membrane at 0.0 V vs Ag/AgCI. Sample injectio11 volu111e100 Jll. Concentrations of injected solutions: 1 - 500 11111011-1; 2 -·400 Ilmol r J; 3 - 300 Ilmolr J; 4 - 200 Jlmol r J; 5 - 100 Jln1011-J; 6 - 50 Ilmoll-J; 7 - 25 Ilmol r J

; 8 - 20 Ilmoll-J; 9 - 15Ilmoll-J; 10- 10 Ilmol r J and 1I - 5 Ilmol r J

RESULTS AND DISCUSSION

Properties of tyrosinase amperometric biosensor

Due to the biocatalytic properties of the used enzylne, the alnperolnetric biosen­sor with tyrosinase responds with different sensitivity to various phenolic cOlnpounds(Fig. 3). The sensors can not be practically used for the selective detertnination ofa sin­gle analyte, but they can be utilized either for the detertnination of the sum of the de­tected phenolic compounds in a given Inatrix, which may be useful for environlnentalanalysis, or for the detection in HPLC with post-colulnn enzylnatic reaction [26-30].

Page 5: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

Detennination o.lphenols vvith tyrosinase al11peronletric biosensor 869

160

120

<1>enc:oc.en(1)»-Q)

>

200 ~ ..

: :

-r1i;

-ViI I

M

80 .~.~

40 ~ ...

II! Io

(5C

....... (1).cc..

Catechol

(;CCD__ .c ....···......·.. 0c. c

CD e Q)c 0 .cg:c c-83.g :>...

." c.:T. .. a- ... "0 .. St- en ..,e g ~ (1)

" II!. 0 :E>.. -c:r • •:J: N 0 0

Figure 3. The relative signall11agnitudefor various phenolic cOlnpounds in flow-injectionamperolnetrywith tyrosinase biosensor and outer polyester membrane. Injected 100 III 100 Jlmol I-I solu­tions

Anotherway to ntilizethetyrqsirtase biosensor in. spite of its limited selectivitymaybe to compose an array ofsnchbiosensors withadditional differentiationin se­lectivity and to elnploytheln inlnulticolnponentlneasurelnents with appropriate pro­cessing oftheexperilnentaldata..The additional differentiation of selectivity ofphenol biosensors can be achieved atleastin two different ways. One way is to use in

suchan array the biosensofsJhathave incorporated different enzymes, which cata­lyze reactions with phel1oliccQmpoLJndsasSllbstrates.orth9ircqimlTIobilized mix­tures e.g. tyrosinase and laccase [31]. In this study another silnpler approach waselnployed, nalnely the use ofbiosensors covered with different lnelnbranes, whichadditionally differentiatein the transport ofanalytes to the itnlnobilizedenzylne layerand the working electrode surface. Because catechol, phenol and l1'l-cresol were theInain phenolic species detected by the tyrosinase based sensor, the flow-injection" ~

anlperolnetric signal for these sp'ecies was exalnined using a biosensor configuration. , '

with 13 different Inelnbranes (Fig. 4). None of the applied lnelnbranes allowed the se- "lective detection of a particular analyte, however, SOlne ofthelTI essentiallylnodifiedthe sensitivity of the biosensor for the three cOlnpounds considered,.Ba~e.donthese.Ineasurelnents, a selection of5 biosensors was Inade, with Inelnbranes listed ill 'Table 1.

Page 6: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

00

-.....

.)

o ~ ~ c:5 I...::::: e" S ~ ~~ ~ ~

~q ~" ~~ ~ ~ ~ ~ ~ <=;"

t-~ e ~ ~ (\ ~" ~ ~

12

0j

E :J

100'

ci w·

.......

....0Q

...:

~0

'"'

~!:

~R

!:.

.E

~~8

02

•"

E•

_e.

'"

41•

'0

'_

.:J..

"1

ft•'.

_.....

v,"0

II

0I

0•

....:.::.

C::J

I.:

~1

10

:'III.

:0

::

::

::

z:

:":<

'41

\0

0:

::e:

.:

o:

:'n

:'E

:~

Do

!i

i:l

i•

!a.

1I~I.:

.::

11

01

::

III

60

::

aI

..9•

l!fj:

::~:

•:

VI

i:..:.

i8.

~:

ii

0i

i:

til..

4141

0I

02

'"

o.

._

.E

"II

'•

4141

412

.....

_41

..

".

...

*1

•::

-::

Ia.:

:8

I:I

CD!

:!

:0

gi

!i

2i

:II

Ii-

:e

:z=

:z:

:=

1=:z.'l)

>4

01

0:

....

I:

:0

'1

1+

0-

:c:

;I

0:

II.

iI

III

I..

..:

(!.:~

:I:

:C:'0

::

a.0

:41

C'?

."::

:~

:0

::J

41:

G.

.!!

·1

'E.:

.··.·2

·:z:

lg

.ee

CD!..

::i

::

ci

OJ·i

.:8.

'0

.:.

-'0

.*

.u.

&i.

.:=

::!~!:=

.0:

a:2

0.b

,,.

I:...

.~Ifit

1!41

2i

~•

'.'*'

•1::1

-.'0

_••

L_

01

:1

g..

:z

z0

-.

"fA

E";;l l

Itl

:'"i·

:O

pr.

g::

lr4

1rAI;:IJ1II::Ital:~:JI'AlIE::r

A./:

;'"

'-:

i.1

'I~i

:.,..

....

..6

78

910

111

21

3

~p

he

no

lfi

lii

m-c

reso

lC

Jca

tech

ol

Fig

ure

4.T

here

lati

vesi

gnal

mag

nitu

defo

rph

enol

,mcc

reso

lan

dca

tech

olin

flow

-inj

ecti

onam

pero

met

ryw

ith

tyro

sina

sebi

osen

sors

wit

hdi

ffer

ento

uter

mem

bra

nes

.In

ject

ed10

0~l

lOO~moll-l

solu

tion

so

fph

eno

ls

Page 7: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

Detennination o/phenols with tyrosinasea111pero/netric bios~l1sor 871

Table 1. Menlbranes used in tyrosinase based anlperODletric biosensors for nlulticonlponentanalysis oflnixtures of phenols

0.4 Celanese Separations Products, USA

0.5 Kone, Finland

0.4 Nucleopore, USA

0.4 Nucleopore, USA

not known Zenith, Italy

[ Menlbrane type Material

I

I

Celgard 2500 polypropylene

Kone 3 celluloseI

Nucleopore PC polycarbonate

Nucleopore PE polyester

Teflon GTT 1210 PTFE

Poresize, pnl

Manufacturer

For each of these 5 biosensors Ineasurelnentsofthe alnperolnetric response in theflow-injection setup were perforlnedin 136luixtures containing three phenolic COln­pounds in different ratios in the concentrationrange froln 0 up to 100 ~t1noll-l . The re­sults of hundred randolnly selected Inixtures were used as the input signals fortraining of the neural network. Theresultsof the other 36 Inixtur~s were treated assan1ples, for which the network after the training phase was used to predict (deter..lnine) ".unknown" concentrationsofanalytes.

Inthe case, where the signal of the biosensor is always equal to the SUIn ofsignalsofindividual ahalytes lTIultiplied by constant selectivitycoefficients independent ofconcentrations and constant in tilne, the processing of data froln Inultico111ponentlneasurelnentis llluch silllplerwith no need to use a neural network. 'The systelTI ex­aillined, however, is not that siluple as it is illustrated by the experilnental dataob­t,fined forbiosensofswitha polypropylene membrane (Celgard2500) (Figs. 5-7).Figure 5 shows individual calibration plots with the largest sensitivity to catechol andthe sillallest for In-cresol. Linear calibration relationships were found for catecholand In-cresol, whereas phenol exhibits a non-linear dependence in the exalnined

180

Phenol

m-Cresol

Catechol160

140

< 120c~ 100.s:.:0')

'm 80L::::s:.C'd 60(l)c..

40

20

00 20 40 60 80 100

Concentration, J.lmOll-1

Figure 5. Calibration plots obtained in flow-injection aIl1perolnetry with tyrosinase biosensor with outerpolypropylenenlell1brane Celgard 2500

Page 8: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

872 M. Trc~ianoH'icz, A. Jagielska, P. RotklelVicz and A. Kierzek "

20 40 eo 80 100

m-Cresol concentration, J.1mQII-1

Figure 6. Calibration plots obtained in tlo'w-injection an1peron1etry with tyrosinase biosensor andCelgard 2500 tnelllbrane for m-cresol in the presence of different concentration of catechol

range of concentration. Figure 6 shows results obtained in two cOlnponent ll1ixtures,where the signal was Ineasured for increasing concentration of 111-cresol at differentlevelsofcatechol in the solutions. At the highest concentration ofcatecholone couldexpect the slnallestchanges of the Inagnitude of the signal with 0 concentration ofm-cresol, but this was not observed. Finally, Figure 7 shows the signal changes ob­served in three cOlnponent Inixtures in the presence of 1O(A} and 1OO(B)' ~unol r,tcatechol. The plots shown in Figure 7A, except for the region with the largest phe1101concentrations are according to expectations, whereas the changes plotted in Figure7B exhibit Inany irregularities. These irregularities observed for biosensors with dif­ferent Inelnbranes were the Inain reason to elnploy the neural network for the process­ing of experilnental data in Inulticolnponent lneasurelnents with biosensors. Oneadditional aspect, that Inay COll1plicate the data processing in this systeln, is a slowdegradation in tilne of the biocatalytic activity of tyrosinase in the biosensor used.

Data processing by artificial neural network

The investigated here analytical case with enzylnatic biosensors, froln the pointof view of data processing is a cOlnplex non-linear systeln. The used ar~hitectureofneural network in this study is a typical three layer perceptron with one hidden layer.In a prelilninary attelnpt several different configurations ofneural network have been .tested. ANN with linear hidden and output layers was not able to learn of given cari­bration data. Satisfactory results have not been obtained also for AN'N with 110n­linear hidden and output layers, for which training process was very slow regardlessvarious ways of scaling training data.

Page 9: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

Detennination (~lphenols lvitli tyrosinase al11per0l11etric biosensor 873

A

140

120<C..: 100.s:::.Q

~ 80.xca:. 60

40

20

0 0 20 40 60 80 100

PhenolCQncentratlon, p.mol [1

B400 -y------------------------

10·.Jlmol.rtphenol

350 L..!::..--__---A-----w;-------4:­

60 p.molr1

•<C.:"cQ •

! 300 I.- --.:.-.....---:.~---------~as:.

250

2000 20 40 60 80 100

m-Cresol concentration, Jlmol [1

Figure 7. Calibration plots for phenol (A) at 10 ~uuol catechol and different concentrations ofm-cresol and for (B) m-cresol at 100 pnl01 r) catechol and different concentrations ofphenolobtained in tlo\v-injection aJUper0111etry with tyrosinase biosensor with outer polypropylene111C111brane eclgard 2500

The best results have been obtained with non-linear hidden layer and linear out­

put layer. The software used in this study for data processing was progratTIlTIed in Clanguage for IBM PC 486DX C0111puter and it was used successfully earlier for the in­terpretation of spectroscopic data [32]. In the neural network feedforward· connec­

tions of neurons were used, signal was lllodified.by sigtTIoidal transfer function and

Page 10: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

874 lvl. Trojano'vllicz, A. Jagielska, P. Rotkielvicz and A. Kierzek

backpropagation learning algorithlll was used. Three-layer network design was ap­plied \-vith architecture shown in Figure 8. In each layer the nlunber ofneurons can bechanged. In this study 5 or 3 inputs corresponded to 5 or 3 biosensors, whereas 3 out­puts to 3 species to be deterlnined. The behavior of the network is approxilnated byadjustlnent of the weights of the connections in the network, which is the learningphase for the network. It starts frOin randolnly initialized weight factors. Using 100Inixtures as standards for the 'learning network and 36 Inixtures as unknown salnples,the effect of several operating paralneters such as nUlnber of learning cycles (nlunberof iterations of the aigorithin over the cOlnplete training set) and rate of learning andrllunber of neurons in the hidden layer were exalnined. Also the effect of decrease ofnlunber of inputs frOITI 5 to 3 was

Input

Hiddenlayer

Figure 8. Schelnatic diagranl ofthree-layer feedforward backpropagation network used in this work [1]

The obtained results for the detern1ination of 3 phenols in 36 salnples were ana­lyzed in terlllS of lnean relative error of deterlnination of a given analyte in all SalTI­pIes, nUlnber of results showing different range of absolute deviation froIll aillount

taken and correlation with alnount taken.Practically no effect was observed in the obtained results regarding nUlnber of

neurons in the hidden layer between 3 and 10, therefore the lnost data processing wascarried out with 3 neurons. Silnilarly no influence vtlas found for different rate oflearning, which was paralnetrized in the range [roill 0.25 to 1.0, and IllOSt often 0.5

was used. The quality of learning was exainined for different non-linearity ofhiddenneurons by changing the slope coefficient for SiglTIoids froin 0.5 to 2.0. The best re­sults were found for values close to 1.0. Certain ilnprovelnent of learning was also

Page 11: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

Detennination o.fphenols vvith (prosinase alnperOlJ1etric biosensor 875

found for Iogarithlnic scaling oftraining data according to expression x == In (x + 0.1).The effect ofnulnber of learning cycles in the range froin 100 to 100000 for the net­"vork with 3 neurons in the hidden layer on nlunber of erroneous results was exaln­ined. Generally the best results were always obtained for catechol and the worst onesfor 111-cresol, which follows the sequence ofsensitivities to these species ofahnost allbiosensors exalnined. The selection of such a criterion is arbitrary, but it SeelTIS to be

90

80A

70

60

50

40

30

20~0 101-

001-..(1)

(l)

> 100...,100n1 100000 1COOO 1000

Q) 9'J:a.-

t:: 00

CO(J) 70

~ED

50

40

3J

20

10

0 - Phenol- Catechol- m-cresol

Figure 9. COlnparison ofthe tnean relative errors ofdetern1ination ofphenol, catechol and n1-cresol mix­tures using data processing with neural network at different epochs for (/\) 3 and (B) 5 neuronsin input layer and 3 neurons in the hidden layer

Page 12: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

876 M.TrojanoH'icz, A. Jagielska, P. Rotkiewicz and A. Kierzek

Phenol.....~

-0 8

E::l,;c 6:J0 •.....c:0

~ ......cCD(.)C00

• ..0 20 40 60 80 100

100

Catechol

80

60

.co

20

00 20 40 60 80 100

Concentration taken, /lmol r1

10

O-t--~--.-~--r-~.....--.----.o 20 40 60 80 100

Figure 10. Correlation plots obtained for FIA dcternlination ofphenol, catechol and m-cresoJ in ll1ixturesusing 5 tyrosinase based atnperolnetric .biosensors with different outer 111C111brane and dataprocessing with neural network with 011e hidden laycr with 3 neurons and 10000 epochs

Page 13: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

Detenninatiol1 o.!,phenob; lvith (rrosinllse (unperolnetric biosensor 877

satisfactory and realistic for the cOluparison ofresultsfrolu analytical pointofview.As the best results were assluued those obtained for 10000 epochs, although fornt-cresol better results were obtained for 100 epochs. The increase of nUlnber of ep­

ochs up to 100000 results in evidently \Norseperforl1lanCe,\vhich is known asover-training of network (overfitting). [5].

A cOlnparison of the results for thesystelll with5 to 3 inputs and variable nunlberof learning cycles is shown by histogralus of the l1lean relative error ofdeterIninationfor each analyte (Fig. 9). In the best configuration with 10000 epochs the error for 5input systern is almost half of that for 3 inputs, except m-cresol,for which errors inboth syste111s were bet\veen 60 and 90%. In the best c~se the luean.relative error vvas35 and 32 % for phenol and catechol, respectively. It is worse result than that reportedfor nlulticoluponentdetenuinations with ion-selective electrodes [12], however, is_comparable to the resultsobtaihed for the determination oftwocornponents inthreecOluponent systelu in X-ray fluorescence spectros.copy; th.en observed errors r~ngedbetween 20 and 290/0 [7].

The correlation.plots obtained foreachanalytea.tthe.optimizedparameters of"data processing are shown in Figure 1o. The valu.es of the correlation coefficients

were 0.88,0.96 and' 0.67 for phenol, catechol and 111-C~esor, respectively.

CONCLUSIONS

The obtained results of data processing by artificial. neural network, except

catechol, can not be considered as fully satisfactory. Ai1alyticaldeterll1inations C,lr­

ried out with such accuracy should be considered as seluiqual1titative, only. T'hey arealso not sufficient to provide- a conclusive diagnosis about the l1lain causes of the ob­served errors in predictions of concentrations by neural network. It is quite probableth~t essential contribution to these effects COlnes [roin the li11lited stability of re-

I sponse ofbiosensors in tinle and toosluall differentiation of the used biosensors indleir sensiti\[ify to particular analytes. The iluprovcnlent of stability· should besearched inalilodificatiol1 of procedure of enzyn1e ilunlobilization, whereas luoresuitable differentiation of sensitivity by,. for instance, coilun10bilization of variousenzynles, luentionedabove.

It is also an open question, whether the neural network used in this investigati'onhas the OptilTIU1TI topology. and transfer function for this pllrpose. In this study alsofOUf layer ANN was applied (with two hidden layers), but such architecture has notprovided iluprovenlent, and additionally learning process was slower than for threelayer ANN. Another essential factor is whether the nUlnber of training data is suffi­

ciently large and properl~selected.In certain applications the l1luuber of objects intraining set was up to 32000 [4]. In the sanle review on chelnical applications ofneu- .

ral networks one can find, that percent of agreeluent or prediction ability for 11105tcases \vas reported in the range frol11 60 to 90, which is close to the results obtained in

Page 14: Flow-Injection.DeterminationofPhenols with Tyrosinase ...beta.chem.uw.edu.pl/chemanal/PDFs/1999/CHAN1999V0044P00865… · Flow-Injection.DeterminationofPhenols with Tyrosinase Amperometric

878 Iv!. Trojano¥vicz, A. Jagielska, P. Rotkiewicz and A. Kierzek

this study. Silnilarly to many other applications, it seelns that also in the case ofarraysof biosensors of differentiated selectivities the use of neural networks opens widepossibilities of their practical analytical application.

Acknowledgment

This work vvas partly supported by COPERNICUS Project, No. CIPA-CT94-0231,/;YJn1 European Conz­111unity and the University o.lWarsaw grant BST - 623/9/99.

REFERENCES

1. TadeusiewiczR.,Sieci neuronowe, Akadenlicka Oficyna Wydawnicza, Warszawa, 1993.2. DUll10nt M. and Perbet M., Neural Networks and Their Applications, Nilnes, 1989.3. Jansson P.A., Anal. Chenz., 63, (1991) 357A.4. Zupari 1. and Gasteiger 1., Anal. Chim. Acta, 248, 1 (1991).5. Bos M., Bos A. and van der Linden W.E., Analyst, 118 (1993) 323.6. Long 1.R., Gregorion V.G. and Genlperline P.J., AnaI.Chem., 62, (1990) 1791.7. Bas A., Bos M. and van der Linden W.E., Anal. Chim. Acta, 277, 289 (1993).8. GoodacreR., Tinllnins E.M., Jones A., Kell D.B., Maddock)., Heginbothonl M.L. and Magee 1.T.,

Anal. Chinz. Acta, 348, 511 (1997).9. Hitznlan B. and Kullick T., Anal. Chim. Acta, 294, 243 (1994).

10. Gobbunl J.V.S., Shelver W.L. and Shelver W.H., J. Liq. Chrom., 18 , 1957 (1995).11. Jimenez 0., Garcia M.A. and Marina M.L., J. Liq. Chromo & ReI. Technol., 20, 731 (1997).12. Bos M., Bos A. and van der Linden W.E., Anal. Chim. Acta, 233, 31 (1990).13. Wehrens R. and Van der Linden W.E., Anal. Chinz. Acta, 334, 93 (1996).14. Hofflleins B.S. and LaufR.J., Ana/usis, 20, 201 (1992).15. Sundgren H. and Lundstrom 1., Sells. Actuators B, 9, 127 (1992).16. Goppert 1. and Rosentiel W., Fresenius J. Anal. Chem., 349,367 (1994).17. Brezmes 1., Ferreras B., Lt"obet E., Vilanova X. and Correig X., Anal. Chim. Acta, 348, 503 (1997).18. Gardner J.W., HinesE.L. and Wilkinson M., Meas.Sci. Techno!., 1,446 (1990).19. Lauf R.1. and Hoftlleins B.S., Fuel. 70,935 (1991).20. Newman A.R., Anal. Chen1., 10, 585A (1991).21. Pearce T.C., Gardner J.W., Frielk S., Bertiett P.N. and Blair N., Ana~vst, 118,371 (1993).22. Hongnlei W., Lishi W., WaniiX., Baogui Z., Chengjun L. and Jianxing F.,Anal. Chem., 69, 699 (1997).23. Ruisanchez I., Lozano 1., LarrechiM.S., Rius F.X. and Zupan 1., Anal. Chim. Acta. 348, 113 (1997).24. de Garcia 1., Saravia M.L.M.F.S., Araujo A.N., Lill1a 1.L.F.C., del Valle M. and Poch Anal. Chim.

Acta, 358, ]43 (1997).25. Hitzl11ann B., Ritzka A., Ulber R., Scheper T. and Schtigerl K., Anal. Chim. Acta. 348, 135 (1997).26. Cohnor M.P., Sanchez 1., Wang 1., Snlyth M.R. and Mannino S., A/1a~vst, 114, 1427 (1989).27. Ortega F, Donlinguez E., Johansson E., Enlneus 1., Gorton L. and Marko-Varga G., 1. Chr()111atog}~, 675,

65 (1994).28. Adeyoju 0., Iwuoha E.I., Sll1yth M.R. and Leech D., Ana(vst, 121, 1885 (1996).29. Wang J., Lu F, Kane S.A., Choi Y.K., Sl11yth M.R. and Rogers K., Electroanalysis, 9, 1102 (1997).30. Politowicz M., Posniak M. and Trojanowicz M., in Buszewski B. (Ed.), Ekoanalityka W ochronie

a!rodowiska, SAR Ponlorze, ToruI1 1998,73.31. Yaropolov A.I., Kharybin A.N., Eillneus 1., Marko-Varga G. and Gorton L., Anal. Chin]. Acta, 308, 137

(1995).32. Kierzek 1., Kierzek A. and Malczewska-Bucko B., Nukleonika, 40, 133 (1995).

Received April 1998Accepted February 1999