Construction of a bitterness prediction model with an electronic
tongue and a trained sensory panel for the assessment of dairy
hydrolysates
J. Newman, N. Harbourne, D. O’Riordan, J.C. Jacquier, & M. O’Sullivan. Email: Jessica.Newman@ucd.ie
Food for Health Ireland, UCD School of Agriculture, (Institute of Food and Health)
University College Dublin, Belfield, D4
Results
Introduction
Experiment 1:
The PCA’s shown in Fig. 2 and 3 demonstrates the e-tongues excellent ability to discriminate
between basic tastes and types of bitter agents with resulting discrimination indexes of 93 and 89,
respectively. It should also be noted in Fig. 2 that the compounds were grouped by basic tastes
e.g. citric acid and tartaric acid, sour.
DI= 89
DI= 93
Na
200
Na
Ca
400
KC
KC
KC
KCKC
KC
KC
QU
QU
QU
QU
CA
CA
Sweet
Ta
Ci
TaTa
Ta
Ta
Ta
Ta
Sour
-1000
0
1000
0
PR
PR
PR
CA
CA
Ci
PR
PR
PR
-400
Ci
Ci Ci
Ci
Ci Ci
Ci
-2000
PR
CA
CA
-200
su
su
su
su
su
su
su
PR
CA
CA
su
-400
-200
PC 2 –
15.7%
Ta
Tr
Tr
TrTrTr
QU
QU
Ca
su
Tr
Tr
TrTr
QU
KC
PC2
PC2 2.76%
Ca
Ca
Ca
Ca
Ca
Ca
Ca
KC
200
400
Na
Na
Na
Na
Na Na
Bitter
QU
Salty
Na
0
Dairy manufacture is one of the Republic of Ireland’s most important industrial sectors; in 2011 over
5,400 million litres of milk was produced and total dairy exports accounted for €2.67 billion (Teagasc
2012). The potential of dairy protein hydrolysates to become part of that market as functional
ingredients is increasingly being researched. They have numerous improved characteristics over unhydrolysed dairy proteins e.g. improved gelation and foaming abilities in food systems and enhanced
nutritional properties in the form of bioactive peptides. However, the hydrolysis process can produce
bitter off-tastes in dairy proteins, limiting their potential use as food ingredients. This is a result of the
alteration of the native proteins to short chained peptides with exposed hydrophobic amino acids
(Ney, 1979). Therefore, it is necessary to screen dairy protein hydrolysates according to their sensory
character prior to application in food products
-500
2000
0
PC1
PC 1
– 78.6%
PC 1- 95.85%
Figure 2. PCA of tongue response to basic taste solutions
Figure 3. PCA of tongue response to PROP (PR), Quinine (QU)
& Caffeine (CA)
Experiment 2:
25
The PLS (Fig.4) displays the predicted
caffeine concentration by the e-tongue Vs.
actual concentration in SMUF. There is a
strong linear correlation R 2 =0.99. The result
suggests that the e-tongue may also be
used to quantify bitter compounds in more
complex solutions.
Predicted caffeine
concentration (mmol/l)
To date the main method for the evaluation of taste has been the trained sensory panel, but in recent
years, a number of electrochemical devices or electronic tongues (e-tongues) have been developed
as an alternative method. The benefits of using such a tool over a trained sensory panel are that it is
rapid, reliable and does not suffer from sensory fatigue. The advantage of the e-tongue is that it is
less time consuming than a sensory panel and can screen potentially toxic or unpleasant samples.
However, it is a relatively new technology and its reliability and accuracy needs to be established.
The e-tongue assessed in this study was the α-Astree e-tongue (Alpha M.O.S., Toulouse, France)
(Fig.1). It is composed of seven lipid/polymer membrane sensors which were developed for food
applications and a Ag/AgCl reference electrode. Each sensor has a different membrane and
depending on each sensor’s selectivity for a taste solution, it will generate electrical potential of
different magnitudes, which are monitored and subsequently analysed using multivariate analysis.
The objective of this study was to construct a bitterness prediction model using the e-tongue for the
assessment of dairy hydrolysates to reduce the reliance on sensory panel analysis.
500
R² = 0.99
20
15
10
5
0
0
5
10
15
20
Actual caffeine concentration (mmol/l)
Figure 4. PLS regression of correlation between caffeine
concentration as predicted by the e-tongue and actual
caffeine concentration in SMUF.
Experiment 3:
Fig. 5A and B show strong correlation between the bitterness values assigned to both whey and
casein hydrolysates by a trained sensory panel and those predicted by the e-tongue, with resultant
R2 values of 0.907 & 0.829 respectively. The robustness of each model was tested using randomly
selected hydrolysate samples as ‘unknowns’, these are denoted on the graphs as red data points.
The predicted values and actual sensory scores are shown in Table 1.
Fig. 5C. is the combination of all the dairy protein hydrolysates into one PLS. There is a strong
correlation between bitterness scores as rated by the sensory panel and by the e-tongue with a
R2=0.8422. The bitterness of one whey sample and one casein sample was predicted by the etongue, again shown as red data points on Fig.5C., the values of which are also shown in Table 1.
Bitterness intensity predicted
by e-tongue
Bitterness intensity as predicted by etongue
y = 1.0016x - 0.2928
R² = 0.907
15
12
9
6
3
0
0
3
6
9
12
15
y = 0.9027x + 0.9627
R² = 0.829
16
14
12
10
8
6
4
2
0
0
2
4
6
8
10
12
14
Figure 1. The electronic tongue with 7 sensor array
Experiment 1: A number of experiments were conducted with the e-tongue to assess its ability to
discriminate between basic taste compounds; salty (10 mmol/l KCl & NaCl), sweet (10 mmol/l
sucrose), sour (10 mmol/l citric acid & tartaric acid) and bitter (1 mmol/l tryptophan & caffeine). 6-npropylthiouracil (PROP) , quinine and caffeine (13 mmol/l) were also analysed to ensure the tongue
could distinguish between a variety of bitter agents. The results are expressed using principle
component analysis (PCA) constructed in the statistical package R version 2.11.1 (The R project for
statistical computing, 2012).
Experiment 2: The ability of the e-tongue to quantify bitterness was assessed using a series of
caffeine solutions (4.119-16.47 mmol/l) solubilised in simulated milk ultra-filtrate (SMUF). The results
were correlated using partial least square regression (PLS) generated using the Alpha M.O.S.
statistical software.
Experiment 3: A variety of casein and whey hydrolysates at a concentration of 10% w/w ranging in
degree of hydrolysis (DH) from 4-60%, were then analysed by a sensory panel (n=10) with more than
70 hours of training. The panellists were required to assign bitterness intensity using the 15 point
scale. Each sample was assessed in triplicate with no more than four samples analysed per session.
The samples were also analysed by the e-tongue. The results were correlated using PLS. The
bitterness of one whey sample and two casein samples were predicted by the e-tongue using the
PLS model. All the results in the analysis of dairy protein hydrolysates were then combined for a
larger PLS model where the bitterness of one whey and one casein sample was predicted.
References
1.Teagasc (2012 ) http://www.teagasc.ie/agrifood/
2.Ney, K. H. (1979) Bitterness of peptides - amino acid- composition and chain- length. Abstracts of
Papers of the American Chemical Society, 115:77.
3.Alpha M.O.S., (2010) α-Astree e-tongue technical notes.
4.Meilgaard. C.G.V., Carr T., Ed.(2000). Sensory Evaluation Techniques. Michigan, CRC Press
Bitterness intensity predicted by etongue
Materials and Methods
Figure 5B. PLS regression of correlation between
bitterness intensity in casein protein hydrolysates as
predicted by the e-tongue the e-tongue vs. assigned by
the sensory panel.
Figure 5A. PLS regression of correlation between
bitterness intensity in whey protein hydrolysates as
predicted by the e-tongue the e-tongue vs. assigned by
the sensory panel.
y = 0.8631x + 1.0856
R² = 0.842
15
12
Table 1. Bitterness intensity of dairy protein
hydrolysates as rated by sensory and predicted by etongue
9
6
Figure
Protein
Hydrolysate
DH
Sensory
E-tongue
Prediction
5.A
Whey
30
6.03
6.32
5.B
Casein
60
6.88
6.58
5.B
Casein
9.6
11.13
12.33
5.C
Whey
7.8
7.33
7.38
5.C
Casein
9.6
11.13
12.3
3
0
0
3
6
9
12
15
BItterness intensity rated by sensory panel
Figure 5C. PLS regression of correlation between
bitterness intensity in Casein & Whey protein
hydrolysates as predicted by the e-tongue vs.
assigned by the sensory panel.
Conclusion
 The e-tongue was able to discriminate between tastant compounds and group them by basic
taste.
 The e-tongue was used to quantify the concentration of a bitter tastant in SMUF, proving that
the e-tongue could be used in more complex systems i.e. real food samples.
 The bitterness values predicted by the e-tongue showed strong correlation with a trained
sensory panel in the evaluation of dairy protein hydrolysates. The PLS models were robust
enough to predict the relative bitterness on the 15 point scale of both whey and casein
samples.
 This study shows the potential for the e-tongue to be used in bitterness screening to reduce the
reliance on time consuming sensory panels.
Acknowledgement
The work described herein was supported by Enterprise Ireland under Grant Number CC20080001