Content uploaded by hélène Soyeurt
Author content
All content in this area was uploaded by hélène Soyeurt on Jan 05, 2016
Content may be subject to copyright.
Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Animal
(2013), 7:2, pp 348–354 &The Animal Consortium 2012
doi:10.1017/S1751731112001218
animal
Validation of fatty acid predictions in milk using mid-infrared
spectrometry across cattle breeds
M. H. T. Maurice-Van Eijndhoven
1,2
-
, H. Soyeurt
3,4
, F. Dehareng
5
and M. P. L. Calus
1
1
Animal Breeding and Genomics Centre, Wageningen UR Livestock Research, PO Box 65, 8200 AB Lelystad, The Netherlands;
2
Animal Breeding and Genomics
Centre, Wageningen University, PO Box 338, 6700 AH Wageningen, The Netherlands;
3
Animal Science Unit, Gembloux Agro-Bio Tech, University of Lie
`ge, 5030
Gembloux, Belgium;
4
National Fund for Scientific Research, 1000 Brussels, Belgium;
5
Valorisation of Agricultural Products, Walloon Agricultural Research Centre,
5030 Gembloux, Belgium
(Received 1 March 2012; Accepted 11 May 2012; First published online 2 July 2012)
The aim of this study was to investigate the accuracy to predict detailed fatty acid (FA) composition of bovine milk by mid-infrared
spectrometry, for a cattle population that partly differed in terms of country, breed and methodology used to measure actual FA
composition compared with the calibration data set. Calibration equations for predicting FA composition using mid-infrared
spectrometry were developed in the European project RobustMilk and based on 1236 milk samples from multiple cattle breeds
from Ireland, Scotland and the Walloon Region of Belgium. The validation data set contained 190 milk samples from cows in the
Netherlands across four breeds: Dutch Friesian, Meuse-Rhine-Yssel, Groningen White Headed (GWH) and Jersey (JER). The FA
measurements were performed using gas–liquid partition chromatography (GC) as the gold standard. Some FAs and groups of
FAs were not considered because of differences in definition, as the capillary column of the GC was not the same as used to
develop the calibration equations. Differences in performance of the calibration equations between breeds were mainly found by
evaluating the standard error of validation and the average prediction error. In general, for the GWH breed the smallest differences
were found between predicted and reference GC values and least variation in prediction errors, whereas for JER the largest
differences were found between predicted and reference GC values and most variation in prediction errors. For the individual
FAs 4:0, 6:0, 8:0, 10:0, 12:0, 14:0 and 16:0 and the groups’ saturated FAs, short-chain FAs and medium-chain FAs, predictions
assessed for all breeds together were highly accurate (validation
R
2
.0.80) with limited bias. For the individual FAs
cis
-14:1,
cis
-16:1 and 18:0, the calibration equations were moderately accurate (
R
2
in the range of 0.60 to 0.80) and for the individual FA
17:0 predictions were less accurate (
R
2
,0.60) with considerable bias. FA concentrations in the validation data set of our study
were generally higher than those in the calibration data. This difference in the range of FA concentrations, mainly due to breed
differences in our study, can cause lower accuracy. In conclusion, the RobustMilk calibration equations can be used to predict
most FAs in milk from the four breeds in the Netherlands with only a minor loss of accuracy.
Keywords: milk, fatty acid, mid-infrared spectrometry, cattle breeds
Implications
Measurement of detailed milk fat composition at individual
cow level is of major interest for the dairy industry because
of the expected relation with human health. Therefore, the
method of analyzing milk fat composition needs to be rapid
and suitable for extensive recording. Our study shows that
mid-infrared spectrometry (MIR) can be used to accurately
predict detailed milk fat composition from different cattle
breeds in the Netherlands.
Introduction
Bovine milk fat consists of a range of different fatty acids
(FAs), both unsaturated fatty acids (UFAs) and saturated
fatty acids (SFAs), and its relatively large amount of SFA
causes some debate about the role of bovine milk in
a healthy diet (Palmquist
et al.
, 2006). Clear variation in
fat content and milk fat composition can be found among
cows (Soyeurt and Gengler, 2008). Milk fat composition
varies with both environmental factors (e.g. feed regime;
Palmquist, 2006) and genetics (Soyeurt and Gengler, 2008;
Stoop
et al.
, 2008). Changing FA composition through the
feed regime or genetic selection requires a precise and
regular measurement.
-
Present address: Wageningen UR Livestock Research, Vijfde Polder 1, 6708 WC
Wageningen, The Netherlands. E-mail: myrthe.maurice-vaneijndhoven@wur.nl
348
To measure the FA composition in milk, several methods can
be used, which differ in throughput level, accuracy, workload
and costs. The most accurate method is gas–liquid partition
chromatography (GC). This largely implemented and regularly
used approach quantifies the concentration of individual FAs in
fat (Gander
et al.
, 1962; Christie, 1998). The major advantage of
GC is the possibility of measuring the individual FA proportions
with high accuracy (Smith, 1961; Christie, 1998) even if the
content of this FA is low. This method, however, is expensive and
time consuming and therefore less suitable for extensive and
regular recording. Another method of analyzing milk fat com-
position, MIR, is rapid and less expensive in case of extensive
use (Wilson and Tapp, 1999; Soyeurt
et al.
, 2006). MIR is routi-
nely used in milk recording schemes to measure lactose, urea,
total fat and protein percentages in bovine milk (Etzion
et al.
,
2004; Bobe
et al.
, 2007). MIR was used by Soyeurt
et al.
(2006
and 2011) and Rutten
et al.
(2009) to estimate calibration
equations predicting the FA concentrations in milk (g/dl of milk)
and milk fat (g/100 g of fat), and these equations were subse-
quently validated. In these studies, the predictions have low
accuracy for FAs that are present in low concentrations, such as
the
trans
and unsaturated 14, 16 and 18 FAs.
Accuracy and bias in calibration equations may also be
affected when there are differences between the samples used
to estimate the calibration equations and the samples for which
FA composition is predicted using the prediction equations.
In Rutten
et al.
(2009), calibration equations were based on
milk samples only from Holstein–Friesian (HF) cows. In this
latter study, analysis of milk samples collected in both winter
and summer indicated that season has a limited effect on
prediction accuracy but generally a large effect on prediction
bias. This indicates that factors causing structural differences
between FA composition of groups of animals such as season
and breed can affect predictability of calibration equations.
The aim of this study was to investigate the accuracy and
bias in predicting detailed FA composition from MIR spectra
of milk from four cattle breeds in the Netherlands, using
calibration equations based on milk samples collected from
Belgian, Irish and Scottish cattle of partly different breeds.
Material and Methods
Calibration equations
In this study, prediction of the composition of 11 individual
FAs and 3 groups of FAs using MIR spectrometry using cali-
bration equations was validated. These calibration equations
were developed in the EU FP 7 project RobustMilk, using a
data set with MIR spectra and GC results of 1236 milk
samples. The methodology used to develop the calibration
equations is explained by Soyeurt
et al.
(2011) for the cali-
bration equations, but it should be noted that in our study
updated versions of the calibrations were used, which are
based on 1236 instead of 517 milk samples.
For all 1236 milk samples, the MIR analysis was performed
using a Fourier-transformed interferogram with a region of
1000 to 5000/cm (MilkoScan FT 6000, Foss Electric, Hillerod,
Denmark). The detailed FA composition of these 1236 milk
samples was obtained using GC realized at the milk laboratory of
the Walloon Agricultural Research Centre (Gembloux, Belgium).
The GC outputs were generated by analyzing methyl esters pre-
pared from milk fat as described in ISO Standard 15 884 (ISO–IDF
(International Organization for Standardization–Interna-
tional Dairy Federation), 2002) and the GC was equipped
with a CPSil-88 column (Varian Inc., Palo Alto, CA, USA) with
a length of 100 m and an internal diameter of 0.25 mm.
The 1236 milk samples were collected from herds in Ire-
land, Scotland and the Walloon Region of Belgium with
purebred and crossbred cows from different breeds, that is,
HF, Jersey (JER), Red and White, Normande, Montbeliarde
and dual-purpose Belgian Blue. This multiple breed and
multiple country composition of the data set was chosen to
cover a wide range of the variability of FA in bovine milk in
order to improve the robustness of the developed calibration
equations. The calibration equations were developed from
three MIR regions located between 926 and 1600/cm, 1712
and 1809/cm and 2561 and 2989/cm. The method used to
relate MIR spectra to FA data was partial least square
regression after a first derivative pre-treatment on spectral
data to correct the baseline drift. A
T
-outlier test was also
used during the calibration process to delete potential GC
outliers. Therefore, the final number of samples included in
each calibration equation varied following the considered
FA. Descriptive statistics of the RobustMilk calibration
equations are given in Table 1. Note that this is an updated
version of the prediction equations described by Soyeurt
et al.
(2011), in the sense that the current prediction equa-
tions are based on ,4.5 times more samples. In addition to
the number of samples included in the calibration data set,
the mean and the standard deviation (s.d.) of the FA content
measured by GC, the standard error of calibration (SEC), the
calibration coefficient of determination (
R
2
c), standard error
of cross-validation (SECV), cross-validation coefficient of
determination (
R
2
cv) and the ratio of s.d. to SECV (RPD) are
shown. The
R
2
c is the square of the correlation coefficient
between the predicted and the reference GC values.
During the development of the updated calibration
equations, a first assessment of the robustness of the pre-
dictions was done by a cross-validation approach to calculate
the
R
2
cv and SECV using the same approach as described by
Soyeurt
et al.
(2011).
Validation data set
Between December 2008 and March 2009 in the Nether-
lands, that is, in the winter season, a total of 190 cows were
sampled once during morning milking. Samples were treated
immediately with 0.03% (w/w) sodium azide to avoid
microbiological growth. Cows belonged to four breeds:
Dutch Friesian (DF; 47 samples from 3 farms), Meuse-Rhine-
Yssel (MRY; 52 samples from 3 farms), Groningen White
Headed (GWH; 45 samples from 3 farms) and JER (46 sam-
ples from 3 farms). The cows were selected by farmers to
reflect variations in age, parity, stage of lactation and ancestry.
On all farms, cows were kept indoors in the studied period
and milked twice a day with conventional milking systems.
Validation of milk fatty acid prediction
349
The number of sampled cows per herd ranged from 6 to 24, and
the selected farms each had between 35 and 120 cows. The
cows were either located at organic or conventional farms. For
each breed samples were collected at one or two organic farms
and the remainder farms were conventional. Differences in FA
composition in milk between the four breeds in this data set
are presented by Maurice-Van Eijndhoven
et al.
(2011).
Briefly, ranges of individual FA content generally overlapped
between breeds, apart from several FAs and groups of FAs of
JER and GWH.
Each milk sample was analyzed using both GC and MIR.
The mean and standard deviation of the FA content of each
of the 11 individual FAs and the 3 groups of FAs obtained
using the GC are given in Table 2. The relative variability,
which wasexamined by calculating the coefficient of variation
(results not shown), between the different FAs was highest
for the
cis
-14:1 (range 31.1 to 35.8) and
cis
-16:1 (range 26.2
to 39.1) and lowest for the 4:0, 6:0, 8:0 and total group
of short-chain FA (SCFA; range 14.6 to 25.0). GC analysis
was performed at the laboratory of Qlip N.V. (Leusden, The
Netherlands). The GC outputs were generated by analyzing
methyl esters prepared from milk fat as described in ISO
Standard 15 884 (ISO–IDF, 2002) and the GC was equipped
with a Varian Fame Select CP 7420 column (Varian Inc., Palo
Alto, CA, USA) with a length of 100 m and an internal dia-
meter of 0.25 mm. The MIR analysis was performed using a
Fourier-transformed interferogram with a region of 1000 to
5000/cm (MilkoScan FT 6000, Foss Electric, Denmark) at the
laboratory of Qlip N.V. (Zutphen, The Netherlands). The
validation data set was independent of the calibration set
Table 1
Descriptive statistics of RobustMilk FA calibration equations and the data used to derive the equations
Trait (g/dl of milk)
n
Mean s.d. SEC
R
2
c SECV
R
2
cv RPD
4:0 1186 0.101 0.030 0.008 0.93 0.008 0.93 3.68
6:0 1189 0.074 0.023 0.005 0.96 0.005 0.96 4.81
8:0 1180 0.048 0.015 0.003 0.96 0.003 0.96 5.00
10:0 1183 0.112 0.036 0.007 0.96 0.008 0.96 4.72
12:0 1180 0.134 0.044 0.009 0.96 0.010 0.95 4.61
14:0 1184 0.448 0.130 0.027 0.96 0.028 0.95 4.70
cis
-14:1 1180 0.040 0.015 0.007 0.80 0.007 0.78 2.13
16:0 1179 1.206 0.424 0.066 0.98 0.068 0.97 6.20
cis
-16:1 1179 0.067 0.023 0.010 0.79 0.011 0.78 2.14
17:0 1167 0.028 0.008 0.002 0.90 0.003 0.89 3.04
18:0 1173 0.375 0.145 0.043 0.91 0.045 0.90 3.24
SFA 1176 2.689 0.785 0.050 1.00 0.051 1.00 15.34
SCFA 1185 0.349 0.104 0.020 0.96 0.020 0.96 5.10
MCFA 1187 2.056 0.645 0.082 0.98 0.086 0.98 7.53
FA 5fatty acid;
n
5number of samples included in the calibration equation; Mean 5mean of gas chromatographic data; s.d. 5standard deviation of gas
chromatographic data; SEC 5standard error of calibration;
R
2
c5calibration coefficient of determination; SECV 5standard error of cross-validation;
R
2
cv 5cross-
validation coefficient of determination; RPD 5the ratio of s.d. to SECV; SFA 5the saturated FAs 4:0 to 22:0 including iso- and ante-iso FAs; SCFA 5short-chain FAs
4:0 to 10:0; MCFA 5medium-chain FAs 12:0 to 16:0.
Table 2
The mean and standard deviation of gas chromatographic measurements of the validation data for all traits of the
individual breeds
Trait (g/dl of milk) GWH (mean 6s.d.) MRY (mean 6s.d.) DF (mean 6s.d.) JER (mean 6s.d.)
4:0 0.131 60.020 0.124 60.024 0.130 60.023 0.171 60.028
6:0 0.093 60.014 0.098 60.019 0.104 60.017 0.133 60.024
8:0 0.059 60.011 0.072 60.015 0.072 60.011 0.088 60.018
10:0 0.138 60.031 0.174 60.046 0.186 60.033 0.224 60.057
12:0 0.184 60.051 0.244 60.066 0.230 60.047 0.273 60.076
14:0 0.563 60.094 0.637 60.140 0.617 60.114 0.782 60.151
cis
-14:1 0.052 60.019 0.055 60.019 0.044 60.014 0.061 60.019
16:0 1.483 60.275 1.472 60.305 2.260 60.442 1.522 60.353
cis
-16:1 0.065 60.017 0.057 60.019 0.058 60.018 0.098 60.028
17:0 0.027 60.008 0.024 60.006 0.023 60.005 0.037 60.007
18:0 0.495 60.150 0.517 60.109 0.524 60.100 0.697 60.118
SFA 3.332 60.503 3.522 60.671 3.563 60.632 4.876 60.817
SCFA 0.436 60.069 0.482 60.101 0.507 60.074 0.637 60.119
MCFA 2.500 60.439 2.621 60.546 2.619 60.537 3.674 60.709
GWH 5Groningen White Headed; MRY 5Meuse-Rhine-Yssel; DF 5Dutch Friesian; JER 5Jersey; FA 5fatty acid; SFA 5the saturated FAs
4:0 to 22:0 including iso- and ante-iso FAs; SCFA 5short-chain FAs 4:0 to 10:0; MCFA 5medium-chain FAs 12:0 to 16:0.
Maurice-Van Eijndhoven, Soyeurt, Dehareng and Calus
350
developed in the RobustMilk project (i.e. different labs for
GC and MIR analysis).
Validation
RobustMilk calibration equations (Table 1) were used to
predict detailed milk composition of the samples recorded in
the validation data set for 11 individual FAs 4:0, 6:0, 8:0,
10:0, 12:0, 14:0,
cis
-14:1, 16:0,
cis
-16:1, 17:0, 18:0 and the 3
groups of FAs, that is, total SFA (SFA 4:0 to 22:0 including
iso- and ante-iso FAs), short-chain FA (SCFA; 4:0 to 10:0) and
medium-chain FA (MCFA; 12:0 to 16:0). SFA 5the saturated
fatty acids 4:0 to 22:0 including iso- and ante-iso FAs;
SCFA 54:0 to 10:0; MCFA 512:0 to 16:0. Owing to the lack
of agreement between the GC analyses of the calibration
and validation data set methods for the long-chain unsatu-
rated FAs and their related FA groups (i.e. total unsaturated,
monounsaturated, polyunsaturated and long-chain FAs),
these FAs and groups of FAs were considered in this study. This
lack of agreement was due to differences in separation of the
long-chain unsaturated FAs during the GC analyses of the
calibration and validation data sets because the capillary col-
umns used were different. For the other FAs, of which the
calibration equations are validated in this study, the separation
during the GC analysis was similar. The FA traits were predicted
on the basis of milk (g/dl) because these predictions are more
accurate than on the basis of fat (g/100 g; Soyeurt
et al.
, 2006;
Rutten
et al.
, 2009; Soyeurt
et al.
, 2011).
The accuracy of the RobustMilk predictions was evaluated
using the root mean squared error of prediction (SEV), the
coefficient of determination (validation
R
2
) and the ratio of
the s.d. of the validation data set to the SEV (RPD
v
). Cali-
bration equations with RPD
v
above 3.0 can be considered as
good predictors (Williams and Sobering, 1993).
The SEV was calculated as
SEV ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
N
i¼1
ð^
yiyiÞ2
n
v
u
u
u
t;
where ^
yiis the predicted value obtained for the sample
i
;
y
i
is the reference GC value of sample
i
;
n
is the number of
samples in the validation set. The approach to calculate SEV
is in line with the approach to calculate SEC and SECV, which
are described by Soyeurt
et al.
(2011).
The prediction bias was assessed using the average pre-
diction error ð^
yiyiÞand slope (
b
1
) of the linear regression,
with the GC values as dependent and the predicted values as
independent variable. To be able to compare the average
prediction error of the calibration equations across traits and
breeds, this measure is expressed as a percentage of the
mean of the gas chromatography values.
Results
For most traits, the SEV was lowest for the predicted FA
contents in GWH milk, except for the individual FA
cis
-14:0
and the group of FA SCFA (Table 3). The SEV for the predicted
FA contents in JER milk was highest for all groups of FAs and
the individual FAs 6:0, 14:0,
cis
-14:0 and 16:0, except for the
individual FAs 4:0, 12:0,
cis
-16:1, 17:0 and 18:0, of which the
SEV was highest for MRY. The validation
R
2
of the predictions
of the individual FAs 4:0, 6:0, 8:0, 10:0, 12:0, 14:0, 16:0 and
for the groups of FAs SFA, SCFA and MCFA for all breeds
were above 0.80 (Table 4). The validation
R
2
for the individual
Table 3
The SEV of 11 FAs and 3 groups of FAs for different dairy breeds
Breed
Trait (g/dl of milk) GWH MRY DF JER All breeds
1
4:0 0.009 0.017 0.011 0.011 0.012
6:0 0.005 0.006 0.006 0.007 0.006
8:0 0.004 0.005 0.006 0.006 0.005
10:0 0.012 0.016 0.025 0.022 0.019
12:0 0.027 0.048 0.036 0.028 0.036
14:0 0.033 0.042 0.037 0.045 0.039
cis
-14:1 0.011 0.010 0.014 0.022 0.015
16:0 0.122 0.201 0.215 0.219 0.192
cis
-16:1 0.023 0.040 0.033 0.028 0.032
17:0 0.010 0.013 0.012 0.010 0.012
18:0 0.106 0.145 0.139 0.137 0.132
SFA 0.061 0.047 0.050 0.130 0.078
SCFA 0.022 0.028 0.028 0.033 0.028
MCFA 0.105 0.171 0.207 0.253 0.190
SEV 5standard error of validation; FA 5fatty acid; GWH 5Groningen White
Headed; MRY 5Meuse-Rhine-Yssel; DF 5Dutch Friesian; JER 5Jersey; SFA 5the
saturated FAs 4:0 to 22:0 including iso- and ante-iso FAs; SCFA 5short-chain FAs
4:0 to 10:0; MCFA5medium-chain FAs 12:0 to 16:0.
1
Breeds total is the SEV of the predictions across the breeds GWH, DF, MRY
and JER.
Table 4
The validation
R
2
of prediction of 11 FAs and 3 groups of FAs
for different dairy breeds
Breed
Trait (g/dl of milk) GWH MRY DF JER All breeds
1
4:0 0.92 0.92 0.89 0.88 0.92
6:0 0.90 0.92 0.88 0.91 0.93
8:0 0.88 0.90 0.88 0.91 0.92
10:0 0.85 0.94 0.89 0.93 0.93
12:0 0.85 0.86 0.80 0.90 0.85
14:0 0.93 0.97 0.93 0.92 0.95
cis
-14:1 0.70 0.76 0.79 0.80 0.64
16:0 0.89 0.90 0.93 0.86 0.93
cis
-16:1 0.56 0.67 0.48 0.59 0.65
17:0 0.15 0.17 0.73 0.24 0.43
18:0 0.80 0.64 0.65 0.58 0.72
SFA 0.99 1.00 1.00 0.98 0.99
SCFA 0.91 0.93 0.93 0.93 0.95
MCFA 0.95 0.97 0.97 0.92 0.96
FA 5fatty acid; GWH 5Groningen White Headed; MRY 5Meuse-Rhine-Yssel;
DF 5Dutch Friesian; JER 5Jersey; SFA 5the saturated FAs 4:0 to 22:0
including iso- and ante-iso FAs; SCFA 5short-chain FAs 4:0 to 10:0;
MCFA 5medium-chain FAs 12:0 to 16:0.
1
Breeds total is the
R
2
of the predictions across the breeds GWH, DF, MRY
and JER.
Validation of milk fatty acid prediction
351
FA 17:0 was lowest over all breeds (0.43). The FA composition
of milk from DF cows was based on the calculated validation
R
2
predicted most accurately with an average
R
2
of 0.84. The
average validation
R
2
of the predicted FA composition of milk
from GWH was generally lowest (0.81). The long-chain FAs 17:0
and 18:0 and medium-chain FAs
cis
-14:1 and
cis
-16:1 showed
the largest variation in validation
R
2
between the breeds. The
RPD
v
was in general lower than the RPD of the cross-validation;
however, a similar trend was observed (Table 5). The RPD
v
is
above 3.0 across all breeds (breeds total) for 6:0; 8:0; 14:0 and
all groups of FAs.
Bias was examined by calculating the average predic-
tion error and the slope of the linear regression with the GC
values as dependent and the predicted values as indepen-
dent variable (Tables 6 and 7). To be able to compare the
average prediction error of the calibration equations
between traits and breeds, the values were expressed as a
percentage of the mean of the gas chromatography absolute
value (Table 6). The bias in terms of average prediction error
for individual breeds is largest for MRY with on average
26.1% followed by JER (25.7%) and smallest for GWH with
on average 2.4%. For all breeds together, the average pre-
diction error is 25.1% with an s.d. of 15.6, which means
that the average difference between the predicted content
using MIR and the reference GC values was 25.1% (nor-
malized to the mean). The average prediction error were
highest for the predicted contents of the individual FAs
cis
-16:1 and 17:0. The
b
1
, which is clearly related to the
R
2
,
does not show unexpected results as
b
1
is generally closer to
1 when the
R
2
is also closer to 1. With a
b
1
value of 1.55, the
Table 5
The RPD
v
1
of 11 FAs and 3 groups of FAs for different dairy
breeds
Breed
Trait (g/dl of milk) GWH MRY DF JER All breeds
2
4:0 2.29 1.44 2.01 2.46 2.41
6:0 3.06 2.94 2.71 3.24 3.86
8:0 2.96 3.14 1.77 3.23 3.53
10:0 2.53 2.89 1.32 2.62 2.76
12:0 1.90 1.36 1.30 2.69 1.91
14:0 2.84 3.30 3.09 3.33 3.80
cis
-14:1 1.74 1.84 1.01 0.85 1.28
16:0 2.25 1.52 2.06 1.61 2.50
cis
-16:1 0.74 0.47 0.55 1.00 0.85
17:0 0.79 0.44 0.40 0.67 0.68
18:0 1.42 0.75 0.72 0.86 1.09
SFA 8.18 14.20 12.72 6.30 11.55
SCFA 3.14 3.66 2.64 3.59 4.29
MCFA 4.16 3.19 2.59 2.80 3.87
RPD
v
5ratio of the standard deviation of the validation samples to the standard
error of prediction of the validation; FA5fatty acid; GWH 5Groningen White
Headed; MRY 5Meuse-Rhine-Yssel; DF 5Dutch Friesian; JER 5Jersey; SFA5the
saturated FAs 4:0 to 22:0 including iso- and ante-iso FAs; SCFA 5short-chain FAs
4:0 to 10:0; MCFA5medium-chain FAs 12:0 to 16:0.
1
Calibration equations with RPD
v
above 3.0 can be considered as good predictors
(Williams and Sobering, 1993. Journal of Near Infrared Spectroscopy 1, 25–32).
2
Breeds total is the RPD
v
across the breeds GWH, DF, MRY and JER.
Table 6
The average prediction error
1
of the predictions of 11 FAs and
3 groups of FAs for different dairy breeds
Breed
Trait (g/dl of milk) GWH MRY DF JER All breeds
2
4:0 24.6 210.7 26.1 23.3 26.4
6:0 20.4 22.9 2.6 0.5 20.1
8:0 20.8 20.4 6.7 20.4 1.5
10:0 0.7 4.9 12.1 7.5 6.3
12:0 6.6 16.8 12.4 4.3 10.3
14:0 3.3 5.1 3.2 21.4 2.6
cis
-14:1 25.5 28.3 223.0 235.9 217.9
16:0 24.5 210.4 211.3 28.5 28.8
cis
-16:1 228.3 254.4 242.5 230.4 239.5
17:0 224.6 242.5 243.5 228.6 235.1
18:0 12.1 22.9 22.1 19.9 19.4
SFA 1.2 0.8 0.9 0.6 0.9
SCFA 20.9 21.8 3.7 0.6 0.4
MCFA 21.1 25.0 26.4 25.4 24.5
Mean
3
23.4 26.1 24.9 25.7 25.1
s.d.
3
10.4 19.7 18.8 15.0 15.6
FA 5fatty acid; GWH 5Groningen White Headed; DF 5Dutch Friesian;
MRY 5Meuse-Rhine-Yssel; JER 5Jersey; SFA 5the saturated FAs 4:0 to 22:0
including iso- and ante-iso FAs; SCFA 5short-chain FAs 4:0 to 10:0;
MCFA 5medium-chain FAs 12:0 to 16:0.
1
The average prediction error calculated as the predicted value minus
the reference gas chromatography values and expressed as percentage of
the mean of the gas chromatography values: (average prediction error/
mean) 3100.
2
All breeds means the average prediction errors across all predictions for
GWH, DF, MRY and JER.
3
The mean and s.d. of all average prediction errors for each breed and all
breeds together.
Table 7
The slope (
b
1
) of the linear regression with the gas chroma-
tography values as dependent and the predicted values as indepen-
dent variable of 11 FAs and 3 groups of FAs for different dairy breeds
Breed
Trait (g/dl of milk) GWH MRY DF JER All breeds
1
4:0 0.95 0.95 0.99 1.16 1.06
6:0 0.90 0.92 0.95 1.07 0.99
8:0 0.97 0.96 1.00 1.06 0.99
10:0 1.04 1.16 1.12 1.14 1.13
12:0 1.27 1.26 1.10 1.14 1.09
14:0 0.93 1.04 1.01 0.91 0.93
cis
-14:1 1.22 1.10 0.91 1.11 0.85
16:0 0.92 0.95 1.03 1.00 0.99
cis
-16:1 0.77 0.71 0.71 0.89 0.88
17:0 0.76 0.37 0.83 0.63 0.80
18:0 1.55 0.90 0.84 0.82 0.98
SFA 0.99 1.00 1.00 1.01 1.00
SCFA 0.93 1.00 0.97 1.08 1.02
MCFA 0.96 0.97 1.01 0.99 0.97
FA 5fatty acid; GWH 5Groningen White Headed; DF 5Dutch Friesian;
MRY 5Meuse-Rhine-Yssel; JER 5Jersey; SFA 5the saturated FAs 4:0 to 22:0
including iso- and ante-iso FAs; SCFA 5short-chain FAs 4:0 to 10:0;
MCFA 5medium-chain FAs 12:0 to 16:0.
1
All breeds means the average prediction errors across all predictions for
GWH, DF, MRY and JER.
Maurice-Van Eijndhoven, Soyeurt, Dehareng and Calus
352
variance of the predicted content of 18:0 for GWH milk
showed the largest underestimation (Table 7). With a
b
1
value of 0.37, the variance of the predicted content of 17:0
for MRY showed the largest overestimation, which indicated
a lack of relation between the true and predicted values also
shown by the
R
2
calculated to be 0.17.
Comparing the descriptive statistics of the GC data, the FA
content in the milk of the validation data set is generally
higher than in the milk of the calibration data set. Especially
JER milk in the validation data set showed higher FA con-
tents, as the mean contents of 6:0, 8:0, 10:0, 12:0, 14:0, SFA,
SCFA and MCFA were outside the 95% confidence interval of
the mean of the calibration data of the calibration data set
(i.e. larger than 2.5 times the standard deviation above the
mean contents).
The performance of the calibration equations to predict
the content of the FAs 16:0 and
cis
-16:1 is also visualized in
Figures 1 and 2. For 16:0, a clear linear pattern is shown in
Figure 1, which result in the high validation
R
2
ranging from
0.86 to 0.93. For 16:1, Figure 2 clearly shows relatively more
deviation of the predicted values. In both figures, especially
predictions for JER are located in a different direction.
Discussion
The aim of this study was to investigate the accuracy of
calibration equations based on milk samples collected from a
population with different origin in terms of country, breed
and methodology used to measure actual FA composition. In
general, FAs with higher content in milk can be predicted
more accurately than milk with a lower FA content (Soyeurt
et al.
, 2006 and 2011; Rutten
et al.
, 2009). In this study,
predictions of FA with high content in milk (.1 g/dl milk)
were also highly accurate (validation
R
2
.0.80); however,
7 of the total 11 FAs with lower content in milk (,1 g/dl
milk) were predicted to be highly accurate by means of
validation
R
2
. Differences in performance of the calibration
equations between breeds were mainly found by evaluating
the SEV and the average prediction error. Results showed on
average for GWH the smallest difference between predicted
and reference values and least variation in prediction errors,
whereas for JER on average the largest differences were
found between predicted and reference values and most
variation in prediction errors.
The RobustMilk calibration equations validated in our
study were updated versions of the calibration equations
reported in Soyeurt
et al.
(2011), in that the calibration data
set was enlarged. Despite this increase in size of the cali-
bration data set, the predictions in our study were in general
less accurate than those of Soyeurt
et al.
(2011). Comparing
both studies, the FA composition in the validation data set of
Soyeurt
et al.
(2011) was generally closer to the FA compo-
sition of the calibration data, whereas FA concentrations in
the validation data set of our study were generally higher
than those in the calibration data. This difference in range
of FA concentrations, mainly due to differences in breed, is
the most likely reason for this lower accuracy. A comparable
difference in accuracy was found by Rutten
et al.
(2009)
when predicting FA composition in winter or summer, using
a calibration equation that was based on winter samples
only. This indicates that differences in FA composition due to
differences in season (in which the feeding regime differs)
are as important as differences due to breed (Rutten
et al.
,
2009). When winter milk samples were used in the calibra-
tion data set to predict FA composition of summer samples,
differences in concentration ranges between the calibration
data set and the validation data set especially affected the
bias (i.e. relative difference in means; Rutten
et al.
, 2009).
In our study, the FAs that showed the largest difference
in mean between our validation and the calibration data
were not necessarily the same FAs as those that showed the
largest bias. For instance, despite a relatively small difference
Figure 1 The predicted content of the individual fatty acid 16:0 based on
mid-infrared spectometry plotted against the reference gas chromatogra-
phy values.
Figure 2 The predicted content of the individual fatty acid
cis
-16:1 based
on mid-infrared spectometry plotted against the reference gas chromato-
graphy values.
Validation of milk fatty acid prediction
353
in concentration of 14:1,
cis
-16:1, 17:0 and 18:0 between
validation and calibration data, those FAs showed the largest
bias (i.e. average prediction error and
b
1
). As
cis
-14:1, 16:1
and 17:0 are present in very low concentrations (,0.01 g/dl
milk), this is the most likely cause of their high bias. Differences
between means of validation and calibration data were
largest for the concentrations of short and medium FAs in
JER milk, which generally had a higher concentration in JER
milk compared with the other breeds. Remarkably, despite
the large difference in concentration, these FAs generally
had accurate predictions. Therefore, it seems that differences
in accuracy and bias are not only caused by differences in
concentration of the individual FAs, but perhaps also by
spectral variability of the milk samples. As indicated by
Soyeurt
et al.
(2011), adding milk samples to the calibration
data to maximize the spectral variability of the samples in
the calibration data set is an effective method to optimize
calibration equations.
The suitability of calibration equations depends on the
application of the predictions. If the primary interest is in
predicting individual FA composition, then highly accurate
and unbiased predictions are important. When the interest is
in predicting differences between individuals or populations
(e.g. for breeding purposes), accurate and unbiased predic-
tions are important; however, less accurate or biased pre-
dictions can still be suitable, especially when multiple
measurements are available per individual. Suitability of
calibration equations, which are to some extent derived
under different conditions, can be evaluated by means of an
external validation as presented in this study.
As the dairy breeding industry is interested in selecting
cows producing milk with a specific FA composition, the
suitability of the calibration equations depends on the
reduction in genetic gain when using MIR information
instead of GC information. As Rutten
et al.
(2010) found, the
possible genetic gain estimated using FA composition
determined by predictions based on MIR was almost equal
to the possible genetic gain estimated using FA composition
determined by GC, in dairy breeding schemes with progeny
testing. The latter result was reached with even moderate
and quite low validation
R
2
’s ranging from 0.53 to 0.77. The
genetic gain estimated by Rutten
et al.
(2010) assumed the
availability of information on large groups of daughters per
sire. Reaching similar gain could be difficult for the Dutch
breeds in our study, as bulls in these breeds have generally
smaller daughter groups. For these Dutch breeds, therefore,
calibration equations that give highly accurate predictions
are necessary to obtain genetic gains similar to the mainstream
cattle breeds.
Conclusion
In conclusion, the RobustMilk calibration equations can be
used to predict the content of most saturated FA in milk
using MIR spectrometry for the breeds GWH, MRY, DF and
JER in the Netherlands with only a minor loss of accuracy
compared with predictions for Holstein cows.
Acknowledgments
This study was financially supported by the Ministry of Agri-
culture, Nature and Food (Programme ‘Kennisbasis Research’,
code: KB-04-002-021 and KB-05-003-041). The 12 involved
farmers are thanked for contributing milk samples to this study.
Furthermore, the RobustMilk project team is thanked for the
use of the calibration equations. The RobustMilk project is
financially supported by the European Commission under the
Seventh Research Framework Programme, Grant Agreement
KBBE-211708. This publication represents the views of the
authors, not the European Commission, and the Commission is
not liable for any use that may be made of the information.
References
Bobe G, Lindberg GL, Freeman AE and Beitz DC 2007. Short communication:
composition of milk protein and milk fatty acids is stable for cows differing in
genetic merit for milk production. Journal of Dairy Science 90, 3955–3960.
Christie WW 1998. Gas chromatography mass spectrometry methods for
structural analysis of fatty acids. Lipids 33, 343–353.
Etzion Y, Linker R, Cogan U and Shmulevich I 2004. Determination of protein
concentration in raw milk by mid-infrared Fourier transform infrared/attenuated
total reflectance spectroscopy. Journal of Dairy Science 87, 2779–2788.
Gander GW, Sampugna J and Jensen RG 1962. Analysis of milk fatty acids by
gas–liquid chromatography. Journal of Dairy Science 45, 323–328.
ISO–IDF (International Organization for Standardization–International Dairy
Federation) 2002. Milk fat – preparation of fatty acid methyl esters. ISO 15884-IDF
184. International Dairy Federation, Brussels, Belgium.
Maurice-Van Eijndhoven MHT, Hiemstra SJ and Calus MPL 2011. Short
communication: milk fat composition of 4 cattle breeds in the Netherlands.
Journal of Dairy Science 94, 1021–1025.
Palmquist DL 2006. Milk fat: origin of fatty acids and influence of nutritional
factors thereon. Springer, New York, NY, USA.
Palmquist DL, Stelwagen K and Robinson PH 2006. Modifying milk composition
to increase use of dairy products in healthy diets – preface. Animal Feed Science
and Technology 131, 149–153.
Rutten MJM, Bovenhuis H and van Arendonk JAM 2010. The effect of the
number of observations used for Fourier transform infrared model calibration for
bovine milk fat composition on the estimated genetic parameters of the
predicted data. Journal of Dairy Science 93, 4872–4882.
Rutten MJM, Bovenhuis H, Hettinga KA, van Valenberg HJF and van Arendonk
JAM 2009. Predicting bovine milk fat composition using infrared spectroscopy
based on milk samples collected in winter and summer. Journal of Dairy Science
92, 6202–6209.
Smith LM 1961. Quantitative fatty acid analysis of milk fat by gas–liquid
chromatography. Journal of Dairy Science 44, 607–622.
Soyeurt H and Gengler N 2008. Genetic variability of fatty acids in bovine milk.
Biotechnologie, Agronomie, Socie
´te
´et Environnement 12, 203–210.
Soyeurt H, Dehareng F, Gengler N, McParland S, Wall E, Berry DP, Coffey M and
Dardenne P 2011. Mid-infrared prediction of bovine milk fatty acids across
multiple breeds, production systems, and countries. Journal of Dairy Science 94,
1657–1667.
Soyeurt H, Dardenne P, Dehareng F, Lognay G, Veselko D, Marlier M,
Bertozzi C, Mayeres P and Gengler N 2006. Estimating fatty acid content in
cow milk using mid-infrared spectrometry. Journal of Dairy Science 89,
3690–3695.
Stoop WM, van Arendonk JAM, Heck JML, van Valenberg HJF and Bovenhuis H
2008. Genetic parameters for major milk fatty acids and milk production traits of
Dutch Holstein–Friesians. Journal of Dairy Science 91, 385–394.
Williams PC and Sobering DC 1993. Comparison of commercial near infrared
transmittance and reflectance instruments for analysis of whole grains and
seeds. Journal of Near Infrared Spectroscopy 1, 25–32.
Wilson RH and Tapp HS 1999. Mid-infrared spectroscopy for food analysis:
recent new applications and relevant developments in sample presentation
methods. TrAC Trends in Analytical Chemistry 18, 85–93.
Maurice-Van Eijndhoven, Soyeurt, Dehareng and Calus
354