www.umb.no 1

(1)

UNIVERSITETET FOR MILJØ- OG BIOVITSKAP

www.umb.no

MILK QUALITY BREEDING VALUE PREDICTION BASED ON FTIR SPECTRA. IGA 2012

1

(2)

Nævdal

(3)

MILK QUALITY

BREEDING VALUE PREDICTION BASED ON

FTIR SPECTRA

Tormod ÅDNØY, Theo ME MEUWISSEN, Binyamin DAGNACHEW Department of Animal and Aquaculture Science (IHA),

University of Life Sciences UMB, Ås, Norvège

(4)

IVERSITETET FOR MILJØ- OG BIOVITSKAP

MILK QUALITY BREEDING VALUE PREDICTION BASED ON FTIR SPECTRA

Tormod ÅDNØY, Theo ME MEUWISSEN, Binyamin DAGNACHEW

 Today:

– fat%, protein% .. in milk samples found by machine prediction from infrared light (FTIR)

– variance structure for fat%, protein%, .. estimated using relationship of animals (A)

– blup breeding values calculated from fat%, protein%, ..

 Propose:

– covariance of FTIR spectra estimated using relationship (A)

(Needs dimension reduction 500-1000 wavelengths –> 8 principal components)

– blup breeding values calculated for FTIR wavelengths

– calculate breeding values for fat%, protein% .. from heritable part of wavelengths (blup)

– more quality measures from FTIR can be included in breeding

K QUALITY BREEDING VALUE PREDICTION BASED ON FTIR SPECTRA. IGA 2012

(5)

www.umb.no

Milk quality

Indirect prediction (IP) Direct prediction (DP)

5

FTIR phenotype

REML,BLUP

FTIR phenotype

FTIR genetic (blup)

FTIR

environmental PCA,REML,BLUP y ˆ

i

u~i

e

i

^u ^~

ⁱ^*

^e

ⁱ^*

PLSi

β^ˆ

PLSi

β^ˆ β^ˆ^PLSⁱ

blup fat%, .. blup fat%, ..

other milk quality

fat%, ..

(6)

MOTIVATION

 We have found genetic variability of goat milk FTIR spectra (Dagnachew & Ådnøy, 2011)

similar to

 Genetic variability of cow milk based on MIR spectra

(Soyeurt et al., 2010)

– Indicated substantial amount of genetic variation – Show some regions are more heritable than others

 Quality in milk is found from FTIR spectra – usually no other info used

 ….Why not use the genetic part of the spectra to predict the genetic part of the traits….?

(7)

www.umb.no

OH group of lactose

Fat A

Carbonyl (C=O)

Amide III (protein)

OPO asym. Stretch (mixture) CO stretch (mixture)

Fat B

C –H of milk fat

Protein (amide II)

HERITABILITIES OF FTIR SPECTRA

(8)

MATERIALS AND METHODS – SAMPLES AND GOATS

 Raw FTIR Spectra data (2007 and 2008)

– TINE (Norwegian dairies) has four D-labs performing routine FTIR analysis on milk samples

 Total of 28,000 milk sample spectra

– 14,869 goats

(Norwegian Dairy Goat Control)

– 271 farms

 Of the FTIR wavelengths: 321 selected, 739 removed

• for physicochemical reasons …

-1 0 1 2

3993.03 3649.668 3306.306 2962.944 2619.582 2276.22 1932.858 1589.496 1246.134

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 Variab les

(9)

www.umb.no

FAT, PROTEIN, LACTOSE% AND FTIR SPECTRA USED

 We used fat%, protein%, and lactose% from Dairy Goat Control

(also found from FTIR spectra, but prediction details unknown to us)

 20,000 FTIR spectra used to find PLS regression models to predict fat, protein and lactose as given above

 8,000 FTIR spectra used in 10 cross validations of

– Direct

^and

– Indirect

methods of predicting breeding values

9

(10)

REGRESSION FTIR-FAT%, .., PCA, VARIANCE COMPONENTS

20000 FTIR spectra used to

 find regression coefficients to predict future fat%, .. etc

 find principal components (PC) of spectra (by PCA)*

 and PC multivariate (co)variance structure (REML):

– additive genetic

– permanent environment – residual

 find variance components of fat%, .. etc (by back solution of principle component covariances using regression coefficients)

– to make comparison of blup values from the two methods on equal variance basis

(11)

www.umb.no

PRINCIPAL COMPONENT ANALYSIS OF FTIR

Available programs for variance component analysis (asreml, Wombat, DMU) will only accept up to 30-40 variables

– Goal: to extract a set of fewer components that explain as much variation as possible of the original variation.

– Result: 8 components explained >99%

– Y are the FTIR spectra for the N samples – observations

– T is values for the new components for the N samples – score matrix

– P tells connection between new components and the spectra – loading matrix – F – error term

‘pls package’ in R on correlation matrix of Y to find scores T considered as new traits

Genetic and Environment Info in goat milk FTIR spectra

N

321 Y N

8 T

321 P N

321 F

(12)

PRINCIPAL COMPONENTS (PC) OF FTIR

Principal components

% variance

explained Variance ratios of total variance Genetic Permanent

environment Residual

1 49.63 0.253 0.135 0.613

2 35.74 0.382 0.214 0.404

3 5.95 0.330 0.199 0.471

4 4.74 0.257 0.227 0.516

5 1.63 0.371 0.197 0.432

6 0.56 0.303 0.163 0.533

7 0.47 0.198 0.192 0.610

8 0.33 0.205 0.161 0.634

Total variance

explained 99.05 %

(13)

www.umb.no

INDIRECT PREDICTION (IP)

– normal today

 predict fat% (using found regression coefficients)

 find blup breeding values for fat% using model:

fat%=Xb+Zu+Qp+e

Xb – fixed effects: herd-testday, kidding season

Zu – random animal breeding value with var(u)=A*σ²_A

Qp – random animal permanent environment var(h)=I*σ²_H e – random residual var(e)=I*σ²_E

(Variance components σ²_A , σ²_H , σ²_E from 20000 spectra.) Blup breeding values for fat% predicted as û

Blup breeding values for protein% and lactose% predicted same way (univariate)

13

(14)

DIRECT PREDICTION (DP)

– proposed new method

 find blup breeding values for principal components of FTIR using multitrait model:

principal components = Xb+Zu+Qp+e Xb – fixed effects: herd-testday, kidding season

Zu – random animal breeding value with var(u)=A*σ²_A

Qp – random animal permanent environment var(h)=I*σ²_H e – random residual var(e)=I*σ²_E

Basis for variance components σ²_A , σ²_H , σ²_E is 20000 spectra, same as for IP, but multivariate version

 Blup breeding values for fat% found from blup values for

principal components using established regression coefficients Blup breeding values for protein% and lactose% predicted same

(15)

www.umb.no

COMPARISON OF IP AND DP – RESULTS

 So now we have blup breeding values for fat%, protein% and lactose% calculated by Indirect Prediction (IP) and Direct

Prediction (DP)

– from same 6000 FTIR spectra

– based on same variance components

Direct Prediction is better than Indirect Prediction of breeding values for milk content when based on FTIR

15

Indirect prediction (IP) Direct prediction (DP) Fat Lactose Protein Fat Lactose Protein

Mean_blupvalues 0.0128 -0.0027 -0.0001 0.0098 -0.0022 -0.0020 STD_blupvalues 0.2925 0.0844 0.0980 0.2923 0.0871 0.1025 Mean PEV 0.0882 0.0062 0.0077 0.0853 0.0059 0.0072

Reduction in mean PEV 3.73% 4.07% 7.06%

Mean accuracy 0.627 0.649 0.647 0.645 0.667 0.679

Relative genetic gain 2.99% 2.77% 4.85%

(16)

DISCUSSION

– Not unexpected that univariate analyses are inferior to mulitvariate

• Multivariate analysis may be better because

correlated traits support each other in predictions – The 8 principal components contain more information

than fat%, .. alone, and this carries over to the genetic prediction

– A trivariate analysis of fat%, protein% and lactose%

gives better results than univariate, but not as good as with Direct Prediction.

(17)

www.umb.no

DISCUSSION

– Finding principle components and particularly estimating their genetic variance is time consuming – but need only be done seldom

• Breeding values for principle components are quick to find

• And so are breeding values of derived traits when regressions are established in calibration

– Breeding values for other traits with genetic information in the FTIR spectra may be derived without estimating variance components for the new traits. Only the

phenotypic regression relation is needed

• Coagulation?

• Fatty acids?

• …

17

(18)

DISCUSSION

 Possibility of using spectra also to detect unwanted changes in environment? – Using deviations from predicted environment.

(19)

www.umb.no

We invite you to:

19

Regional IGA conference:

Goat Milk Quality

Tromsø, Norway

4 - 6 June 2013

(20)

Milk quality

Indirect prediction (IP) Direct prediction (DP)

FTIR phenotype

REML,BLUP

FTIR phenotype

FTIR genetic (blup)

FTIR

environmental PCA,REML,BLUP y ˆ

i

u~i

e

i

^u ^~

ⁱ^*

^e

ⁱ^*

PLSi

β^ˆ

PLSi

β^ˆ β^ˆ^PLSⁱ

blup fat%, .. blup fat%, ..

other milk quality

www.umb.no 1

MILK QUALITY

BREEDING VALUE PREDICTION BASED ON

FTIR SPECTRA

MILK QUALITY BREEDING VALUE PREDICTION BASED ON FTIR SPECTRA

Milk quality

Indirect prediction (IP) Direct prediction (DP)

FTIR phenotype

REML,BLUP

FTIR phenotype

FTIR genetic (blup)

FTIR

environmental PCA,REML,BLUP y ˆ

e

u ~

e

blup fat%, .. blup fat%, ..

fat%, ..

MOTIVATION

HERITABILITIES OF FTIR SPECTRA

MATERIALS AND METHODS – SAMPLES AND GOATS

– 14,869 goats

– 271 farms

FAT, PROTEIN, LACTOSE% AND FTIR SPECTRA USED

– Direct

– Indirect

REGRESSION FTIR-FAT%, .., PCA, VARIANCE COMPONENTS

PRINCIPAL COMPONENT ANALYSIS OF FTIR

N

321

Y N

8 T

321

P N

321

F

PRINCIPAL COMPONENTS (PC) OF FTIR

INDIRECT PREDICTION (IP)

DIRECT PREDICTION (DP)

COMPARISON OF IP AND DP – RESULTS

DISCUSSION

DISCUSSION

DISCUSSION

We invite you to:

Regional IGA conference:

Goat Milk Quality

Tromsø, Norway

4 - 6 June 2013

Milk quality

Indirect prediction (IP) Direct prediction (DP)

FTIR phenotype

REML,BLUP

FTIR phenotype

FTIR genetic (blup)

FTIR

environmental PCA,REML,BLUP y ˆ

e

u ~

e

blup fat%, .. blup fat%, ..

fat%, ..

^u ^~

^e

^u ^~

^e