Improving Prediction Accuracy of Meat

Tenderness in Nelore Cattle using

Artificial Neural Networks

F.B. Lopes1,2, C.U. Magnabosco1, T.L. Passafaro3, L.F.M. Mota2,3, M.G. Narciso4, G.J.M.

Rosa3, F. Baldi2

1Embrapa Cerrados, BR-020, 18, Brasília, DF, Brazil2São Paulo State University - Júlio de Mesquita Filho (UNESP), Department of Animal Science, Jaboticabal, SP, Brazil3University of Wisconsin-Madison, Department of Animal Sciences, WI, USA4Embrapa Rice and Beans, Santo Antônio de Goiás, GO, Brazil

Dubrovnik, Croatia, 2018

AFC, AP, 3P, STAY and SP

RFI, FC, FE

Rib eye area, rib fat thickness

Pre and post

weaning weight

W120

and

W210

FERTILITY

MATERNAL

ABILITY

GROWTHCARCASS

QUALITY

FEED

INTAKE

Selection Criteria in Brazilian Beef Cattle2

++++= mpeZmZaZXy 321

++= aZXy 1

Fixed Effects

Maternal Effect

Maternal Permanent

Environment effect

Animal ModelBreeding Value

Additive Genetic

Effect

3

Animal ModelBreeding Value for Meat Tenderness

++= aZXy 1Additive

Genetic

Effect

Contemporary

GroupSex

4

Genomic Analyses

I can see

much

tenderness

in your

future…

5

Quality Control

✓ Redundant position

✓ X, Y and MT Chromosome

✓ Minor Allele Frequency < 5%

✓ Deviation from HWE (p<10-6)

✓ Linkage Disequilibrium > 0.8

575 samples and 219,863 SNP

6

Training and Validation Population

7

2006 2014

Training Validation

510 samples and 219,863 SNP

2002 2010 201765 samples and 219,863 SNP

Bayesian Regression Models

Bayes A

( )( )

( )0,2~

,~

,0~|

22

22

22

−−

−

e

j

jjj

Sdf

Ng

Bayes B

( )

( )

( )0,2~

,~

0

,0~|

22

22

2

22

−

=

−

−

e

j

j

jjj

Sdf

Ng

Bayes Cπ

( )( )

( )

( )0,2~

,~

1,0~

,0~|

0

22

22

22

−

=

−

−

e

g

ggj

j

Sdf

Uniform

Ng

g

Bayesian Lasso

( )( )

( )0,2~

~

,0~|

22

2

22

−−

e

j

jjj

DE

Ng

Bayesian Ridge Regression

( )( )

( )0,2~

~

,0~|

22

2

22

−−

e

j

jjj

N

Ng

8

▪ Genomic prediction

▪ Successful examples

Artificial Neural Network

9

• Delivery high prediction accuracy

• Can handle large number of variables

• Can capture non-linear relationshipbetween predictors and outcomevariable

• No assumption about the distributionof predictors and output variable

• Biological interpretation

• Overfitting

Artificial Neural Network

Input Layer

Hidden Layer

Output Layer

10

Deep Neural Network

SNP WBSF

Input Layer

Hidden Layer

Output Layer

Topology

✓ 1 to 4 hidden layers

➢ 10, 35, 75, 105 and 250 neurons

✓ Rectifier and Maxout activation

✓ Dropout (50%)

✓ Quadratic loss function

✓ 10K Epochs

✓ ADADELTA adaptive learning

rate algorithm

➢ Rho: 0.99

➢ Epsilon: 1e-08

✓ Stopping criterion: 1e-06

11

Genomic Prediction Ability

12

GEBVPedigree / Genomic

Information

MatingOptimum Contribution

Selection

Animal Breeding Framework13

Thank you!camult@gmail.com