GpNet: Locally connected Network for Genomic prediction

Model summary

We propose a GpNet, a deep learning network for genomic prediction of beef cattle. GpNet can estimate genetic effects of each LD-block consisting of locally adjacent two or more single nucleotide polymorphisms (SNP) by locally connected layer.

1. Locally connected layer

Locally connected layer is inspired by causal convolution (Oord, et al., 2016) and local convolution (Taigman, et al., 2014). By using locally connected layer, the network cannot violate the order of SNP: the SNP at the n-position cannot depend on any of SNPs to the 5'-end (n+1,n+2,...,E).

2. Network Structure

GpNet consists of the stacks of locally connected layer. Both skip connection (He, et al., 2016) and relu activation (Nair and Hinton, 2010) are used throughout the network to enable training of much deeper model. GpNet can be scaled by the different layer depth d and stack number s.

our models were implemented by tensorflow 2.3 and keras

Implementation

1. Preparing

If you want to use the pre-trained weight, the weight file must exist as weights.h5 in the path where main.py is located. The weights file attached at this repository is for GpNet (d: 4, s: 1) and for predicting GEBV of carcass-weights, trained by the 8000 Korean native cattle.

1.1. Configuration

You can set the configuration of GpNet at main.py.

'batch_size' : batch size,
'epochs' : train epochs
'lr' : train learning rate
'depth' : GpNet #N of locally connected layer depth
'stack' : GpNet #N of locally connected layer stacks
'gv' : genetic variance of phenotype
'rv' : residual variance of phenotype
'device' : GPU number to use
'data_load' : A or G (A : array , G : generator) #If you use more than 10000 individuals, we recomand G for memory saving.

1.2. Input Data format

Our code requires the .raw data format of plink. See https://www.cog-genomics.org/plink2/formats#raw for more details.

Caution: Our code does not allow for variant missing

2. Execution

2.1. Train GpNet

python main.py --raw [raw] --out [out] --mode train

1. [raw] : Path of input raw data.
2. [out] : Directory name for saving the weights, history and prediction results.

example: python main.py --raw data/sample.raw --out results --mode train

2.2. Test GpNet

python main.py --raw [raw] --out [out] --mode test

1. [raw] : Path of input raw data.
2. [out] : Directory name for saving the prediction results.

example: python main.py --raw data/sample.raw --out results --mode test

2.3. Genomic prediction with GpNet

1. Set configuration of GpNet at main.py.
2. Prepare train and test sets as separate raw files.
3. Train GpNet with train set: python main.py --raw [TRAIN RAW] --out [OUT] --mode train.
4. Move weights.h5 file from the [OUT] to the path where main.py is located.
5. Doing genomic prediction with pre-trained weights: python main.py --raw [TEST RAW] --out [OUT2] --mode test.

Reference

• Oord, A.v.d., et al. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499 2016.
• Taigman, Y., et al. Deepface: Closing the gap to human-level performance in face verification. In, Proceedings of the IEEE conference on computer vision and pattern recognition. 2014. p. 1701-1708.
• He, K., et al. Deep residual learning for image recognition. In, Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. p. 770-778.
• Nair, V. and Hinton, G.E. Rectified linear units improve restricted boltzmann machines. In, Icml. 2010.