TrG2P: A transfer learning based approach to improve genomic prediction

TrG2P is a novel strategy to improve genomic prediction (GP) performance based on transfer learning. TrG2P aims to get knowledge from the source task's learning system to improve the target GP task. Two main functions were designed in this approach, including "train.py" and "predict.py", respectively. Three processes were applied by "train.py" function, including pretraining, fine-tuning and building fusion model. The "predict.py" function is used to predict genomic estimated breeding values (GEBVs) for new accessions.

---

Environment

This package was builded with the followed packages:

python 3.9
tensorflow 2.11.0
keras 2.11.0
scikit-learn 1.0.2
pandas 1.5.3
numpy  1.24.2

Data format

genotype

• The genotype file need to contain all markers in a "*.txt" file, including m accessions and n markers.

	markers
accessions	1	0	0	-1
	0	1	1	0
	-1	1	0	-1

phenotype

• The phenotype file need to contain mutiple traits in a "*.csv" file, and the header is the trait's name.

	traits
accessions	GHID	PH	PnN	LA
	A1257	118.8	12.9	21.1
	A1258	125	11.8	27.6
	A1302	117.7	12.5	26

Parameter list

train.py

Parameter	Type	Default	Description
seed	int	42	Value of random and cv seeds
output	directory	None	Override automatic results directory name
cv_folds	int	5	Number of folds to use in cross-validation
geno	directory	./data/geno.txt	Training genotype *.txt file (m x n matrix)
pheno	directory	./data/pheno.csv	Training phenotype *.csv file (m x i vector)
st	character, n=2	None	Source domain traits, n=2
tt	character, n=1	None	Target domain trait, n=1
channel	int	1	The number of out channels in convolutional layer
kernel	int	10	The kernel size in convolutional layer
stride	int	10	The stride in convolutional layer
unit	int	64	The units in fully connected layers
dropout	float	0.3	The dropout in fully connected layers
pretrain_epoch	int list	40	Pretraining epoch
pretrain_batch	int list	40	Pretraining batch size
pretrain_lr	float list	0.001	Pretraining learning rate
ft_epoch	int list	40	Fine-turning epoch
ft_batch	int list	40	Fine-turning batch size
ft_lr	float list	0.001	Fine-turning learning rate
fu_epoch	int list	40	Fusion epoch
fu_batch	int list	40	Fusion batch size
fu_lr	float list	0.001	Fusion learning rate
direct_epoch	int list	40	Direct prediction epoch
direct_batch	int list	40	Direct prediction batch size
direct_lr	float list	0.001	Direct prediction learning rate

predict.py

Parameter	Type	Default	Description
output	directory	None	Override automatic results directory name
geno	directory	./data/geno.txt	Training genotype *.txt file (m x n matrix)
model	directory	None	Model *.h5 file

Demo

Modeling

python train.py -s 42 --output results --cv_folds 5 --geno ./data/rice299_geno2.txt --pheno ./data/rice299_phe.csv --st YPP GW --tt YLD --pretrain_epoch 60 --pretrain_batch 20  --pretrain_lr 0.001 --ft_epoch  60 --ft_batch 20 --ft_lr 0.001 --fu_epoch  60 --fu_batch 20  --fu_lr  0.001 --direct_epoch 40 --direct_batch 40 --direct_lr 0.001
## epoch, batch, lr can be a list

11 files will be outputed by train.py program, including:

• direct_*.h5, Direct prediction model for the target trait based on CNN.
• direct_res.csv, Direct predictive ability for the target trait based on CNN.
• pretrain_*.h5, Pretraining model for the source traits based on CNN, 2 models will be outputed.
• pretaining_res.csv, Pretraining predictive ability for the source trait based on CNN.
• fine_tuned_model_*.h5, Fine-tuned model for the target traits based on transfer learning, 2 models will be outputed.
• fine_tuning_res.csv, Fine-tuning predictive ability for the target trait based on transfer learning.
• fusion_model_+.h5, Fusion model for the target traits based on transfer learning.
• fusion_res.csv, Fusion model's predictive ability for the target trait based on transfer learning.
• results_summary.csv, The summary of all results for the target trait.

Predicting with the trained model

python predict.py  --output results  --geno "./data/rice299_geno2.txt"  --model ./results/fine_tuned_model_PH.h5
##  for "--geno", a new genotype file should be used, here just an example.

As a result, 1 files will be outputed by predict.py program,

• prediction_res.csv, The predictive values with known genotype.