Prior Knowledge Dual-Path CNN

Research from:

State Key Laboratory of Tree Genetics and Breeding, Chinese Academy of Forestry, Beijing 100091, China
Research Institute of Subtropical Forestry, Chinese Academy of Forestry, Hangzhou 311400, China
State Key Laboratory of Tree Genetics and Breeding, Nanjing Forestry University, Nanjing 210037, China

---

PKDP (Prior Knowledge Dual-Path CNN) is a dual-path convolutional neural network framework designed to enhance genomic selection (GS) by integrating genome-wide association study (GWAS) results with genome-wide minor-effect markers.

---

Installation

# Create a new conda environment
conda create -n PKDP_env python=3.8

# Activate the environment (works on all platforms)
conda activate PKDP_env

# Install PKDP
git clone https://github.com/aiPGAB/PKDP-GS.git
cd ./PKDP-GS
chmod +x ./PKDP.py

# Install dependencies
pip install -r requirements.txt

Requirement

• Python >= 3.8
• PyTorch >= 1.10.0
• NumPy >= 1.20.0
• Pandas >= 1.3.0
• Matplotlib >= 3.4.0
• Seaborn >= 0.11.0
• Scikit-learn >= 1.0.0
• SciPy >= 1.7.0
• Optuna >= 2.10.0

Options and usage

Training

python ./PKDP.py train -h

Required Parameters

Parameter	Description
--train_phe	Path to the training phenotype file
--geno	Path to the genotype file
--output_path	Directory to save outputs

Optional Parameters

Parameter	Description	Default Value
--pnum	Phenotype column index or name	First column
--prefix	Prefix for output files	Timestamp
--batch_size	Batch size for training	32
--epochs	Number of training epochs	100
--optuna_trials	Number of Optuna trials for hyperparameter tuning	50
--device	Device to use (cuda or cpu)	cuda:0
--optimizer	Optimizer type (Adam, SGD, AdamW)	Adam
--early_stop	Enable early stopping	False
--prior_features	Prior knowledge features (space-separated IDs)	None
--prior_features_file	Path to a file with one prior feature ID per line	None
--adjust_encoding	Adjust genotype encoding from {0,1,2} to {-1,0,1}	False

Usage

python ./PKDP.py train \
               --train_phe demo/train_phe.csv \
               --geno demo/train_geno.csv \
               --output_path results/ \
               --prior_features_file ./demo/prior_features.txt

Notes

In the train model, the training set will automatically calculate hyperparameters through cross-validation. After obtaining the best hyperparameters, the model will be trained using the entire training set. This mode outputs the trained model file, which can be used for prediction in the predict model. For genotypes with values 0/1/2, it is recommended to set --adjust_encoding.

Prediction

python ./PKDP.py predict -h

Required Parameters

Parameter	Description
--geno	Path to the genotype file
--model_path	Path to the trained model file
--output_path	Directory to save outputs

Optional Parameters

Parameter	Description	Default Value
--test_phe	Path to the testing phenotype file (optional)	None
--pnum	Phenotype column name	First column
--prefix	Prefix for output files	Timestamp
--device	Device to use (cuda or cpu)	cuda:0
--adjust_encoding	Adjust genotype encoding from {0,1,2} to {-1,0,1}	False
--prior_features	Prior knowledge features (space-separated IDs)	None
--prior_features_file	Path to a file with one prior feature ID per line	None

Usage

python ./PKDP.py predict \
               --geno demo/test_geno.csv \
               --test_phe demo/test_phe.csv \
               --prior_features_file ./demo/prior_features.txt \
               --model_path results/best_model.pth --output_path predictions/

Notes

• When no prior features are provided, the model will only utilize the main convolutional path.
• It is recommended to use the --prior_features_file parameter instead of the --prior_features parameter to specify prior features.
• The --pnum parameter can be used to specify the phenotype column to predict.
• During model training, samples with NA values in the phenotype will be automatically ignored, so there is no need to manually remove samples with NA values.
• The order of SNPs in the prediction should match the order used during training.
• Input phenotype data format: see ./demo/demo_phenotypes.csv
• Input genotype data format: see ./demo/demo_genotypes.csv
• During model training, please pay close attention to adjusting the following hyperparameters, as they significantly impact model performance:
  • --main_channels: Number of channels in the main network.
  • --prior_channels: Number of channels in the prior knowledge network.
  • --learning_rate: Learning rate.
  • --batch_size: Batch size.

Version

v0.0.1

• Initial release of PKDP.
• Added dual-path CNN architecture.
• Integrated support for prior knowledge features.
• Implemented training with cross-validation and hyperparameter optimization.
• Added visualization tools for training progress and predictions.

v0.0.8

• In the previous version, the optional parameter test_phe was only used for evaluating the predictive capability after model training. It has been removed in the new version.

Citation

Han F, Gao M, Zhao Y, Bi C, et al. Improving genomic selection accuracy using a dual-path convolutional neural network framework: a terpenoid case study. Unpublished.