DRANetSplicer

1. Environment setup

All experiments were implemented using TensorFlow and Keras on the Linux system with a single NVIDIA GA102 [GeForce RTX 3090 Ti] 24GB GPU, and if you would like to reproduce our models, we recommend that you use Anaconda to build a python virtual environment.

1.1 Create and activate a new virtual environment

conda create -n DRANetSplicer python=3.8
conda activate DRANetSplicer

1.2 Clone the project and install requirements

git clone https://github.com/XueyanLiu-creator/DRANetSplicer
cd DRANetSplicer
python3 -m pip install --editable .

1.3 View the help file for running the program.

python run_code.py -h

2. Re-train (Skip this section if you just want to use trained models)

2.1 Data processing

Please see the sample data at /data/dna_sequences/. If you are trying to re-train DRANetSplicer with your own data, please process you data into the same format as it. Our program automatically performs one-hot encoding of the data under /data/dna_sequences/ and saves it in data/encode_datasets/.

2.2 Model Training

python run_code.py \
    --organism_name oryza \
    --splice_site_type donor \
    --train \
    --num_train_epochs 20 \
    --batch_size 64 \
    --learning_rate 0.01 \
    --verbose 1 \
    --report

3. Prediction

You can run the following code to get predictions using the model we provide or a model you have re-trained. If you want to use your re-trained models to predict the data, remove the --use_our_trained_models option.

python run_code.py \
    --organism_name oryza \
    --splice_site_type donor \
    --use_our_trained_models DRANetSplicer1 \
    --test \
    --batch_size 64 \
    --verbose 1 \
    --report

With the above command, the trained or re-trained DRANetSplicer model will be loaded from models/trained_models/ , and makes prediction on the *_test.npz file that saved in data/encode_datasets/ and save the prediction result at resluts/.

4. Visualization

Visualiazation of DRANetSplicer consists of 2 steps. Calculate contribution scores and Plot. Calculate with only one model (For example, DRANetSplicer1). If you want to use your re-trained models for visualization, remove the --use_our_trained_models option. The method of visualization can be either deep_shap or grad_cam.

python run_code.py \
    --organism_name oryza \
    --splice_site_type donor \
    --use_our_trained_models DRANetSplicer1 \
    --visualization deep_shap

With the above command, the trained or re-trained DRANetSplicer model will be loaded from models/trained_models/ , and calculates contribution scores on the *_test.npz file that saved in data/encode_datasets/ and save the result at visualization/figure/.