N-PIP

N-PIP (Nannochloropsis - Promoter Interpretable Prediction) : a novel interpretable deep learning framework for promoter identification in Nannochloropsis

Required Packages

• python 3.6
• tensorflow-gpu 1.10.0
• keras 2.2.0
• numpy 1.14.5
• matplotlib 3.2.2
• scikit-learn 0.23.2
• scipy 1.5.4

Scrambler code reference:

Linder, Johannes, et al. "Interpreting neural networks for biological sequences by learning stochastic masks." Nature machine intelligence 4.1 (2022): 41-54.

http://www.github.com/johli/scrambler

Run N-PIP with trained parameter

To start making predictions, you can place your fasta file in the promoter_data folder, noting that all sequences must be 1,000 kb in length.

python run_npip.py strain_name fasta_file save_figs

Command-line arguments:

• strain_name can be set to CCMP525, CCMP526, CCMP529, CCMP531, CCMP537, IMET1, determine the trained parameters of which strain to use.

• fasta_file can be set to your fasta file name, such as CCMP525.fa.

• save_figs can be set to True or False, noting that if the number of samples is large, it may take a long time to save figures.

Output files will be generated in the npip_output folder:

• label.npy (predictive confidence)
• importance_scores.npy
• importance_pwm.npy
• figures (optional)

Retrain N-PIP

If you want to repeat the training steps in the paper, or retrain with your own dataset, please follow these steps:

1. Train the predictor - ADNPPro (this step will run five-fold cross-validation).

   python run_adnnpro_5fold.py pos_data_path neg_data_path n_5fold save_name

   Command-line arguments:

   • pos_data_path can be set to your positive dataset path, such as promoter_data\CCMP525.fa.
   • neg_data_path can be set to your negative dataset path, such as promoter_data\randomCCMP525.fa.
   • n_5fold determine the repeat number of five-fold cross-validation.
   • save_name determine the h5 file name you want to set. Models and validation results will be saved in the adnppro_trained_model folder.

2. Train the interpreter - Scrambler.

   python run_scrambler_train.py pos_data_path load_name save_name

   Command-line arguments:

   • pos_data_path can be set to your positive dataset path, such as promoter_data\CCMP525.fa. In our work, only positive dataset are used to train the interpreter.
   • load_name should be set same to the save_name in the previous step.
   • save_name also determine the h5 file name you want to set. Models will be saved in the scrambler_trained_model folder.