Inferring structural and dynamical properties of gene networks from data with deep learning

We have annotated the code in detail to help the reader understand the code.

Take the MISA task as an example

There are 8 files under "MISA" file, including

5 python code files: data_generation.py, train.py, Edge_removal.py, predict_stable_state.py and monotonicity_f.py;

2 image files: The_role_X1.png and f1_f2_X2.png;

1 Parameters_saved.pickle, save the DNN parameters after training

ps: 1 data.pickle after we run data_generation.py, which has been divided into three parts, training set (80%), validation set (10%) and test set (10%);

We should first run data_generation.py to prepare the data for training, we can get a data.pickle after about 60 seconds.

Then we can run train.py for training, of note, we include "epochs" and "sub_epochs" in our parameters, each epoch contains "sub_epochs" training. We apply the DNN model to the validation set after each epoch, i.e., we perform a validation after training "sub_epochs" times (Default "sub_epochs" = 10).

After we run train.py, we can get the Parameters_saved.pickle.

Then, we can use this DNN model do everything we want, including inferring the structural of gene networks by Edge_removal.py (corresponding The_role_X1.png can be obtained), getting the monotonicity of the synthesis rate f with respect to the variable x by monotonicity_f.py (corresponding f1_f2_X2.png can be obtained), and predicting the steady states by predict_stable_state.py.

Other files "Four_dimensional" and "Oscillation" focus on the inference of gene regulatory networks. And in "Four_dimensional" file, We additionally test the performance of DNN on network inference when it has 3 hidden layers.