MPRA-BPNet

---

Step 0:

config.py: contains all the models. Needed by almost all the scripts
create_one_hot_seq.py: create one hot encoded sequences for all the following analyses

---

Step 1: Train the new top layer (regression) and get the prediction.

The output file is named "cv_free_MPRA.csv" bacause the prediction was not made using cross_val_predict but predicted using the entire "bpnet_bottleneck_feat" as input and the top layer as the model.
Snakefile_new_top_layer (requires util_cross_val.py)

---

Step 2: Reconstruct the model by combining the top layer with the bottleneck model.

Snakefile_reconstruction (requires util_cross_val.py and reconstruction_util.py)

---

Step 3:

plot.py (to get spearman's correlation)

---

Step 4: Run DeepLIFT and tfmodisco on the reconstructed models

Snakefile_modisco: I made some modifications to plot_weights() in modisco/visualization/viz_sequence.py. Specifically, I added a name parameter, which is the name of the image generated, and plt.show() was changed to plt.savefig(name) to save the image.
modisco_pattern_generator.py