Juillermo / saliencyPSSP

Code for the paper "Saliency Maps on CNNs for Protein Secondary Structure Prediction"

This repository contains the code for the paper:

Saliency Maps on CNNs for Protein Secondary Structure Prediction

The convolutional neural network (CNN) for protein secondary structure prediction (PSSP) has been built and trained using the framework lasagne4bio developed by Vanessa Jurtz, and can be found in the secondary_proteins_prediction folder of this repository. My only additions to this folder are the CNN architectures pureConv.py, HpureConv.py and pssmConv.py, which are located at secondary_proteins_prediction/configurations. These architectures were employed for the experiments in our paper and they correspond to the main architecture (pureConv.py) and variations predicting H values using aminoacids from one side (left or right) of a position only (HpureConv.py), or with only pssm input (pssmConv.py). The last two modified architectures also alter the training-prediction-evaluation pipeline, so I also had to make modifications in the files run for training, making predicitions, and evaluating models; resulting in the new files Hpuretrain.py and pssmtrain.py for training the HpureConv and pssmConv respectively, the new files predict_side.py and pssmpredict.py for making predictions on them, and eval_avrg_side.py for evaluating predictions on the HpureConv model. Note that diagrams of the architecture of pureConv.py (in jpeg and dia formats) are also included in the secondary_proteins_prediction/configurations folder.

The code related to the creation of the saliency maps is located at the folder test_saliency. Some preliminary exploration of the code can be found at the jupyter notebook Exploring Saliencies.ipynb on the root directory. A short description of each file is included next:

• saliency.py: functions to produce the saliency maps (this is done in batches)
• saliency_aggregation.py: functions to aggregate the saliency maps along different dimensions
• reparer.py: scans saliency maps already generated or repairs the ones that had problems
• results.py: generates figures, statistics, and processed data
• results_paper.py: similar to results.py, but specific to the figures included in the paper
• utils.py: various basic utility functions and data

The protein dataset can be downloaded from here.

Training models and evaluating them

In the secondary_proteins_prediction/ folder:

1. Train model

   python train.py pureConv

   (or alternatively Hpuretrain.py HpureConv / pssmtrain.py pssmConv)

2. After which use debug_metadata.py to find the best training epochs, found by validation performance

   python debug_metadata.py <]topX> <metadata_path>

3. Use predict.py to gather predictions

   python predict.py <metadata_path> [subset=test]

   (or predict_side.py / pssmpredict.py)

4. Use eval_avrg.py to evaluate the model and combining several predictions.

   python eval_avrg.py <predictions_path> [subset=test]

   (or eval_avrg_side.py for HpureConv models)

Computing saliencies from a trained model

1. Compute saliencies via saliency.py*.

2. Exploring saliencies.ipnb is a good way to get familiarised with other functions. Some of the functions found there may need data files containing different aggregations of the saliencies. These can be produced via saliency_aggregation.py.

3. Figures from the paper can be reproduced with the functions from results_paper.py, while other results not included in the paper can be found in results.py.

*Note that Computing saliencies requires a big amount of computational power, so this process may take very long times.

---

Installation (from the README of the lasagne4bio framework)

Please refer to lasagne's for installation and setup of GPU environment on an Ubuntu 14.04 machine.

Dependencies

All code was written in python programming language version 2.7. Neural networks are implemented using the lasagne library, please find installation instructions here: https://lasagne.readthedocs.io/en/latest/user/installation.html.

The libraries used in this code are:

• argparse
• cPickle
• csv
• datetime
• gc
• glob
• gzip
• importlib
• itertools
• lasagne
• math
• matplotlib
• numpy
• operator
• os
• platform
• scipy
• sklearn
• string
• subprocess
• sys
• theano
• time