MIC prediction and feature selection

MIC Prediction and FeatureSelection using XGBoost for Neural Networks

Data

contig fasta files were collected from Klebsila Pneumoniae (KPN).

Algorithms

There are two algorithms that will be used for prediction:

1. Neural Network
2. Gradient Boosted Forests (More specifically, XGBoost)

Testing your own data

If you would like to test your own data with our best NN model, follow these steps. You should be in the repository's main directory to call step 2.

1. Collect Fastas and MIC CSV file (put them in data folder)
2. Run python script to run test pipeline

Because testing the NN model does not require changing any of the files, the base file for script was made into the test pipeline. This pipeline will run the same things that the following 4 lines would run, but automatically.

1. Run python get_format_data.py test
2. Run python nn.py files to generate NN test file.
3. Run python nn.py test to test NN model
4. Run python nn_analysis.py to generate error rate CSV and MIC heatmap.

Steps to reproduce

Note that all python run commands below must be done within the script directory.

1. Collect Fastas and MIC CSV file (put them in data folder)
2. Run python get_format_data.py
3. Zip up train.libsvm and put that zip file in an S3 bucket
4. Put xgboost.ipynb in Sagemaker, modify to point to train.zip in S3, and run all cells
5. Collect trained XGBoost model and control XGBoost model and put them in output/models
6. Run python xgboost_test.py to get feature_importance.csv and test results for XGBoost models
7. Use feature_importance.csv to modify nn.py with most important features (all features with importance >=10 in sorted order)
8. Run python nn.py files to generate NN training, validation, and testing files. Take note of the console output to update the nn.py file.
9. Run python nn.py cv to do grid search using 10-fold cross validation to determine best learning rate, architecture combination. Update BEST_LEARNING_RATE, BEST_ARCH_SIZE, and BEST_DROPOUT_PROB at top of nn.py when finished.
10. Run python nn.py train and note the file name of the best model hdf5 file. Modify nn.py with that name (delete all other hdf5 files)
11. Run python nn.py test to test NN model
12. Run python nn_analysis.py to generate error rate CSV and MIC heatmap.

For more information/detail on each step, visit the respective folders/script files.