A reusable machine learning pipeline that can be used for binary, multiclass or multilabel classification. Current output is trained on genetic data where the features are genomic annotations and the target variable is the presence of a particular trait (T2D, Lipids, CAD, BMI, etc).
Clone repository or download .yml file. Ensure anaconda is installed. To install an anaconda environment from a .yml through the command line:
Create environment:
conda env create -f t2d_model.yml
Verify all packages were installed with:
conda list
Activate anaconda environment:
source activate t2d_model
Once the environment is activated, you can use the jupyter notebooks to run through the workflow. Change directory into src/notebooks/ and execute jupyter notebook from the command line.
Note: If you don't want to use anaconda, you will need to ensure that all packages in requirements.txt are installed in your virtual environment or globally if you are not using a virtual environment.
-
Feature Selection: Configure filepath and then use to create and save feature-selected datasets. Currently configured to store files in an S3 bucket, change this to store to your bucket or a local drive by editing the outpath and replacing
write_to_S3functions withwrite_to_local. -
Clustering: Exploratory clustering of data to detect patterns prior to classification
-
One Vs Rest Model Evaluation: Runs multiple sklearn models on the training data and compares their performance on test data. Configure filepaths for input data before running.
-
Multilabel Model Evaluation: Similar to One Vs Rest Notebook except for multilabel problems.
-
Trait vs Trait: Used to compare two distinct sets of traits (e.g, blood traits vs metabolic traits) in order to determine whether a machine learning classfier can learn to distinguish between the two sets, provided an aggregated dataset.
Evaluator.py is a helper class for plotting and evaluating model performance. Documentation for methods is contained in the python file.