Genomic Eye Color Classification with Machine Learning

Kevin Fang, Curoverse 2017

Using machine learning and the Arvados Lightning project, we were able to predict eye color to 95% accuracy.

This project is reliant on the following Python libraries: scikit-learn, pandas, matplotlib, numpy, scipy. In addition, the tile searches cannot be run on non-UNIX machines as it requires the system grep and cat commands.

To download the NumPy arrays and assembly files needed for the project, set the Arvados API tokens and run ./download_dependencies.sh - this downloads the tiled data, names, information, and assembly files into the appropriate folders.

Running the classifier

First, clone the GitHub repository with git clone --recursive https://github.com/kevin-fang/lightning-eye-classifier. The --recursive is important as the tile-searching script is in a submodule.

There are three ways to run the classifier. A Dockerfile has been provided in docker/ as well as instructions for running through Docker.

Run the Python scripts by themselves:

1. Navigate to src/.
2. Generate the classifier with python generateLeftClassifier.py
3. Save the coefficients with python saveCoefs.py
4. Search for each tile in python tileSearch.py

Run the Arvados workflow in arvados_impl/

Run the interactive Python notebooks:

1. First, navigate to notebooks/
2. Open the IPython session using jupyter notebook
3. Open leftEyeClassifier.ipynb, saveCoefs.ipynb, and tileSearch.ipynb.
4. Run leftEyeClassifier.ipynb first, and set whether to exclude or include hazel. This will generate the classifier and save it in svc.pkl.
5. Then run saveCoefs.ipynb. This will open the svc.pkl classifier and will serialize the learned coefficients in coefs.pkl.
6. Finally, run tileSearch.ipynb. This will open coefs.pkl and search for each tile.

Details

The classifier is able to predict the blue eye color to approximately 95% accuracy when the hazel color is excluded. Otherwise, it is able to reach 88% accuracy.

The classifier is able to find that eye color is reliant on base pairs 28,264,893 to 28,265,118, which is consistent with the HERC2 gene, responsible for eye color (https://ghr.nlm.nih.gov/gene/HERC2#location, https://link.springer.com/article/10.1007%2Fs00439-007-0460-x)