Resources:

• README.md: this file.
• data1.7z and data2.7z: zipped data files.

source codes:

• train_all.py: running anomaly detection using all normal data for training.
• train_10_to_90.py: running anomaly detection using 10,20,..,90 percent of normal data for training.

folder data/:

• data files extracted from data1.7z and data2.7z

folder figs/:

• visualization of normal versus tumor data points detected by the proposed method.

folder results/:

Performance of the proposed method on different cancers, using either all or parts of normal data for training. Measured in F1, Precision, Recall, Specificity, Accuracy, and AUC.

Step-by-step running:

0. Installing Python libaries needed

• Install sklearn: pip install scikit-learn

1. Using all normal data for training

Running

python train_all.py

This returns results/result.csv, containing the performance of the proposed model on different cancers, using all normal data for training. This code also returns a visualization of normal versus tumor data points detected by the proposed method.

2. Using parts of normal data for training

Running

python train_10_to_90.py

This returns results/result_x0.csv, where x ranges from 1 to 9, containing the performance of the proposed model on different cancers, using 10,20,..,90 percent of normal data for training.

Cite:

If you use this code in your work please cite our paper as follows:

@article{quinn2018cancer,
  title={Cancer as a tissue anomaly: classifying tumor transcriptomes based only on healthy data},
  author={Quinn, Thomas and Nguyen, Thin and Lee, Samuel and Venkatesh, Svetha},
  journal={Frontiers in Genetics},
  year={2019}
}