Python 3.5 or above

Pytorch 1.0.1 or above

Scikit-learn 0.19.0

1. Download L1000 expression profiles (i.e., input features) from https://drive.google.com/file/d/1lBJdTRSHtw16FIN45mmoEW-dJp53D2O4/view?usp=sharing. Put it in folder L1000.

2. To perform cross-validation, run the following command: python train.py [cell_line] [test scenario] [semi-supervised loss weight] [dnn layer] [hidden dimension] [l2 weight] [n fold] [n repetition]

   Here, the range of cell line is [A549, HT29, A375].

   The range of test scenario is [edge, gene]. "edge" represents "split pair" and "gene" "split gene" settings (see our paper for more details).

   [semi-supervised loss weight] should be a nonnegative value represents the BPR loss weight (i.e., the semi-supervised learning loss weight).

   [dnn layer] should be a nonnegative integer represents the depth of the neural network.

   [hidden dimension] should be a nonnegative integer represents the number of hidden units of the neural network.

   [l2 weight] should be a nonnegative value represents the weight of L2 regularization term.

   [n fold] should be a nonnegative integer represents the n-fold cross-validation you want to perform.

   [n repetition] should be a nonnegative integer represents the the number of repetition of cross validation under different random seeds you want to perform.

3. A two-column list of AUROC (first column) and AUPR (second column) of cross-validation results will be saved.

If you have any questions or comments, please feel free to email Fangping Wan (wanfangping92[at]gmail[dot]com) and/or Jianyang Zeng (zengjy321[at]tsinghua[dot]edu[dot]cn).