BeatAML CTD^2 DREAM Challenge: Example 1

Example implementation of a solution to subchallenge 1 of the BeatAML CTD^2 DREAM challenge. This example uses gene expression to train a RidgeRegression model for each inhibitor to predict AUC.

To Train a model

• Run Jupyter with docker run -p 8888:8888 -v "$PWD:/home/jovyan" jupyter/scipy-notebook
  • Stdout will include a URL to open the notebook
• Go through the steps in index.ipynb
  • The model will be stored in model/ in two files: pkl_1.csv and pkl_2.csv
  • Read more about the model below

To Run Your Model on Training Data

This model can be run on the same data it was trained on, to test whether the Dockerfile works:

SYNAPSE_PROJECT_ID=<...>
docker build -t docker.synapse.org/$SYNAPSE_PROJECT_ID/sc1_model .
docker run -v "$PWD/training/:/input/" -v "$PWD/output:/output/" docker.synapse.org/$SYNAPSE_PROJECT_ID/sc1_model 

Submitting to Synapse DockerHub

SYNAPSE_PROJECT_ID=<...>
docker login docker.synapse.org
docker build -t docker.synapse.org/$SYNAPSE_PROJECT_ID/sc1_model .
docker push docker.synapse.org/$SYNAPSE_PROJECT_ID/sc1_model

The Model

One Ridge Regression model is trained for each inhibitor to predict AUC. The only input is gene expression (rnaseq.csv).

Specifics:

• The 1000 most variable genes are used for training
• The log2(cpm) values are normalized per-specimen
• The z-score is computed for each gene
• Ridge Regression is trained using hold-one-out cross-validation to predict AUC

On-Disk Representation

The trained model is stored in two "pickles": pkl_1 and pkl_2:

• pkl_1: has one row per gene included in the model and N+3 columns (N is the number of inhibitors):
  • gene: Include this gene's expression in the linear fit.
  • gene_mean: The mean expression in the training data (to compute z-score).
  • gene_std: The standard deviation of expression in the training data (to compute z-score).
  • : The Ridge Regression weight coefficient for this gene for inhibitor.
• pkl_2: one row per inhibitor and two columns:
  • inhibitor: The inhibitor name.
  • intercept: The Ridge Regression intercept.