This is a reference implementation for Black Box FDR (BB-FDR). BB-FDR is an empirical-Bayes method for analyzing multi-experiment studies when many covariates are gathered per experiment. It performs two stages of selection:
Stage 1: Fit a (black box) neural network model using the covariates for each experiment. The NN is then used to select the significant outcomes at a given false discovery rate (FDR).
Stage 2: For each covariate, fit a (black box) gradient boosting classifier to predict the probability of a covariate given other covariates. The classifier is then used to perform a conditional randomization test (CRT) to determine whether that mutation is significantly associated with differential response.
Note that the code is setup to be run easily on a cluster, in case one has hundreds of such studies to analyze. The model is checkpointed frequently in order to allow for preemption without losing the progress.
We use pytorch to fit the neural network prior model and scikit-learn to fit our gradient boosting trees.
We provide an example of using BB-FDR to analyze dose-response data from the Genomics of Drug Sensitivity in Cancer. Seepython/cancer.py for details.
If you use this code, please cite:
@inproceedings{tansey:etal:icml:2018:bbfdr, title={Black Box {{FDR}}}, author={Tansey, W. and Wang, Y. and Blei, D. B. and Rabadan, R.}, booktitle={International Conference on Machine Learning (ICML'18)}, year={2018}}