src

The code

One-time scripts

• download.py: downloads the human GEO data sets and their corresponding papers/abstracts.
• manual_inspect.py and manual_inspect_helper.py: manual inspection CLI tool for GEO data sets. For each data set it displays its title, abstract, description, and summary statistics. Then it displays info of different sample annotations/tasks and the correlation matrix between tasks. It then takes user inputs to decide what tasks to accept/reject and how to remap tasks to binary tasks.
• prepare_train.py: keep only data sets with certain gene expression attributes and sufficient sample size, then split the data into training and testing group.
• sample_distributions.py: scrape the pages of ArrayExpress, GEO, and dbGaP for data set sample size.

Implementations

• l1000.py: holds the L1000 genome data.
• geo_dataset.py: implements the Dataset class, which holds and prepares training, validation, and testing matrices.
• geo_models.py: implements different architectures: autoencoder, constraint autoencoder, variational autoencoder, multi-task model, and multi-task model with attention.
• geo_train.py: implements the training loop.
• geo_test.py: implements the testing loop.

Experiments

• Graphics_for_thesis.ipynb: contains code for analysis of all results and for plotting all figures.
• run_autoencoder.py: train and test multiple autoencoder models.
• run_multitask.py: train and test multiple multi-task models.
• test_baseline.py: create the baseline results with logistic regression.
• test_autoencoder.py: retest multiple autoencoder models.
• test_multitask.py: retest multiple multi-task models.