mm-cell_lines

Deep representation learning for predicting progression-free survival in multiple myeloma patients.

keywords: multiple myeloma, representation learning, RNA-seq

Dependencies/Startup

In order to run this code, you will need to install the ml-mmrf library.

You will also need to download the following files from the MMRF researcher gateway:

• CoMMpass_IA15_FlatFiles.tar.gz (unzipped)
• MMRF_CoMMpass_IA15a_E74GTF_Salmon_Gene_TPM.txt
• MMRF_OS_PFS_ASCT.csv
• MMRF_OS_PFS_non-ASCT.csv

Build the MMRF data by navigating to ml_mmrf/core and running the following two commands:

python build_mmrf_dataset.py ––fdir [YOUR FOLDER] --outcomes_type pfs --ia_version IA15 --recreate_splits True

python build_mmrf_dataset.py ––fdir [YOUR FOLDER] --outcomes_type pfs --ia_version IA15 --recreate_splits False

To use the CCLE data, you will need the following files from the DepMap dataset:

• sample_info.csv
• CCLE_expression.csv
• Achilles_gene_effect.csv*
• sanger-dose-response.csv

*Due to a recent renaming, you may need to download CRISPR_gene_effect.csv, instead.

Reproducing Experiments

The core experiments for this project are found in 3 separate Jupyter notebooks, each of which can be run end-to-end. More detailed comments are available within each Jupyter notebook.

• Autoencoders.ipynb: Trains and evaluates various autoencoders and downstream classifiers, and writes the embeddings to the .h5 files in autoencoder_embeddings/ (used by downstream notebooks).
• CCLE_transfer.ipynb: Trains and evaluates classifiers for IC50 prediction-based transfer experiments.
• Cell_perturbation.ipynb: Loads and evaluates classifiers for gene perturbation-based transfer experiments.

The requirements.txt file for conda environments used to run each of these sets of experiments are found in requirements/.

Trained versions of our best models are also available in trained_models.