Deep Learning Project

Predicting splicing behaviour based on combined local sequence and long-range 3D genome information.

Repo Setup and Package Installation

First, please clone the repo in Euler. Then, navigate to main folder and run source ./init_leonhard.sh

Then, install torch-geometry as explained here: https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html Make sure to check your torch and cuda versions as explained, in order to not run into any issues.

Then, navigate to src folder and run:

pip install -e splicing/

Important: Structure of Repository and Commands

The entire project revolves around two required arguments for every command:

• Window size (-ws): 1000 or 5000
• Context length (-cl): 80 or 400

At almost every step, from data preprocessing to model training, these arguments are required. Separate datasets, models, etc. are created for each combination of these settings.

The rest of this guide takes you through the -ws 5000 -cl 80 workflow, just change the values to run for a different data configuration.

Data Pipeline

Setup

First, create a data directory for the project. Then, download raw data from https://polybox.ethz.ch/index.php/s/YZc2gt5jmQZ4cpy, and extract the raw_data folder within the data directory.

Then, open src/splicing/splicing/config.yaml and set the DATA_DIRECTORY variable at the top to match the path of the directory you just created.

Euler Commands

All commands should be run from src/splicing/splicing. Please run the preprocessing/create_datafile.py commands in a sequential manner, as we have occasionally encountered errors when running the three commands concurrently.

bsub -R "rusage[mem=16000]" python preprocessing/create_datafile.py -g train -p all -ws 5000 -cl 80

bsub -R "rusage[mem=16000]" python preprocessing/create_datafile.py -g valid -p all -ws 5000 -cl 80

bsub -R "rusage[mem=16000]" python preprocessing/create_datafile.py -g test -p all -ws 5000 -cl 80

bsub -R "rusage[mem=16000]" python preprocessing/create_dataset.py -g train -p all -ws 5000 -cl 80

bsub -R "rusage[mem=16000]" python preprocessing/create_dataset.py -g valid -p all -ws 5000 -cl 80

bsub -R "rusage[mem=16000]" python preprocessing/create_dataset.py -g test -p all -ws 5000 -cl 80

Graph Creation

You now need to create the graph using the Hi-C values that need to be downloaded before. The graph only depends on the window size and not on the context length. So you will only need to run this once for the 5000 window size and once for the 1000 window size

The hic data can be downloaded here: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE63525

filename: GSE63525_GM12878_combined_intrachromosomal_contact_matrices.tar.gz

(Gene Expression Omnibus (GEO) accession number for the Hi-C data sets reported in this paper is GSE63525.)

Use the following commands to download the Hi-C data:

cd {your specified data_dir}/raw_data/hic

wget https://ftp.ncbi.nlm.nih.gov/geo/series/GSE63nnn/GSE63525/suppl/GSE63525%5FGM12878%5Fcombined%5Fintrachromosomal%5Fcontact%5Fmatrices%2Etar%2Egz

tar -xf GSE63525_GM12878_combined_intrachromosomal_contact_matrices.tar.gz

The hic datafolder 5kb_resolution_intrachromosomal (for 5000 window size) or 1kb_resolution_intrachromosomal (for 1000 window size) should now be located inside the folder: {your specified data_dir}/raw_data/hic/GM12878_combined/

After the download is complete and the data is in the right place, run the following command to create the graph:

bsub -R "rusage[mem=32000]" -W 4:00 python preprocessing/create_graph.py -ws 5000

The command relies on the file graph_windows_5000.bed, that is created in the create_datafile command from above. You should now have four graph related .pkl files in {your specified data_dir}/processed_data

Model Training

Pretrain the nucleotide representation CNN

bsub -R "rusage[mem=48000,ngpus_excl_p=1]" -W 04:00 python main.py -pretrain -modelid base -ws 5000 -cl 80

-modelid enables you to train multiple models with the same -ws and -cl without overwriting.

Export nucleotide representations using saved pretrained model

bsub -R "rusage[mem=64000,ngpus_excl_p=1]" python main.py -save_feats -load_pretrained -modelid base -mit 10 -ws 5000 -cl 80

-mit specifies which model iteration (epoch) to load. Default number of epochs for pretraining is 10.

Train Graph & Full model

bsub -R "rusage[mem=64000,ngpus_excl_p=1]" -W 24:00 python main.py -finetune -modelid base -ws 5000 -cl 80 -adj_type both

bsub -R "rusage[mem=64000,ngpus_excl_p=1]" -W 24:00 python main.py -finetune -modelid base -ws 5000 -cl 80 -adj_type hic

bsub -R "rusage[mem=64000,ngpus_excl_p=1]" -W 24:00 python main.py -finetune -modelid base -ws 5000 -cl 80 -adj_type constant

bsub -R "rusage[mem=64000,ngpus_excl_p=1]" -W 24:00 python main.py -finetune -modelid base -ws 5000 -cl 80 -adj_type none

-adj_type: specifies which type of graph to use: hic, constant, both, none