Genome mapped to Image (GENIUS) Project

Mapping entire genome information to the image and learning the features of the genome as image.

Step 1: Organize raw data in data/raw_data folder \

• Check example input

Step 2: Map genome data to images

• First we enter the directory with the scripts for making images \
• We run the script for every image transofrmation we want: \
• For example: \
  •   TP53 images need to remove the TP53 gene from the list while making images therefore we put --tp53 1 \
  •   If we want to created shuffled image with random position of genes we use --shuffle 1, otherwise 0 \
• You can run this code in parallel if you have enough memory and CPU power

cd src/image_to_picture
python3 make_images.py --output data/Metastatic_data --tp53 0 --shuffle 0
python3 make_images.py --output data/Metastatic_data --tp53 0 --shuffle 1
python3 make_images.py --output data/TP53_data --tp53 1 --shuffle 1
python3 make_images.py --output data/TP53_data --tp53 1 --shuffle 0

• If you want to build Chromosome organized images use the following code

python3 make_images_by_chr.py --output data/TP53_data --tp53 1 --shuffle 0
python3 make_images_by_chr.py --output data/Metastatic_data --tp53 0 --shuffle 0

• make_images and make_image_by_chr will also make a manifest file which you need for training. Those files link image path with sample ID and response variables

Step 3: Prepare for Training

• Take AutoEncoder.AE.AE class if you are working with Chromosome Organized Images
• Take AutoEncoder.AE.AE_Square if you are working with rectangular Images
• If you are predicting numerical response use train_num_response.py and for categorical use train.py
• Write config file which tells the training which hyperparameters to use and which folders to use

Step 4: Training the model

• In this step we train the model for predicting metastatic disease by running the following code:

python3 train.py config/metastatic

• In this step we train the model for predicting wGII (numeric value) by running the following code:

python3 train_num_response.py config/wGII

Step 5: Inspect the model performance

• If you are happy with model performance continue with the next step
• If you do not "trust" the model because it does not predict outcome variable with high certainty, try tweaking the model parameters or hyperparameters.

Step 6: Integrated Gradients (IG)

• The training script outputs model with the best loss in .pb format
• This file is needed for IG step
• We run the following script to get the gradients of a model

cd ../../inference
python3 analyze_network.py

• This script will produce csv files representing attribution for every input

Step 7: Encode genome data

• We can also use trained model to encode entire genome image to a vector of length 128
• This can be used in further analysis
• We run the following script to exstract encoded genomes

python3 encode_genome_image.py

Step 8: Downstream Analysis

• Downstream analysis of attribution genes
• Downstream analysis of encoded genomes