aikseng/Kras-Nrf2-Dataset

KRAS/NRF2 Classification

We aim to classify individual cells into one of four categories based on their Kras and Nrf2 mutation status: Kras-wildtype / Nrf2-wildtype, Kras-mutant / Nrf2-wildtype, Kras-wildtype / Nrf2-mutant, or Kras-mutant / Nrf2-mutant.

First, we generate labeled data by analyzing aligned sequencing reads to determine the mutation status of Kras and Nrf2 for each cell. Using this labeled data and gene read counts as features, we train four binary neural network classifiers: two classifiers for Kras status and two for Nrf2 status.

Each Kras classifier is paired with one Nrf2 classifier to form two classifier pairs. Each pair independently assigns one of the four mutation classes to a cell. We then compare the classifications from both pairs and retain only the cells where both pairs agree, discarding those with conflicting predictions.

Label: 0 for wild type, 1 for mutant Groups: counts_ctrl, counts_2-1, counts_3-2, counts_4-3

Files

• parse_kras.py: get labeled kras data
• parse_nrf2.py: get labeled nrf2 data
• model.py: neural network training and prediction

Training

python model.py --do_train --do_preprocessing

1. Get data into train_[gene_type]_[folder].csv files (Ex. train_nrf2_counts_ctrl.csv), where each csv file has rows of barcode,label
2. All steps after are in model.py. Run get_train_data() if you want to use a new set of training data in csv files from step 1
3. Include --do_train if you want the model to train otherwise exclude if you want the model to only predict
   • If --do_train, decide whether you want to set the --do_preprocessing, which will shuffle the data Note: you should also make sure labels are equal (equal number of wild type and mutant labels) so the model isn't biased
   • When predicting, decide which models you want to use to predict (set parameter in torch.load to the .pth file of the model)
4. When training, the best model will be saved in the top directory as model_[gene_type]_checkpoint.pth. If you want to save your model, rename and move this to the saved_models folder
5. When predicting all test data, the results will be saved in the top directory as predict_results.mat. Rename and move this to the saved_models folder