hms-dbmi/MOMA

MOMA

MOMA: The Multi-omics Multi-cohort Assessment (MOMA) Platform

Pei-Chen Tsai, Tsung-Hua Lee, Kun-Chi Kuo, Fang-Yi Su, Tsung-Lu Michael Lee, Eliana Marostica, Tomotaka Ugai, Melissa Zhao, Mai Chan Lau, Juha P. Väyrynen, Marios Giannakis, Yasutoshi Takashima, Seyed Mousavi Kahaki, Kana Wu, Mingyang Song, Jeffrey A. Meyerhardt, Andrew T. Chan, Jung-Hsien Chiang, Jonathan Nowak, Shuji Ogino, Kun-Hsing Yu. Histopathology Images Predicted Multi-Omics Aberrations and Prognoses in Colorectal Cancer Patients. Nature Communications. 2023 Apr 13;14(1):2102. Paper

Requirements

• Survival prediction
  • Python==3.6.0
  • tensorflow==2.4.0
  • lifelines
  • scipy
  • statistics
  • matplotlib
• Multi-omics characterization
  • Python==3.6.0
  • torch==1.6.0
  • torchvision==0.7.0
  • scikit-learn
  • numpy
  • smooth-topk
  • opencv-python
  • tqdm

Dataset

• Survival prediction: TCGA-COAD and TCGA-READ
• Multi-omics characterization: TCGA-COAD and TCGA-READ
• Interpretation: Dataset(NCT-CRC-HE-100K) provided by Kather et al

Data Preprocessing

• Tiling : Modify from github Deepslide, or you can download the processed dataset provided by Kather et al.
• Tumor detection : Resnet50
• Color normalization : Modify from github HEnorm_python

Feature Extraction

You can use any pre-trained CNN model (like our multi-omics characterization task) or train model on our own (like our survival prediction task) to extract each patch's features.

Data Preparation

• Survival Prediction

  • Color normalization
  • Create a dataframe

    # Survival dataframe
    data = {
        'bcr_patient_barcode' : patient id,
        'vital_status' : overall survival status or disease free status,
        'Days' : overall survival days or disease free days
        '0' : pathology image feature (dimension 1)
        '1' : pathology image feature (dimension 2)
        ...
        'n' : pathology image feature (dimension n)
    }
    
    df = pd.DataFrame(data)

• Multi-omics characterization

  • XXX_id can be patient’s ID or slide’s ID, which is depending on your task. And please be sure that the patch_name in features pickle file and in cluster pickle file is the same.

  • Sample file

    # Patch features pickle
    {
      'patch_name' : array([latent feature]),
      'patch_name' : array([latent feature]),
      ...
    }

    # Cluster pickle file
    {
      XXX_id: {
        'patch_name' : cluster label,
        'patch_name' : cluster label,
        ...
      },
      XXX_id: {
        'patch_name' : cluster label,
        'patch_name' : cluster label,
        ...
      },
    }

    # Label pickle file
    {
      XXX_id: class,
      XXX_id: class,
      ...
    }

• Interpretation

  • Create a dataframe

    # Interpretation dataframe
    data = {
        'fig' : fig name,
        'folder' : file path,
        'class' : 0~n (class num)
      
    }
    
    df = pd.DataFrame(data)

Usage

• Survival prediction

  • Both overall survival prediction and disease free prediction use the same .ipynb file

• Multi-omics characterization

  • Sample Command

    # Training
    python3 Train.py --level patient --hidden_dim 512 --encoder_layer 6 --k_sample 3 --tau 0.5 --save_path 'path/to/save/' --label 'path/to/label pickle file' --use_kather_data True --epoch 60 --lr 3e-4 --evaluate_mode kfold --kfold 5

    # Validation
    python3 Validation.py --level patient --hidden_dim 512 --encoder_layer 6 --k_sample 3 --tau 0.5 --save_path 'path/to/save/' --label 'path/to/label pickle file' --use_kather_data True

    --level                 slide or patient level
    --hidden_dim            The dimension in the Transformer encoder
    --encoder_layer         The layers of the Transformer encoder
    --k_sample              The top-k and bottom-k for the instance selection
    --tau                   The smoothness term for smoothSVM
    --use_kather_data       Using the data provided by kather et al. or not
    --save_path             Model weights save path
    --label                 Path to label pickle file
    --lr                    Learning rate
    --epoch                 Training epochs
    --evaluate_mode         Kfold or holdout test
    --kfold                 The number of fold

• Interpretation

  • It's not recommended to test data with low foreground and background ratio