Overview

This repository contains the code for the following manuscript:

“Spatially-aware Graph Neural Networks and Cross-level Molecular Profile Prediction in Colon Cancer Histopathology: A Retrospective Multicentre Cohort Study”，volume 4, number 11, pages: e787-795, The Lancet Digital Health, 2022.

https://doi.org/10.1016/S2589-7500(22)00168-6

Introduction

Digital whole-slide images are a unique way to assess the spatial context of the cancer microenvironment. Exploring these spatial characteristics will enable us to better identify cross-level molecular markers that could deepen our understanding of cancer biology and related patient outcomes. In this study, we proposed a graph neural network approach that emphasises spatialisation of tumour tiles towards a comprehensive evaluation of predicting cross-level molecular profiles of genetic mutations, copy number alterations, and functional protein expressions from whole-slide images. In addition, we introduced a transformation strategy that converts whole-slide image scans into graph-structured data to address the spatial heterogeneity of colon cancer. Furthermore, we developed and assessed the performance of the model on The Cancer Genome Atlas colon adenocarcinoma [TCGA-COAD] and validated it on two external datasets (ie, The Cancer Genome Atlas rectum adenocarcinoma [TCGA-READ] and Clinical Proteomic Tumor Analysis Consortium colon adenocarcinoma [CPTAC-COAD]). We also predicted microsatellite instability and result interpretability. Finally, we showed that spatially connected graph models enable molecular profile predictions in colon cancer and are generalised to rectum cancer. After further validation, our method could be used to infer the prognostic value of multiscale molecular biomarkers and identify targeted therapies for patients with colon cancer.

The pipeline and data statistics

Repo Structure

├── functions
    ├── model_architecture.py
    ├── model_utils.py
    └── utils.py
├── 1. Data_preprocessing
    ├── 1-1-1.prepare_molecular_label.py
    ├── 1-2-1.segment_tiles.py
    ├── 1-2-2.tumor_detection.py
    ├── 1-2-3.tumor_prediction.py
    ├── 1-3-1.tile_normalization.py
    ├── 1-3-2.feature_extraction.py
    ├── 1-4-1.generate_adj_similar.py
    └── readme.md
├── 2. Prediction_on_TCGA_dataset
    ├── 2-1.coad_cross_validation_all_task
    ├── 2-2.read_validation_all_task
    └── readme.md
├── 3. Validation_on_CPTAC_dataset
    ├── 3-1.cptac_coad_validation_all_task
    ├── 2-2.read_validation_all_task
    └── readme.md

Dependencies

Pytorch 1.6.0

torch-geometric 1.6.1

Torchvision 0.7.0

openslide 1.1.1

Pillow 6.2.0

numpy 1.16.4

pandas 0.25.1

scikit-image 0.15.0

scikit-learn 0.21.3

Pillow 6.2.0

h5py 2.8.0

Data download

Download the FFPE whole slide images from GDC portal (https://portal.gdc.cancer.gov/) for colon adenocarcinoma (TCGA-COAD), rectum adenocarcinoma (TCGA-READ), and CANCER portal (https://wiki.cancerimagingarchive.net/display/Public/CPTAC-COAD) colon adenocarcinomacolon adenocarcinoma (CPTAC-COAD).

Download corresponding omics data (gene mutation, cna, expression) from cBioPortal (https://www.cbioportal.org/) for each dataset.

Usage

Step 1. Data preprocessing

Using the code under 1. Data_preprocessing to perform

1. Patch extraction and molecular profile label preparation.

Get gene mutation, CNA, MSI, and protein information of selected slides
```
python 1-1-1.prepare_molecular_label.py
```

2. Patch extraction and selection

Extract patches
```
python 1-2-1.segment_tiles.py
```
Tumor detection model training
```
python 1-2-2.tumor_detection.py
```
Tumor patch prediction (selection by tumor detection model)
```
python 1-2-3.tumor_prediction.py
```

3. Color normmalization and feature extraction

Color normalization
```
python 1-3-1.tile_normalization.py
```
Extract features for each patch
```
python 1-3-2.feature_extraction.py
```

4. Graph construction

Generate adjacency matrix infomation (construct graph edge). For each node, we save its connected neighbor nodes.
```
python 1-4-1.generate_adj_similar.py
```

Step 2. Model training for molecular profile prediction on TCGA-COAD, and validating on TCGA-READ

Using the code under 2. Prediction_on_TCGA_dataset

1. (On TCGA-COAD) Gene mutation and CNA : model training and prediction

We use 10 fold cross-validation on TCGA-COAD dataset.
This code could be used for all tasks, including gene mutation, CNA, MSI, and protein.
```
CUDA_VISIBLE_DEVICES=1,2,3,4 python 2-1.coad_cross_validation_all_task.py
```

2. Validation model on TCGA-READ
【 Please define the task type before run the code 】

We train the model on TCGA-COAD dataset, and validate model on TCGA-READ datast.
This code could be used for all tasks, including gene mutation, CNA, MSI, and protein.
```
CUDA_VISIBLE_DEVICES=1,2,3,4 python 2-2.read_validation_all_task.py
```

Step 3. Model validation for molecular profile prediction on CPTAC-COAD

Using the code under 3. Validation_on_CPTAC_dataset

1. Validation model on CPTAC-COAD
【Please define the task type before run the code 】

We train the model on TCGA-COAD dataset, and validate model on CPTAC-COAD datast.

This code could be used for predicting gene mutation, CNA, and MSI.

CUDA_VISIBLE_DEVICES=1,2,3,4 python 3-1.cptac_coad_validation_all_task.py

Citation

@article{DING2022e787,
title = {Spatially aware graph neural networks and cross-level molecular profile prediction in colon cancer histopathology: a retrospective multi-cohort study},
journal = {The Lancet Digital Health},
volume = {4},
number = {11},
pages = {e787-e795},
year = {2022},
issn = {2589-7500},
doi = {https://doi.org/10.1016/S2589-7500(22)00168-6},
url = {https://www.sciencedirect.com/science/article/pii/S2589750022001686},
author = {Kexin Ding and Mu Zhou and He Wang and Shaoting Zhang and Dimitri N Metaxas}
}

Luoxd1996/Review_Molecular_profile_prediction_GNN