Lung Cancer Subtyping using GANs 🧬

This repository contains the code for our project on:

• lung cancer subtyping using GANs (Subtype-GAN [1]) - implemented in PyTorch.
• and unsupervised learning of image segmentation based on differentiable feature clustering.

Our objective is to classify lung cancer subtypes based on multi-omics data, and the resulting subtype classifications are used to plan treatment and determine prognosis. We use differentiable ensamble methods during preprocessing and feature clustering to improve the accuracy and efficiency of our GAN model.

Subtype-GAN architecture as in [1] plus a classification head.

🔗 Check the final project presentation.

Problem description

Cancer Suptyping 🧬 Describes the smaller groups that a type of cancer can be divided into, based on certain characteristics of the cancer cells. These characteristics include how the cancer cells look and specific gene expressions. It is important to know the subtype of a cancer in order to plan treatment and determine prognosis.

Goal 🎯 Due to the diversity and complexity of multi-omics data, it is challenging to develop integrated clustering algorithms for tumor molecular subtyping. Our objective is to classify lungs cancer subtypes given multi-omics data exploiting the power of Generative adversarial networks.

Dataset

Data downloaded from GDC portal [4], with manifest files.

• Multi-omics considered: mRNA, miRNA, meth and images
• Suptypes considered: TCGA-LUSC, TCGA-LUAD, CPTAC-3, TCGA-BRCA, TCGA-KIRC.
• Preprocessing (NA, log2, scaler)
  • Feature selection
  • Dimensionality reduction
• Early integration approach

Unsupervised Learning of Image Segmentation Based on Differentiable Feature Clustering

Since images comes without label the method in [2] has been implemented to extract the cancer areas of the images. Consequently the image information are flattened and added as an input.

Example of unsupervised segmentation on pathological images of adenocarcinoma (left) and squamous cell carcinoma (right).

The following table reports the classification accuracy on the test dataset. The best model found is the one that use the SubtypeGAN plus the classification layer. Moreover, the best preprocessing method overall is the ensemble method to choose the best features for each omic.

Preprocessing	Model	Classifier + Encoder	Only Classifier
K Best	0.9792	0.9594	0.9422
PCA	0.9744	0.8887	0.8758
Ensemble Method	0.9805	0.9615	0.9391

References

[1] Yang, Hai et al. “Subtype-GAN: a deep learning approach for integrative cancer subtyping of multi-omics data.” Bioinformatics (2021)
[2] Kim, Wonjik et al. “Unsupervised Learning of Image Segmentation Based on Differentiable Feature Clustering.” IEEE Transactions on Image Processing 29 (2020): 8055-8068.
[3] Wang, S. et al. Comprehensive analysis of lung cancer pathology images to discover tumor shape and boundary features that predict survival outcome (2018)
[4] GDC Dataset: https://portal.gdc.cancer.gov/