NeuroSyd/latent-space-discovery

Noble self-supervised adversarial auto-encoder is proposed to extract biologically relevant genes from cancer transcriptomes.

Extracting Biologically Relevant Genes using AFExNet from Cancer Transcriptomes [Paper]

In this project, we introduce neural network based adversarial autoencoder (AAE) model to extract biologically-relevant features from RNA-Seq data. We also developed a method named TopGene to find highly interactive genes from the latent space. AFExNet in combination with TopGene method finds important genes which could be useful for finding cancer biomarkers.

Getting Started

The following instructions will get you a copy of the project up and running on your local machine for development and testing purposes. See the instruction below:

Prerequisites

The following libraries are required to reproduce this project:

1. Keras (2.0.6)

2. Keras-adverserial (0.0.3)

3. Tensorflow (1.13.1)

4. Scikit-Learn (0.20.3)

5. Numpy (1.16.3)

6. Imbalanced-Learn (0.4.3)

Supports both Python 2.5.0 and Python 3.5.6

Directory Layout

├── results
│   ├── saved_results
│   │   ├── Gene_Analysis_Breast_Cancer.xlsx
│   │   ├── Gene_Analysis_UCEC.xlsx
│   ├── AAE
│   │   ├── aae_encoded.tsv
│   │   ├── aae_sorted_gene.tsv
│   │   ├── aae_weight_distribution.png
│   │   ├── aae_weight_matrix
│   ├── PCA
│   ├── ... # add LDA, SVD etc
├── data
│   ├── data will be stored here
├── feature_extraction
│   ├── AAE
│   │   ├── aae_encoder.h5
│   │   ├── aae_decoder.h5
│   │   ├── aae_discriminator.h5
│   │   ├── aae_history.csv
│   ├── PCA
│   ├──VAE
│   ├── ...
├── README.md
├── figures
│   ├── saved_figures
│   │   ├── Olfactory__Transduction_pathway.png
└── .gitignore

Usage

Run the following to extract features using different autoencoders

main.py

And run the following to extract features when PCA, NMF, FastICA, ICA, RBM etc. are used

main_pca.py

Gene ontology of molecular function was performed using DAVID 6.7 https://david-d.ncifcrf.gov/

More regarding gene ontology http://geneontology.org/docs/ontology-documentation/

Proposed Architecture

Datasets

• cBioPortal - Cancer Genomics Datasets
• Breast Invasive Carcinoma (TCGA, Cell 2015) - Clinical information is used to label various molecular subtypes

Breast Invasive Carcinoma (BRCA)

Molecular Subtypes	Number of Patients	Label
Luminal A	304	0
Luminal B	121	1
Basal & Triple Negetive	137	2
Her 2 Enriched	43	3

Total Number of Samples (Patients)	Total Number of Features (Genes)
605	20439

• Details about Molecular Subtypes of Breast Cancer

Validation Data

• Uterine Corpus Endometrial Carcinoma (TCGA, Nature 2013) - Clinical information is used to label various molecular subtypes.

Uterine Corpus Endometrial Carcinoma (UCEC)

Molecular Subtypes	Number of Patients	Label
Copy Number High	60	0
Copy Number Low	90	1
Hyper Mutated (MSI)	64	2
Ultra Mutated (POLE)	16	3

Total Number of Samples (Patients)	Total Number of Features (Genes)
230	20482

• Details about Molecular Subtypes of Endometrial Cancer

Contribution

If you want to contribute to this project and make it better, your help is very welcome. When contributing to this repository please make a clean pull request.

Acknowledgments

• The proposed architecture is inspired by https://github.com/bstriner/keras-adversarial