bit-whacker / cgga

Cancer subtype identification by consensus guided graph autoencoder

Cancer Subtype Identification by Consensus Guided Graph AutoEncoders(CGGA)

😃Update

We have uploaded the processed cancer datasets under the cancer_datasets/ directory. Each cancer dataset consists of three omics and could be loaded into the Matlab directly.

Method Description

CGGA is a computational framework that can effectively and reliably uncover cancer subtypes. It mainly consists of two steps. First, for each omic, a new feature matrix is learned by using graph autoencoders, which can incorporate both structure information and node features during the learning process. Second, a set of omic-specific similarity matrices as well as a consensus matrix is learned based on the features obtained in the first step. The learned omic-specific similarity matrices are then fed back to the graph autoencoders to guide the feature learning. By iterating the two steps above, our method obtains a final consensus similarity matrix for cancer subtyping.

Requirements

>= MATLAB 2014b.

Usage

To run our algorithm, please load the script 'CGGA.m' into your MATLAB programming environment and click 'run'. Users can also run the script in standard command-line mode, where you should input the following commands for each function, respectively:

matlab -nodisplay -nodesktop -nosplash -r "CGGA;exit;"

All the cancer datasets used in the code can be directly downloaded at http://acgt.cs.tau.ac.il/multi_omic_benchmark/download.html.

Parameters

There are three parameters in our algorithm that users can tune according to their own needs, i.e. lambda, the number of neighbors k and the number of layers in CGGA. The default values for lambda and k are fixed to 1e-5 and 15, respectively. The number of layers in CGGA is set to 2 and user can specify a larger value to construct a deeper graph autoencoder.

Input and Output Directories

To change the input file directory, please refer to the 'dataDir' variable in the processTCGAdata.m. For output file directory, please refer to the 'outDir' variable in the same script.

Contact

For any questions regarding our work, please feel free to contact us: [email protected].

Cite Our Paper

Cheng Liang, Mingchao Shang, Jiawei Luo. Cancer subtype identification by consensus guided graph autoencoders. Bioinformatics, 2021, doi: 10.1093/bioinformatics/btab535.