Multi-omics Cluster Variational Autoencoders (MCluster-VAEs)

In this study, we propose Multi-omics Clustering Variational Autoencoders (MCluster-VAEs), a novel end-to-end deep learning-based method, to perform clustering using multi-omics data. MCluster-VAEs describes multi-omics data using a new probabilistic model with a global discrete latent variable considered as the clusters. Using Variational Bayes approach, we derive a unified single-model architecture and a novel objective function, to fit the multi-omics data and infer the posterior probability of the clustering assignments. The results on TCGA database with four kinds of omics demonstrate that MCluster-VAEs performs significantly better than 10 other methods. In particular, MCluster-VAEs achieves an unsupervised clustering accuracy of about 0.82 and adjusted Rand index of about 0.78 on the data with known labels, an increase of 6%-75% compared to other methods.

Overview of MCluster-VAEs. (A) Two deep learning-based multi-omics clustering pipelines. (B) The data generative model. (C) The posterior variational inference path. (D) The architecture of MCluster-VAEs model for the experiments using mRNA, miRNA, CNA and methylation.

Dependencies

1. Download and install miniconda.

2. Create an python environment and install required packages.

conda create -n mcluster && conda activate mcluster
bash environment.sh

The environment can also be created by create env -f enviroment.yml

Running

1. Prepare data. Example data (Pan Cancer) is in release and can be download to test.

2. Modify the configuration. The all configurations are in conf/, are stored in yaml format and managed by hydra module. The modifying part is the data path and output path, like following:

   # config.yaml
   hydra:
       ...
       run:
           # indicate the root path contained outputs
           dir: output_path/
   dataset:
       # indicate the root path contained datasets
       root: data_path/
       ...

   The more configuration can be found in conf/, including data pipeline (conf/dataset/), model architecture (conf/model/) and training setttings (conf/train/). These configuration can be added in conf/config.yaml, too. The rules of hydra configuration can be found in hydra documents.

3. Run script.

   python main.py

   The results will be in output_path/

contact

[email protected] or [email protected]