Integrated Multi-omics Analysis Using Variational Autoencoders: Application to Pan-cancer Classification

Xiaoyu Zhang, Jingqing Zhang, Kai Sun, Xian Yang, Chengliang Dai, Yike Guo

Abstract

Different aspects of a clinical sample can be revealed by multiple types of omics data. Integrated analysis of multi-omics data provides a comprehensive view of patients, which has the potential to facilitate more accurate clinical decision making. However, omics data are normally high dimensional with large number of molecular features and relatively small number of available samples with clinical labels. The "dimensionality curse" makes it challenging to train a machine learning model using high dimensional omics data like DNA methylation and gene expression profiles. Here we propose an end-to-end deep learning model called OmiVAE to extract low dimensional features and classify samples from multi-omics data. OmiVAE combines the basic structure of variational autoencoders with a classification network to achieve task-oriented feature extraction and multi-class classification. The training procedure of OmiVAE is comprised of an unsupervised phase without the classifier and a supervised phase with the classifier. During the unsupervised phase, a hierarchical cluster structure of samples can be automatically formed without the need for labels. And in the supervised phase, OmiVAE achieved an average classification accuracy of 97.49% after 10-fold cross-validation among 33 tumour types and normal samples, which shows better performance than other existing methods. The OmiVAE model learned from multi-omics data outperformed that using only one type of omics data, which indicates that the complementary information from different omics datatypes provides useful insights for biomedical tasks like cancer classification.

Introduction

In the era of precision medicine, "omics" data—such as genomics, transcriptomics, and epigenomics—offers a critical window into the molecular drivers of disease. However, researchers face a significant "dimensionality curse": a single patient sample might contain hundreds of thousands of molecular features, while the number of labeled clinical samples available for training remains relatively small.

Traditional approaches often rely on manual feature selection or linear dimensionality reduction techniques like Principal Component Analysis (PCA), which can overlook complex, non-linear patterns hidden across the genome. Furthermore, most existing models analyze only one type of data (e.g., gene expression alone), missing the complementary biological insights available when multiple layers of data are integrated simultaneously.

Methods

The researchers developed OmiVAE, an end-to-end deep learning architecture designed to handle the scale and complexity of multi-omics data. The study utilized the The Cancer Genome Atlas (TCGA) pan-cancer dataset, consisting of 9,081 samples across 33 tumor types and normal controls. Each sample featured a combined total of 450,804 molecular features, including RNA-Seq gene expression profiles (58,043 features) and DNA methylation profiles (392,761 features).

OmiVAE Architecture: The model features a split encoder/decoder structure across two GPUs, utilizing chromosome-specific encoding for DNA methylation and a task-oriented classifier branch.

Key components of the OmiVAE methodology include:

  • Task-Oriented Feature Extraction: Unlike standard Variational Autoencoders (VAEs) that focus solely on data reconstruction, OmiVAE includes a classification branch. This forces the model to learn a "latent space" (a compressed representation) specifically optimized for distinguishing between cancer types.
  • Two-Phase Training: The model first undergoes an unsupervised phase to learn data compression and natural hierarchical clustering, followed by a supervised phase where the classifier is activated to maximize diagnostic accuracy.
  • Chromosome-Specific Encoding: To manage the massive DNA methylation dataset, the authors split the data into 23 blocks based on chromosomes. This captures intra-chromosomal relationships while preventing the model from being overwhelmed by parameters.
  • Model Parallelism: Due to the high parameter count (~700 million), the encoder and decoder were split across two separate GPUs (GPU 0 and GPU 1).

Results and Key Insights

OmiVAE achieved a remarkable average classification accuracy of 97.49% across 34 classes (33 cancers plus normal tissue) using 10-fold cross-validation. A primary finding was the "Power of And": the model performed significantly better when using both DNA methylation and gene expression data than when using either datatype in isolation (97.49% vs. 96.37% for single-omics).

Comparison of Latent Spaces: OmiVAE (left) demonstrates clear, task-oriented clustering of cancer types in a 2D latent space, significantly outperforming the traditional PCA approach (right).

In visualization tasks, OmiVAE demonstrated a superior ability to organize biological data. In the unsupervised latent space, the model automatically grouped samples by organ system (e.g., kidney cancers) and then by specific subtypes (e.g., KIRC vs. KIRP). When compared to baselines, OmiVAE + SVM (84.40% accuracy) significantly outperformed PCA + SVM (30.13%) and surpassed popular manifold learning methods like t-SNE and UMAP. Furthermore, because OmiVAE learns a parametric mapping function, it can classify new, unseen samples instantly without needing to be retrained.

Discussion and Future Work

While OmiVAE is highly accurate, its computational intensity—requiring significant GPU resources for its 700 million parameters—remains a consideration for widespread adoption. Additionally, like many deep learning models, it functions as a "black box," making it challenging to pinpoint the specific genes or methylation sites driving a classification.

Future research could focus on improving model interpretability to identify novel biomarkers and integrating additional data types, such as proteomics or medical imaging, to further enhance diagnostic precision in oncology.

Summary updated

Abstract

Different aspects of a clinical sample can be revealed by multiple types of omics data. Integrated analysis of multi-omics data provides a comprehensive view of patients, which has the potential to facilitate more accurate clinical decision making. However, omics data are normally high dimensional with large number of molecular features and relatively small number of available samples with clinical labels. The "dimensionality curse" makes it challenging to train a machine learning model using high dimensional omics data like DNA methylation and gene expression profiles. Here we propose an end-to-end deep learning model called OmiVAE to extract low dimensional features and classify samples from multi-omics data. OmiVAE combines the basic structure of variational autoencoders with a classification network to achieve task-oriented feature extraction and multi-class classification. The training procedure of OmiVAE is comprised of an unsupervised phase without the classifier and a supervised phase with the classifier. During the unsupervised phase, a hierarchical cluster structure of samples can be automatically formed without the need for labels. And in the supervised phase, OmiVAE achieved an average classification accuracy of 97.49% after 10-fold cross-validation among 33 tumour types and normal samples, which shows better performance than other existing methods. The OmiVAE model learned from multi-omics data outperformed that using only one type of omics data, which indicates that the complementary information from different omics datatypes provides useful insights for biomedical tasks like cancer classification.
BibTeX
@inproceedings{zhang2019integrated,
  author = {Xiaoyu Zhang and Jingqing Zhang and Kai Sun and Xian Yang and Chengliang Dai and Yike Guo},
  title = {{Integrated Multi-omics Analysis Using Variational Autoencoders: Application to Pan-cancer Classification}},
  booktitle = {BIBM},
  year = {2019},
  url = {https://openreview.net/forum?id=xKr79HmGjqk},
}
Figure