OmiTrans: Generative Adversarial Networks Based Omics-to-omics Translation Framework

Xiaoyu Zhang, Yike Guo

TL;DR

OmiTrans is a novel deep learning framework that applies Generative Adversarial Networks (GANs) to "translate" one type of biological data into another—specifically predicting gene expression from DNA methylation. By treating omics data like an image-to-image translation task, it achieves state-of-the-art accuracy and demonstrates a remarkable ability to generalize to entirely new, unseen datasets.

What Problem Does This Paper Address?

In modern precision medicine, researchers collect "multi-omics" data: different layers of biological information ranging from the genome (DNA) and epigenome (chemical modifications like methylation) to the transcriptome (RNA/gene expression). While these layers are deeply interconnected, collecting all of them for every patient is prohibitively expensive and technically challenging.

The challenge is omics-to-omics translation: can we computationally generate one type of data (e.g., RNA-Seq) from another (e.g., DNA methylation)? Previous attempts using linear regression or simple neural networks often lacked the "capacity" to capture the complex, genome-wide correlations between these layers. OmiTrans seeks to fill this gap by using the high-capacity generative power of GANs.

Methods

The authors introduce OmiTrans, a framework based on the Conditional Generative Adversarial Network (cGAN).

  • The Architecture: The framework consists of a Generator (which tries to create realistic gene expression profiles from methylation data) and a Discriminator (which tries to distinguish between real pairs of data and the "fake" pairs created by the generator).
  • Two Flavors: The authors tested two versions:
  • FC-based: Uses standard fully-connected layers.
  • CNN-based: Uses a 1D U-Net architecture. To make this work, they used "location alignment," sorting biological features by their physical position on the chromosomes to mimic the spatial structure found in images.
  • The Loss Function: Unlike standard GANs that only focus on "realism," OmiTrans uses a hybrid loss function. It combines the adversarial loss (to ensure the data looks biologically plausible) with a distance constraint (L1 or MSE) to ensure the generated values are numerically accurate compared to the ground truth.

Datasets

  • TCGA (The Cancer Genome Atlas): A massive pan-cancer database. The authors used 9,081 samples containing both DNA methylation and RNA-Seq data for training and primary evaluation.
  • BTM (Brain Tumour DNA Methylation): An independent dataset (GSE109381) used specifically to test the model’s generalization. This is a "blind test" to see if a model trained on general cancer data can work on a specific, previously unseen dataset.

Key Insights and Ideas

  • FC beats CNN (for now): Interestingly, the fully-connected (FC) version outperformed the CNN version. In computer vision, CNNs are king because pixels have local correlations. In biology, while "location alignment" helps, a gene might be regulated by a methylation site very far away on the 1D sequence. The FC layers were better at capturing these long-range, non-spatial dependencies.
  • Adversarial "Faithfulness": The GAN doesn't just predict averages; it learns the underlying distribution of the data. This results in "reconstructed" profiles that retain the complex hierarchical clustering patterns found in real biological samples.
  • Efficiency: While traditional methods like LASSO regression took 154 hours to train on high-end cloud instances, OmiTrans completed its training in just 12 hours on a standard local workstation.

Key Results

  • Superior Accuracy: OmiTrans-FC achieved the best performance across nine different metrics, including a Mean Square Error (MSE) of 0.1097, significantly outperforming the previous state-of-the-art method, TDimpute (MSE 0.1141).
  • High Correlation: The model achieved a sample-wise $R^2$ score of 0.9453, meaning the generated gene expression profiles are nearly indistinguishable from the real ones.
  • Downstream Utility: To prove the generated data is useful, the authors ran a cancer classification task. Using "fake" data generated by OmiTrans, they achieved an AUC of 0.9948, which is nearly identical to the 0.9973 achieved using real data.
  • Generalization: When the TCGA-trained model was applied to the independent Brain Tumour dataset, it successfully reconstructed the multi-level hierarchical patterns of the original data, proving it learned fundamental biology rather than just memorizing the training set.

Limitations and Future Work

  • Scope of Omics: This paper focused primarily on the DNA methylation to gene expression pipeline. The authors note that the framework is designed to be flexible, but its performance on other pairs (like proteomics or metabolomics) remains to be tested.
  • The "Spatial" Problem: While the FC-based model won here, the authors suggest that more advanced architectures (like Graph Neural Networks) might be the next step to better capture the 3D physical interactions of the genome that 1D CNNs miss.
  • Data Imputation: A promising future direction is using OmiTrans to fill in missing data in incomplete clinical studies, potentially rescuing valuable samples that would otherwise be discarded.

Summary updated

Abstract

With the significant development of high-throughput experimental technologies, different types of omics (e.g., genomics, epigenomics, transcriptomics, proteomics, and metabolomics) data are produced from clinical samples at an unprecedented speed. The correlations between different types of omics data attract enormous research interest. In contrast, the study on genome-wide omics-to-omics translation (i.e., generation and prediction of one type of omics data from another type of omics data) is almost blank. Generative adversarial networks and variants are among the most state-of-the-art deep learning technologies, which have shown great success in image-to-image translation, text-to-image translation, etc. Here we propose OmiTrans, a deep learning framework that adopted the idea of generative adversarial networks to achieve omics-to-omics translation with promising results. OmiTrans can faithfully reconstruct genome-wide gene expression profiles from DNA methylation data with high accuracy and excellent model generalisation, as demonstrated in the experiments.
BibTeX
@inproceedings{zhang2022omitrans,
  author = {Xiaoyu Zhang and Yike Guo},
  title = {{OmiTrans: Generative Adversarial Networks Based Omics-to-omics Translation Framework}},
  booktitle = {BIBM},
  year = {2022},
  url = {https://openreview.net/forum?id=53s4aJRyRZ},
}

Loading related content...

Figure