Abstract
With the rapid development of high-throughput experimental technologies, different types of omics (e.g., genomics, epigenomics, transcriptomics, proteomics, and metabolomics) data can be produced from clinical samples. The correlations between different omics types attracts a lot of research interest, whereas the study on genome-wide omics data 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 their variants are one of 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 proposed OmiTrans, a deep learning framework that adopts the idea of generative adversarial networks to achieve omics-to-omics translation with promising results. OmiTrans was able to faithfully reconstruct gene expression profiles from DNA methylation data with high accuracy and great model generalization, as demonstrated in the experiments.
Introduction
In modern precision medicine, researchers collect multi-omics data to understand the complex layers of biological information, ranging from the genome and epigenome to the transcriptome. While these layers are deeply interconnected, the cost and technical difficulty of collecting all omics types for every patient are significant barriers. OmiTrans addresses this by treating omics-to-omics translation as a generative task, specifically focusing on predicting gene expression (RNA-Seq) from DNA methylation data. Unlike previous linear models, OmiTrans leverages the high-capacity generative power of Generative Adversarial Networks (GANs) to capture complex, genome-wide correlations.
Methods
The OmiTrans framework is based on a Conditional Generative Adversarial Network (cGAN) architecture. It consists of two primary components: a Generator ($G$) and a Discriminator ($D$). The Generator attempts to create realistic gene expression profiles ($y'$) from input DNA methylation data ($x$), while the Discriminator learns to distinguish between real pairs $(x, y)$ and generated "fake" pairs $(x, y')$.
The authors evaluated two architectural variants: a fully-connected (FC) version and a 1D CNN-based version using a U-Net structure. To adapt CNNs for biological data, "location alignment" was used to sort features by chromosomal position. The training process utilizes a hybrid loss function combining adversarial loss with a distance constraint (L1 or MSE) to ensure both biological plausibility and numerical accuracy.
Results and Evaluation
The framework was trained on 9,081 samples from The Cancer Genome Atlas (TCGA) and evaluated for its ability to reconstruct gene expression profiles. OmiTrans-FC emerged as the superior model, achieving a Mean Square Error (MSE) of 0.1097, outperforming existing methods like TDimpute.
Key performance metrics include:
- High Correlation: A sample-wise $R^2$ score of 0.9453 indicates that generated profiles are highly similar to real ones.
- Downstream Utility: In a cancer classification task, data generated by OmiTrans achieved an AUC of 0.9948, nearly matching the 0.9973 achieved with real data.
- Generalization: When tested on an independent Brain Tumour dataset (GSE109381), the TCGA-trained model successfully reconstructed hierarchical patterns, proving it learned fundamental biological relationships rather than just memorizing training data.
Discussion
Interestingly, the fully-connected architecture outperformed the CNN-based version. This suggests that while spatial proximity on a chromosome is important, gene regulation often involves long-range dependencies that FC layers are better equipped to capture. OmiTrans also demonstrated significant computational efficiency, completing training in 12 hours on a standard workstation compared to over 150 hours for traditional regression methods on cloud instances.
Future work will explore the framework's applicability to other omics pairs, such as proteomics or metabolomics, and investigate advanced architectures like Graph Neural Networks to better model the 3D physical interactions of the genome.
Summary updated