Author Response: GENIUS: GEnome traNsformatIon and spatial representation of mUltiomicS data

Mateo Sokač, Asbjørn Kjær, Lars Dyrskjøt, Benjamin Haibe-Kains, Hugo J. W. L. Aerts, Nicolai J. BirkbakOpenalex Unknown 2023 · DOI · OpenAlex

Abstract

The application of next-generation sequencing (NGS) has transformed cancer research, enabling the generation of multi-layered molecular data—genomics, transcriptomics, and methylomics—from the same samples. Integrating these multi-omics datasets remains a significant challenge, often addressed through ad-hoc statistical methods that ignore the spatial orientation of the genome and suffer from stringent p-value corrections. We present GENIUS (GEnome traNsformatIon and spatial representation of mUltiomicS data), a deep learning framework that transforms multi-omics data into images where genes are spatially connected pixels. By leveraging advanced image analysis techniques, GENIUS extracts relevant biological information to predict cancer progression. Applying GENIUS to The Cancer Genome Atlas (TCGA) datasets, we demonstrate its ability to predict metastatic cancer from primary tumors and identify driver genes through model inference, providing a proof of concept for multi-omics data transformation and analysis without traditional statistical correction.

Introduction: The Multi-Omics Integration Challenge

In modern cancer research, scientists frequently collect multiple types of molecular data from a single tumor sample, such as gene expression, DNA methylation, and somatic mutations. Traditionally, this "multi-omics" data is stored in massive tables and analyzed using standard statistical methods. However, this approach faces two primary limitations:

  1. Loss of Genomic Context: Tabular representations treat genes as independent variables, ignoring their physical positioning on chromosomes and their 3D interactions within the nucleus. Genes are not just spreadsheet entries; they are physically positioned and interact in 3D space.
  2. The Multiple Testing Trap: Testing thousands of genes for associations requires severe statistical corrections (like p-value adjustments) to avoid false positives. This often "throws the baby out with the bathwater," masking subtle but biologically significant signals.

Methodology: From Genome to Image

To bypass these hurdles, the authors developed GENIUS (GEnome traNsformatIon and spatial representation of mUltiomicS data). The core innovation is mapping tabular multi-omics data onto a 198x198 pixel grid. Each pixel corresponds to a gene, while different "channels" (analogous to RGB channels in digital photography) represent distinct omics layers: expression, methylation, mutation, and copy-number variations.

The GENIUS workflow: from multi-omics data transformation to model inference and biological discovery.

Crucially, the genes are ordered based on Hi-C sequencing data, which captures how chromosomes fold and interact in 3D space. This ensures that genes physically proximal in the nucleus are also adjacent in the "genome image," preserving essential biological context that tabular models miss.

Model Architecture

The GENIUS architecture utilizes a four-part Convolutional Neural Network (CNN) designed for both feature extraction and clinical prediction:

  • Encoder: Compresses the multi-channel genomic image into a 128-dimensional "latent vector" (a mathematical summary of the genome).
  • Decoder: Reconstructs an optimized version of the image to ensure the latent space captures essential features.
  • Extractor: Employs convolutional layers to identify complex patterns within the reconstructed genomic landscape.
  • Predictor: A fully connected network that outputs clinical predictions, such as the likelihood of metastasis.
The GENIUS model architecture, featuring an autoencoder (Encoder/Decoder) for feature optimization and a Predictor for clinical outcomes.

To make the model interpretable, the authors used Integrated Gradients (IG). This technique assigns an "attribution score" to every pixel (gene), allowing researchers to see exactly which biological features the AI used to make its prediction.

Results and Validation

The study primarily utilized data from The Cancer Genome Atlas (TCGA), covering 2,332 tumor samples across six cancer types: bladder, uterine, stomach, ovarian, kidney, and colon.

Performance and Biological Discovery

  • Predictive Accuracy: GENIUS achieved a validation AUC of 0.87 for predicting whether a primary tumor would eventually become metastatic.
  • Spatial Significance: Comparative tests showed that Hi-C-informed layouts consistently outperformed random or flat vector layouts, confirming that genomic "geography" contains vital predictive information.
  • External Validation: In independent bladder cancer cohorts (UROMOL and the Mariathasan cohort), 7 out of the top 10 genes identified by GENIUS were significantly associated with patient survival.
  • Driver Gene Identification: The model identified 152 top genes driving metastasis. It highlighted genes like TOP3A (linked to chromosomal instability) and KRT17 (linked to cytoskeletal signaling) as major drivers of aggressive bladder cancer.

Discussion and Future Directions

GENIUS represents a significant proof of concept for multi-omics data transformation and analysis without the need for traditional statistical correction. While the results are promising, several areas for growth remain. The current 198x198 grid provides a relatively "coarse" mapping; future versions could integrate more granular data, such as Topologically Associating Domains (TADs), to better capture 3D genomic architecture. Additionally, while the model performed well on TCGA data, further testing is needed to establish generalizability across rarer cancer types and non-malignant diseases. Finally, while computationally more intensive than linear regression, the framework's ability to integrate five data types simultaneously offers a powerful new tool for biological discovery.

Summary updated

Abstract

The application of next-generation sequencing (NGS) has transformed cancer research. As costs have decreased, NGS has increasingly been applied to generate multiple layers of molecular data from the same samples, covering genomics, transcriptomics, and methylomics. Integrating these types of multi-omics data in a combined analysis is now becoming a common issue with no obvious solution, often handled on an ad-hoc basis, with multi-omics data arriving in a tabular format and analyzed using computationally intensive statistical methods. These methods particularly ignore the spatial orientation of the genome and often apply stringent p-value corrections that likely result in the loss of true positive associations. Here, we present GENIUS (GEnome traNsformatIon and spatial representation of mUltiomicS data), a framework for integrating multi-omics data using deep learning models developed for advanced image analysis. The GENIUS framework is able to transform multi-omics data into images with genes displayed as spatially connected pixels and successfully extract relevant information with respect to the desired output. Here, we demonstrate the utility of GENIUS by applying the framework to multi-omics datasets from the Cancer Genome Atlas. Our results are focused on predicting the development of metastatic cancer from primary tumors, and demonstrate how through model inference, we are able to extract the genes which are driving the model prediction and likely associated with metastatic disease progression. We anticipate our framework to be a starting point and strong proof of concept for multi-omics data transformation and analysis without the need for statistical correction.
BibTeX
@inproceedings{sokac2023author,
  author = {Mateo Sokač and Asbjørn Kjær and Lars Dyrskjøt and Benjamin Haibe-Kains and Hugo J. W. L. Aerts and Nicolai J. Birkbak},
  title = {{Author Response: GENIUS: GEnome traNsformatIon and spatial representation of mUltiomicS data}},
  booktitle = {Openalex Unknown},
  year = {2023},
  doi = {10.7554/elife.87133.2.sa3},
}
Figure