# CoFormerSurv: Collaborative transformer for multi-omics survival analysis

## TL;DR

CoFormerSurv is a new method that uses two connected Transformers to better predict patient survival using multi-omics data, improving precision medicine.

## Contribution

CoFormerSurv introduces a dual Transformer framework that collaboratively models cross-omics interactions and sample relationships for survival analysis.

## Key findings

- CoFormerSurv outperforms existing survival prediction models on real-world datasets.
- The inter-omics Transformer captures complementary information across different omics.
- The inter-sample graph Transformer enhances feature learning by modeling sample relationships.

## Abstract

In the field of biomedicine, advances in high-throughput sequencing have generated vast amounts of high-dimensional multi-omics data. Survival analysis methods with multi-omics data can comprehensively uncover the heterogeneity and complexity of diseases from multiple perspectives, thereby improving prognostic predictions for patients, which is critical for developing personalized treatment strategies in precision medicine. Recently, Transformer architecture has emerged as a dominant paradigm in multiple domains. However, due to the inherent challenges in modeling right-censored data, it remains unclear how to effectively utilize Transformer architecture in multi-omics survival analysis to fully extract complementary information across different omics for improving survival prediction performance. In this work, we propose an innovative collaborative Transformer framework for multi-omics survival analysis, namely CoFormerSurv, with two consecutive Transformer architectures including an inter-omics Transformer and an inter-sample graph Transformer. The inter-omics Transformer learns multiple meaningful feature interactions by multi-head self-attention mechanism to capture and quantify complementary information across different omics, while the inter-sample graph Transformer integrates structural information from the fused multi-omics graph into the Transformer architecture, enabling more effective exploration of neighborhood relationships among samples. The two kinds of Transformer architectures can work collaboratively to generate more comprehensive multi-omics features for improving the Cox-PH model performance in survival analysis. Experimental results on multiple real-world datasets show that our proposed method outperforms both single-Transformer architectures and existing survival prediction models by simultaneously exploring complementary information from inter-omics and cross-sample perspectives.

We propose CoFormerSurv, a collaborative Transformer framework that improves survival prediction with multi-omics data. CoFormerSurv method consists of two complementary components: an inter-omics Transformer that models cross-omics interactions, and an inter-sample graph Transformer that learns the neighborhood relationships among multi-omics samples. By integrating these two perspectives, our dual Transformer architecture enables more comprehensive feature learning and superior performance compared to existing approaches.