Core Concepts
Introducing the Path-GPTOmic framework for improved cancer survival outcome prediction by addressing challenges in multi-modal algorithms.
Abstract
The Path-GPTOmic framework aims to enhance cancer survival outcome prediction by combining pathology images and genomic data. Challenges in existing multi-modal algorithms are addressed, such as overlooking gene-gene interactions and imbalance between modalities during training. The framework regulates the embedding space of a foundation model, scGPT, to adapt it for bulk RNA-seq data. It also proposes a gradient modulation mechanism to ensure both modalities are adequately trained. Evaluation on TCGA datasets shows significantly improved survival prediction accuracy.
Stats
Evaluated on two TCGA datasets.
Achieved substantially improved survival prediction accuracy.
Quotes
"Our contributions can be summarized in three key aspects."
"In this paper, we propose a new multi-modal Path-GPTOmic framework."
"To address the first challenge, we seek to learn a smooth latent space for bulk RNA-seq embeddings."