POMP: Pathology-omics Multimodal Pre-training Framework for Cancer Survival Prediction

Suixue Wang, Shilin Zhang, Huiyuan Lai, Weiliang Huo, Qingchen ZhangIJCAI 2025 · DOI

TL;DR

This paper introduces POMP, a novel multimodal pre-training framework that aligns pathology images and multi-omics data before fusion to improve cancer survival prediction. By leveraging contrastive learning and a unique pre-training strategy, POMP achieves state-of-the-art results across six cancer datasets, demonstrating the power of cross-modal alignment for better clinical insights.

What Problem Does This Paper Address?

Predicting cancer survival is crucial for personalized medicine, helping clinicians tailor treatments. Traditionally, this relied on clinical features, which are often laborious and insufficient. The advent of high-throughput sequencing and computational pathology has enabled the integration of multi-omics data (like RNA, miRNA, DNA methylation) and pathology images (whole slide images or WSIs). However, existing multimodal learning methods often struggle to effectively capture the complex interactions between these data types. A key challenge is the lack of proper alignment between modalities before they are fused, leading to suboptimal performance. This paper aims to bridge this gap by proposing a framework that explicitly aligns pathology and omics information from the same patient in a shared latent space.

Methods

The core of POMP is a multimodal pre-training framework designed to learn robust representations from both pathology images and multi-omics data. The architecture consists of three main components:

  • Pathology Encoder: Utilizes Vision Transformers (ViTs) to embed image patches and a Pathology-Transformer to capture inter-patch relationships within a WSI.
  • Omics Encoder: Processes multi-omics data through a shared linear network and a 2-layer Transformer to learn intrinsic and interactive omics information.
  • Multimodal Encoder: Employs an omics-guided cross-attention module, a Transformer, and a residual module to fuse the encoded representations.

What makes POMP novel is its pre-training strategy, which involves three self-supervised tasks:

  1. Pathology-Omics Contrastive Learning (POC): This is the key alignment mechanism. It pulls representations of pathology images and omics data from the same patient closer in the latent space while pushing dissimilar pairs apart, using cosine similarity.
  2. Pathology-Omics Matching (POM): A binary classification task to predict whether a given pathology-omics pair is from the same patient.
  3. Masked Omics Modeling (MOM): A reconstruction task where parts of the omics data are masked, and the model learns to predict them using the remaining omics and pathology information.

During fine-tuning for survival prediction, POMP leverages the cosine similarity between the pathology and omics representations (learned during POC) as the survival risk score, a departure from traditional methods.

Datasets

The study evaluates POMP on six cancer datasets from The Cancer Genome Atlas (TCGA) portal:

  • Colon adenocarcinoma (COAD)
  • Hepatocellular carcinoma (LIHC)
  • Stomach adenocarcinoma (STAD)
  • Breast invasive carcinoma (BRCA)
  • Lung adenocarcinoma (LUAD)
  • Lower grade glioma (LGG)

Each dataset includes pathological images and multi-omics data (RNA-Seq, miRNA, DNA methylation). Patients with missing data or very short survival times (<30 days) were excluded. The experiments use 5-fold cross-validation. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Alignment is Crucial: The paper strongly advocates for aligning multimodal data before fusion. This explicit alignment, particularly through contrastive learning, significantly improves the model's ability to capture cross-modal interactions compared to simply concatenating or attending to features.
  • Cosine Similarity as a Risk Score: Using the cosine similarity between aligned pathology and omics representations as the survival risk score is an effective and intuitive approach. This method aligns with the contrastive learning objective and outperforms other fusion-based risk calculation methods.
  • Pre-training Tasks Synergize: While individual pre-training tasks contribute, their combination, especially POC and MOM, yields the best results. The MOM task, which initially seemed to struggle to converge on its own, becomes beneficial when combined with other tasks, highlighting the importance of a comprehensive pre-training strategy.
  • Omics Data Dominance (in some cases): Interestingly, in some datasets, omics data alone performed better than pathology images alone, suggesting that omics might carry more direct prognostic information in certain cancer types. However, multimodal integration consistently outperformed any single modality.

Key Results

POMP consistently achieved state-of-the-art performance across all six TCGA cancer datasets. The paper reports significant improvements in the concordance index (C-index) over existing methods. For example, POMP outperformed previous best methods by margins of 0.4% to 2.0% on various datasets. The Kaplan-Meier curves also showed statistically significant differences in survival distributions between high-risk and low-risk groups identified by POMP. Ablation studies confirmed the superiority of multimodal integration over unimodal approaches and validated the effectiveness of the proposed cosine similarity-based survival risk calculation.

Limitations and Future Work

The authors acknowledge several areas for future work:

  • Expanding Data Modalities: Integrating additional omics data types (e.g., somatic mutations, copy number alterations) and clinical data to further enrich the model's understanding.
  • Domain-Specific Pre-training: Pre-training the model on specific cancer types (e.g., pancreatic cancer) to learn more prior knowledge about modality relationships.
  • Advanced Gene Selection: Exploring more sophisticated strategies for selecting important genes from multi-omics data to enhance the omics modality's contribution.

Summary updated

BibTeX
@inproceedings{wang2025pomp,
  author = {Suixue Wang and Shilin Zhang and Huiyuan Lai and Weiliang Huo and Qingchen Zhang},
  title = {{POMP: Pathology-omics Multimodal Pre-training Framework for Cancer Survival Prediction}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/869},
}
Figure