Transfer Learning via Optimal Transportation for Integrative Cancer Patient Stratification

Ziyu Liu, Wei Shao, Jie Zhang, Min Zhang, Kun HuangIJCAI 2021 · DOI

Introduction

Accurately predicting survival outcomes for early-stage cancer patients remains a significant challenge in oncology. While multi-view analysis—integrating diverse data types like genomics and histopathology—offers a more holistic view of a patient's condition, it is often hampered by data scarcity and unreliable labels. This paper introduces MVTOT (Multi-view Transfer Learning via Optimal Transport), an unsupervised framework designed to improve patient stratification by leveraging knowledge from data-rich cancer types to inform predictions in data-poor ones.

The MVTOT Framework

MVTOT addresses the limitations of traditional domain adaptation by focusing on the underlying geometric structure of multi-modal data. The framework integrates three key components:

  1. Multi-view Integration: Utilizing Semi-Non-negative Matrix Factorization (Semi-NMF), MVTOT decomposes data from multiple sources into a shared latent feature space, capturing the commonalities between genomics and imaging.
  2. Optimal Transport (OT) for Alignment: Unlike methods that only match statistical means, MVTOT uses OT to define a distance between feature distributions across different cancer types (domains). This approach preserves the geometric relationships within the data, leading to more precise alignment.
  3. Joint Optimization: The model iteratively refines feature dictionaries and domain couplings to minimize cross-cancer discrepancies while maintaining the integrity of each specific dataset.
Figure 1: Visualization of feature alignment. (a) Standard Domain Adaptation via OT vs. (b) the proposed MVTOT framework, showing superior cluster preservation and alignment.

Experimental Evaluation

The researchers validated MVTOT using both synthetic benchmarks and real-world data from The Cancer Genome Atlas (TCGA). The TCGA analysis focused on three early-stage cancers: Breast invasive carcinoma (BRCA), Kidney renal papillary cell carcinoma (KIRP), and Lung squamous cell carcinoma (LUSC). For each patient, the model integrated gene expression data with morphological features extracted from histopathology images.

Synthetic Performance

On synthetic datasets designed to simulate multi-view clusters, MVTOT achieved accuracy (ACC) scores between 0.91 and 0.97. This significantly outperformed baseline methods such as TCA (0.34-0.49) and mSDA (0.33), demonstrating its robustness in recovering latent structures.

Clinical Stratification Results

The primary goal was to stratify patients into high- and low-risk groups. MVTOT demonstrated exceptional performance in cross-cancer transfer scenarios. For instance, transferring knowledge from Breast cancer (BRCA) to Kidney cancer (KIRP) resulted in a highly significant p-value of 0.0062. Even more striking was the Lung (LUSC) to Breast (BRCA) transfer, which achieved a p-value of 0.0004.

Figure 3: Kaplan-Meier survival curves for different cross-cancer transfer scenarios (a-f). MVTOT successfully stratifies patients into high-risk (red) and low-risk (black) groups across most pairs.

Key Insights

  • Geometric Robustness: By treating alignment as a resource allocation problem via Optimal Transport, MVTOT proves more effective than traditional metrics like Maximum Mean Discrepancy (MMD) for complex biological data.
  • Cross-Cancer Synergy: The success of the model confirms that different cancer types share fundamental latent biological features that can be exploited through unsupervised transfer learning.
  • Multi-modal Advantage: Integrating histopathology and genomics consistently provided better stratification than single-modality models, highlighting the complementary nature of these data views.

Limitations and Future Directions

While highly effective, MVTOT showed limitations in certain transfers, such as Kidney to Lung ($p=0.0758$), indicating that some cancer pairs may lack sufficient biological overlap for this specific feature set. Future work will explore extending the model to include more than two data views (e.g., proteomics) and investigating "many-to-one" transfer learning to further enhance predictive power.

Summary updated

BibTeX
@inproceedings{liu2021transfer,
  author = {Ziyu Liu and Wei Shao and Jie Zhang and Min Zhang and Kun Huang},
  title = {{Transfer Learning via Optimal Transportation for Integrative Cancer Patient Stratification}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/380},
}
Figure