acceptodds
Under review as a conference paper at ICLR 2027

Global-to-Instance Structural Alignment for Pathway-based Multi-view Learning in Multimodal Cancer Survival Prediction

Abstract

Multimodal cancer survival prediction aims to leverage complementary information from histopathological images and gene expression to estimate patient-specific prognostic risks. Recent approaches organize genomic information into biological pathways, forming a multi-view learning framework where each pathway serves as an individual view jointly modeled with the shared histopathological modality. However, these methods often treat pathways as independent sources, overlooking complex inter-view relationships among biological pathways. We investigate pathway structures from a population perspective and identify heterogeneous cross-pathway correlations and pathway-survival associations across cancer cohorts. Motivated by these observations, we propose StructAlign, a global-to-instance alignment framework that incorporates population-level pathway structures into patient-specific multi-view representation learning. StructAlign introduces cross-pathway and pathway-survival alignment objectives to preserve inter-pathway dependencies and guide pathway aggregation toward survival-associated patterns. Extensive experiments across multiple cancer cohorts demonstrate that StructAlign improves survival prediction and interpretability through consistent pathway structure preservation, survival-aware view selection, and biologically meaningful attribution.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Cancer Survival Prediction, Multi-view Learning, Structural Alignment

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_91kmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

StructAlign uses cohort-level pathway statistics to guide multimodal cancer survival prediction from histopathology and genomic data. CPA aligns similarities among patient-specific pathway representations with population-level pathway correlations, while PSA guides pathway attention using survival associations. Experiments across six cancer cohorts support the method’s predictive effectiveness and the contribution of both regularizers.

Critical strengths

S1. Clear rationale and complementary regularization objectives.

CPA constrains relationships between pathway representations, while PSA guides pathway aggregation. This provides a coherent and plausible approach to incorporating cohort-level statistical knowledge into patient-specific multimodal learning. The formulation is clear, and the paper specifies that the guidance is computed exclusively from training data.

S2. Empirical support for predictive effectiveness.

The comparative experiments support improved survival prediction, and the component ablations support a useful contribution from each regularizer within the proposed architecture. Together, these findings support the method’s empirical value.

Critical weaknesses

W1. Missing comparisons of structural and attention properties with and without regularization.

The ablation study evaluates predictive performance, whereas the structural and attention analyses are reported only for the full model. Consequently, the experiments do not establish how much CPA and PSA change these properties, or whether such changes accompany predictive improvements. In particular, the model without either regularizer may already exhibit some structural consistency or survival-associated attention preferences. Demonstrating these properties in the full model alone does not establish their incremental improvement due to the proposed regularizers. This limits the evidence supporting the paper’s structural alignment contribution.

W2. Qualitative evidence does not quantify the strength or consistency of the claimed properties across patients.

The paper computes similarity matrices for individual patients and then averages them for visualization. However, it does not report patient-level alignment errors or their distribution. Similarity between the average matrix and the target can conceal substantial variation across patients. Likewise, the attention analysis visualizes pathway preferences for uncensored patients ordered by observed event time, but does not quantify the strength or uncertainty of the attention–survival association. These visualizations provide suggestive evidence, but are insufficient to assess the magnitude and consistency of the claimed properties across patients.

Questions for authors
  1. On the same held-out patients, how do the four configurations—neither regularizer, CPA only, PSA only, and both—compare in structural consistency and survival-associated attention preferences? To what extent does the model without either regularizer already exhibit these properties?
  2. Could the authors report patient-level structural alignment error distributions and quantitative measures of the attention–survival association, and clarify whether these analyses use training, validation, or test patients?
  3. Across cohorts and cross-validation folds, do improvements in these properties accompany improvements in C-index? Are there cases where prediction improves without a corresponding improvement in structural alignment or attention preferences?
Reviewer_BjzPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes StructAlign, a global-to-instance structural alignment framework for multimodal cancer survival prediction using histopathological whole-slide images and genomic information. The main motivation is that biological pathways are not independent but exhibit population-level correlations and heterogeneous associations with patient survival, which are not explicitly modeled by existing pathway-based multi-view learning methods. StructAlign incorporates two complementary alignment objectives. First, Cross-Pathway Alignment (CPA) encourages patient-specific pathway representations to preserve population-level correlations among pathways derived from GSVA pathway activity scores. Second, Pathway-Survival Alignment (PSA) uses pathway-specific Cox regression coefficients and observed survival times to construct patient-specific supervision for pathway importance weights. These objectives are integrated with a pathway-based multimodal learning framework that employs cross-attention between pathway features and histopathological representations, followed by attention-weighted pathway aggregation for survival prediction. Experiments on six TCGA cancer cohorts, including BLCA, BRCA, CESC, GBMLGG, LUAD, and UCEC demonstrate the effectiveness of the proposed method.

Critical strengths
  1. StructAlign introduces a global-to-instance learning framework that incorporates population-level pathway correlations and survival associations into patient-specific multimodal representations, an interesting global-to-instance structural alignment framework for pathway-based multimodal survival prediction.

  2. The proposed method is evaluated on six TCGA cancer cohorts and achieves the highest reported C-index among the compared methods on all six datasets, showing consistent performance improvements.

Critical weaknesses
  1. The PSA objective uses univariate Cox regression to identify risk-promoting and protective pathways. It then constructs a patient-specific guidance distribution using the observed survival time. My main concern therefore is the PSA objective introduces outcome-derived supervision that may undermine the interpretation of patient-specific pathway importance.

  2. The CPA objective aligns within-patient cosine similarities between pathway representations with population-level Pearson correlations computed from pathway activities across patients. However, these quantities capture different relationships: population-level co-variation does not necessarily imply similarity between pathway-specific representations within each individual patient. Consequently, enforcing direct agreement may suppress patient-specific heterogeneity or impose inappropriate structural constraints. Moreover, the structural consistency visualizations largely compare learned representations against the same correlation structure used as the training objective, providing limited independent evidence of biological validity. Although the ablation results demonstrate the benefit of including CPA, they do not establish whether the improvements result specifically from meaningful pathway correlations rather than generic representation regularization.

Questions for authors

Q1: How does PSA distinguish biologically meaningful pathway preferences from explicitly supervised attention patterns?

Q2: Why should population-level pathway correlations match within-patient representation similarities?

Q3: Are the improvements attributable to genuine biological priors or auxiliary regularization?

Q4: How robust are the improvements under different patient splits and censoring patterns?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.