acceptodds
Under review as a conference paper at ICLR 2027

PHISA: Polarized Hypergraphs Induce Spectral Adaptation of Biomedical LLMs for Survival Prediction

Abstract

Tissue organization and molecular information provide complementary perspectives that jointly inform cancer survival prediction. However, existing multimodal survival methods struggle to characterize the diverse relational patterns of tissue interactions, weakening the connection between tissue organization and molecular representation learning. To address this limitation, we propose PHISA, which encodes diverse interaction tendencies into a structural representation of patient tissue to guide molecular representation learning with pretrained biomedical knowledge for survival prediction. Specifically, PHISA first forms a polarized hypergraph by learning phase orientations for regions within each hyperedge, such that relative phases encode cooperative and antagonistic tendencies. The resulting polarized hypergraph governs the propagation of representations across tissue regions, with global structural coordinates and local responses capturing tissue structure. PHISA then aggregates the global and local structural components into a patient tissue descriptor. This structural descriptor modulates spectral coefficients along fixed principal singular directions of pretrained biomedical LLM weights. Finally, PHISA combines tissue representations with molecular representations refined by spectral adaptation conditioned on the patient tissue descriptor for survival prediction. Experiments on four TCGA cohorts show that PHISA achieves a mean C-index of 0.750, representing a 6.1% relative improvement over the strongest evaluated baseline. Ablations show predictive gains from polarized tissue modeling, with additional benefit from conditioning molecular refinement on patient tissue structure. Further analyses reveal spatial variation in tissue region importance and significant survival differences between groups defined by predicted risk scores.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multimodal Learning, Survival Analysis, Computational Pathology, Hypergraph Learning, Biomedical Language Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PGB6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents PHISA for multimodal cancer survival prediction. It models relations among histology regions with a complex-valued polarized hypergraph and adapts BioMistral to patient-level transcriptomic information through tissue-conditioned spectral adaptation (TCSA). Results are reported on four TCGA cohorts using survival discrimination and calibration-related metrics, with component ablations and comparisons among adaptation strategies.

Critical strengths

The paper presents PHISA for multimodal cancer survival prediction. It models relations among histology regions with a complex-valued polarized hypergraph and adapts BioMistral to patient-level transcriptomic information through tissue-conditioned spectral adaptation (TCSA). Results are reported on four TCGA cohorts using survival discrimination and calibration-related metrics, with component ablations and comparisons among adaptation strategies

Critical weaknesses
  1. The biological interpretation of learned phases is not directly validated. Eq. (2) learns complex orientations from region and hyperedge features without supervision on cellular cooperation or antagonism. Theorem 2 establishes a mathematical response under the restrictive opposing-feature assumption, but does not show that learned phase differences correspond to actual tissue interactions. This leaves an important part of the proposed motivation empirically untested.
  2. The adaptation comparison does not fully isolate the proposed mechanism. Table 3 compares patient-structure-conditioned TCSA against task-level LoRA, PiSSA, and SVFT, changing both the conditioning information and the adaptation rule. Although Table 5 reports favorable paired fold-level tests, the large READ gain is obtained from 172 patients with 83.72% censoring (28 observed events). The five fold-level observations do not by themselves establish patient-level uncertainty or external reproducibility.
Questions for authors
  1. Is there independent quantitative evidence relating learned phase differences to annotated tissue-cell interactions or spatial-molecular patterns, beyond the idealized algebraic result?
  2. Does TCSA outperform a parameter-matched patient-conditioned LoRA or another conditional adapter? How much of the gain depends specifically on BioMistral?
  3. How robust are the READ findings under patient-level resampling, alternative survival splits, and independent-cohort evaluation, particularly given the limited number of observed events?
1 comment
AuthorsOpenReview ↗

Follow up on our responses

Dear Reviewer PGB6,

Thank you again for your thoughtful review and constructive suggestions. We have carefully responded to your concerns by providing quantitative evidence relating learned phase differences to cellular spatial patterns (W1/Q1), comparing TCSA with Cond. Adapter under the same patient tissue descriptor and PEFT parameter budget (W2/Q2), examining BioMistral and Mistral under matched adaptation settings (W2/Q2), and evaluating READ robustness through patient resampling and alternative survival splits (W2/Q3).

We would sincerely appreciate it if you could review our responses when convenient and let us know whether these additional explanations and results have helped address your concerns. We would be happy to discuss any remaining questions or provide further clarification.

Thank you very much for your time and consideration.

Authors of Submission #27354

Reviewer_51xdOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PHISA for multimodal cancer survival prediction using WSIs and genomic data. The method introduces polarized hypergraphs to capture cooperative and antagonistic relationships among tissue regions, and uses the resulting structural representations to guide spectral adaptation of a pretrained biomedical LLM for genomic representation refinement. Experiments on four TCGA cohorts show promising performance improvements over existing methods.

Critical strengths
  1. The paper provides detailed theoretical analyses of the proposed methods, with seemingly solid mathematical formulations and proofs.
  2. The experimental results are promising, with consistent improvements over the baselines. The authors also conduct extensive ablation studies and additional analyses.
Critical weaknesses
  1. The methodological motivation is not sufficiently clear, and there is a disconnect between the theory and its application. The paper spends considerable space explaining how the mathematics works, but not enough explaining why these mathematical constructions are needed for survival prediction. For example, my understanding is that polarized hypergraphs aim to address feature cancellation when conventional hypergraphs aggregate opposing regional features. Although the second paragraph of the Introduction provides some biomedical motivation, the Hypergraph Learning subsection in Related Work does not clearly discuss the technical limitations of existing methods for survival prediction. Section 3.1 then directly introduces the mathematical formulation without sufficient motivation. The same concern applies to TCSA in Section 3.2.
  2. The Method section is difficult to follow. The paper combines techniques from Hypergraph Learning and LLM Parameter-Efficient Adaptation, which may not be familiar to readers working on WSI-based survival prediction. I suggest adding necessary preliminaries and simplifying the mathematical presentation. Some intermediate equations and mathematical details could be moved to the appendix, leaving more space for explaining the intuition behind the methods.
  3. The experiments lack a unimodal decomposition of PHISA itself. The unimodal baselines in Table 1 are not equivalent to evaluating PHISA's own components separately. This is particularly important given the multiple complex modules proposed. I suggest adding: (a) WSI-only: Standard graph/hypergraph vs. the proposed PHISA tissue encoder, to isolate the benefit of polarization. (b) Genomics-only: Standard genomic encoder vs. BioMistral-based refinement, to evaluate the contribution of the LLM. (c) Cross-modal adaptation: With the same encoders, compare simpler conditional fusion methods with TCSA to justify the necessity of spectral conditioning.

I am open to raising my score if the authors can adequately address my concerns in their rebuttal.

Questions for authors
  1. What exactly goes wrong when conventional hypergraphs are applied to common histopathological structures? Why would these problems affect survival prediction?
  2. Why are both Macro and Micro structural encodings needed? How do these representations correspond to meaningful tissue structures related to survival prediction?
  3. Why is spectral adaptation of a biomedical LLM necessary? Have the authors compared it with a pretrained LLM without adaptation, a genomic encoder without an LLM, or a simpler Transformer encoder?
  4. Appendix H (Figure 10) shows that most of the spectral update comes from patient-independent static recalibration, while the tissue-conditioned component is relatively small. However, Table 3 shows additional gains from patient-specific conditioning. How can such small patient-specific modifications lead to noticeable improvements in survival prediction?
  5. What is the significance of Proposition 4? How does this theoretical result relate to the survival prediction task?
Reviewer_bA2VOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PHISA, a framework for multimodal cancer survival prediction that integrates whole-slide histopathology images (WSIs) with transcriptomic profiles. The method has two core components. (1) A "polarized hypergraph" that learns a complex phase for each (tissue region, hyperedge) pair in the incidence matrix, using relative phases to encode cooperative versus antagonistic interaction tendencies within tissue regions; regional representations are propagated over this complex-valued structure, and global structural coordinates (Krylov-subspace spectral coordinates) together with local closed-walk responses are aggregated into a "patient tissue descriptor." (2) Topology-Coherent Spectral Adaptation (TCSA): conditioned on the patient tissue descriptor, PHISA modulates spectral coefficients within the fixed principal singular subspace of the attention projections of a pretrained BioMistral-7B, achieving parameter-efficient (approx. 36.9K PEFT parameters), tissue-conditioned refinement of molecular representations, which is then fused with the pathology representation for discrete-interval survival prediction.

Critical strengths

1. Rigorous ablations and causal validation; a complete chain of evidence. The ablation chain — DS-IASE (0.720) → unconditional refinement (0.728) → full conditioning (0.750) → shuffled descriptors (0.717) — together with the descriptor-substitution sensitivity analysis under a fixed model (Figure 2), constitutes strong evidence that the tissue descriptor must match the patient and acts through spectral modulation during molecular refinement. Such direct verification of a patient-level conditioning mechanism is rarely seen in comparable multimodal survival prediction work, and it carries the greatest weight in my assessment.

2. A self-consistent polarized hypergraph design with a closed theory–empirics loop. Theorem 2 provides a clear mechanistic account of how relative phases convert feature cancellation into non-zero responses, and the phase-synchronization analysis in Appendix G empirically examines the spectral bound of Proposition 3 rather than leaving it decorative. The method is conceptually distinct from prior directed/sheaf hypergraph work (encoding cooperation/antagonism rather than direction), and the PEFT efficiency narrative (36.9K vs. 27.3M parameters) is complete.

Critical weaknesses

1. The central claim of "biomedical LLM knowledge" is not isolated, which threatens the paper's core narrative. BioMistral is a text-pretrained LLM; gene expression is projected into its embedding space and passed through a frozen transformer for residual refinement, yet the paper includes no control against a non-biomedical LLM of the same scale (e.g., Mistral-7B), making it impossible to separate the contribution of biomedical pretraining knowledge from that of an arbitrary high-capacity encoder. If the conclusions hold without the "biomedical" qualifier, the title and motivation overstate the contribution. This directly affects the credibility of the claimed contributions.

2. The main performance gain concentrates in the smallest, most heavily censored READ cohort, raising concerns about robustness and generalization. The READ C-index jumps from 0.773 to 0.875 (+0.102), while the other three cohorts improve by only 0.019–0.026. The bulk of the mean improvement is driven by a single small cohort, and no independent external validation is provided (acknowledged by the authors in the limitations). This weakens the strength of the "consistently better on four TCGA cohorts" claim and leaves readers unable to tell whether the gains come from the method itself or from chance fitting on that cohort.

Questions for authors
  1. [Clarification request, materially affects my assessment] Please add a control experiment using a non-biomedical LLM of the same scale (e.g., Mistral-7B) under the identical TCSA pipeline. If biomedical pretraining knowledge is indeed the key factor, I would expect it to perform significantly better than this control; if the results are close, the "Biomedical LLMs" phrasing in the title and motivation needs revision, and my rating would be adjusted accordingly.

  2. [Clarification request] The +0.102 C-index gain on READ is much larger than on the other cohorts. Please comment on the stability of this gain: beyond fold-level paired comparisons, have you examined sensitivity under different hyperparameters/hypergraph constructions? Are there systematic differences in hypergraph scale, number of hyperedges, or the phase network architecture on READ versus the other cohorts?

  3. [Clarification request] The cos(·) modulation in TCSA can flip the sign of a spectral coefficient, which is equivalent to reversing the direction of the corresponding singular vector. What does this mean in SVD semantics? After training, what fraction of spectral coefficients actually changed sign? Was a non-negative parameterization (e.g., sigmoid scaling) considered as an alternative?

Comments on the paper

2 comments
AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Revised Manuscript

Dear Area Chair (AC) and Reviewers,

We appreciate the AC's coordination of the review process and the reviewers' constructive feedback.

Our detailed rebuttal was submitted on November 13, providing clarifications and justifications supported by additional experimental evidence. The manuscript has also been revised accordingly.

As the discussion period ends on November 18, we would greatly appreciate your feedback on our rebuttal and revised manuscript. We remain available and would be happy to discuss and address any remaining concerns.

Best regards,

Authors of Submission #27354

AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

Thank you for your careful reviews and constructive feedback. We are pleased that all three reviewers have responded positively to our rebuttal. In particular, Reviewer 51xd acknowledged that our responses had "adequately addressed the experimental concerns" and raised the score from 2 to 3. Reviewer PGB6 confirmed that the new results and clarifications "addressed the main concerns and strengthened confidence", and also raised the score from 2 to 3. Reviewer bA2V considered "all three questions largely addressed" and maintained the positive rating (3).

We would also like to highlight PHISA's central contribution. Our work addresses a key challenge in multimodal cancer survival prediction: capturing diverse interactions among tissue regions and leveraging patient-specific tissue organization to guide genomic representation learning. It combines polarized hypergraph modeling with tissue-conditioned spectral adaptation of a pretrained biomedical LLM, achieving a mean C-index of 0.750 across four TCGA cohorts, a 6.1% relative improvement over the strongest evaluated baseline.


As the discussion draws to a close, we summarize the main concerns addressed and the additional evidence provided during the rebuttal:

(1) Further analyses strengthen the evidence for polarized tissue modeling. Theoretical clarification and matched propagation comparisons support the preservation of regional feature responses that conventional aggregation may cancel (51xd–Q1/W1). Additional structural encoding comparisons confirm the complementary benefits of Macro and Micro (51xd–Q2), while independent nuclei annotations reveal associations between learned phase differences and cellular spatial patterns (PGB6–Q1/W1).

(2) Additional analyses reinforce the predictive value of tissue-guided spectral adaptation. Patient-matched tissue conditioning improves prediction over unconditioned and shuffled alternatives (0.750 vs. 0.728/0.717). Descriptor substitution further demonstrates its influence on genomic refinement (51xd–Q4). Proposition 4 precisely characterizes weight perturbations through changes in spectral coefficients (51xd–Q5), while coefficient sign analysis and a comparably performing sigmoid alternative clarify cosine modulation (bA2V–Q3).

(3) New finer-grained ablation and comparison experiments clarify the distinct contributions of PHISA's key components. Unimodal evaluations independently support the benefits of tissue modeling and biomedical LLM refinement (51xd–W3). Matched comparisons demonstrate TCSA's advantage over alternative adaptation methods (51xd–Q3/W3; PGB6–Q2/W2) and an additional predictive benefit associated with biomedical pretraining (bA2V–Q1/W1; PGB6–Q2/W2).

(4) Additional robustness checks support PHISA's larger gain on the TCGA-READ cohort. Patient bootstrap analyses, alternative survival splits, and neighborhood sensitivity tests consistently preserve PHISA's advantage over MRePath, the strongest competing baseline (bA2V–Q2/W2; PGB6–Q3/W2). These findings support within-cohort robustness, while independent external validation remains necessary.


The corresponding revisions have been incorporated into the Related Work section, Sections 3–4, and the Appendices, with revised text highlighted in blue for easy reference.

We sincerely appreciate the reviewers' time and constructive engagement throughout the discussion.

Best regards,

Authors of Submission #27354

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.