Pathway-Structured Privileged Distillation for Deployable Computational Pathology

arXiv:2606.02877 · cs.CV · Submitted 2026-06-01 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Pathway-Structured Privileged Distillation for Deployable Computational Pathology".

Jane: Here's the gist of the paper. The gist The Mixture of Pathway Experts (MoPE) is a knowledge-distillation framework that reframes multimodal learning as privileged distillation for histology-only inference, consistently improving WSI-only inference performance relative to baseline methods across diverse public benchmarks and two independent breast cancer cohorts Problem Setting and Motivation The study addresses a deployment asymmetry where RNA profiles can inform model development, but routine use often depends on H&E slides alone <ref:2606.02877#pg3>.

Tom: This asymmetry is compounded by partial observability, as histology cannot be expected to reconstruct the full transcriptomic state, meaning molecular supervision for histology-only deployment should exploit this partial correspondence <ref:2606.02877#pg3>.

Jane: The goal is not to reconstruct molecular profiles from images, but to transfer biologically meaningful structure that is reliably expressed through tissue morphology while preserving modality-specific information that remains unobservable in H&E sections <ref:2606.02877#pg3>.

Tom: Existing knowledge distillation methods often treat molecular representations as fully transferable, which leaves the central question of how to transfer molecular supervision when histology provides only a partial observability of molecular state <ref:2606.02877#pg3>.

Title and authors: Tom: Moving on to who wrote this. The paper is called "Pathway-Structured Privileged Distillation for Deployable Computational Pathology," and the authors are Yongxin Guo, Hao Lu, Onur C. Koyun, Muhammet F. Demir, and Metin N. Gurcan >

Jane: That's a solid team of researchers from Wake Forest University. When you look at that title, "Pathway-Structured Privileged Distillation," it sounds very technical but hints at how they are organizing the knowledge transfer >

Lu: The structure itself suggests they are using pathways as the backbone for this distillation process rather than just dumping raw gene expression data into a model >

Tom: Yeah, and that's where the paper gets specific about what they’re doing: encoding RNA-derived pathways as pathway tokens during training >

The paper's summary: Jane: So, what is the actual mechanism they use to handle that partial observability we talked about earlier? What does MoPE actually do during training?

Tom: During training, they encode the RNA profiles into pathway tokens, and then the whole-slide image branch uses pathway-indexed experts to process those image features >

Lu: The key innovation here is how they transfer that privileged information. They use memory-usage distillation to encourage the RNA pathway tokens and the pathway-indexed experts to share a similar memory basis in certain ways >

The paper's improvements: Jane: What are the specific improvements they claim? Because I’ve heard knowledge distillation before, and I want to know how this differs from what we usually do with molecular supervision.

Tom: They show that MoPE consistently improved WSI-only inference performance relative to baseline methods across diverse public benchmarks and two independent breast cancer cohorts >

Lu: Across five biomarker classification tasks, MoPE showed the strongest WSI-only point estimates compared to the strongest histology-only baseline, AttMIL-MoE, which had an average AUC of eighty-two point three six percent versus seventy-nine point eight four percent >

Conclusion: Jane: So we’ve seen these performance gains in various settings. Where does this leave us regarding how we use molecular data with histology images? What's the final word on this whole pathway-structured distillation idea?

Tom: The conclusion is that MoPE successfully addresses deployment asymmetry by framing molecularly supervised histology prediction as a partial-observability problem >

Lu: They establish a biologically grounded distillation paradigm for partial-observability settings, where RNA supervision improves histology-only models without treating it like a fully reconstructable image target >

Meng: From an engineering side, the paper shows that you can leverage molecular data without needing RNA at inference, which is huge for deployment because it simplifies the pipeline >

Lalam: I’m confident in this direction. The results show that using pathway supervision organized through this distillation process is more useful when it’s structured by pathways than just a raw whole-transcriptome embedding >

Jane: It sounds like a solid framework, but before we wrap up, what are the real limits? What does the paper say they can't do?

Lu: They flag that they require paired WSI-RNA data during training, which means development is currently restricted to cohorts with sufficient multimodal annotation >

Tom: And they also state that pathway expert readouts are descriptive and need validation against spatially resolved molecular measurements before you can claim biological truth >

Yongxin Guo, Hao Lu, Onur C. Koyun, Muhammet F. Demir, Metin N. Gurcan

School of Medicine, Wake Forest University

cs.CV

Submitted: 2026-06-01

Updated: 2026-10-16

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: The gist The Mixture of Pathway Experts (MoPE) is a knowledge-distillation framework that reframes multimodal learning as privileged distillation for histology-only inference, consistently improving

Key concepts

Knowledge Distillation: A technique where a large, complex model (the teacher, trained with both image and RNA data) transfers its learned knowledge to a smaller model (the student). In MoPE, this transfer focuses on teaching the WSI branch how to predict outcomes using the structured biological information from the RNA profiles.
Pathway-Indexed Experts: These are specialized components within the WSI branch, where each expert is trained specifically on one of the 50 Hallmark biological pathways. These experts learn to look for specific visual patterns (morphology tokens) related to that pathway, allowing the model to focus on biologically relevant tissue structures.
Memory-Usage Distillation: This is a method used during training where RNA pathway tokens and pathway-indexed experts are encouraged to share a similar memory basis. Instead of forcing direct feature identity, this encourages the WSI branch to learn a soft, shared representation of molecular supervision, making the transfer more robust for partial observability.

Terminology

Distillation — 4× in this paper · explained in 2 episodes
Benchmark — 1× in this paper · explained in 2 episodes
Mixture of Pathway Experts — 1× in this paper
Multimodal Learning — 1× in this paper · explained in 2 episodes

Summary

The gist The Mixture of Pathway Experts (MoPE) is a knowledge-distillation framework that reframes multimodal learning as privileged distillation for histology-only inference, consistently improving WSI-only inference performance relative to baseline methods across diverse public benchmarks and two independent breast cancer cohorts

Problem Setting and Motivation

The study addresses a deployment asymmetry where RNA profiles can inform model development, but routine use often depends on H&E slides alone <ref:2606.02877#pg3>. This asymmetry is compounded by partial observability, as histology cannot be expected to reconstruct the full transcriptomic state, meaning molecular supervision for histology-only deployment should exploit this partial correspondence <ref:2606.02877#pg3>. The goal is not to reconstruct molecular profiles from images, but to transfer biologically meaningful structure that is reliably expressed through tissue morphology while preserving modality-specific information that remains unobservable in H&E sections <ref:2606.02877#pg3>. Existing knowledge distillation methods often treat molecular representations as fully transferable, which leaves the central question of how to transfer molecular supervision when histology provides only a partial observability of molecular state <ref:2606.02877#pg3>.

MoPE Architecture and Mechanism

MoPE is a pathway-structured knowledge distillation framework for WSI-only prediction of biomarker and survival endpoints from routine H&E images <ref:2606.02877#pg4>. During training, paired RNA profiles are encoded as pathway tokens, and the WSI branch uses pathway-indexed experts to process image features <ref:2606.02877#pg4>. MoPE transfers privileged information through memory-usage distillation, where RNA pathway tokens and pathway-indexed experts are encouraged to use a shared memory basis in similar ways rather than being forced into direct feature identity <ref:2606.02877#pg4>. At inference, the RNA branch is removed and predictions are made from WSIs alone <ref:2606.02877#pg4>. The WSI branch maps patch-level image features into pathway-indexed morphology tokens using pathway-indexed experts, where each expert corresponds to one Hallmark pathway and contains L learnable morphology slots <ref:2606.02877#pg4>. These slots allow one pathway-indexed expert to collect multiple candidate visual patterns before producing a pathway-level WSI token <ref:2606.02877#pg4>. The resulting token e wsi k summarizes WSI morphology evidence for pathway k, and these are aggregated into a slide representation zslide after the slot gate first selects relevant morphology patterns within each pathway expert, and the expert gate then selects relevant pathway-level summaries at the slide level <ref:2606.02877#pg12>.

Performance and Generalizability

MoPE consistently improved WSI-only inference performance relative to baseline methods across five biomarker classification tasks and five survival prediction tasks from The Cancer Genome Atlas (TCGA). Across five biomarker classification tasks, MoPE showed the strongest WSI-only point estimates in the displayed task-level comparisons, with an average AUC of 82.36%, compared with 79.84% for the strongest histology-only baseline (AttMIL-MoE) and 79.03% for the strongest knowledge distillation baseline, G-HANet <ref:2606.02877#pg5>. When aggregating effect sizes across tasks, MoPE improved AUC by a mean of 2.0 percentage points over AttMIL-MoE (95% CI, 0.4–3.6; P = 0.009) and by 2.7 percentage points over G-HANet (95% CI, 1.2–4.2; P = 0.0002) <ref:2606.02877#pg5>. For survival prediction across five TCGA cohorts, MoPE achieved an average WSI-only C-index of 70.96%, compared with 67.48% and 68.81% for the strongest histology-only and knowledge-distillation baselines, respectively <ref:2606.02877#pg5>. Furthermore, in two independent external patient cohorts (OSUWMC and Dartmouth) on the Oncotype DX risk prediction, MoPE maintained consistent discrimination for the ODX prediction, achieving an AUC of 80.89% (95% CI, 77.93–83.85) in OSUWMC and 80.45% (95% CI, 75.73–85.16) in Dartmouth <ref:2606.02877#pg5>.

Interpretability and Inspection

Pathway expert behavior reflects incomplete RNA-WSI correspondence, showing that pathway expert usage was non-uniform and task dependent <ref:2606.02877#pg7>. The top-selected experts accounted for 35.7–51.7% of total usage across tasks, compared with the 10% expected under uniform use of 50 Hallmark pathway experts <ref:2606.02877#pg7>. Human-audited inspection provided bounded, auditable evidence about where the model assigned attention and why pathway-indexed hypotheses could fail <ref:2606.02877#pg9>. In a correctly predicted BRCA-ODX case, the highest-ranked experts corresponded to G2M checkpoint, apical junction and PI3K AKT MTOR SIGNALING pathways <ref:2606.02877#pg9>. The human expert audit found moderate morphological accuracy, image support and pathway compatibility while noting slot-level ambiguity <ref:2606.02877#pg9>.

Limitations and Future Directions

Several limitations remain, including the requirement for paired WSI-RNA data during training, which restricts development to cohorts with sufficient multimodal annotation <ref:2606.02877#pg10>. Pathway expert readouts are descriptive and require validation against spatially resolved molecular measurements before they can support biological claims <ref:2606.02877#pg10>. The external evaluation was retrospective and focused on one molecularly informed breast cancer endpoint, suggesting that prospective and broader multi-institutional studies are needed <ref:2606.02877#pg11>. Future work should compare WSI-side pathway readouts against spatial transcriptomics, multiplex immunohistochemistry or related assays <ref:2606.02877#pg11>.

Conclusion

MoPE addresses a deployment asymmetry by framing molecularly supervised histology prediction as a partial-observability problem and using pathway-structured privileged distillation to improve WSI-only prediction without requiring RNA at inference. The results support a biologically organized distillation paradigm for partial-observability settings, in which RNA supervision improves histology-only models without being treated as a fully reconstructable image target <ref:2606.02877#pg5>. This framework offers a way to use molecular data without assuming that histology contains all molecular information while making the boundary between prediction, observability and biological interpretation explicit <ref:2606.02877#pg10>.

How it works

The model is trained on paired H&E whole-slide image (WSI) and RNA profiles structured by the Hallmark 50 pathways <ref:2606.02877#pg4>.
The WSI branch maps patch-level image features into pathway-indexed morphology tokens using K = 50 pathway-indexed experts, where each expert corresponds to one Hallmark pathway and contains L learnable morphology slots <ref:2606.02877#pg4>.
The RNA branch encodes bulk expression profiles as pathway tokens using the Hallmark gene sets, and is trained with a masked pathway reconstruction objective <ref:2606.02877#pg13>.
Distillation in the latent space uses a memory bank as a shared latent reference system for soft distillation, defined by the loss LKD = 1/K sum k DKL probrna k probwsi k <ref:2606.02877#pg4>.
At inference, the RNA branch is removed and predictions are made from WSIs alone <ref:2606.02877#pg4>.

Results Summary

MoPE showed consistent classification benefits across evaluated endpoints under WSI-only deployment, with an average AUC of 82.36% compared to baselines <ref:2606.02877#pg5>. It also achieved a mean WSI-only C-index of 70.96% for survival prediction across five TCGA cohorts <ref:2606.02877#pg5>. The framework's performance was validated in two independent external breast cancer cohorts, maintaining high discrimination for the Oncotype DX risk prediction task <ref:2606.02877#pg5>.

Ablation Analysis

Ablations support the design rationale; replacing memory-usage distillation with direct cosine alignment between matched RNA and WSI pathway embeddings reduced mean classification AUC from 82.36% to 80.03% <ref:2606.02877#pg7>. Removing pathway supervision resulted in a mean classification AUC decrease to 79.70%, consistent with privileged RNA information being more useful when organized through a pathway vocabulary <ref:2606.02877#pg7>.

Evaluation Protocol

For external evaluation, probability calibration was performed using Platt scaling, and decision-curve analysis was used to compute net benefit across thresholds <ref:2606.02877#pg19>. Uncertainty intervals were estimated by patient- or slide-level bootstrap resampling <ref:2606.02877#pg19>. The human expert audit assessed whether pathway-indexed model readouts were visually plausible and identified failure modes.

Data Availability

The use of data from The Ohio State University Wexner Medical Center (OSUWMC) was approved by the Ohio State University Cancer Institutional Review Board with a waiver of informed consent and HIPAA authorization <ref:2606.02877#pg21>. Imaging data from the TCGA project can be accessed at https://portal.gdc.cancer.gov <ref:2606.02877#pg21>. The corresponding omics data are available via the UCSC Xena browser at https://xenabrowser.net/datapages/ <ref:2606.02877#pg21>. The Dartmouth Breast Cancer Recurrence Risk Dataset is available from its original source <ref:2606.02877#pg21>.

Code Availability

The source code is provided to reviewers as an attached ZIP file alongside the manuscript and will be deposited in a public repository before publication <ref:2606.02877#pg21>.

Author Contributions

Y.G.

Improvements for AI systems

Improve WSI-only inference performance by using memory-usage distillation instead of direct feature matching, as this encourages the WSI branch to learn a softer correspondence to pathway-structured molecular information. This allows for the transfer of RNA pathway tokens and pathway-indexed experts through a shared memory basis, which is crucial when histology provides only a partial observability of molecular state.
Enable more interpretable model behavior by utilizing pathway-indexed readouts, as these create an auditable intermediate layer between slide-level prediction and local image evidence. This allows for inspection of model behavior in a way that is not just direct validation of pathway activation, but rather a means to assess whether highlighted tissue regions support a morphology-linked hypothesis.
Enhance robustness against poor image quality by leveraging slot diversity regularization, as this helps the model represent heterogeneous evidence by preventing collapsing each pathway into a single visual prototype. This ensures that the model can better handle complex histological variations when aggregating information into the final slide representation.
Improve generalization across external settings by utilizing pathway-structured privileged distillation, which has been shown to improve WSI-only prediction in two independent external patient cohorts and maintain consistent discrimination for endpoints like Oncotype DX risk prediction. This suggests the framework is effective even when molecular testing is unavailable, delayed or selectively ordered.
Establish a biologically grounded interpretation of model features by using pathway comparisons to separate different regimes, as the analysis showed that WSI-enriched pathways had stronger histology-derived scores than omics-derived scores in this comparison, providing a coordinate system for inspecting WSI branch gate patterns.