acceptodds
Withdrawn from ICLR 2027

Can Protein-Derived Knowledge Improve Pathology Foundation Models?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Molecularly guided pathology foundation models (PFMs) exploit transcriptomic or proteomic information to enrich whole-slide image (WSI) representations, yet effectively leveraging large standalone molecular corpora remains challenging. First, existing molecular foundation models encode protein sequences or single-cell states, not the patient-level bulk expression profiles paired with WSIs. Second, because cross-modal supervision is restricted to paired WSI–omics samples, knowledge from standalone molecular corpora reaches the pathology encoder only indirectly, creating a paired-support bottleneck. To address these challenges, we propose a three-stage framework that decouples proteomic knowledge acquisition from cross-modal transfer, yielding ProSlide, a slide-level hierarchical pathology foundation model. First, to close the modality gap, we pretrain a Proteomic Foundation Encoder (PFE) on 12,695 sample-level bulk protein profiles using virtual profile generation and expression-space multi-view pretraining. Second, we pretrain ProSlide, a patch–region–slide encoder, to predict protein expression from paired WSI–protein samples. Third, to relax the paired-support bottleneck, we introduce Prot2Path, a cross-modal relational distillation objective. For each paired sample, it aligns the similarity distributions of the WSI and its protein profile over a shared, frozen bank of PFE-encoded paired and standalone profiles. We evaluate ProSlide on 12 downstream tasks across breast, lung, and renal cancers. Despite being pretrained with only 2,229 WSIs and 12,695 sample-level protein profiles, ProSlide achieves the highest mean accuracy and AUC within each cancer group.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computational Pathology, Pathology Foundation Models, Proteomic Representation Learning, Cross-Modal Knowledge Distillation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5mRrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ProSlide, a hierarchical whole-slide representation model that incorporates molecular knowledge through three stages: protein-feature encoder pretraining, paired WSI-protein supervision, and Prot2Path relational distillation using paired anchors and a bank of paired and standalone protein profiles. The evaluation covers twelve downstream tasks, with component ablations, data-efficiency comparisons, and selected pathologist-annotated qualitative examples.

Critical strengths
  1. Using standalone molecular profiles to supplement limited paired WSI-protein supervision addresses a real data bottleneck. The staged architecture gives a concrete method for transferring protein-derived relational information to slide features.
  2. The evaluation spans twelve tasks, and Table 1 reports the highest mean ACC and AUC for ProSlide in each of the three task groups. The paper also includes ablations and pathologist-annotated qualitative examples, although task-level outcomes vary.
Critical weaknesses
  1. The incremental role of standalone profiles remains incompletely isolated. Stage III uses a paired protein profile to define each teacher query, while standalone profiles appear as similarity references; standalone profiles also support Stage I pretraining. Table 4 does include paired-only and all-profile bank comparisons, but the bank composition and sampling schemes differ, and results are mixed across metrics. It is therefore unclear how much benefit comes specifically from standalone references under matched bank size and training conditions.
  2. The independent novelty and broad generalization claims need stronger evidence. The approach combines established paired supervision and KL-based relational distillation. Although the task-group averages in Table 1 favor ProSlide, Table S10 includes counterexamples: Task 4 AUC is 76.00 for ProSlide versus 81.96 for GigaPath, and Task 11 accuracy is 53.51 versus 61.45 for PRISM. These do not negate the aggregate gains, but they qualify claims of consistent superiority. Differences in pretrained patch encoders and auxiliary protein data also complicate resource-matched attribution.
Questions for authors
  1. Holding Stage I pretraining and bank size fixed, what is the effect of replacing standalone references with paired-only, shuffled, or source-restricted profiles?
  2. With the same patch features, auxiliary data, and compute, does Prot2Path outperform simpler paired alignment and capacity-matched relational distillation?
  3. Are gains statistically reliable at the patient level for individual tasks, and do they extend beyond the cross-cohort lung-subtyping and external protein-retrieval settings already reported?
3 comments
AuthorsOpenReview ↗

Responses to W1 and Q1: Isolating the contribution of standalone profiles

We thank the reviewer for recognizing the practical paired-data bottleneck and the breadth of our evaluation, and for suggesting controls to clarify attribution and generalization. The revision adds factorial pretraining/bank ablations, random-reference controls, simpler molecular teachers, biological prediction tasks, and a matched UNI2 experiment. We address related weaknesses and questions together below; table references follow the revised manuscript.

W1 & Q1: Separating teacher pretraining from reference-bank expansion.

Standalone profiles contribute through two distinct routes: learning the PFE representation space in Stage I and supplying molecular references during Stage III. Table 6 separates these factors. Rows vary PFE pretraining data, while columns vary the reference bank. Each cell reports balanced ACC / AUROC (%), averaged equally across the three cancer task groups: | PFE pretraining data | Bank: paired only | Bank: paired + standalone | | --- | ---: | ---: | | Paired only | 62.96 / 76.57 | 63.04 / 76.64 | | Paired + standalone | 64.22 / 77.24 | 64.68 / 77.70 |

Comparing rows shows that broader teacher pretraining improves downstream performance with either bank composition held fixed. Comparing columns evaluates adding standalone references with the teacher fixed. Bank expansion provides a larger improvement with the full-source teacher, and the highest mean ACC and AUROC occur when standalone profiles are used in both stages. These results distinguish the benefit learned during teacher pretraining from the additional benefit of reference-bank participation.

W1 & Q1: Reference content with the teacher fixed.

Table 6 additionally retains the paired references and replaces the standalone reference embeddings with non-informative random embeddings. This replacement control targets reference content rather than the addition of reference entries. With the same full-source teacher, the AUROC results (%) are: | Reference bank | Renal | Breast | Lung | | --- | ---: | ---: | ---: | | Paired only | 76.61 | 77.65 | 77.47 | | Paired + random references | 75.87 | 77.32 | 76.43 | | Paired + standalone references | 76.67 | 78.53 | 77.90 |

Real standalone references outperform both controls across the three groups. The paired-only condition tests removal of standalone references, whereas the random-reference condition tests replacement of their content. These controls address different aspects of the reviewer's attribution question.

W1 & Q1: Source contribution.

Table 6 also examines excluding CPPA from teacher pretraining. Entries below are mean AUROC (%) across the three cancer task groups: | PFE pretraining data | Bank: paired only | Bank: paired + standalone | | --- | ---: | ---: | | Paired + standalone, excluding CPPA | 76.71 | 76.81 | | Paired + standalone, all sources | 77.24 | 77.70 |

Including CPPA increases mean AUROC from 76.71 to 77.24 with a paired-only bank and from 76.81 to 77.70 with a paired-plus-standalone bank, supporting broader molecular pretraining. This is a teacher-pretraining source ablation, rather than a source-restricted bank experiment with Stage I fixed.

AuthorsOpenReview ↗

Responses to W2 and Q2: Contribution and controlled transfer comparisons

W2 & Q2: Shared molecular supervision and the role of Prot2Path.

Our contribution is a framework for acquiring protein knowledge beyond the WSI-paired cohort and transferring it into a slide encoder. PFE learns sample-level molecular representations from paired and standalone profiles; protein-expression supervision initializes the slide encoder; Prot2Path makes standalone molecular states explicit references during transfer. The contribution lies in this use of standalone molecular knowledge, rather than in introducing KL divergence itself.

Both direct alignment and Prot2Path transfer information from the same PFE teacher. Tables 4 and 5 compare them within the ProSlide framework using the same frozen patch-feature backbone. The following scores are averaged equally across the three cancer task groups (%): | Training configuration | Mean balanced ACC | Mean AUROC | | --- | ---: | ---: | | Stage II only | 62.88 | 75.98 | | Stage II + direct alignment | 63.88 | 77.98 | | Stage II + Prot2Path | 64.68 | 77.70 |

Both transfer objectives improve aggregate performance over Stage II, indicating that the benefit of PFE supervision is not confined to one alignment objective. Prot2Path extends paired alignment with soft relational supervision: Table S12 assigns 56.82% of teacher probability to the matched protein reference and distributes the remaining 43.18% across other profiles. This retains a paired molecular anchor while teaching relationships to additional molecular states. We adopt Prot2Path for this explicit participation of standalone references and its higher balanced accuracy across all three cancer groups.

Q2: Relational and simpler-teacher controls.

Table 5 includes paired-bank Soft KL, which uses the same type of relational objective and PFE representation space but restricts references to paired profiles. Prot2Path improves mean ACC and AUROC in each cancer group relative to this control. The frozen reference bank adds molecular targets without introducing another trainable encoder. These within-framework comparisons are more direct evidence about the transfer design than comparisons with separately pretrained slide models.

Table S11 further compares teacher representations under Soft KL. The reported AUROCs (%) are: | Teacher representation | Renal | Breast | Lung | | --- | ---: | ---: | ---: | | Raw Protein | 75.18 | 76.36 | 75.49 | | MLP-AE | 76.01 | 77.49 | 75.95 | | PFE | 76.67 | 78.53 | 77.90 |

PFE obtains higher mean ACC and AUROC than both simpler teachers in all groups, supporting its learned representations as a useful source of supervision beyond the original measurements.

W2 & Q2: Dependence on the pretrained patch encoder.

Table S13 replaces CONCH v1.5 with frozen vision-only UNI2 while keeping the slide aggregation architecture, training data, and evaluation protocol unchanged. Each cell reports balanced ACC / AUROC (%): | Patch encoder | Stage | Renal | Breast | Lung | | --- | --- | ---: | ---: | ---: | | CONCH v1.5 | II | 58.07 / 75.86 | 69.22 / 76.69 | 61.34 / 75.39 | | CONCH v1.5 | III | 60.85 / 76.67 | 70.02 / 78.53 | 63.17 / 77.90 | | UNI2 | II | 59.57 / 75.88 | 69.41 / 78.00 | 61.34 / 76.07 | | UNI2 | III | 60.78 / 76.39 | 70.92 / 78.58 | 63.42 / 78.08 |

Stage III improves both metrics across all three cancer groups with either patch encoder. This extends the observed transfer benefit beyond CONCH v1.5 and provides a controlled comparison within each backbone, using the same auxiliary molecular data.

With UNI2, ProSlide achieves slightly higher mean balanced accuracy and a similar mean AUROC compared with CONCH v1.5. We retain CONCH v1.5 for the main experiments to match the patch encoder used by several key baselines in our evaluation, including FEATHER, TITAN, and CARE (Table S9), thereby reducing differences attributable to the patch-feature backbone.

AuthorsOpenReview ↗

Responses to W2 and Q3: Statistical evidence and scope of generalization

W2 & Q3: Overall trends and individual-task outcomes.

No evaluated foundation model leads on every task in Table S10. We acknowledge the examples highlighted by the reviewer: ProSlide trails GigaPath in Task 4 AUROC and PRISM in Task 11 balanced accuracy. Sections 4.2 and D.3 discuss endpoint-dependent performance. Our aggregate result is the highest observed mean balanced ACC and AUROC within each of the three selected cancer task groups, rather than leadership on every endpoint.

Figure S1 provides complementary context: ProSlide achieves strong aggregate performance with 2,229 WSIs for slide-level pretraining, and its inference GPU memory usage is lower than TITAN, PRISM, and GigaPath across the evaluated patch counts, and lower than CARE at larger patch counts. These results characterize the combination of predictive performance, slide-level data requirements, and inference memory efficiency.

Q3: Individual-task paired confidence intervals against FEATHER.

To assess the reliability of the gains on individual tasks, we compare ProSlide with FEATHER using paired 95% confidence intervals. Both models use CONCH v1.5 patch features in our evaluation. Differences below are ProSlide minus FEATHER, in percentage points (pp). Task numbers follow Table S10. | Task | Balanced ACC difference [95% CI] | AUROC difference [95% CI] | | --- | --- | --- | | Task 1 | +6.77 [6.55, 6.99] | +2.68 [2.58, 2.78] | | Task 2 | +8.13 [7.96, 8.30] | +2.61 [2.57, 2.65] | | Task 3 | +2.25 [1.97, 2.53] | +9.46 [8.92, 10.00] | | Task 4 | +6.45 [6.06, 6.84] | +8.10 [7.87, 8.33] | | Task 5 | -0.32 [-0.56, -0.08] | +0.77 [0.59, 0.95] | | Task 6 | +4.89 [4.65, 5.13] | +4.53 [4.32, 4.74] | | Task 7 | +7.85 [7.70, 8.00] | +8.83 [8.72, 8.94] | | Task 8 | +0.50 [0.44, 0.56] | +0.81 [0.73, 0.89] | | Task 9 | +3.47 [3.38, 3.56] | +3.08 [2.99, 3.17] | | Task 10 | +0.63 [0.46, 0.80] | -0.11 [-0.25, 0.03] | | Task 11 | -1.77 [-1.84, -1.70] | +2.34 [2.16, 2.52] | | Task 12 | +9.47 [9.15, 9.79] | +7.17 [6.82, 7.52] |

The reported intervals lie entirely above zero on 10 of 12 tasks for balanced ACC and 11 of 12 tasks for AUROC, with both metrics favoring ProSlide on 9 tasks. These positive differences span renal, breast, and lung molecular prediction as well as histological subtyping, extending beyond cross-cohort lung subtyping. Balanced ACC favors FEATHER on Tasks 5 and 11, while the Task 10 AUROC interval includes zero. These individual-task comparisons support the overall advantage over FEATHER while identifying the endpoints where the evidence differs.

W2 & Q3: Evidence beyond lung subtyping and protein retrieval.

Table S13 shows Stage III improvements in both metrics across renal, breast, and lung task groups with both patch encoders. This supports transfer beyond the lung-subtyping examples, although individual endpoints remain variable.

We also add prediction of RNA-derived tumor microenvironment states on three external CPTAC cohorts (Table 3). Patho-Bench labels group gene-expression-derived ESTIMATE scores into low, medium, and high categories reflecting stromal and immune infiltration. Under the same patient splits and logistic-regression protocol, the AUROCs (%) are: | Cohort | Patients | Raw Protein | Protein-BERT | MLP-AE | PFE | | --- | ---: | ---: | ---: | ---: | ---: | | CPTAC-BRCA | 94 | 69.33 | 66.15 | 70.49 | 73.13 | | CPTAC-LSCC | 88 | 71.01 | 80.45 | 78.69 | 82.01 | | CPTAC-LUAD | 73 | 57.61 | 65.99 | 65.60 | 69.83 |

PFE achieves the highest reported mean ACC and AUROC in all three cohorts. This provides biological prediction evidence beyond near-identity retrieval by using independently measured RNA-derived labels. The task evaluates the molecular teacher; the UNI2 and Stage II/III comparisons separately assess transfer to the slide encoder. Together, these experiments broaden the evidence across molecular endpoints and patch encoders while keeping the scope of each evaluation explicit.

Reviewer_3H5kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper investigates whether standalone bulk proteomic profiles can provide additional knowledge for pathology foundation models when paired WSI-omics data are limited. The authors propose a three-stage framework consisting of: (1) a Proteomic Foundation Encoder (PFE), pretrained on 12,695 RPPA profiles using expression-space augmentations and self-distillation; (2) a hierarchical WSI encoder, ProSlide, pretrained to predict protein expression from paired WSIs; and (3) Prot2Path, a relational distillation objective that aligns WSI-derived and protein-derived similarity distributions over a shared reference bank containing paired and standalone proteomic profiles.

The resulting WSI encoder is evaluated on 12 classification tasks across renal, breast, and lung cancers. ProSlide achieves the highest average balanced accuracy and AUROC within each cancer group, with additional ablations examining protein supervision, PFE pretraining, and reference-bank construction.

Critical strengths
  1. Well-motivated problem formulation and practical cross-modal framework

The paper addresses a critical limitation in computational pathology: while large-scale molecular datasets are increasingly available, matched whole-slide images (WSIs) are scarce, restricting the utility of unpaired molecular data. Decoupling proteomic feature extraction from cross-modal distillation offers a clean, practical strategy for leveraging standalone molecular references. The proposed reference-bank distillation mechanism allows unpaired proteomic knowledge to guide WSI representation learning without requiring molecular measurements at downstream test time.

  1. Evaluation across multiple downstream tasks and competitive baselines

The authors evaluate ProSlide on 12 downstream tasks across renal, breast, and lung cancers, covering molecular biomarker prediction and histological classification, including cross-cohort settings. Comparisons with several representative slide-level pathology foundation models provide useful empirical context. The inclusion of detailed task-level results in the supplementary file also allows readers to examine performance variations across different endpoints.

Critical weaknesses
  1. Insufficient isolation and potential self-matching bias of standalone proteomic references

While overcoming the paired-support bottleneck via standalone molecular profiles is highlighted as the primary novelty, its standalone contribution is modest and methodologically confounded. In Table 4, adding standalone references to paired-bank KD yields marginal AUROC gains on renal, breast, and lung tasks (0.060.06, 0.880.88, and 0.430.43 percentage points, respectively), while direct alignment actually outperforms Prot2Path on two cancer groups. More critically, because the reference bank retains all 1,3381,338 paired training profiles—including the query's own proteomic target—alongside only 256256 sampled standalone profiles, the distillation loss (at a sharp temperature of 0.050.05) may be trivially dominated by self-matching. To substantiate the core novelty claim, the authors must disentangle teacher pretraining from reference-bank expansion using a controlled factorial ablation, evaluate performance after removing self-matches and same-patient duplicates, and report the empirical probability mass assigned to standalone entries during training.

  1. Unconvincing biological validity and downstream statistical superiority

The evaluation protocol does not sufficiently prove that the model captures deep biological structure rather than surface-level dataset artifacts. The PFE concatenates tumor samples, cell lines, and CPPA perturbation profiles into a small 3838-protein panel; without source-stratified ablations or harmonization details, it remains unclear whether the model learns biological states or source-specific batch effects. Furthermore, the masked-profile retrieval in Table 2 provides weak representation validation, as queries differ from gallery profiles by only one to three measurements out of 3838—a task likely solvable by raw measurement similarity or PCA baselines. Finally, downstream performance claims are overrepresented: ProSlide achieves top AUROC on only two of the 1212 individual tasks in Table S10, with several molecular endpoints favoring prior methods. Given the absence of formal statistical significance testing and the surprisingly tiny standard deviations reported across repeated patient-level splits, the claim of general superiority is not adequately supported.

Questions for authors

While the paper addresses an interesting problem, several critical points regarding causal attribution, representation learning, and evaluation rigor need to be addressed to strengthen the submission.

  1. To clarify how much improvement is causally attributable to standalone proteomic data, the authors should separate the impact of molecular teacher pretraining from reference-bank augmentation. Specifically, a 2×22 \times 2 ablation experiment would be nice to verify varying whether PFE is trained on paired-only versus paired-plus-standalone profiles, and whether the distillation bank contains paired-only versus paired-plus-standalone references. Including a control bank with non-informative reference embeddings would also help confirm whether the gains arise from meaningful biological relationships rather than simple memory expansion.

  2. The effectiveness of the reference-bank objective must be validated in the absence of trivial self-matching shortcuts. The authors should report the empirical probability mass assigned to the query's own reference, the entropy of teacher similarity distributions, and the mass assigned to standalone profiles. To demonstrate true generalization, it is essential to evaluate performance after removing the query’s own profile and any same-patient duplicates from the reference bank.

  3. The authors should clarify how biological signals are distinguished from source-specific batch effects, particularly regarding how absolute abundance values and treatment-induced changes are harmonized between baseline cell-line data and CPPA perturbation profiles. Source-stratified ablations and benchmarks against raw normalized similarity, PCA, and lightweight molecular baselines would better demonstrate representation quality. Evaluating independent prediction of held-out protein or pathway states would offer far stronger biological validation than near-identity profile retrieval.

7 comments
AuthorsOpenReview ↗

Responses to W1, Q1, and Q2

We thank the reviewer for recognizing the motivation and practical value of our framework and for suggesting concrete experiments to clarify its contributions. We have added controlled ablations, teacher-distribution diagnostics, and biological prediction experiments. All table and section references below refer to the revised manuscript.

W1 & Q1: Separating teacher pretraining from reference-bank expansion.

We have added the requested 2 × 2 ablation to separate the effects of standalone data in PFE pretraining and in the distillation reference bank. Rows vary the teacher's pretraining data, and columns vary the bank composition. Each cell reports mean balanced ACC / AUROC (%), averaged equally across the renal, breast, and lung task groups using the results in Table 6. | PFE pretraining data | Bank: paired only | Bank: paired + standalone | | --- | ---: | ---: | | Paired only | 62.96 / 76.57 | 63.04 / 76.64 | | Paired + standalone | 64.22 / 77.24 | 64.68 / 77.70 |

The two factors have distinguishable effects. Comparing rows within either column shows that including standalone data in PFE pretraining improves downstream performance even when the reference bank is held fixed. In particular, the improvement with a paired-only bank demonstrates a contribution through teacher pretraining that does not require bank expansion. Comparing columns within either row isolates the effect of adding standalone references while holding the teacher fixed. Bank expansion yields a larger improvement with the teacher pretrained on paired-plus-standalone data than with the paired-only teacher, indicating that its utility depends on the teacher's pretraining. The highest mean ACC and AUROC are obtained when standalone data are used in both stages, supporting their complementary roles in learning the molecular representation space and providing additional references during cross-modal distillation.

Table 6 also includes the suggested non-informative reference control: paired references are retained, while standalone references are replaced with random embeddings. With the same full-source teacher, the AUROC results (%) are: | Reference bank | Renal | Breast | Lung | | --- | ---: | ---: | ---: | | Paired only | 76.61 | 77.65 | 77.47 | | Paired + random references | 75.87 | 77.32 | 76.43 | | Paired + standalone references | 76.67 | 78.53 | 77.90 |

Real standalone references outperform both controls across all three cancer groups, while random references perform worse than the paired-only bank. This supports the importance of reference content beyond simply increasing the number of entries, without establishing a specific biological mechanism.

W1 & Q2: Self-matching and the empirical teacher distribution.

Self-exclusion in Table S12 removes the query's own reference and all other same-patient records. Other paired and standalone values are mean probability masses per reference, whereas entropy is computed over the full teacher distribution. | Configuration | Self mass (%) | Other paired mean mass (%) | Standalone mean mass (%) | Entropy (nats) | ACC (%) | AUROC (%) | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Self-included | 56.82 | 0.030 | 0.012 | 2.3212 | 64.68 | 77.70 | | Self-excluded | — | 0.069 | 0.031 | 4.1078 | 64.38 | 76.95 |

The self-reference provides a matched molecular anchor: the query protein has maximal similarity to its own reference, so a high target probability is expected at the chosen temperature. Its 56.82% mass emphasizes alignment with the corresponding molecular state, while the remaining 43.18% supervises relationships to other profiles. The objective thus combines paired alignment with a distributed relational target.

On a per-reference basis, other paired profiles receive approximately 2.5 times the mass of standalone profiles when self-references are retained, and 2.2 times after exclusion. This is consistent with closer molecular neighborhoods within the paired cohort than across the heterogeneous standalone sources. It does not by itself identify the biological origin of these differences. PFE supplies graded reference weights, allowing the slide encoder to preserve similarities and differences between molecular states rather than treating all non-matching profiles equally. The benefit of real standalone references is supported separately by their advantage over random references in Table 6.

Excluding the self-reference from both distributions increases entropy from 2.3212 to 4.1078 nats, producing a more distributed target. The resulting ACC/AUROC is 64.38/76.95, compared with 64.68/77.70 when the matched anchor is retained. Thus, removing self-matching changes the target structure without eliminating downstream predictive performance. Together, these results support combining the paired molecular anchor with PFE-guided relationships to other profiles, while quantifying the effect of removing that anchor.

AuthorsOpenReview ↗

Responses to W2 and Q3

W2 & Q3: Source handling and scaling self-supervised pretraining.

As described in Appendix A.1–A.3, proteins and phosphorylation sites are mapped to a shared 38-target panel, with distinct sites retained separately. Missing values use a learnable mask, and external profiles are filtered for protein coverage. We retain each resource's processed measurements during dataset construction and standardize valid values using protein-specific statistics during PFE training. This provides consistent feature identities, missing-value handling, and numerical scaling. CPPA represents treatment-induced changes; the other resources provide baseline abundance.

Expanding beyond paired profiles is motivated by the potential of self-supervised learning to benefit from greater data scale and diversity. Standalone profiles expose PFE to molecular states beyond the limited WSI-paired cohort, potentially improving its representations for distillation. Table 6 tests this motivation with broader pretraining corpora. Each entry below is mean AUROC (%) across the three cancer task groups: | PFE pretraining data | Bank: paired only | Bank: paired + standalone | | --- | ---: | ---: | | Paired only | 76.57 | 76.64 | | Paired + standalone, excluding CPPA | 76.71 | 76.81 | | Paired + standalone, all sources | 77.24 | 77.70 |

With either bank configuration, mean AUROC improves when pretraining includes standalone profiles without CPPA and improves further with CPPA, supporting broader molecular pretraining.

W2 & Q3: Biological validation beyond near-identity retrieval and comparison with simpler molecular representations.

We agree that recovering a profile after masking only a few measurements is insufficient on its own to establish biological representation quality. We have added a complementary prediction experiment using RNA-derived tumor microenvironment states (Section 4.3, Table 3). The Patho-Bench labels categorize gene-expression-derived ESTIMATE scores into low, medium, and high states reflecting stromal and immune infiltration. This evaluates prediction of an independently measured molecular endpoint, rather than retrieval of the original protein profile.

Table 3 compares PFE, standardized raw protein measurements, Protein-BERT, and a lightweight MLP autoencoder on the same patient splits and logistic-regression protocol. The reported AUROCs are: | Cohort | Patients | Raw Protein | Protein-BERT | MLP-AE | PFE | | --- | ---: | ---: | ---: | ---: | ---: | | CPTAC-BRCA | 94 | 69.33 | 66.15 | 70.49 | 73.13 | | CPTAC-LSCC | 88 | 71.01 | 80.45 | 78.69 | 82.01 | | CPTAC-LUAD | 73 | 57.61 | 65.99 | 65.60 | 69.83 |

Table 3 reports the highest mean ACC and AUROC for PFE in all three cohorts, supporting an association between its protein representations and transcriptome-derived microenvironment states. This is complementary biological prediction evidence, rather than a mechanistic validation of individual pathways.

We additionally compare simpler teacher representations in the cross-modal transfer setting. Under the same Soft KL objective, Table S11 reports the following AUROCs (%): | Teacher representation | Renal | Breast | Lung | | --- | ---: | ---: | ---: | | Raw Protein | 75.18 | 76.36 | 75.49 | | MLP-AE | 76.01 | 77.49 | 75.95 | | PFE | 76.67 | 78.53 | 77.90 |

PFE achieves higher mean AUROC and balanced accuracy than both simpler teachers in all three cancer groups. These results support the value of its learned representations for cross-modal distillation beyond using standardized raw measurements or a lightweight autoencoder.

AuthorsOpenReview ↗

Response to W2: Overall performance, efficiency, and statistical interpretation

W2: Task-level performance and the overall evaluation trend.

No evaluated foundation model leads on every task in Table S10. Different endpoints favor different representations, so we assess the overall pattern across tasks alongside the individual results. ProSlide achieves the highest AUROC on two of the twelve tasks and the highest mean balanced ACC and AUROC within each of the renal, breast, and lung task groups. This indicates competitive performance across the selected task groups, while Sections 4.2 and D.3 acknowledge endpoints where other models perform better.

The practical value of this performance should also be considered alongside pretraining data requirements and inference memory usage. Figure S1a shows that ProSlide achieves the highest aggregate performance among the compared models with only 2,229 WSIs for slide-level pretraining. This supports the value of proteomic supervision when paired slide data are limited.

Figure S1b further shows lower inference GPU memory usage than TITAN, PRISM, and GigaPath across the evaluated patch counts, and lower usage than CARE at larger patch counts. This supports processing large slides with more modest GPU memory requirements. Together, these results demonstrate a favorable combination of aggregate predictive performance, limited slide-level pretraining data, and inference memory efficiency. Our argument concerns this overall combination rather than a claim to lead every individual task.

W2: Statistical interpretation and paired confidence intervals.

As described in Section 4.1 and Appendix D.1, downstream evaluations use repeated patient-level splits, with 5 or 50 repetitions depending on the task. Table 1 reports the arithmetic mean of task-level standard deviations, which describes evaluation variability rather than uncertainty in a performance difference.

To assess the comparison with direct alignment, we have computed 95% paired confidence intervals using the repeated downstream evaluation results, pairing the two methods on the same split. Differences below are Prot2Path minus direct cosine alignment, in percentage points; ACC denotes balanced accuracy. | Cancer group | Metric | Mean difference | 95% paired CI | | --- | --- | ---: | --- | | Renal | ACC | +0.78 | [0.19, 1.37] | | Renal | AUROC | -1.19 | [-2.59, 0.21] | | Breast | ACC | +0.80 | [-0.08, 1.68] | | Breast | AUROC | -0.16 | [-0.95, 0.63] | | Lung | ACC | +0.83 | [0.14, 1.52] | | Lung | AUROC | +0.51 | [-0.57, 1.59] |

The paired results provide stronger evidence for Prot2Path's balanced-accuracy gains than for an AUROC advantage over direct alignment. Viewed alongside Table 6, the factorial ablation separates the contributions of standalone data through teacher pretraining and reference-bank expansion, while this paired analysis assesses the transfer objective. These confidence intervals concern repeated downstream evaluations, rather than independent pretraining runs. Our broader comparison with other foundation models remains a statement about the highest observed mean ACC and AUROC within each cancer task group.

Reviewer_vYgnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ProSlide, a pathology model trained using protein expression supervision and distillation from a proteomic encoder. Prot2Path incorporates unpaired protein profiles through a reference bank. The motivatin is reasonable, and the downstream results are competitive. However, the experiments do not convincingly establish the advantage of the proposed distillation method or adequately validate the proteomic representation it relies on.

Critical strengths

Using unpaired molecular data to improve slide representations addresses a practical limitation of paired multimodal training .

Evaluation covers twelve tasks across three cancer groups includes external cohorts and provides useful comparisons between all of the training stages.

Critical weaknesses

A. The benefit of Prot2Path over simpler alignment is quite unclear. In Tab. 4, direct cosine alignment achieves high AUC for renal tasks (77.86 vs 76.67) and breast tasks (78.69 vs 78.53). Prot2Path improves balanced accuracy, but the results are mixed across metrics. Adding standalone references to paired-bank distillation improves AUC by only 0.06. These comparisons need uncertainty estimates and repeated pretraining runs to establish whether the additional machinery provides a reliable benefit. Tab. 1 avg task level standard deviations which does not quantify uncertainty in the aggregate performance difference.

The proteomic encoder needs more convincing validation. The retrieval experiment masks only one to three of the 38 protein measurements and asks the model to identify the original profile. This primarily tests robustness to limited masking , it provides little evidence that the encoder learns useful biological info or relationships beyond those already present in the input vector. Raw-profile similarity and simpler learned encoders are important missing controls both for retrieval and downstream distillation. The data construction also deserves closer examination eg. App. A combines baseline protein abundance with treatment-induced changes from CPPA. These represent different quantities, and protein-wise standardization alone does not establish that their similarities are biologically comparable.

Questions for authors

A. Does Prot2Path outperform direct alignment across independent pretraining runs? with paired confidence intervals for the downstream differences.

B. How much of the downstream improvement remains when using raw protein profiles or a small autoencoder as the teacher? C. What happens when CPPA is excluded and how is overlap between the TCPA derived CCLE profiles and the separate CCLE collection handled?

D. What does the learned proteomic teacher contribute beyond the original measurements? The retrieval task masks only one to three of 38 targets before retrieving the original profile. How well does a distance-based baseline perform using the remaining observed measurements? More importantly does pfe outperform distillation based on standardized protein profiles or a simple learned representation under the same downstream protocol? Without these controls, the necessity of the proteomic foundation encoder remains unresolved.

E. As written, the reference bank contains all paired training profiles, including the query’s own profile, and the teacher temperature is 0.05. Is the selfreference excluded when constructing the target?? If not please report the probability mass assigned to the selfreference and to standalone profiles along with an ablation removing the selfreference. This would establish weather the obj. uses the proposed molecular relationships or mainly rewards identifying the paired profile.

F. How dependent are the reported gains on conch1.5? Please repeat the protein supervision and full prot2path experiments using a frozen vision only encoder such as uni 2, keeping the slide aggregation architecture, training cohort, and evaluation protocol matched. This would establish whether the benefit of proteomic distillation extends across patch encoder. Comparisons with existing slide models which also change the aggregation architecture and training data, do not resolve this question.

4 comments
AuthorsOpenReview ↗

Responses to W1 and Q-A: Alignment, standalone references, and statistical evidence

We thank the reviewer for recognizing the practical motivation and breadth of the evaluation, and for identifying the controls needed to assess the molecular teacher and transfer objective. The revision adds simpler teacher baselines, controlled pretraining/reference-bank ablations, biological prediction tasks, self-reference exclusion, and a matched UNI2 experiment. References below follow the revised manuscript. We group the alignment concerns as W1 and the proteomic representation/data concerns as W2.

W1 & Q-A: Shared molecular supervision and the role of relational transfer.

Our central motivation is to learn from standalone proteomic data and transfer that molecular knowledge to a pathology encoder. Both direct alignment and Prot2Path transfer information from the same PFE teacher. Tables 4 and 5 show their effects relative to Stage II, before teacher distillation. Results below are averaged equally across the three cancer task groups (%): | Training configuration | Mean balanced ACC | Mean AUROC | | --- | ---: | ---: | | Stage II only | 62.88 | 75.98 | | Stage II + direct alignment | 63.88 | 77.98 | | Stage II + Prot2Path | 64.68 | 77.70 |

Both objectives improve aggregate performance over Stage II. For each metric, either objective's improvement over Stage II exceeds the difference between the two objectives. This suggests that the benefit of PFE-derived supervision is not confined to a single alignment objective. Table S11 further isolates teacher quality: under the same Soft KL objective, PFE outperforms raw measurements and an MLP autoencoder, supporting the value of its learned molecular representations.

Table S12 helps explain the relationship between the two transfer objectives. The self-reference receives 56.82% of the teacher probability, retaining the paired protein as the main molecular anchor; the remaining 43.18% supervises relationships to other profiles. Prot2Path therefore extends paired alignment with soft relational supervision. It matches a reference-based probability distribution rather than directly optimizing the cosine similarity of the paired vectors. We adopt Prot2Path because it combines this paired anchor with explicit participation of standalone references and achieves higher balanced accuracy across all three cancer groups. Its contribution is this broader route for molecular knowledge transfer, supported by the reference-bank controls below.

W1: Separating teacher pretraining from reference-bank expansion.

Standalone data affect both the teacher's representation space and the references available during distillation. Table 6 separates these factors. Each cell below reports mean balanced ACC / AUROC (%), averaged equally across the three cancer task groups: | PFE pretraining data | Bank: paired only | Bank: paired + standalone | | --- | ---: | ---: | | Paired only | 62.96 / 76.57 | 63.04 / 76.64 | | Paired + standalone | 64.22 / 77.24 | 64.68 / 77.70 | Comparing rows isolates the effect of expanding teacher pretraining while holding the bank composition fixed. Comparing columns isolates reference-bank expansion with the teacher fixed. Broader teacher pretraining improves both metrics under either bank configuration, and bank expansion provides a larger improvement with the full-source teacher. The highest means occur when standalone profiles are used in both stages, supporting their complementary roles.

Table 6 further compares real and random references with the same full-source teacher. AUROC (%) is shown below: | Reference bank | Renal | Breast | Lung | | --- | ---: | ---: | ---: | | Paired only | 76.61 | 77.65 | 77.47 | | Paired + random references | 75.87 | 77.32 | 76.43 | | Paired + standalone references | 76.67 | 78.53 | 77.90 | Real references outperform both controls, while random references perform worse than the paired-only bank.

AuthorsOpenReview ↗

Response to W1 and Q-A: Paired confidence intervals

W1 & Q-A: Uncertainty in the comparison with direct alignment.

We have computed 95% paired confidence intervals for the performance differences between Prot2Path and direct cosine alignment using the repeated downstream evaluation results. Each task uses 5 or 50 patient-level evaluation splits, depending on its protocol, with the two methods paired on the same split. Differences are defined as Prot2Path minus direct alignment, in percentage points; ACC denotes balanced accuracy. | Cancer group | Metric | Mean difference | 95% paired CI | | --- | --- | ---: | --- | | Renal | ACC | +0.78 | [0.19, 1.37] | | Renal | AUROC | -1.19 | [-2.59, 0.21] | | Breast | ACC | +0.80 | [-0.08, 1.68] | | Breast | AUROC | -0.16 | [-0.95, 0.63] | | Lung | ACC | +0.83 | [0.14, 1.52] | | Lung | AUROC | +0.51 | [-0.57, 1.59] |

The paired results provide stronger evidence for Prot2Path's balanced-accuracy gains than for an AUROC advantage over direct alignment. Viewed alongside Table 6, these findings clarify two aspects of the framework: the factorial ablation separates the contributions of standalone data through teacher pretraining and reference-bank expansion, while the paired comparison assesses the choice of transfer objective. Together, they support the use of a teacher learned from broader proteomic data and its transfer through standalone-reference-based relational supervision. The confidence intervals concern repeated downstream evaluations; they do not measure variability across independent pretraining runs.

AuthorsOpenReview ↗

Responses to W2, Q-B, Q-C, and Q-D: Proteomic teacher and data sources

W2 & Q-B & Q-D: What does PFE contribute beyond raw measurements and a small autoencoder?

Table S11 adds the requested controls under the same Soft KL distillation objective and downstream evaluation protocol. Stage II is the protein-supervised slide encoder before distillation. Each cell reports balanced ACC / AUROC (%): | Teacher / training configuration | Renal | Breast | Lung | | --- | ---: | ---: | ---: | | Stage II only | 58.07 / 75.86 | 69.22 / 76.69 | 61.34 / 75.39 | | Raw Protein + Soft KL | 57.98 / 75.18 | 69.49 / 76.36 | 61.18 / 75.49 | | MLP-AE + Soft KL | 59.23 / 76.01 | 68.99 / 77.49 | 62.32 / 75.95 | | PFE + Soft KL | 60.85 / 76.67 | 70.02 / 78.53 | 63.17 / 77.90 |

Raw-profile distillation produces mixed changes relative to Stage II, while the autoencoder improves AUROC in all three groups. PFE yields the highest mean ACC and AUROC throughout. Averaging across the three groups, AUROC changes from 75.98 for Stage II to 75.68 with Raw Protein, 76.48 with MLP-AE, and 77.70 with PFE. Thus, simpler representations retain some utility, but they do not reproduce the full improvement obtained with PFE. This directly tests the teacher representation in the intended transfer setting.

W2 & Q-D: Biological prediction beyond masked-profile retrieval.

Masking one to three measurements primarily assesses robustness to limited missingness; we therefore complement retrieval with prediction of RNA-derived tumor microenvironment states (Table 3). Patho-Bench labels categorize gene-expression-derived ESTIMATE scores into low, medium, and high states reflecting stromal and immune infiltration. All representations use the same patient splits and logistic-regression protocol. Table 3 reports the following AUROCs (%): | Cohort | Patients | Raw Protein | Protein-BERT | MLP-AE | PFE | | --- | ---: | ---: | ---: | ---: | ---: | | CPTAC-BRCA | 94 | 69.33 | 66.15 | 70.49 | 73.13 | | CPTAC-LSCC | 88 | 71.01 | 80.45 | 78.69 | 82.01 | | CPTAC-LUAD | 73 | 57.61 | 65.99 | 65.60 | 69.83 |

The reported PFE results lead in mean ACC and AUROC across these cohorts. These independently measured RNA-derived targets complement the distillation controls by testing biological prediction beyond recovery of a nearly unchanged protein profile.

W2 & Q-C: Source handling and exclusion of CPPA.

As described in Appendix A.1–A.3, protein identities and phosphorylation sites are mapped to a common 38-target panel, distinct sites are retained separately, missing values use a learnable mask, and external profiles are filtered by protein coverage. Source measurements retain their processed scales during dataset construction, with protein-specific standardization applied during PFE training. These operations establish consistent input identities and numerical scaling. CPPA represents treatment-induced changes rather than baseline abundance, so numerical standardization is not an assertion that the underlying quantities are biologically identical.

Expanding the corpus is motivated by the potential of self-supervised learning to benefit from more data and broader molecular variation. Our virtual profile generation follows a related motivation: expression-space augmentations can provide useful training variation without assigning each synthetic profile a specific biological interpretation. Table 2 shows that increasing the virtual batch size from B_v = 0 to 64 improves both retrieval and downstream distillation performance, supporting the utility of structured data augmentation for PFE pretraining. Table 6 evaluates CPPA's contribution under both bank configurations. Entries below are mean AUROC (%) across the three cancer task groups: | PFE pretraining data | Bank: paired only | Bank: paired + standalone | | --- | ---: | ---: | | Paired only | 76.57 | 76.64 | | Paired + standalone, excluding CPPA | 76.71 | 76.81 | | Paired + standalone, all sources | 77.24 | 77.70 |

Including CPPA increases mean AUROC from 76.71 to 77.24 with a paired-only bank and from 76.81 to 77.70 with a paired-plus-standalone bank. Table 6 also reports higher balanced accuracy in all three cancer groups, supporting the practical utility of including perturbation profiles in teacher pretraining.

Q-C: TCPA/CCLE overlap.

The TCPA contribution consists of CCLE RPPA500 profiles distributed through TCPA, so overlap with the separate CCLE collection is possible. We have performed cross-resource deduplication to remove overlapping records between these two collections before PFE pretraining, preventing the same record from being included twice through different sources.

AuthorsOpenReview ↗

Responses to Q-E and Q-F: Self-reference exclusion and patch-encoder dependence

Q-E: Self-reference probability and its role in the objective.

The default bank retains the query's own protein reference. In Table S12, self-exclusion masks the query's own reference and all other same-patient records in both teacher and student distributions before normalization. Other paired and standalone values are mean probability masses per reference; entropy is computed over the full teacher distribution. ACC/AUROC are 12-task macro-averages (%). | Configuration | Self mass (%) | Other paired mean mass (%) | Standalone mean mass (%) | Entropy (nats) | ACC (%) | AUROC (%) | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Self-included | 56.82 | 0.030 | 0.012 | 2.3212 | 64.68 | 77.70 | | Self-excluded | — | 0.069 | 0.031 | 4.1078 | 64.38 | 76.95 |

The query protein has maximal similarity to its own reference, so a high target probability is expected at the chosen temperature. This reference provides a matched molecular anchor for the corresponding slide. Its 56.82% probability emphasizes the paired molecular state, while the remaining 43.18% supervises relationships to other profiles. The target therefore combines paired alignment with distributed relational supervision.

On a per-reference basis, other paired profiles receive approximately 2.5 times the mass of standalone profiles with self-references retained, and 2.2 times after exclusion. This is consistent with closer molecular neighborhoods within the paired cohort than across heterogeneous standalone sources. PFE supplies graded reference weights so that the slide encoder can preserve similarities and differences between molecular states. The downstream utility of real standalone references is supported separately by the real-versus-random comparison in Table 6.

Removing the self-reference produces a more distributed teacher target, with entropy increasing from 2.3212 to 4.1078 nats. The corresponding mean ACC and AUROC decrease by 0.30 and 0.75 percentage points, respectively, quantifying the effect of retaining the paired molecular anchor within the relational objective.

Q-F: Matched evaluation with the frozen vision-only UNI2 encoder.

We have added the requested patch-encoder experiment in Table S13. CONCH v1.5 is replaced with frozen UNI2 while keeping the slide aggregation architecture, training data, and evaluation protocol unchanged. Stage II uses protein-expression supervision; Stage III adds Prot2Path. Each cell reports balanced ACC / AUROC (%): | Patch encoder | Training stage | Renal | Breast | Lung | | --- | --- | ---: | ---: | ---: | | CONCH v1.5 | Stage II | 58.07 / 75.86 | 69.22 / 76.69 | 61.34 / 75.39 | | CONCH v1.5 | Stage III | 60.85 / 76.67 | 70.02 / 78.53 | 63.17 / 77.90 | | UNI2 | Stage II | 59.57 / 75.88 | 69.41 / 78.00 | 61.34 / 76.07 | | UNI2 | Stage III | 60.78 / 76.39 | 70.92 / 78.58 | 63.42 / 78.08 |

Prot2Path improves both ACC and AUROC in all three cancer groups with either patch encoder. With UNI2, the mean ACC/AUROC across groups increases from 63.44/76.65 to 65.04/77.68. This matched comparison shows that the observed benefit of proteomic distillation extends to a vision-only patch encoder and is not confined to CONCH v1.5.

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

Related papers

Discussion (0)

Sign in to comment.