acceptodds
Under review as a conference paper at ICLR 2027

AutoMarker: Outcome-Guided Discovery of Multiscale Biomarker Candidates

Abstract

Clinical outcomes are observed at the patient level, whereas associated biomarker candidates emerge across cells, local microenvironments, and tissue organization. AutoMarker introduces an outcome-guided framework for discovering multiscale biomarker candidates directly from pathology representations. Protein-aligned morphology representations organize tissue into four coordinated views capturing cellular phenotypes, neighborhood composition, tissue organization, and context-dependent cellular states. Adaptive local divergence search identifies representation-space regions enriched in different outcome groups, while sign-consistent graph refinement consolidates local responses into coherent candidate components. WSI-level abundance comparisons, pathological characterization, and spatial back-tracing connect candidates to tissue context. Across four cohorts, we evaluate multiscale candidate discovery, including independent external validation in TCGA-COAD of candidates discovered and frozen in Orion-CRC. Comparative and ablation experiments demonstrate the contribution of adaptive search and graph refinement, while scalability experiments support mining at the million-entity scale.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
computational pathology, weakly supervised learning, spatial biomarker discovery, whole-slide image analysis, multimodal representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XTnNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AutoMarker addresses outcome-guided discovery of pathology biomarker candidates rather than outcome prediction alone. It learns protein-aligned morphology representations from paired H&E–mIF images, constructs cell, microenvironment, tissue, and tissue-conditioned cell views, and uses adaptive representation-space neighborhoods to identify outcome-enriched responses. Direction-consistent graph refinement consolidates these responses into candidate components with WSI abundance and tissue back-tracing. Orion-CRC candidates are frozen and evaluated in TCGA-COAD; two private cohorts provide additional within-cohort analyses. The paper also reports search/refinement ablations, patient-level sensitivity analyses, and computational scalability.

Critical strengths
  1. A useful formulation of outcome-guided biomarker discovery. The framework jointly specifies candidate membership, scale, enrichment direction, and spatial support. Connecting clinical outcome differences to inspectable cellular and tissue patterns gives the discovery task a concrete and interpretable output.

  2. An adaptive approach to candidate localization. Selecting neighborhood scales from outcome-conditioned divergence trajectories provides a motivated alternative to a single global scale. The stopping and refinement ablations support practical improvements in search efficiency and candidate redundancy.

  3. A traceable discovery and validation workflow. The method consolidates local responses into candidates and maps them back to supporting tissue. Freezing candidates before external evaluation strengthens the evidence for transfer, while H&E-only target inference increases the workflow's practical applicability.

Critical weaknesses
  1. The advantage over established differential abundance methods remains unclear. DA-seq is discussed but not experimentally compared, and the separate Milo analysis does not provide a matched discovery comparison. Since outcomes guide candidate selection, stronger discovery-set significance alone cannot establish better discovery quality. A comparison using the same representations and patient splits, assessed through external replication, would better isolate the search contribution.

  2. The value of mIF supervision is insufficiently isolated. H&E-only external associations do not establish accurate protein prediction or improvement over ordinary morphology features. A matched alignment ablation and held-out channel-prediction assessment are missing. A strong H&E encoder such as Virchow2 or UNI would also provide a useful tissue-level baseline for assessing the added complexity.

  3. The complementarity of the representation views is not demonstrated. Cellular, microenvironmental, and tissue candidates may reflect overlapping composition differences. Finding associations in each view does not show that the hierarchy adds information. A focused comparison against cell-only discovery and simple regional pooling would clarify whether the additional views yield distinct, externally reproducible findings.

  4. Abundance is defined differently during discovery and validation. Discovery counts retained query-center component members, whereas external evaluation counts entities covered by frozen neighborhoods. This acknowledged difference complicates interpretation of cross-cohort replication. Applying the external assignment rule back to the discovery cohort would establish whether the principal associations persist under a consistent measurement.

  5. The main external results need stronger patient-level reporting. Primary tests treat WSIs as independent observations, while the patient-aggregation sensitivity analysis covers only a separate cell-level workload. Patient-level external inference and reporting the complete frozen candidate family, with effect directions and multiplicity adjustment, would clarify the reliability and scope of validation.

Questions for authors
  1. Under matched representations and patient splits, does AutoMarker improve external discovery replication over DA-seq or Milo?
  2. What discovery benefit comes specifically from mIF alignment and the additional representation views?
  3. Do the principal external associations persist with consistent abundance definitions, patient-level inference, and complete-family reporting?
5 comments

1. How Does AutoMarker Compare with Established Differential Abundance Methods?

We extended our matched comparison to DA-seq (Zhao et al., 2021), Milo (Dann et al., 2022), and Kompot (Otto et al., 2025). AutoMarker achieved the highest membership-reconstruction Jaccard (0.8994) and the lowest minimum external BH q under common transfer (0.0171). Under native transfer, AutoMarker and Kompot shared the lowest external BH q (0.0076).

MetricDA-seqMiloKompotAutoMarker
Reconstruction Jaccard ↑ (common kNN)0.56690.63570.85850.8994
Best external BH q ↓ (native)0.15360.53160.00760.0076
Best external BH q ↓ (common)0.30710.64530.75810.0171
Discovered candidates85220

The comparison uses matched CellID-19 representations from 23,000 Orion-CRC cells (23 patients), followed by evaluation on 162 TCGA-COAD patients. Each method retains its native discovery procedure. We examine both native transfer and a common source-trained kNN assignment protocol, with patient-level external abundance testing.

AutoMarker's distinguishing capability is that its discovery output directly includes frozen adaptive center–radius geometry, enabling deterministic assignment of unseen representations without training an additional membership model.

Ground-truth discovery quality. AutoMarker achieved the highest ground-truth-best-match Jaccard (0.138) among the evaluated configurations, outperforming fixed-k (0.113) and fixed-radius (0.086). Moreover, removing sign-consistent graph pruning reduced recovery at Jaccard ≥ 0.10 from 63.3% to 3.9%, demonstrating its contribution to structured phenotype discovery. The semi-synthetic benchmark used Orion-CRC embeddings, three geometries, four signal conditions, ten seeds, and patient-disjoint splits.

Discovery-ranked recovery without oracle selection. AutoMarker achieved the highest median held-out Top-1 Jaccard (0.1122), compared with DA-seq (0.0040) and Milo (0.0015). The corresponding Top-5 union Jaccard values were 0.1356, 0.1080, and 0.0050, respectively. Unlike ground-truth-best matching, this evaluation selects candidates using discovery-based ranking without oracle information, assessing their recovery under the evaluated frozen-transfer protocols.

These experiments support AutoMarker's adaptive search, direction-aware refinement, and transferable membership. Full protocols will appear in the supplement.

2. Does mIF Supervision Improve Protein Prediction Beyond Morphology Features?

Yes. Three experiments demonstrate the benefits of mIF supervision for cell-level protein decoding, alignment objectives, and patch-level representations.

(1) Held-out cell-level protein decoding. CellID-19 achieved the highest Pearson (0.2435), Spearman (0.2723), and lowest MAE (0.2929) among the evaluated representations.

RepresentationPearson ↑Spearman ↑MAE ↓
CTransPath-7680.19450.20340.2969
CTransPath PCA-190.18850.20000.2951
CellID-190.24350.27230.2929

This evaluation used 40,192 matched H&E–mIF cells from 23 Orion-CRC patients, covering 16 protein channels, with identical patient-disjoint Ridge probes.

(2) Controlled alignment ablation. Combining cell-level and patch-level supervision achieved the highest Pearson (0.1809) and lowest MAE (0.2817), outperforming the evaluated morphology-only controls on both metrics.

ConfigurationPearson ↑MAE (log1p) ↓
CTransPath + Ridge0.12250.2986
PCA-16 + Ridge0.14790.2964
Cell loss only0.15920.2944
Patch loss only0.17780.2825
Cell + patch losses0.18090.2817

Using an identical frozen CTransPath backbone, we evaluated seven patient-disjoint folds and three matched seeds (63 training runs). All mIF-supervised configurations improved protein-decoding point estimates over morphology-only controls.

(3) Foundation encoder comparison. CellID-19 achieved the highest Pearson (0.4826) and lowest MAE (0.2188) among the evaluated representations, compared with UNI (0.4354/0.2461) and Virchow2 (0.4770/0.2340).

This separate patient-disjoint evaluation used 466 protein-labeled Orion-CRC patches and a shared Ridge probing protocol. The measurements assess protein decodability from patch-aggregated targets.

These results support mIF supervision for learning compact, protein-aligned representations useful for outcome-guided phenotype discovery.

3. Do the Hierarchical Views Discover Complementary Phenotypes Beyond Cell-Only Representations?

We thank the reviewer for this insightful suggestion. The hierarchical views retrieved distinct phenotype candidates beyond Cell-only discovery, and a Tissue-level candidate demonstrated additional external clinical relevance. Matched comparisons with Cell-only discovery and simple regional mean pooling revealed differences in both candidate membership and patient-level abundance patterns.

(1) Candidate distinctiveness. Environment, Tissue, and Tissue–Cell showed median maximum Cell-overlap Jaccard values of 0.0019, 0.0984, and 0.2038, respectively, with corresponding maximum absolute abundance correlations of 0.2545, 0.4391, and 0.4626.

ViewMedian max Cell Jaccard ↓Median max |r| ↓
Mean pooling0.00120.2387
Environment0.00190.2545
Tissue0.09840.4391
Tissue–Cell0.20380.4626

These results were obtained through matched patient-disjoint Orion-CRC discovery and frozen candidate assignment. We assessed spatial membership overlap using maximum Cell-candidate Jaccard and patient-level abundance redundancy using maximum absolute correlation, summarized by their medians.

The four views reflect distinct levels of tissue pathology: Cell captures individual phenotypes, Environment local cellular context, Tissue higher-order morphology, and Tissue–Cell cells conditioned on tissue context. Their purpose is to discover complementary pathological phenotypes across tissue scales, rather than merely improve outcome prediction through feature fusion.

(2) External clinical relevance. One frozen Tissue candidate demonstrated a direction-consistent association in TCGA-COAD (Cliff's δ = −0.264, 95% CI [−0.446, −0.087]; BH q = 0.0183).

Importantly, this association persisted after adjustment for an Orion-frozen Cell-only score (adverse OR = 5.09, 95% CI [1.58, 16.34]). In a further prescreening control adjusting for both the Cell-only score and regional mean-pooling information, the Tissue candidate remained associated with outcome (OR = 4.99, 95% CI [1.60, 15.57]).

Together, these findings provide evidence that hierarchical representations retrieve distinct phenotype populations and can reveal externally associated tissue-level information beyond the evaluated cellular and regional-pooling summaries, supporting AutoMarker's pathology-motivated multiscale discovery design.

4. Do the Principal Associations Persist Under Consistent Abundance Definitions?

Sorry for the confusion. 35/47 candidates (74.5%) preserved their outcome-association directions when we reapplied the frozen external membership rule to Orion-CRC.

The reviewer correctly observes that Eqs. 12 and 20 count different entities. This reflects an intentional distinction between phenotype-core discovery and transferable phenotype coverage.

Eq. 12 measures the retained query-center component identified through outcome-guided search and graph refinement, characterizing the discriminative phenotype core. Eq. 20 uses the union of frozen adaptive center–radius neighborhoods to assign broader populations of phenotypically related entities.

The distinction is pathology-motivated: identifying a representative outcome-associated phenotype and quantifying the prevalence of its broader population address different but related analytical objectives.

Following the reviewer's suggestion, we preserved the original query centers, radii, and component structures and reapplied the frozen Coverage definition to the discovery cohort without repeating outcome-guided search.

The observed 74.5% direction preservation indicates that most evaluated candidates retained their outcome-association directions under the alternative membership definition.

We will clarify the distinction between Center and Coverage abundances and report the corresponding candidate-level measurements in the supplementary material. Together, these analyses make the discovery-to-transfer measurement relationship explicit while preserving AutoMarker's intended two-stage design.

5. How Reliable Are the External Results Under Patient-Level Evaluation?

We thank the reviewer for emphasizing patient-level inference and complete-family reporting. Two AutoMarker candidates retained significant external associations after BH, Bonferroni, and permutation-based MaxT correction, with both remaining significant across all 162 leave-one-patient-out evaluations and after clinical covariate adjustment.

MetricCandidate 1Candidate 2
Cliff's δ (Favorable − Adverse)0.3890.458
95% CI[0.189, 0.576][0.287, 0.616]
BH q0.008940.00136
Bonferroni p0.017880.00136
MaxT permutation p0.01800.0019
LOPO Bonferroni significance162/162162/162
Clinically adjusted adverse OR0.4440.293
Clinically adjusted Bonferroni p0.04050.000914

(1) Patient-level inference and complete-family correction. Both candidates remained significant after BH and Bonferroni correction across the complete family of 47 frozen Cell candidates, including non-estimable entries. We evaluated 162 independent TCGA-COAD patients (36 favorable, 126 adverse) under the specified 24/60-month prognostic criteria. Abundances were first calculated per WSI and then averaged within patients, ensuring one observation per patient.

(2) Permutation and influence robustness. Both candidates retained MaxT-adjusted significance (p = 0.0180 and 0.0019) and remained Bonferroni-significant in 162/162 LOPO evaluations. We performed 19,999 patient-label permutations, preserving outcome-group sizes and jointly permuting the complete candidate family. LOPO analyses recalculated the full family after each patient deletion.

(3) Clinical covariate adjustment. Both associations remained significant after adjustment for pathological stage, age, and sex in 157 complete-case patients, with adjusted adverse-prognosis ORs of 0.444 and 0.293, respectively.

Together, these results demonstrate robust patient-level external associations under complete-family multiplicity correction, permutation calibration, patient-influence analysis, and clinical adjustment. Complete candidate-level statistics and protocols will be provided in the supplementary material.

Reviewer_5GhqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose AutoMarker, an outcome-guided framework for discovering multiscale biomarker candidates from histopathology images. The method uses protein-aligned H&E representations at four levels (cell, microenvironment, tissue, and tissue-conditioned cell), followed by adaptive neighborhood expansion based on outcome-associated log-fold changes and graph-based refinement. The discovered candidates are characterized through WSI-level abundance differences and spatial back-tracing. The authors evaluate the method across colorectal, breast, and bile-duct cancer cohorts, including independent external validation on TCGA-COAD.

Critical strengths
  • I like the overall problem formulation. Instead of training a model for outcome prediction and interpreting attention maps, the authors directly search for outcome-associated phenotypes.

  • The external validation is a strong aspect of the paper. Candidates discovered in Orion-CRC are frozen and evaluated on 462 TCGA-COAD slides, with several showing significant associations and consistent enrichment directions

Critical weaknesses
  • Table 3 uses p-value and number of significant candidates for comparison, but stronger associations on the discovery set does not mean better biomkarer discovery. Is there a benchmark the authors can test on with known outcome-associated populations?

  • In eq 12, the query components are used for the abundance, but eq 20 uses union on expanded neighborhood? They are not the same entities if I am not mistaken. Can you freeze the membership to the discovery cohort and see candidate rankings?

  • When doing adaptive selection, how robust are these associations and candidate boundaries across samples/seed? Orion-CRC is quite small so it might be good to see how this behaves

Overall I think the paper addresses an interesting problem, and I like the idea of directly discovering outcome-associated phenotypes across different tissue scales. The concern for me is that the evaluation relies on association statistics from discovery without external validation of baselines. I still lean towards acceptance in the current state, but would like if the authors can improve on the metrics.

Questions for authors

Mentioned along with weaknesses.

4 comments

1. Can AutoMarker Be Evaluated Against Known Outcome-Associated Populations?

We thank the reviewer for this insightful suggestion. AutoMarker achieved the highest ground-truth recovery performance, with a best-match held-out Jaccard of 0.138 and a recovery rate of 63.3% at Jaccard ≥ 0.10, outperforming the evaluated fixed-neighborhood configurations.

MethodBest-match Jaccard ↑Recovery at J ≥ 0.10 ↑
Fixed-radius0.08646.9%
Fixed-k0.11356.1%
AutoMarker0.13863.3%

To directly evaluate discovery accuracy beyond association p-values, we constructed a controlled semi-synthetic benchmark using real Orion-CRC embeddings with known differential-region memberships. The benchmark covers three geometries, four signal conditions, ten seeds, and patient-disjoint discovery/evaluation splits, enabling direct assessment of precision, recall, Jaccard, and region recovery. Reported Jaccard values represent the best ground-truth match among discovered candidates.

In the neighboring-region setting, where adjacent populations exhibit opposite outcome-enrichment directions, AutoMarker also achieved higher median seed-averaged Jaccard than each evaluated fixed-neighborhood configuration across all three non-null signal strengths.

Furthermore, removing sign-consistent graph pruning reduced recovery from 63.3% to 3.9% at Jaccard ≥ 0.10, demonstrating its contribution to recovering structured phenotype populations.

These experiments provide direct quantitative evidence of the effectiveness of AutoMarker's adaptive search and graph refinement, extending evaluation beyond discovery-set statistical significance. Complete benchmark protocols and results will be provided in the supplementary material.

2. How Do Candidate Rankings Change Under the Frozen Neighborhood Membership Definition?

Sorry for the confusion. 35/47 candidates (74.5%) preserved their outcome-association directions when we reapplied the frozen neighborhood-membership rule to the Orion-CRC discovery cohort.

The reviewer correctly observes that Eq. 12 and Eq. 20 count different entities. This reflects AutoMarker's intentional distinction between phenotype-core discovery and transferable phenotype coverage.

During discovery, Eq. 12 measures retained query-center components, characterizing the outcome-enriched phenotype cores identified through adaptive search and graph refinement. For external assignment, Eq. 20 measures the union of frozen center–radius neighborhoods, capturing broader populations of phenotypically related entities.

This distinction is pathology-motivated: identifying a discriminative phenotype core and quantifying its broader prevalence address related but different analytical objectives.

Following the reviewer's suggestion, we preserved the original query centers, adaptive radii, and component structures, then recomputed abundances using the same Coverage definition employed for external assignment.

The observed 74.5% direction preservation supports consistency between the two definitions for the majority of evaluated candidates. We will provide paired Center/Coverage abundances, association statistics, and candidate rankings in the supplementary material, clarifying their respective roles and implications for cross-cohort interpretation.

3. How Robust Are Candidate Associations and Boundaries Across Patients and Sampling Seeds?

We thank the reviewer for this valuable suggestion. AutoMarker demonstrated strong outcome-direction consistency among matched candidates, reaching 100% under patient subsampling, 98.3% under leave-one-patient-out (LOPO), and 89.5% across cell-sampling perturbations. We also identified two broad-scale phenotype components with highly reproducible boundaries.

PerturbationDirectional agreement
Patient subsampling100%
LOPO98.3%
Cell sampling89.5%

To address the reviewer's concern regarding the limited Orion-CRC cohort, we conducted 20 stratified patient-subsampling experiments, 23 LOPO evaluations, and 9 independent cell-sampling perturbations. We separately assessed association-direction consistency and geometric candidate recovery.

The directional agreement rates above are calculated among successfully matched candidates.

For geometric reproducibility, two broad-scale phenotype components were recovered across all 9/9 cell-sampling perturbations, with LOPO recovery reaching 22/23 and 23/23 evaluations, respectively, at Jaccard ≥ 0.5.

These results demonstrate that AutoMarker can consistently recover outcome-associated directions and broad-scale phenotype structures under patient and sampling perturbations. Localized candidates exhibited greater boundary sensitivity, highlighting the distinction between directional stability and exact geometric reproducibility.

Complete robustness analyses and candidate-level results will be provided in the supplementary material.

4. How Does AutoMarker Compare with Existing Methods Under External Evaluation?

We sincerely appreciate the reviewer's encouraging assessment and suggestion to strengthen external baseline comparisons. AutoMarker achieved the highest membership-reconstruction Jaccard (0.8994) and the lowest minimum external BH-adjusted q under common transfer (0.0171) among the four evaluated methods. Under native transfer, AutoMarker and Kompot shared the lowest external BH-adjusted q (0.0076).

MetricDA-seqMiloKompotAutoMarker
Reconstruction Jaccard ↑ (common kNN)0.56690.63570.85850.8994
Best external BH q ↓ (native transfer)0.15360.53160.00760.0076
Best external BH q ↓ (common transfer)0.30710.64530.75810.0171
Discovered candidates85220

Following the reviewer's suggestion, we conducted matched comparisons with three differential abundance approaches: DA-seq (Zhao et al., 2021), Milo (Dann et al., 2022), and Kompot (Otto et al., 2025).

All four methods used the same 23,000 Orion-CRC cells from 23 patients, identical CellID-19 representations, and matched outcome labels. Their discovered candidates were evaluated on the same 162 independent TCGA-COAD patients using patient-level abundance measurements and multiple-testing correction.

We evaluated both native transfer, retaining each method's source-defined assignment procedure, and common transfer, applying a shared source-trained kNN assignment framework.

Importantly, AutoMarker directly preserves its adaptive center–radius geometry, allowing deterministic assignment of discovered phenotypes to unseen representations without fitting an additional transfer model. These results demonstrate AutoMarker's advantages in the evaluated membership-reconstruction and external-association metrics, extending comparison beyond discovery-set significance.

Complete experimental protocols, transfer comparisons, and candidate-level results will be provided in the supplementary material.

Thanks for the rebuttal. I keep my score.

Thank you for your time and thoughtful consideration. We sincerely appreciate your constructive feedback, which has helped us improve our work.

Reviewer_PNi5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AutoMarker, an outcome-guided framework for discovering multiscale biomarker candidates directly from pathology representations. Protein-aligned morphology representations organize tissue into four coordinated views capturing cellular phenotypes, neighborhood composition, tissue organization, and context-dependent cellular states. This framework is evaluated in four cohorts.

Critical strengths
  1. Searching from the representation space is a new aspect and can be falsified on a new cohort, which attention maps can't.
  2. The candidate (view, set of center radius pairs, enrichment direction, per WSI abundance, and source coordinates) is a good choice.
Critical weaknesses
  1. The machine learning contributation is weak and unclear. The paper's contributions (a formulation of outcomeguided biomarker candidate discovery, adaptive representation-space search, and a traceable refinement and validation workflow) are more meaningful in biology/pathology area but there is no novelty in machine learning or algorithms.
  2. No evaluation of the machine learning part. The machine learning part of this paper (alignment with protein MIFs) is not evaluated. We don't know the model can predict proteins better after the alignment.
  3. This paper only evaluates the difference of a biomarker from two slide groups. That's not enough for a machine learning paper.
Questions for authors

What is this paper's contributation in machine learning? Can you show more metrics and compare with groundtruth? Can this framework apply to other pathology foundation model embeddings?

11 comments

1. What Is the Machine Learning Contribution of AutoMarker?

Sorry for the confusion. AutoMarker's ML contribution is to discover and explicitly define outcome-associated phenotype populations from weak patient-level supervision—not merely to validate an encoder through biomarker testing. Rather than predicting an outcome and interpreting influential regions afterward, AutoMarker uses outcome information to determine which phenotype populations are retrieved, where their boundaries lie, and how they are organized.

CellID is an integral part of this framework: paired H&E–mIF training produces a compact, 19-dimensional protein-aligned representation with molecular-channel semantics. The Cell, Environment, Tissue, and Tissue–Cell views then describe phenotype organization at complementary pathological scales. On these spaces, three algorithmic mechanisms turn weak labels into structured candidates:

(i) Adaptive boundary inference. Fixed neighborhood sizes can fragment or merge phenotype populations of different scales. AutoMarker tracks outcome-conditioned LogFC over expanding neighborhoods and selects query-specific radii through direction-normalized curvature stopping. Outcome labels therefore influence candidate geometry, not simply post-discovery association.

(ii) Direction-consistent structural refinement. Neighboring phenotypes may have opposite outcome associations. AutoMarker prunes sign-inconsistent edges in a graph of enriched queries and consolidates compatible responses into structured candidate components, combining phenotype proximity with enrichment direction.

(iii) Explicit transferable membership. Each component retains its adaptive center–radius geometry. Its frozen neighborhood union assigns new entities without reselecting boundaries or rerunning outcome-guided discovery, yielding reusable phenotype definitions with spatial traceability.

AutoMarker introduces a distinct machine learning paradigm for outcome-guided structured phenotype discovery. Its central methodological contribution is to move beyond learning representations for outcome prediction toward directly discovering the phenotype structures associated with clinical outcomes.

2. Does mIF Alignment Improve Protein Prediction?

Yes. CellID-19 improved held-out protein decoding over CTransPath, the joint alignment objective achieved the best ablation results, and CellID attained the highest patch-level Pearson and lowest MAE among the evaluated foundation representations. We used matched H&E–mIF measurements and patient-disjoint evaluation rather than inferring alignment quality from clinical associations.

(i) CellID-19 achieved the highest Pearson (0.2435), Spearman (0.2723), and lowest MAE (0.2929) among the evaluated representations. We tested 40,192 matched cells from 23 Orion-CRC patients and 16 channels using identical patient-disjoint Ridge probes:

RepresentationPearson ↑Spearman ↑MAE ↓
CTransPath-7680.19450.20340.2969
CTransPath PCA-190.18850.20000.2951
CellID-190.24350.27230.2929

This demonstrates that CellID retains more linearly decodable protein-associated information in only 19 dimensions than the evaluated morphology-only baselines.

(ii) The combined cell- and patch-level alignment achieved the highest Pearson (0.1809) and lowest MAE (0.2817): in our controlled ablation. Using the same frozen CTransPath backbone, seven patient-disjoint folds, and three matched seeds (63 runs), we compared:

ConfigurationPearson ↑MAE (log1p) ↓
CTransPath + Ridge0.12250.2986
PCA-16 + Ridge0.14790.2964
Cell loss only0.15920.2944
Patch loss only0.17780.2825
Cell + patch losses0.18090.2817

All mIF-supervised configurations improved protein-decoding point estimates over morphology-only controls.

(iii) CellID-19 achieved the best results on both evaluated patch-level metrics: Pearson = 0.4826 and MAE = 0.2188, compared with UNI (0.4354/0.2461) and Virchow2 (0.4770/0.2340). This patient-disjoint comparison used 466 protein-labeled patches from 23 Orion-CRC patients.

These results support AutoMarker's representation design: a compact, protein-aligned phenotype space that retains useful molecular information while providing interpretable coordinates for large-scale adaptive discovery. The reported metrics assess Ridge-probe protein decodability; cell- and patch-level evaluations use different targets.

3. How Is AutoMarker Evaluated Beyond Two-Group Differences and Against Ground Truth?

We appreciate the reviewer's concern and would like to clarify an important aspect of AutoMarker's methodological design. Outcome-guided phenotype discovery is itself a central machine learning contribution of AutoMarker, rather than merely a downstream evaluation of CellID. Patient-level outcome labels directly guide the discovery of previously unknown phenotype populations, including the inference of their adaptive boundaries, structural organization, and transferable membership.

We evaluate these components separately. W2 reports held-out protein decoding and controlled mIF-alignment ablations. For structured discovery, the original study includes independent TCGA-COAD frozen-candidate evaluation, algorithmic ablations, and pathological back-tracing.

AutoMarker achieved the highest ground-truth recovery performance, with a best-match held-out Jaccard of 0.138 and a recovery rate of 63.3% at Jaccard ≥ 0.10, outperforming the evaluated fixed-neighborhood configurations.

MethodBest-match held-out Jaccard ↑Recovery at J ≥ 0.10 ↑
Fixed-radius0.08646.9%
Fixed-k0.11356.1%
AutoMarker0.13863.3%

To directly evaluate discovery accuracy beyond association p-values, we constructed a controlled semi-synthetic benchmark using real Orion-CRC embeddings with known differential-region memberships. The benchmark covers three geometries, four signal conditions, ten seeds, and patient-disjoint discovery/evaluation splits, enabling direct assessment of precision, recall, Jaccard, and ground-truth recovery.

Furthermore, removing sign-consistent graph pruning reduced recovery from 63.3% to 3.9%, demonstrating its contribution to structured phenotype discovery.

These results provide direct quantitative evidence that AutoMarker's adaptive search and structural refinement contribute to recovering outcome-associated phenotype populations, rather than merely identifying statistically significant differences between patient groups.

4. Can AutoMarker Be Applied to Other Pathology Foundation Model Embeddings?

Yes. UNI and Virchow2 embeddings were successfully substituted into AutoMarker's unchanged Tissue-level discovery-and-assignment pipeline. We evaluated their native and PCA-reduced forms alongside the original representation—five encoder configurations across three patient-disjoint folds—without modifying adaptive search, graph refinement, or frozen membership assignment.

These experiments demonstrate that the discovery mechanism is compatible with alternative pathology foundation representations. CellID remains an integral component of the proposed framework because its compact, 19-dimensional protein-aligned space provides explicit molecular semantics for interpreting discovered phenotypes.

As shown in W2, CellID also achieved the highest patch-level Pearson (0.4826) and lowest MAE (0.2188) among the evaluated representations.

Complete experimental protocols, quantitative comparisons, and additional results will be provided in the supplementary material.

Thank you for your reply. I still think the machine learning/deep learning contribution and novelty of this paper are limited for ICLR. ICLR (The International Conference on Learning Representations) is top tier conference for artificial intelligence, representation learning, or deep learning. All papers should have clear contributation for deep learning area. As the author mentioned, this paper's main contributation is to discover and explicitly define outcome-associated phenotype populations from weak patient-level supervision—not merely to validate an encoder through biomarker testing. This is more interesting for biology or pathology area than machine learning/deep learning area. In detail, all methods are a combination of existing ideas and the thresholds are hand-set. There is no theoretical analysis to prove the generalization of this new framework or metric. Even in pathology area, the results are shown only in COAD, but not in more common cancer types, such as breast cancer and lung cancer. Overall, this is a good paper but it is not a good fit for ICLR conference, considering the limited novelty of deep learning and generalization to other cohorts. I would recommend submitting to other conferences like RECOMB, ISMB, or other journals.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the Program Chairs, Area Chairs, and reviewers for constructive feedback. We address methodological novelty and generality (PNi5); discovery quality (PNi5, 5Ghq, XTnN); mIF supervision and encoder dependence (PNi5, XTnN); and robustness and external validation (5Ghq, XTnN). Details appear in reviewer-specific replies.

Q1. Does AutoMarker Merely Combine Existing Methods with Manually Chosen Thresholds? (PNi5)

AutoMarker directly aligns with ICLR 2027's scope of "applications to physical sciences (physics, chemistry, biology, etc.)", addressing a fundamental machine learning challenge in biological discovery.

A central methodological innovation is our outcome-to-phenotype discovery paradigm itself. Unlike conventional approaches that first define cellular populations or neighborhoods and subsequently evaluate their clinical associations, AutoMarker works backward from patient-level outcomes to discover the underlying multiscale pathological phenotypes. This enables biomarker discovery without local phenotype annotations, while automatically determining candidate-specific boundaries, enrichment directions, and spatial support.

As stated in our Introduction, AutoMarker "jointly defines candidate membership, scale, enrichment direction, and spatial support." By integrating outcome-conditioned adaptive search, direction-consistent refinement, and frozen geometric assignment, our framework transforms weak patient-level supervision into explicit, interpretable, and transferable biomarker definitions, rather than merely producing statistical associations with predefined populations.

Candidate boundaries are not manually chosen. Appendix B specifies that radius grids are shared per view; fixed support and graph parameters are algorithmic controls, but candidate radii are inferred from outcome-conditioned trajectories (Eqs. 5–9). Direction-normalized stopping gave 100% radius agreement under label reversal, versus 32.55% without normalization. Substituting HNSW for KD-tree retained all 13 regions (top-region Jaccard 0.999999).

The same algorithm operates across four views and colorectal prognosis, breast cancer metastasis, and bile-duct neoadjuvant response (MPR/nMPR). These cohorts test adaptability across clinical endpoints.

Q2. How Is Discovery Quality Established Beyond Association Statistics? (PNi5, 5Ghq, XTnN)

On an Orion-based semi-synthetic benchmark, ground-truth best-match Jaccard was 0.138 versus fixed-k 0.113 and fixed-radius 0.086. Recovery at Jaccard ≥ 0.10 reached 63.3% versus 56.1%/46.9%, falling to 3.9% without sign pruning. In oracle-free discovery ranking, held-out Top-1 Jaccard was 0.1122 versus DA-seq 0.0040 and Milo 0.0015.

Against DA-seq, Milo, and Kompot, AutoMarker led common-kNN reconstruction (Jaccard 0.8994) versus 0.5669/0.6357/0.8585. Minimum external BH q under common transfer was 0.0171 versus 0.3071/0.6453/0.7581; native transfer tied Kompot at 0.0076.

Q3. What Do mIF Supervision, Multiple Views, and Encoder Substitution Add? (PNi5, XTnN)

CellID-19 improved patient-disjoint cell-level protein-decoding Pearson over CTransPath (0.2435 vs. 0.1945). Joint cell-plus-patch alignment gave the best shared-backbone ablation (Pearson 0.1809, MAE 0.2817). On patch-level probes, CellID reached Pearson 0.4826 / MAE 0.2188, versus UNI 0.4354/0.2461 and Virchow2 0.4770/0.2340.

To address PNi5 directly, we ran Tissue-level biomarker discovery with UNI, Virchow2, and CTransPath on 16,724 shared patches (24 WSIs/23 patients; three patient-disjoint folds). All seven configurations generated nominally associated candidates with unchanged algorithms. CellID-19 had the highest discovery yield: 36/489 (7.4%), versus UNI-PCA19 9/210, Virchow2-PCA19 22/526, and CTransPath-native 6/204.

Other hierarchical views yielded complementary phenotypes. A frozen Tissue candidate had direction-consistent TCGA-COAD association (BH q=0.0183), persisting after Cell-only (adverse OR 5.09) and mean-pooling (OR 4.99) adjustment.

Q4. How Robust Are Candidate Definitions and External Associations? (5Ghq, XTnN)

Among matched candidates, outcome-direction agreement reached 100% under patient subsampling, 98.3% under leave-one-patient-out, and 89.5% under cell sampling. Two broad-scale candidates were recovered in 9/9 sampling runs and 22/23 and 23/23 leave-one-patient-out runs at Jaccard ≥ 0.5. Frozen Coverage retained directions in 35/47 (74.5%) Orion candidates.

Among 162 TCGA-COAD patients, two of 47 frozen Cell candidates survived complete-family Bonferroni (p=0.01788/0.00136) and MaxT (p=0.0180/0.0019) adjustment. Both remained significant in all 162 leave-one-patient-out checks and after clinical-covariate adjustment.

We thank all reviewers and chairs. Protocols and candidate-level results appear in detailed replies and supplementary analyses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.