Pith. sign in

REVIEW 4 major objections 5 minor 70 references

Spatial proteomics guided by H&E-based AI reveals recurrence-risk niches in triple-negative breast cancer

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read H&E-based AI recurrence heatmaps localize molecular niches, and a 13-protein composite adds prognostic signal in TNBC.

desk verdict A genuinely novel workflow—AI risk heatmaps as coordinates for spatial proteomics—with suggestive biology, but the headline complementary-prognosis claim rests on very few events and in-sample selection; worth a careful referee, not acceptance as is. read the letter →

arxiv 2608.03145 v1 pith:A3O4QMZR submitted 2026-08-04 cs.AI q-bio.QM

Yesung Cho , Ji Hwan Park , Chanil Kim , Hyewon Kim , Honglan Li , Yumin Lee , Geongyu Lee , Sujeong Hong
show 22 more authors
Seong Min Park Yoonyoung Lee Hee Sool Rho Sumin Lee Amos Chungwon Lee Changhwan Lee Hwanyoung Shim Hyunwook Kim Hyeji Shin Sanha Park Jihoon Yu Yoon Hee Shin Sooheon Kim Hyunjin Park Seung Min Park Sangwan Kim Yujung Kim Sung-Im Do Eun-Young Kim Dongmyung Shin Jongbae Park In-Gu Do
This is my paper · ORCID
classification cs.AIq-bio.QM
keywords triple-negativebreastcancerrecurrencepredictionwhole-slideimagingdeeplearningspatialproteomicstumormicroenvironmentmassspectrometrybiomarkerdiscovery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that the recurrence-risk heatmaps produced by an H&E-based deep-learning model label real, spatially localized biology rather than arbitrary statistical patterns. By using those heatmaps as coordinates to physically isolate high- and low-risk tumor patches and profiling them with mass-spectrometry proteomics, the authors found a concordant molecular contrast in both profiled recurrence patients: mitotic programs dominate high-risk patches, while immune and antigen-presentation programs dominate low-risk patches. They then built a 13-protein composite from that contrast and showed that adding it to the image risk score improved recurrence discrimination, increasing the out-of-bag C-index from 0.679 to 0.739. This matters because it turns an opaque prediction into a spatially explicit molecular map that can guide sampling and biomarker discovery in TNBC.

What carries the argument

The load-bearing object is the patch-level recurrence-associated risk score (RRS), assigned by a weakly supervised patch classifier to roughly 150-micrometer H&E patches and then remapped to slide coordinates to form recurrence-risk heatmaps. Distribution-based aggregation, specifically a histogram of the ten highest-scoring patches followed by Lasso regression, converts patch scores into a patient-level risk score. The second mechanism is AI-guided spatial isolation: laser-based microdissection physically cuts out RRS-defined tumor patches for mass-spectrometry proteomics, yielding the spatial protein contrasts that seed the 13-protein tumor composite.

What would settle it

Profile AI-defined high- and low-risk tumor regions across, say, 20 or more recurrence patients spanning the full risk-score range; if the concordant direction (mitotic enrichment in high-risk regions, immune and antigen enrichment in low-risk regions) does not reappear in the majority, or if adding the 13-protein composite to the H&E risk score does not beat the H&E-only C-index of 0.679 in an independent cohort, the central claim would be falsified.

Watch

Extended reading notes

Core claim

The central discovery is that outcome-trained patch-level risk scores on H&E slides separate the tumor into spatially distinct molecular states. In a 156-patient development/test design, aggregating the top-scoring patches with a histogram reached an AUC of 0.77 and a C-index of 0.77, and heatmaps showed high- and low-risk patches coexisting within the same tumor compartment. Guided by those heatmaps, the authors isolated 46 tumor regions from two recurrent patients; high-risk regions were concordantly enriched in mitotic and cell-cycle programs, low-risk regions in immune and antigen-presentation programs. The concordant proteins were condensed into a 13-protein tumor composite whose bulk-t

Load-bearing premise

The spatial molecular contrast is derived from only two deliberately selected recurrence patients, so the claim that AI-defined high- versus low-risk regions carry a generalizable mitotic-versus-immune biology assumes those two cases stand in for TNBC as a whole.

Editorial extensions

If this is right

  • The ten highest-scoring patches aggregated as a 20-bin histogram distinguish recurrent from non-recurrent TNBC patients with an AUC of 0.77 and a C-index of 0.77 in an independent test cohort.
  • High- and low-risk patches coexist within the same tumor compartment, so the risk heatmap captures intratumoral heterogeneity beyond tissue-compartment identity.
  • Spatially isolated high-risk tumor regions are enriched in mitotic programs and low-risk regions in immune and antigen-presentation programs, with the contrast concordant across the two profiled patients.
  • Adding the 13-protein tumor composite to the H&E risk score raises the out-of-bag C-index from 0.679 to 0.739 and improves time-dependent discrimination at 3 and 5 years.
  • A transcript-based version of the composite stratifies recurrence-free survival in an independent TNBC cohort.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Extension: if the mitotic-versus-immune contrast replicates in larger cohorts, outcome-guided spatial proteomics could be applied to other tumor types whose recurrence drivers are poorly understood.
  • Extension: because each patch inherits its patient's recurrence label, part of the risk signal may reflect patient-level confounders; a cross-validated design with patient-level splitting would test how much of the heatmap is genuinely local.
  • Extension: the temporal complementarity, with H&E strong early and the protein composite stronger later, suggests a combined score could be evaluated as a dynamic surveillance tool.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents an outcome-informed spatial pathology framework for triple-negative breast cancer (TNBC). A patch-level H&E classifier trained with weak patient-level labels produces recurrence-risk heatmaps; histogram-plus-top-k aggregation yields a patient-level risk score with AUC and C-index of 0.77 in a 49-patient test cohort. Bulk proteomics of the test cohort associates high-risk status with cell-cycle/genome-maintenance programs and low-risk status with immune programs. Within-slide analysis shows high- and low-risk patches coexisting in the same tissue compartments, with distinct nuclear/architectural features. Using AI heatmaps as guides, the authors physically isolate and profile 46 tumor regions from two recurrence patients by mass spectrometry, reporting concordant mitotic vs immune/antigen-presentation programs. A 13-protein composite derived from these spatial contrasts is evaluated in an expanded 96-patient cohort, where adding it to the H&E score improves an out-of-bag C-index from 0.679 to 0.739; a transcript-based version of the composite stratifies recurrence in the independent METABRIC TNBC cohort.

Significance. This is a conceptually novel and technically ambitious integration of outcome-trained deep learning with laser-capture-mass-spectrometry spatial proteomics. If the central claims hold, the framework would convert black-box H&E risk predictions into physically addressable molecular hypotheses, a valuable step for computational pathology. The paper also contains a substantial set of internal controls: patch-level distribution analyses, tissue-compartment deconvolution, bulk proteomic enrichment, and an external transcript-level validation. These strengths make the core idea compelling even though the quantitative incremental-value claim is not yet established at the evidence level presented.

major comments (4)
  1. [Section 6, Fig. 5e–f] The central claim that the spatially derived 13-protein composite adds prognostic information beyond H&E morphology rests on a C-index increase from 0.679 to 0.739 in an expanded cohort with only 12 recurrence events. No confidence intervals are reported for either C-index or the difference, despite the authors acknowledging the modest event count. Moreover, the expanded cohort includes the original 49-patient test cohort, which contains the two discovery patients (P1, P2) and the bulk samples used to characterize risk groups; the composite's constituent markers and their directions were selected using those same samples. The OOB bootstrap procedure does not make this a held-out evaluation. The authors should report CIs, a bootstrap test of the increment, and ideally a sensitivity analysis with the discovery patients excluded, or temper the claim accordingly.
  2. [Section 5, Fig. 4b] The spatial discovery is based on two deliberately selected patients: P1 (greatest within-slide variance) and P2 (highest patient-level risk). This selection maximizes the chance of finding contrasting molecular programs, but it does not establish that those programs reflect a generalizable high- vs low-risk axis. The paper itself reports limited protein-level overlap between the two patients (21 overlapping DEPs, of which 19 concordant), with substantial inter-patient heterogeneity. The text in the abstract and Results states that spatial profiling 'revealed a concordant molecular contrast,' but this is a process-level enrichment in two extreme cases, not a population-level discovery. The authors should explicitly frame this as hypothesis-generating and provide more patients or external spatial validation before making generalizable claims.
  3. [Methods, Model development; Results, Section 2] The patch-level classifier is trained with each patch inheriting the patient's recurrence label. This weak supervision scheme means patch scores may reflect patient-level confounders (e.g., staining batch, treatment differences, tumor size) rather than localized recurrence drivers. The claim that high- and low-risk patches coexist within the same tissue compartment depends on the assumption that patch-level scores are biologically local. Although tissue-compartment analysis partially addresses composition, the paper does not report any analysis of within-patient patch-score variability against between-patient variability, nor any correction for known confounders that differ between development and test cohorts (Table 1: tumor size, T stage, neoadjuvant/adjuvant chemotherapy, radiotherapy are significantly different). This omission leaves open the possibility that the observed 'intratumor
  4. [Section 6, Fig. 5c–d] The METABRIC validation is transcript-based, whereas the discovered signature is protein-based and measured by DIA-MS. The transcript-level stratification of RFS supports the biological direction of the markers, but it does not validate the protein composite itself, the DIA-MS assay, the precursor selection, or the incremental prognostic value over H&E. Given that the C-index improvement is already estimated on a small, partly overlapping cohort, the external validation should be at the protein level (or clearly labeled as only supporting biological plausibility) for the incremental-value claim to be credible.
minor comments (5)
  1. [Abstract / Section 6] 'Out-of-bag' (OOB) is used without definition in the abstract and main text; define the procedure when first used, since it is not a standard term for all readers.
  2. [Results, Section 6, Fig. 5e] The C-index comparison is presented with point estimates only; adding bootstrap confidence intervals or a distribution plot would make the precision of the estimates visible.
  3. [Results, Section 5, Fig. 4e] The standardization of protein abundances 'using the corresponding low-risk regions as the reference' is described informally; a precise formula in the main text or Supp. Methods would improve reproducibility.
  4. [Results, Section 2] Minor grammar issues: 'Patients with recurrence had higher mean fractions' contains a duplicate space; also the sentence beginning 'Patches from patients with recurrence were relatively more represented...' could be rephrased for clarity.
  5. [References] References 24 and 30 are dated 2026; if these are preprints or in-press articles, the citation style should be made consistent with the journal's guidelines.

Circularity Check

1 steps flagged · score 5.0 of 10

The 13-protein composite's expanded-cohort validation is partly in-sample: the cohort includes the two spatial-discovery patients and the bulk test cohort used to select the markers, so the C-index improvement is not fully independent.

  1. fitted input called prediction [Results Section 6 (Fig. 5a); Results Section 5 (Fig. 4b); Methods, 'AI-guided Spatial Tissue Isolation']
    "This cohort comprised the original test cohort (n = 49) and 4 7 additional independent patients... we quantified 33 precursors representing 13 spatial tumor signatures that were consistently detected in both the spatial and bulk test-cohort datasets. ... Among recurrence patients in the test cohort for whom adjacent FFPE sections were available, two cases were selected: P1, which showed the greatest within-slide variance in patch-level risk scores, and P2, which had the highest overall patient-level risk score."

    The 13-protein tumor composite is derived from spatial proteomic contrasts in P1 and P2 (both in the test cohort) and filtered by detectability in the bulk test-cohort dataset. The 'expanded cohort' used for all patient-level validation—Kaplan-Meier, Cox, OOB C-index improvement from 0.679 to 0.739, and time-dependent AUC—contains the original test cohort, including P1/P2 and the bulk test-cohort samples used for the detectability filter. Therefore the marker set is not independent of the evaluation cohort: the composite's prognostic performance is partly in-sample for the cases that generated the marker set. METABRIC provides external transcript-level support for the biological direction, but it does not validate the protein-level composite or the incremental gain over the H&E score, so t

full rationale

The H&E patch-level model derivation is not circular: it is trained on the development cohort and evaluated on a held-out test cohort, and the bulk proteomic associations are descriptive rather than predictive. The main circularity is in the validation of the 13-protein spatial composite. The protein markers were discovered from two test-cohort recurrence patients and filtered using the bulk test-cohort dataset, while the 'expanded cohort' used to estimate the composite's prognostic value and the incremental C-index gain includes those same patients and that same bulk cohort. Thus the expanded-cohort evaluation is not an independent test of the protein composite; the reported improvement may be inflated by in-sample marker selection. The METABRIC transcript-based analysis is genuinely external and supports the biological axis, but it does not validate the protein-level DIA-MS composite or the incremental value over the H&E risk score. No load-bearing self-citation or uniqueness-imported-from-authors pattern was found: citations to the authors' prior G2L and SLACS work are methodological, not evidentiary for the central claim. The paper's own limitation statement about the modest number of recurrence events further reinforces that the incremental C-index is fragile, though that is a statistical limitation rather than circularity per se.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The central claims rest on noisy weak labels for patch risk, the representativeness of two selected recurrence patients, the assumption that protein-level directions carry over to transcript-level METABRIC data, and standard survival-statistics assumptions at very small event counts. The main free parameters are the trained network and aggregation model weights, the tissue-classifier fine-tuning, and the hand-specified 13-protein composite.

free parameters (5)
  • ConvNeXt-Base patch-level recurrence classifier weights = Trained on 107 patients, 13 recurrence events; not enumerated
    The patch-level RRS is the output of a network fit to inherited patient labels; it drives all heatmaps and downstream risk groups.
  • Hist. + Top-k aggregation hyperparameters and Lasso coefficients = Top-10 patches, 20-bin histogram, Lasso
    Chosen by development-cohort model selection; patient-level risk score is a Lasso fit on the histogram.
  • TIGER tissue compartment classifier fine-tuning = ConvNeXt-Base initialized from H-optimus-0
    Compartment assignments define which patches are available for spatial selection.
  • DIA-MS FDR and protein retention thresholds = FDR <= 1%; 6,609 proteins retained
    Determines which proteins enter the DEP and composite analyses.
  • 13-protein tumor composite scoring rule = Equal-weight, direction-aligned sum of 13 proteins
    Hand-specified aggregation; details are in missing supplementary methods.
assumptions (5)
  • domain assumption Weak supervision: each patch inherits its patient's recurrence status (patch-level label = patient-level label)
    Methods: 'each retained patch inherited the recurrence status of its corresponding patient.' Assumes the patch-level RRS can be learned from these noisy labels and that high-scoring patches reflect localized recurrence drivers.
  • domain assumption The two selected recurrence patients (P1 and P2) are representative of AI-defined high/low risk biology in TNBC
    Results Section 5: cases selected by extreme criteria; n=2. The concordant 19-protein signal is assumed to generalize beyond these two cases.
  • domain assumption Concordance between protein abundance and transcript abundance for the 13 markers in METABRIC
    Results Section 6 uses transcript-level METABRIC data to validate a protein-derived composite, assuming the directionally aligned transcript signature reflects the same biology.
  • domain assumption ESTIMATE deconvolution gene sets remain valid when applied to FFPE DIA-MS proteomic abundance
    Results Section 3 uses ESTIMATE immune and tumor-purity scores on proteomic data; ESTIMATE was developed for transcriptomic data.
  • standard math Standard survival-analysis assumptions (Cox proportional hazards, OOB bootstrap, time-dependent AUC estimators) hold at the observed event counts
    Statistical analysis section; the small event counts (7 and 12) make asymptotic inference fragile but the methods are standard.
invented entities (2)
  • AI-defined recurrence-risk niches independent evidence
    purpose: Spatial units on the H&E slide used to select tissue regions for MS proteomic profiling.
    The niches are coordinate-level heatmap constructs; their biological meaning is supported by the spatial proteomic contrast and partly by METABRIC, though not by a separate imaging cohort.
  • 13-protein tumor composite independent evidence
    purpose: Patient-level recurrence-risk score intended to complement the H&E risk score.
    The transcript-based version stratifies RFS in the external METABRIC cohort, providing a handle outside the discovery data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatial proteomics guided by H&E-based AI reveals recurrence-risk niches in triple-negative breast cancer." pith.science (2026). https://pith.science/paper/A3O4QMZR

@misc{pith2026260803145,
  author       = {Pith},
  title        = {Pith review of: Spatial proteomics guided by H&E-based AI reveals recurrence-risk niches in triple-negative breast cancer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A3O4QMZR}},
  note         = {Machine review of arXiv:2608.03145}
}
read the original abstract

Deep learning models can predict cancer recurrence from H&E stained slides, but the localized molecular states underlying these predictions remain largely obscured. Here, we developed an outcome informed spatial pathology framework in TNBC that integrates AI generated recurrence risk heatmaps with mass spectrometry based spatial proteomics. In a cohort of 156 patients, distribution based aggregation of high scoring patches achieved an AUC of 0.77 and a C-index of 0.77 in an independent test cohort. Bulk proteomics associated high image derived risk with cell cycle and genome maintenance programs and low risk with immune activation. High and low risk patches coexisted within the same tumor compartment and displayed distinct nuclear and architectural features, revealing intratumoral heterogeneity beyond tissue compartment identity. We then used the heatmaps as coordinate level guides to physically isolate and profile 46 AI defined tumor regions from two recurrence patients. Spatial proteomic profiling revealed a concordant molecular contrast across both patients: mitotic programs were enriched in high risk regions and immune and antigen presentation programs in low risk regions. A 13 protein composite derived from these spatial contrasts showed a trend toward poorer recurrence-free survival with increasing scores in an expanded cohort, while the corresponding transcript based composite stratified recurrence free survival in the independent METABRIC TNBC cohort. Integrating the protein composite with the H&E derived risk score improved the out of bag C-index from 0.679 to 0.739 and enhanced time dependent discrimination at 3 and 5 years. Together, these findings define a new role for outcome trained AI models as spatially explicit experimental guides that connect prognostic morphology with localized molecular states and advance biologically grounded, multiscale biomarker discovery in TNBC.

Figures

Figures reproduced from arXiv: 2608.03145 by the authors.

Figure 1
Figure 1. TNBC Recurrence prediction framework and performance of the H&E [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. Tissue-compartment distribution and morphological heterogeneity of AI-derived patch-level risk scores in TNBC. (a) UMAP projection of DINO-derived patch features colored by the AI-derived patch-level recurrence-associated risk score (RRS). (b) UMAP projection of the same patches colored by tissue compartment, including tumor, tumor-associated stroma (TAS), necrosis, and immune compartments. (c) Sankey plot showing t… view at source ↗
Figure 4
Figure 4. Spatial proteomic profiling of AI-defined high- and low-risk tumor regions in TNBC. (a) Schematic overview of the spatial proteomics workflow. Regions of interest (ROIs) were selected using AI-derived recurrence-risk heatmaps, isolated by infrared pulse laser–based cell sorting, collected, and subjected to LC–MS/MS-based data-independent acquisition (DIA) proteomic analysis. (b) Distribution of tumor and immune patc… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Cohort-level evaluation of the spatial proteomics-derived tumor composite and H&E AI risk score in TNBC. (a) Schematic overview of the expanded-cohort evaluation. The original test cohort (n = 49) was combined with 47 additional patients, yielding an expanded cohort of…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 69 canonical work pages

  1. [1]

    Ltd., Seoul, Republic of Korea

    OmixAI Co. Ltd., Seoul, Republic of Korea

    work page

  2. [2]

    Ltd., Seoul, Republic of Korea

    Oncocross Co. Ltd., Seoul, Republic of Korea

    work page

  3. [3]

    Department of Biomedical Sciences, Kyung Hee University., Seoul, Republic of Korea

    work page

  4. [4]

    Department of Pathology, Kangbuk Samsung Hospital, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea

    work page

  5. [5]

    Ltd., Seoul, Republic of Korea

    Meteor Biotech Co. Ltd., Seoul, Republic of Korea

    work page

  6. [6]

    College of Pharmacy, Ewha Womans University., Seoul, Republic of Korea

    work page

  7. [7]

    Department of Surgery, Kangbuk Samsung Hospital, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea * These authors contributed equally to this study. + Corresponding author: Dongmyung Shin, Jongbae Park and Ingu Do Abstract Deep learning models can predict cancer recurrence from H&E stained slides, but the localized molecular states und...

    work page

  8. [8]

    Baseline Characteristics of the Study Cohort Image Developing (n =

    Clinicopathological characteristics of TNBC cohorts Table 1. Baseline Characteristics of the Study Cohort Image Developing (n =

    work page

Show all 70 references
  1. [9]

    Salgado, R. et al. The evaluation of tumor -infiltrating lymphocytes in breast cancer: recommendations by an International TILs Working Group 2014. Ann. Oncol. 26, 259–271 (2015)

    2014

  2. [10]

    TNBC Recurrence prediction framework and performance of the H&E -based AI model

    H&E image AI analysis of TNBC Fig 1. TNBC Recurrence prediction framework and performance of the H&E -based AI model. (a) Overview of the patch-level recurrence prediction pipeline. H&E whole-slide images (WSIs) from the TNBC cohort were partitioned into 150- μm patches and su...

  3. [11]

    Bulk proteomic characterization of AI-derived risk groups in TNBC

    TNBC Bulk proteomics analysis Fig 2. Bulk proteomic characterization of AI-derived risk groups in TNBC. (a) Schematic overview of the bulk proteomic workflow. FFPE TNBC tissue specimens were subjected to LC–MS/MS-based data-independent acquisition (DIA) proteomic analysis. (b)...

  4. [12]

    Morphological characterization of high- and low-risk patches Fig. 3. Tissue -compartment distribution and morphological heterogeneity of AI -derived patch -level risk scores in TNBC. (a) UMAP projection of DINO -derived patch features colored by the AI -derived patch-level rec...

  5. [13]

    Spatial proteomic profiling of high- and low-risk patches Fig. 4. Spatial proteomic profiling of AI-defined high- and low-risk tumor regions in TNBC. (a) Schematic overview of the spatial proteomics workflow. Regions of interest (ROIs) were selected using AI-derived recurrence...

  6. [14]

    Validation using spatial proteomics -derived markers and H&E AI risk score Fig. 5. Cohort -level evaluation of the spatial proteomics -derived tumor composite and H&E AI risk score in TNBC. (a) Schematic overview of the expanded-cohort evaluation. The original test cohort (n =...

  7. [15]

    D., Marron, J

    Couture, H. D., Marron, J. S., Perou, C. M., Troester, M. A. & Niethammer, M. Image analysis with deep learning to predict breast cancer grade, ER status, histologic subtype, and intrinsic subtype. npj Breast Cancer 4, 30 (2018)

    2018

  8. [16]

    & Gianni, L

    Bianchini, G., De Angelis, C., Licata, L. & Gianni, L. Treatment landscape of triple -negative breast cancer—expanded options, evolving needs. Nat. Rev. Clin. Oncol. 19, 91–113 (2022)

    2022

  9. [17]

    C., Lin, N

    Garrido-Castro, A. C., Lin, N. U. & Polyak, K. Insights into molecular classifications of triple- negative breast cancer: improving patient selection for treatment. Cancer Discov. 9, 176–198 (2019)

    2019

  10. [18]

    Dent, R. et al. Triple-negative breast cancer: clinical features and patterns of recurrence. Clin. Cancer Res. 13, 4429–4434 (2007)

    2007

  11. [19]

    Von Minckwitz, G. et al. Definition and impact of pathologic complete response on prognosis after neoadjuvant chemotherapy in various intrinsic breast cancer subtypes. J. Clin. Oncol. 30, 1796–1804 (2012)

    2012

  12. [20]

    Lehmann, B. D. et al. Identification of human triple -negative breast cancer subtypes and preclinical models for selection of targeted therapies. J. Clin. Invest. 121, 2750–2767 (2011)

    2011

  13. [21]

    Burstein, M. D. et al. Comprehensive genomic analysis identifies novel subtypes and targets of triple-negative breast cancer. Clin. Cancer Res. 21, 1688–1698 (2015)

    2015

  14. [22]

    Keren, L. et al. A structured tumor–immune microenvironment in triple negative breast cancer revealed by multiplexed ion beam imaging. Cell 174, 1373–1387.e19 (2018)

    2018

  15. [23]

    Risom, T. et al. Transition to invasive breast cancer is associated with progressive changes in the structure and composition of tumor stroma. Cell 185, 299–310.e18 (2022)

    2022

  16. [24]

    Greenwald, N. F. et al. Temporal and spatial composition of the tumor microenvironment predicts response to immune checkpoint inhibition in metastatic TNBC. Nat. Cancer 7, 435– 450 (2026)

    2026

  17. [25]

    Denkert, C. et al. Tumour-infiltrating lymphocytes and prognosis in different subtypes of breast cancer: a pooled analysis of 3,771 patients treated with neoadjuvant therapy. Lancet Oncol. 19, 40–50 (2018)

    2018

  18. [26]

    Campanella, G. et al. Clinical-grade computational pathology using weakly supervised deep learning on whole-slide images. Nat. Med. 25, 1301–1309 (2019)

    2019

  19. [27]

    Lu, M. Y. et al. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat. Biomed. Eng. 5, 555–570 (2021)

    2021

  20. [28]

    Echle, A. et al. Deep learning in cancer pathology: a new generation of clinical biomarkers. Br. J. Cancer 124, 686–696 (2021)

    2021

  21. [29]

    Mobadersany, P. et al. Predicting cancer outcomes from histology and genomics using convolutional networks. Proc. Natl Acad. Sci. USA 115, E2970–E2979 (2018)

    2018

  22. [30]

    van Rijthoven, M. et al. Analysis of computational tumor -infiltrating lymphocytes in breast cancer from the results of the TIGER challenge. Nat. Commun. 17, 6480 (2026)

    2026

  23. [31]

    He, B. et al. Integrating spatial gene expression and breast tumour morphology via deep learning. Nat. Biomed. Eng. 4, 827–834 (2020)

    2020

  24. [32]

    Schmauch, B. et al. A deep learning model to predict RNA -Seq expression of tumours from whole slide images. Nat. Commun. 11, 3877 (2020)

    2020

  25. [33]

    Jones, E. A. et al. Imaging mass spectrometry -based molecular histology differentiates microscopically identical and heterogeneous tumors. J. Proteome Res. 12, 1847–1855 (2013)

    2013

  26. [34]

    & Mann, M

    Aebersold, R. & Mann, M. Mass-spectrometric exploration of proteome structure and function. Nature 537, 347–355 (2016)

    2016

  27. [35]

    Giesen, C. et al. Highly multiplexed imaging of tumor tissues with subcellular resolution by mass cytometry. Nat. Methods 11, 417–422 (2014)

    2014

  28. [36]

    Angelo, M. et al. Multiplexed ion beam imaging of human breast tumors. Nat. Med. 20, 436– 442 (2014)

    2014

  29. [37]

    Lewis, S. M. et al. Spatial omics and multiplexed imaging to explore cancer biology. Nat. Methods 18, 997–1012 (2021)

    2021

  30. [38]

    Consistent with this framework, our bulk proteomic analysis identified a cohort -specific risk gradient, with cell-cycle and genome-maintenance programs enriched in the high-risk group, fibroblast proliferation, FGFR signaling, oxidative -stress responses, and angiogenic progr...

    2004

  31. [39]

    Liu, Y. et al. High-spatial-resolution multi-omics sequencing via deterministic barcoding in tissue. Cell 183, 1665–1681.e18 (2020)

    2020

  32. [40]

    Mund, A. et al. Deep Visual Proteomics defines single -cell identity and heterogeneity. Nat. Biotechnol. 40, 1231–1240 (2022)

    2022

  33. [41]

    Ilse, M., Tomczak, J. M. & Welling, M. Attention -based deep multiple instance learning. In Proc. 35th International Conference on Machine Learning (ICML) 2127–2136 (PMLR, 2018)

    2018

  34. [42]

    Shao, Z. et al. TransMIL: transformer based correlated multiple instance learning for whole slide image classification. In Adv. Neural Inf. Process. Syst. 34, 2136–2147 (Curran Associates, 2021)

    2021

  35. [43]

    B., Vernardis, S

    Demichev, V., Messner, C. B., Vernardis, S. I., Lilley, K. S. & Ralser, M. DIA -NN: neural networks and interference correction enable deep proteome coverage in high throughput. Nat. Methods 17, 41–44 (2020)

    2020

  36. [44]

    Yoshihara, K. et al. Inferring tumour purity and stromal and immune cell admixture from expression data. Nat. Commun. 4, 2612 (2013)

    2013

  37. [45]

    Pan, S. et al. LCP1 correlates with immune infiltration: a prognostic marker for triple-negative breast cancer. BMC Immunol. 25, 42 (2024)

    2024

  38. [46]

    Oquab, M. et al. DINOv2: learning robust visual features without supervision. Transactions on Machine Learning Research (2024)

    2024

  39. [47]

    & Melville, J

    McInnes, L., Healy, J. & Melville, J. UMAP: uniform manifold approximation and projection for dimension reduction. Preprint at https://arxiv.org/abs/1802.03426 (2018)

    2018 arXiv

  40. [48]

    Hartigan, J. A. & Hartigan, P. M. The dip test of unimodality. Ann. Stat. 13, 70–84 (1985)

    1985

  41. [49]

    Lee, A. C. et al. Spatial epitranscriptomics reveals A -to-I editome specific to cancer stem cell microniches. Nat. Commun. 13, 2540 (2022)

    2022

  42. [50]

    Pereira, B. et al. The somatic mutation profiles of 2,433 breast cancers refine their genomic and transcriptomic landscapes. Nat. Commun. 7, 11479 (2016)

    2016

  43. [51]

    Lehmann, B. D. et al. Multi-omics analysis identifies therapeutic vulnerabilities in triple - negative breast cancer subtypes. Nat. Commun. 12, 6276 (2021)

    2021

  44. [52]

    & O'Brien, C

    Francavilla, C. & O'Brien, C. S. Fibroblast growth factor receptor signalling dysregulation and targeting in breast cancer. Open Biol. 12, 210373 (2022)

    2022

  45. [53]

    Forero, A. et al. Expression of the MHC class II pathway in triple-negative breast cancer tumor cells is associated with a good prognosis and infiltrating lymphocytes. Cancer Immunol. Res. 4, 390–399 (2016)

    2016

  46. [54]

    Elston, C. W. & Ellis, I. O. Pathological prognostic factors in breast cancer. I. The value of histological grade in breast cancer. Histopathology 19, 403–410 (1991)

    1991

  47. [55]

    Abel, J. et al. AI-powered quantification of nuclear morphology in cancers enables prediction of genome instability and prognosis. npj Precis. Oncol. 8, 134 (2024)

    2024

  48. [56]

    Conklin, M. W. et al. Aligned collagen is a prognostic signature for survival in human breast carcinoma. Am. J. Pathol. 178, 1221–1232 (2011)

    2011

  49. [57]

    Saltz, J. et al. Spatial organization and molecular correlation of tumor-infiltrating lymphocytes using deep learning on pathology images. Cell Rep. 23, 181–193 (2018)

    2018

  50. [58]

    Goltsev, Y. et al. Deep profiling of mouse splenic architecture with CODEX multiplexed imaging. Cell 174, 968–981 (2018)

    2018

  51. [59]

    Xiong, G. et al. Collagen prolyl 4-hydroxylase 1 is essential for HIF-1α stabilization and TNBC chemoresistance. Nat. Commun. 9, 4456 (2018)

    2018

  52. [61]

    Sijts, A. et al. The role of the proteasome activator PA28 in MHC class I antigen processing. Mol. Immunol. 39, 165–169 (2002)

    2002

  53. [63]

    Yuan, Y. et al. Quantitative image analysis of cellular heterogeneity in breast tumors complements genomic profiling. Sci. Transl. Med. 4, 157ra143 (2012)

    2012

  54. [64]

    Hilsenbeck, S. G. et al. Time-dependence of hazard ratios for prognostic factors in primary breast cancer. Breast Cancer Res. Treat. 52, 227–237 (1998)

    1998

  55. [65]

    Sgroi, D. C. et al. Prediction of late distant recurrence in patients with oestrogen -receptor- positive breast cancer: a prospective comparison of the breast -cancer index (BCI) assay, 21 - gene recurrence score, and IHC4 in the TransATAC study population. Lancet Oncol. 14, 1...

    2013

  56. [66]

    Cheng, Q. et al. A signature of epithelial -mesenchymal plasticity and stromal activation in primary tumor modulates late recurrence in breast cancer independent of disease subtype. Breast Cancer Res. 16, 407 (2014)

    2014

  57. [67]

    Sparano, J. A. et al. Clinical and genomic risk for late breast cancer recurrence and survival. NEJM Evid. 3, EVIDoa2300267 (2024)

    2024

  58. [68]

    A threshold selection method from gray -level histograms

    Otsu, N. A threshold selection method from gray -level histograms. IEEE Trans. Syst. Man Cybern. 9, 62–66 (1979)

    1979

  59. [69]

    & Xie, S

    Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T. & Xie, S. A ConvNet for the 2020s. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 11976– 11986 (2022)

    2022

  60. [70]

    Saillard, C. et al. H -optimus-0. https://github.com/bioptimus/releases/tree/main/models/h- optimus/v0 (2024)

    2024

  61. [71]

    Cho, Y. et al. G2L: from giga-scale to cancer-specific large-scale pathology foundation models via knowledge distillation. Preprint at https://arxiv.org/abs/2510.11176 (2025)

    2025

  62. [107]

    The cohort was divided into a development set of 107 patients and a test set of 49 patients (Table 1 and Supplementary Table S1)

    Test (n = 49) p-value Recurrence, n (%) 13 (12.1%) 7 (14.3%) 0.797 ‡ Time-to-recurrence, months (mean ± SD) * 12.4 ± 8.1 16.6 ± 8.6 0.275 † Age, years (mean ± SD) 56.5 ± 10.6 55.0 ± 13.3 0.435 † Histology type, n (%) 0.260 § Invasive ductal carcinoma 89 (83.2%) 44 (89.8%) Inva...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.