Pretreatment DCE-MRI Resolves Response Quality Within Pathologic Endpoints in Neoadjuvant Breast Cancer

Dattatreya Kantha; Murray H. Loew

Source-linked AI summary

Pretreatment DCE-MRI Resolves Response Quality Within Pathologic Endpoints in Neoadjuvant Breast Cancer

Dattatreya Kantha, Murray H. Loew

arXiv:2608.22097v1Aug 22, 2026eess.IVcs.CVcs.LGq-bio.QM

TL;DR

pCR is clinically important but binary, so it cannot distinguish biologically different complete responses or identify complete responders who later recur. This paper tests pretreatment DCE-MRI entropy alongside pathology and finds a four-tier response-quality framework that enriches for recurrence risk without replacing pCR or RCB.

Problem
pCR is a reassuring binary endpoint, yet 5–15% of complete responders recur and current pathology provides no residual-disease-triggered signal for identifying them.
Method
The study crosses pretreatment structural entropy with pathologic response across three cohort systems to construct and evaluate a four-tier response-quality framework, with molecular analyses examining associated immune architecture.
Results
The framework spanned a 4.1-fold recurrence range in I-SPY1 and a 7.7-fold range at response extremes, while structurally adverse complete responders were enriched for recurrence risk across external analyses.
Takeaways & Limitations
Pretreatment MRI entropy adds response-quality resolution within pCR and RCB, identifying a recurrence-enriched group for prospective validation rather than determining recurrence or replacing pathology.
Takeaways & Limitations
Validated recurrence endpoints were unavailable for public I-SPY2 data, and the smallest pooled responder analysis had 33 patients and 10 events, requiring larger prospective within-pCR validation before clinical use.

Abstract

from arXiv · show

Pathologic complete response (pCR) is a strong neoadjuvant endpoint, yet 5-15% of complete responders recur and clinical/genomic variables do not reliably identify them. We tested whether pretreatment dynamic contrast-enhanced MRI entropy - intratumoral enhancement heterogeneity - resolves response quality hidden within pCR and residual cancer burden (RCB). Across four cohorts (1,200 patients), a prespecified entropy threshold defined favorable and adverse structural states. Crossing structure with pathology yielded a four-tier framework spanning 4.1-fold recurrence in I-SPY1 and 7.7-fold at response extremes. In I-SPY2, 55 of 219 complete responders (25.1%) were structurally adverse, pretreatment. In an external HER2-positive responder synthesis (I-SPY1 pathology-confirmed pCR plus UCSF best-response proxy; n = 33, 10 events), adverse structure was associated with higher recurrence risk (HR = 2.87, 95% CI 1.38-5.96) capturing 7 of 10 recurrences, enriching rather than determining risk. In a HER2-positive RCB-0 subset, recurrence was 12.5% with favorable and 80.0% with adverse structure; Firth Cox regression preserved the association (HR = 8.13, 95% CI 1.71-49.21; n = 21, 6 events). In Duke (n = 908; 76 events), favorable structure remained independently associated with lower distant-recurrence risk (adjusted HR = 0.61, 95% CI 0.41-0.91). RNA linked favorable structure to a directionally reproduced immune-architecture program among non-overlapping patients within ISPY2; EMT-pathway enrichment was favorable-side, while the adverse tier contained a broadly immune-depleted substate. Yet full-cohort RNA models weakly discriminated structural state and did not recover continuous entropy. Pretreatment MRI therefore does not replace pCR or RCB; it reveals response-quality differences that these endpoints compress and identifies a recurrence-enriched group for prospective validation.

1. Introduction

pCR is clinically valuable but binary, so it cannot distinguish biologically durable from fragile complete responses. The study proposes combining pretreatment structural entropy with pathology to resolve this hidden response-quality variation.

5–15% of complete responders recur despite pCR, which reports no residual invasive disease but not biological durability or recurrence risk.
Residual disease can trigger validated adjuvant escalation, whereas recurrent complete responders lack a residual-disease signal for further treatment decisions.
Enhancement entropy measures the distributional heterogeneity of pretreatment tumour enhancement and provides information not captured by pathology.
Crossing structural entropy with pathologic response creates four tiers that add resolution within pCR while retaining RCB’s residual-disease grading.

2. Methods

The study analyzed publicly available, de-identified data across four cohorts totaling 1,200 neoadjuvant patients, assigning each cohort a distinct discovery, validation, or transportability role. Pathology endpoints included pCR and RCB.

I-SPY2 provided discovery data, I-SPY1 external framework validation, UCSF pooled HER2-positive responder support, and Duke prognostic transportability validation.
1,200 neoadjuvant patients were assembled from I-SPY2, I-SPY1, UCSF, and Duke cohorts for distinct analytical roles.
All analyses used publicly available de-identified datasets from TCIA and associated clinical resources.
pCR was defined as no residual invasive disease in the breast and axillary lymph nodes, while RCB quantified residual invasive cancer after neoadjuvant therapy.
Table 1 organizes the study cohorts according to their analytical roles.

2.2 DCE-MRI feature extraction

Pretreatment DCE-MRI radiomics quantified tumour enhancement heterogeneity from voxel-level SER maps. Baseline first-order Shannon entropy was the primary structural feature and was evaluated across treatment timepoints and cohort systems.

DCE-MRI feature extraction: Radiomic features were extracted from SER-derived DCE-MRI maps within tumour regions defined by functional tumour volume segmentation.
DCE-MRI feature extraction: Baseline first-order Shannon entropy of the SER distribution was the primary structural feature used throughout the study.
DCE-MRI feature extraction: Entropy_T0 uses the voxel-value histogram rather than voxel adjacency, distinguishing it from spatial-texture entropy features.
DCE-MRI feature extraction: T0, T1, T2, and T3 represented pretreatment, early-treatment, inter-regimen, and presurgical imaging timepoints.
Four-tier framework: The four-tier framework crossed pathologic response with favourable or adverse structural state defined by baseline entropy.
DCE-MRI feature extraction: Low entropy indicates concentrated, relatively homogeneous enhancement, whereas high entropy indicates a broader mixture of enhancement behaviours.

2.4 Deep learning pipeline

A frozen ResNet-18 representation from baseline DCE-MRI was tested for independent recovery of radiomic structural adversity. Logistic regression predicted the radiomic label, while external validation assessed representation compatibility before outcome analysis.

Deep learning pipeline: A frozen ImageNet-pretrained ResNet-18 encoded 512-dimensional features from single axial baseline DCE-MRI slices with peritumoural context.
Deep learning pipeline: Logistic regression with bootstrap stability-based feature selection predicted radiomic-defined structural adversity from frozen CNN features.
Deep learning pipeline: A four-cell consensus taxonomy crossed radiomic and deep-learning classifications in 61 pCR patients with both analyses available.
External validation: External validity was assessed conditionally on commensurability between structural representations across cohorts.
External validation: For Duke, the frozen I-SPY2 threshold was applied directly after Entropy_Tumor was classified as regime-compatible.

2.6 Responder recurrence methodology

Responder recurrence analyses were restricted to HER2-positive patients and evaluated structural refinement using locked thresholds, audited cohorts, and survival models designed for small samples.

HER2-positive restriction provided the pCR prevalence and recurrence-event density needed for within-responder survival evaluation.
The I-SPY1 RCB analysis included 138 patients and 42 RFS events across RCB-0 through RCB-III, applying the locked MAD-z threshold unchanged.
The audited I-SPY1 HER2-positive pCR cohort narrowed from 30 raw patients to 21 in the native MAD Cox model without losing recurrence events.
Analyses used standard nonparametric, correlation, categorical, and survival tests, with effect sizes and a fragility index reported across comparisons.
Pooled responder estimates used random-effects meta-analysis, while Firth penalized Cox regression supplied small-sample hazard estimates with profile-likelihood confidence intervals.
Quality-adjusted pCR rates multiplied raw pCR rates by structural retention rates across three treatment-mechanism classes, with bootstrap pairwise comparisons.

2.8 Molecular cohort and pathway methodology

Molecular analyses combined I-SPY2 expression data with structural classification and complementary pathway-scoring frameworks to assess biological context across structural states.

Expression data were available for 699 of 701 I-SPY2 patients, including 220 pCR and 479 non-pCR patients with structural classification.
Pathway enrichment used 50 MSigDB Hallmark gene sets, with ssGSEA scores and full-rank pathway-shift testing comparing Tier 1 and Tier 2 transcriptomic distributions.
ORA tested the top 500 differentially ranked genes using Fisher’s exact test with Benjamini-Hochberg FDR correction, while gene-level analysis used Mann-Whitney U tests.
Grouped entropy-split analysis tested pathway-score separation between high- and low-entropy groups in the pooled pCR cohort.
The grouped and full-rank estimators addressed different questions: cohort-level two-group coherence versus pathway-direction structure along the ranked transcriptome.
An exploratory baseline analysis matched 704 MRI–RNA subjects and classified 524 as structurally favourable and 174 as structurally adverse under the locked operator.

2.9 Duke adjuvant and Duke 27 methodology

Duke analyses modeled entropy alongside clinical and adjuvant-treatment variables, while subgroup and external-cohort designs defined structural and molecular comparison sets.

Duke adjuvant methodology: Duke sequential Cox modeling adjusted entropy for histologic grade, nodal stage, molecular subtype, and four adjuvant treatment modalities.
Duke 27 methodology: The Duke 27 subgroup comprised HER2-positive patients with neoadjuvant chemotherapy and strict pCR, excluding residual DCIS or LCIS from the primary definition.
I-SPY1 framework methodology: I-SPY1 framework analysis used 120 eligible patients from 153 with structural and outcome data, crossing binary pCR with binary structural classification.
I-SPY1 framework methodology: The bilateral I-SPY1 framework restricted analysis to RCB-0 and RCB-3 response extremes in 70 patients.
Molecular cohort methodology: The I-SPY1 BPE-matched molecular cohort contained 83 patients, including 23 with full Tier 1/Tier 2 assignments, and was distinct from the 19-patient GSE22226 recurrence bridge.
Molecular cohort methodology: The locked-nine panel combined seven immune/inflammatory Hallmark pathways with apoptosis and EMT to form a structured 7+2 architecture.

2.13 GSE22226 transcriptomic bridge methodology

The GSE22226 bridge linked I-SPY1 HER2-positive pCR imaging, transcriptomic, and recurrence data using cross-platform harmonization and predefined pathway analyses.

GSE22226 supplied 149–150 baseline expression samples on Agilent G4502A arrays across GPL1708 and GPL4133 platforms.
The imaging-outcome crosswalk captured 16 of 21 patients from the locked I-SPY1 HER2-positive pCR cohort.
Probe mapping selected the most variable probe per gene, harmonized platforms, and scored pathways with ssGSEA using 50 Hallmark sets plus the locked-nine immune panel.
The MAD-z reconciliation projected the I-SPY2 entropy threshold onto the I-SPY1 expression cohort using a locked threshold of 0.056217.
Transcriptomic recurrence analyses used a primary HER2-positive pCR cohort of 19 patients with 4 RFS events and an all-pCR sensitivity cohort of 36 patients with 6 events.
The hypothesis sub-analysis prespecified five explanations for broad immune depletion in recurrence-enriched Tier 2, including stromal-hypoxic exclusion and cytotoxic deficits.

3. Results

Pretreatment structural entropy was independent of pCR status but identified response-quality differences within complete responders. External cohorts linked adverse structure to recurrence enrichment, including a broadly immune-depleted substate, while the framework was evaluated across distinct discovery and validation systems.

In I-SPY2, entropy distributions were indistinguishable between pCR and non-pCR patients, yet 55 of 219 complete responders (25.1%) were structurally adverse before therapy.This indicates that pretreatment structure captures variation within pCR rather than simply predicting the binary pathologic endpoint.
The external responder synthesis found higher recurrence risk for adverse structure, with pooled HR = 2.87, while enrichment was stronger inside RCB-0.The adverse complete-response state was invisible to 27 conventional clinical variables.
The study used three cohort systems with distinct roles: I-SPY2 for framework discovery and biology, I-SPY1 for external recurrence validation, and System C for pathology-confirmed within-pCR inference.System A made no recurrence claim because its internal event indicator was not a validated recurrence endpoint.

3.1 Pretreatment DCE-MRI defines a four-tier response-quality framework that adds a structural axis orthogonal to pathologic response

Pretreatment DCE-MRI entropy defines a structural axis that is orthogonal to pathologic response and persists despite convergent volumetric response. Crossing this axis with pathology produces response-quality tiers whose recurrence patterns transport across cohorts, while molecular and treatment analyses provide contextual biological and mechanistic associations.

3.1.1 The structural entropy axis reproduces across representations and institutions: Entropy was orthogonal to pCR in I-SPY2 (Mann–Whitney p = 0.966; r_b = 0.002), and adding ten structural features to FTV reduced AUC by 0.037.The null replicated across external institutions and molecular subtypes.
3.1.2 Four-tier framework definition and within-axis prognostic structure: The four-tier framework classified 25.1% of I-SPY2 pCR patients as structurally adverse (55 of 219) and produced a 9.5% → 73.3% recurrence gradient at response extremes.The framework used the I-SPY2 training-partition 75th-percentile threshold, q75 = 3.284.
3.1.2 Four-tier framework definition and within-axis prognostic structure: Matched pCR examples showed nineteen-fold entropy differences despite near-identical tumour burden, and treatment-matched patients showed opposite structural profiles under the same regimen.The burden-matched pair had FTV difference 0.028 cc, with entropy 0.22 versus 4.26.
3.1.3 Persistence, deep learning convergence, and cross-pipeline transport: The structural state persisted through treatment, with within-pCR retention of 81% (Spearman ρ = 0.671) and global structural-state persistence of 88.2%.Tier 2 patients were absent from the persistently favourable low-to-low trajectory class.
3.1.3 Persistence, deep learning convergence, and cross-pipeline transport: Within pCR, entropy remained strongly separated throughout treatment while volumetric trajectories did not significantly separate structural groups.Timepoint entropy comparisons were significant from T0 through T3, whereas all volumetric Mann–Whitney p-values were ≥ 0.219.
3.1.3 Persistence, deep learning convergence, and cross-pipeline transport: In HER2-positive RCB-0 patients, recurrence was 12.5% with favourable structure versus 80.0% with adverse structure despite convergent longest-diameter collapse.The favourable group contained 2 of 16 recurrences and the adverse group 4 of 5.
3.1.3 Persistence, deep learning convergence, and cross-pipeline transport: A frozen I-SPY2 projection associated favourable structure with lower distant-recurrence risk in Duke (HR = 0.466), while UCSF showed a 22.6-month restricted-mean-survival advantage.These findings support external transport of the structural state across cohorts.
3.1.4 External cohort transport: HER2-targeted therapy had the highest quality-adjusted pCR rate (37.7%), and favourable-structure retention was 83.3% versus 67.6% for neratinib.Entropy predicted pCR strongly in dual-antibody HER2 therapy (AUROC = 0.847) but was essentially uninformative for neratinib (AUROC = 0.517).

3.2 The structural axis corresponds to distinct immunological and clearance-mechanism states

Pretreatment structural entropy separates complete responders into reproducible immune–architecture states and reflects therapy-mechanism differences. The adverse state includes an immune-depleted substate, while transcriptomic models remain weak at recovering the continuous structural axis.

3.2.3 Transcriptomic recovery of the structural axis: Bulk RNA weakly discriminated locked structural state (out-of-fold AUROC = 0.537) and failed to reconstruct continuous Entropy_T0 (out-of-fold R² = −0.070; Spearman ρ = −0.005).
3.2.1 Structurally favourable pCR is immune-engaged: The prespecified nine-pathway panel separated structurally favourable Tier 1 from adverse Tier 2 complete response, with coordinated favourable-side pathway organisation.The panel included seven immune/inflammatory pathways plus apoptosis and EMT.
3.2.2 Entropy biology depends on response state: Higher entropy tracked lower immune-pathway activity in HER2-positive non-pCR, but showed no meaningful continuous entropy–pathway gradient across HER2-positive pCR.Within pCR, 0/9 pathways were FDR-significant and maximum |ρ| = 0.072; within non-pCR, IFN-γ showed ρ = −0.289 and FDR = 0.016.
3.2.1 Structurally favourable pCR is immune-engaged: 9/9 pathways reproduced directionally in a non-overlapping I-SPY2 HER2-positive pCR subset (n = 68), including IFN-γ response at FDR = 2.3 × 10⁻¹³.
3.2.4 Adverse complete response is estrogen-programmed and harbours a broadly immune-depleted substate: Tier 2 was estrogen-response enriched and bimodal, containing an immune-depleted Cluster 1 and an immune-engaged Cluster 2 rather than a uniformly immune-cold state.Cluster 1 comprised approximately 40% of Tier 2, while Cluster 2 comprised approximately 60%.
3.2.5 Clearance-mechanism states: Dual-antibody HER2 therapy produced structurally favourable complete response (structure-selectivity AUROC = 0.847), unlike neratinib (0.517; interaction p = 0.037).HER2-targeted therapy retained 83.3% of complete responses as structurally favourable.
3.2.6 External structural validation: In an external I-SPY1 cohort, the architecture composite separated Tier 2 from Tier 1, whereas background parenchymal enhancement did not.The Tier 2 composite was −0.661 versus +0.150 in Tier 1 (Mann–Whitney p = 0.037; 8/9 pathways concordant).

3.3 Structurally adverse pCR concentrates within-pCR recurrence risk that is biologically resolved at the substate level

Pretreatment structural adversity adds prognostic resolution within pCR and RCB-0, concentrating recurrence risk without determining recurrence. Across external cohorts, this signal was strongest in HER2-positive complete responders and preferentially tracked distant failure.

Pooled responder recurrence: HR = 2.87 (95% CI 1.38–5.96) associated structural adversity with recurrence in 33 external HER2-positive responders, capturing 7 of 10 recurrences.The pooled synthesis combined I-SPY1 pathology-confirmed pCR with a UCSF best-response proxy; adverse structure enriched risk rather than determining it.
RCB-0 concentration: 4 of 5 structurally adverse HER2-positive RCB-0 patients recurred versus 2 of 16 favourable patients, with Firth Cox HR = 8.13 (95% CI 1.71–49.21).The analysis included 21 patients and 6 events; its fragility index was 1, so the result was treated as supportive rather than standalone.
RCB-0 concentration: 44.4% versus 9.1% recurrence occurred in structurally adverse versus favourable patients among 42 all-subtype RCB-0 cases, with Cox HR = 5.50 (95% CI 1.23–24.66).RCB-0 remained structurally heterogeneous, supporting non-redundancy between residual burden and pretreatment structure.
Distant-metastasis specificity: Duke structural phenotype predicted distant but not local recurrence, with DRFS HR = 0.466 and LRFS p = 0.361.The pattern was consistent with systemic-failure biology, while recurrence events in UCSF and HER2-positive patients were predominantly distant.
Within-pCR stratification: 18 of 31 dramatic volumetric responders (58.1%) were structurally adverse despite near-complete shrinkage, showing that response magnitude did not determine response quality.Mean longest-diameter decline was 89.3% in adverse versus 93.8% in favourable cases, with median 100% in both groups.
Within-pCR stratification: Approximately 21% of HER2-positive I-SPY2 pCR patients remained structurally adverse across increasingly restrictive pathology, volumetric, genomic, and treatment criteria.Because I-SPY2 lacked a validated recurrence endpoint, this finding indicates phenotype persistence rather than recurrence stratification.

3.4 Confounder closure and treatment context

Across four cohorts, the structural phenotype showed no evidence of association with tested baseline clinicogenomic variables, while remaining independently prognostic after adjustment for adjuvant treatment and clinical factors. In a HER2-positive Duke pCR subgroup, treatment intensity—not entropy—predicted membership, and no distant recurrences occurred during follow-up.

Baseline clinicogenomic closure: All fifteen baseline clinicogenomic variables showed no evidence of association with the structural phenotype at α = 0.05.The battery included age, receptor-defined subtype, and genomic risk among other variables.
Duke adjuvant closure: HR = 0.610 (95% CI 0.41–0.91; p = 0.015) for the structural signal after adjustment for nodal stage, grade, subtype, and four adjuvant modalities.The eight-variable Duke Cox model included 908 patients and 76 distant-recurrence-free-survival events.
Duke treatment context: Zero distant recurrences occurred among 27 HER2-positive Duke patients achieving strict pCR over 3.85 years median follow-up.All patients received anti-HER2 therapy, and 19/27 (70.4%) received adjuvant radiation.
Duke treatment context: OR = 10.38 (p = 0.004) for treatment-stack intensity predicting subgroup membership, whereas entropy was not predictive (p = 0.677).The subgroup was not enriched for structurally favourable entropy and had higher median tumour volume than the comparison group.
Confounder closure: Baseline structural-immune associations were interpreted as independent of tested confounders, with the driver/confounder map targeting zero untested cells for major clinical variables.This closure supports the structural-immune axis as indexing pCR quality and within-pCR recurrence.

3.5 Clinical synthesis

The four-tier framework integrates pretreatment structural state with pathologic response to distinguish clinically interpretable response-quality states. It identifies recurrence-enriched complete responders and links favourable structure to immune engagement, while remaining a hypothesis-generating architecture rather than validated clinical guidance.

Framework: Systems A–C define four clinically interpretable response-quality states integrating structural measurement, biological identity, and clinical consequence.The framework maps structural state across pCR and residual-disease categories.
Tier 1: favourable pCR: Tier 1 favourable pCR showed an immune-engaged programme, with IFNγ replication FDR = 2.32 × 10⁻¹³ and allograft rejection replication FDR = 1.54 × 10⁻².System B recurrence was 11.5% in the full-cohort framework and 9.5% in the bilateral framework.
Tier 2: adverse pCR: Tier 2 adverse pCR contained 55 I-SPY2 patients and had an external pooled responder HR of 2.87, with recurrence predominantly distant, often fatal, and early.RNA decomposition identified broadly immune-depleted and immune-engaged substates, but did not validate Cluster 1-specific recurrence risk.
Residual disease tiers: Tier 3 favourable residual disease had 29.7% recurrence in the full-cohort framework, while Tier 4 adverse residual disease reached 47.8%.These values are from the pathology-confirmed System B full-cohort framework.
Interpretation and boundary: The structural axis functions as a gateway rather than a complete biological classifier, and the gateway-plus-resolver model lacks direct within-cohort outcome validation.The model is prospectively testable but is not current clinical guidance.

4. Discussion

The discussion presents pretreatment structural entropy as a response-quality axis that complements pCR and RCB, revealing clinically relevant heterogeneity within pathologic categories. Clinical, molecular, and treatment-context analyses support the association while retaining observational, sample-size, and validation limits.

4. Discussion: 4.1-fold recurrence separation in I-SPY1 and 7.7-fold separation at response extremes support clinically distinct response-quality tiers.The framework distinguishes structurally favourable pCR from structurally adverse pCR and organised or aggressive residual disease.
4. Discussion: A four-tier framework adds structural response quality to pCR and RCB, with cross-cohort prognostic and biological support beyond residual disease quantity.The axis is orthogonal to pCR, persists across institutions and treatment contexts, and complements RCB rather than replacing it.
4.2 Drug efficacy and treatment context: Drug mechanism influences pCR quality, offering a candidate explanation for why higher pCR rates do not always yield larger survival gains.The proposed explanation is a tier-composition disconnect: mechanisms may increase structurally adverse pCR without increasing durable disease control.
4.2 Drug efficacy and treatment context: Intensive multimodality therapy was associated with zero distant recurrences among 27 HER2-positive Duke pCR patients despite heterogeneous baseline structural profiles.Treatment-stack intensity, rather than baseline entropy, predicted membership in the durable-response subgroup.
4.4 Biological interpretation: Favourable structure was associated with an immune-architecture programme, whereas adverse Tier 2 contained a broadly immune-depleted substate.Within pCR, IFNγ and EMT pathway shifts were prominent; however, patient-level EMT scores did not reproduce the favourable-side relationship.
4.4 Biological interpretation: Bulk RNA provided biological context but weakly reconstructed structural state and failed to recover the continuous imaging phenotype in held-out analyses.The molecular findings support layered association rather than a one-way causal cascade, and do not establish immune depletion as the mechanism of recurrence.

Extended Data Figures

The extended data figures provide complementary checks on structural-phenotype convergence, multivariable prognostic stability, treatment context, causal architecture, response tiers, and molecular non-reconstruction.

Extended Data Figure 1: AUROC = 0.738 with permutation p = 0.007 was the stability-aware deep-learning estimate, while 87.5% of signal was non-entropy.The in-sample resubstitution AUROC was 0.980 and was included only for contrast.
Extended Data Figure 2: The locked Duke Cox model included structural phenotype, clinicopathologic variables, and four adjuvant treatment modalities for distant recurrence-free survival.The analysis covered 908 patients and 76 events.
Extended Data Figure 3: All 27 HER2-positive Duke pCR patients received neoadjuvant anti-HER2 therapy, while 24/27 received adjuvant anti-HER2 therapy.Adjuvant radiation was given to 19/27 and mastectomy was performed in 16/27.
Extended Data Figure 4: The causal-architecture map relates variable classes to pCR occurrence, pCR quality, recurrence within pCR, and local or distant Duke recurrence.Colors distinguish drivers, tested null results, partial contributions, and untested relationships.
Extended Data Figure 5: The four-tier framework partitions pCR into structurally favourable and adverse states and non-response into organised and aggressive residual-disease states.The adverse pCR tier has a bimodal distribution containing approximately 40% immune-depleted and 60% immune-engaged patients.
Extended Data Figure 6: Pooled out-of-fold structural-state recovery was AUROC = 0.537 with permutation p = 0.112, while RNA models failed to reconstruct continuous Entropy_T0.Sparse and PCA-based models yielded OOF R² = −0.070 and −0.036.