PaperlayerView exact revision on arXiv

Source-linked AI summary

Transcriptome-informed multi-modal AI for predicting neoadjuvant therapy response from breast cancer biopsies

Jungkyu Park, Dhruva Biswas, Joseph Cappadona, Cerise Tang, Ken G. Zeng, Bartosz Machura, Chuwen Liu, Paolo Tarantino, Coral Omene, Francisco J. Esteva, Rohit Bhargava, Marcin Braun, Kamila Paździerz, Jakub Czerwiński, Hanna Romańska-Knight, Albert Grinshpun, Bareket Daniel, Michele Buchinger, Frederick Howard, Piotr Wysocki, Brie Chun, Freya Schnabel, Rich Caruana, Jan Witowski, Krzysztof J. Geras

arXiv:2610.03693v1cs.AI

TL;DR

Small neoadjuvant therapy cohorts limit deep-learning prediction of breast cancer response. NEO infers gene expression from biopsy histology and combines it with clinical variables to predict pCR, outperforming histopathological biomarkers across its evaluation cohorts.

  • Problem

    Small neoadjuvant therapy cohorts with recorded response limit development of deep-learning predictors of treatment response.

    Source: §1 · p. 2

  • Method

    NEO infers transcriptome-wide expression from biopsy H&E slides, compresses it to five components, and combines these with baseline clinical variables to predict pCR.

    Source: §2.2 · p. 4

  • Results

    Across nine evaluation cohorts, NEO outperformed computational TIL biomarkers and pathologist-assessed Ki-67.

    Source: §1 · p. 2, §3 · p. 12

  • Takeaways & Limitations

    NEO estimates pCR from diagnostic biopsy slides and baseline clinical variables without adding tissue, an assay, or a manual read.

    Source: §3 · p. 12

    Because evaluation cohorts were observational and followed local practice, this study alone cannot inform specific treatment decisions.

    Source: §3 · p. 13

Abstract

from arXiv · show

Scarcity of labeled data limits development of deep learning biomarkers in oncology. We develop a two-stage AI model predicting pathological complete response (pCR) to neoadjuvant therapy in breast cancer. The first stage learns the transcriptome from histopathology using 8,742 patients across 32 cancer types, corroborated by pathologist review and spatial agreement with measured expression. This simplifies the second stage to predicting pCR from inferred expression and clinical variables. Developed using 1,080 patients (five cohorts) and evaluated in 1,412 patients (nine cohorts), the model achieves a pooled AUROC of 0.79 (95% CI, 0.73-0.85), discriminating responders within molecular subtypes. It outperforms histopathological biomarkers, remaining stable across intratumoral sampling and with minimal biopsy tissue. Ablations show transcriptome-wide inference improves discrimination over clinical variables alone or one-stage pathology models, and robustness by avoiding genomic assays' gene selection constraints. These results indicate that biologically informed compression may generalize to data-sparse applications in precision oncology.

1 Introduction

Predicting neoadjuvant therapy response is challenged by limited labeled data, biomarkers that may incompletely capture response biology, and unstable assessments. The paper proposes learning transcriptome-wide features from histopathology before predicting pCR from inferred expression and clinical variables.

1 Introduction

  • About one third of patients do not achieve pCR, motivating identification before treatment of patients at high risk of residual disease.For these patients, neoadjuvant therapy can mean months of systemic treatment and delayed definitive surgery.Source: §1 · p. 2
  • Patient-level prediction is constrained by small datasets, biomarkers with limited expressivity, and instability from tumor sampling bias and observer variability.Biopsies typically sample less than 1% of a tumor, and gene-assay risk categories can vary across sampled regions.Source: §1 · p. 2
  • The two-stage strategy learns morphology-to-expression relationships from paired pan-cancer data, then predicts response from inferred biology in smaller neoadjuvant cohorts.The authors motivate gene expression as an intermediate representation because it relates to both tissue morphology and treatment response.Source: §1 · p. 2
  • NEO combines pre-treatment biopsy H&E with baseline clinical variables, using MORPHEUS to infer expression of 14,773 genes from morphology.MORPHEUS was developed using 8,742 patients across 32 TCGA projects; NEO was developed in 1,080 patients across five cohorts and evaluated in 1,412 across nine independent cohorts.Source: §1 · p. 2

2 Results

The results characterize MORPHEUS’s recovery of gene-expression signals from H&E and NEO’s transcriptome-informed pCR prediction, including performance across cohorts and molecular subtypes, stability across slides and tissue subsamples, and the role of expression summarization.

2.1 Multi-national development and evaluation datasets

  • The study assembled 14 cohorts across seven countries, using five cohorts and 1,080 patients for development and nine independent cohorts and 1,412 patients for evaluation.The evaluation set had an overall pCR rate of 37.3%; pooled performance estimates used a random-effects model to account for cohort heterogeneity.Source: §2.1 · p. 4

2.2 NEO: a multi-modal AI model for predicting neoadjuvant therapy response

  • NEO combines baseline clinical variables with five compressed components of MORPHEUS-inferred expression, which is predicted from H&E slides using patch embeddings and multiple-instance learning.Only the final pCR-prediction step was trained on response data.Source: §2.2 · p. 4

2.3 External evaluation and benchmarking of NEO

  • NEO achieved a pooled AUROC of 0.792 (95% CI, 0.732–0.852) across nine external cohorts, with above-chance discrimination in every cohort and subtype-specific AUROCs of 0.699, 0.669, and 0.762.The subtype AUROCs correspond to TNBC, HER2+, and HR+/HER2−, respectively; calibration slopes did not differ significantly from 1 in any stratum.Source: §2.3 · p. 4, Figure 2 · p. 5
  • NEO outperformed computational TIL biomarkers and Ki-67 on shared patients, and its score remained associated with pCR after adjustment for baseline clinical variables and grade.The pooled AUROC was 0.791 (95% CI, 0.732–0.850) versus 0.516–0.583 for the four computational biomarkers; against Ki-67, AUROCs were 0.868 versus 0.648. The adjusted odds ratio was 1.46 per 0.1-unit increase (95% CI, 1.12–1.89).Source: §2.3 · pp. 4–5, Figure 2 · p. 5

2.4 MORPHEUS: H&E-inferred expression recovers biologically meaningful programs

  • On unseen CPTAC cases, MORPHEUS achieved mean per-type median expression correlations of 0.20 (0.13–0.25), rising to 0.52 (0.44–0.60) for each cancer type’s 1,000 best-predicted genes.Expression was recovered above zero in every evaluated cohort; TCGA validation correlations were higher but likely optimistic because that subset informed model selection.Source: §2.4 · p. 6
  • MORPHEUS preferentially recovered immune, stromal, and context-dependent proliferative pathways, and pathologists and spatial transcriptomics supported the morphological and spatial grounding of selected predictions.In zero-shot spatial analyses, TAGLN and IL7R predictions correlated with measured expression at Spearman ρ = +0.54 and ρ = +0.65, respectively.Source: §2.4 · pp. 6, 8, Figure 3 · p. 7

2.5 Integration of H&E-inferred expression in pCR prediction: performance, explainability, and design

  • With clinical variables, transcriptome-informed prediction reached AUROC 0.785 versus 0.740 for direct image-based prediction and 0.746 for clinical variables alone.The transcriptome-informed model’s five components had group importance 0.28 versus 0.44 for clinical variables; NEO scores also associated with proliferative and immune programs and negatively with estrogen-response programs.Source: §2.5 · p. 8, Figure 4 · p. 9
  • Random 1,000-gene subsets reached AUROCs of 0.679–0.699, close to 0.689 for all 14,773 genes, while deliberate gene selection showed no consistent advantage over random selection.Across reduction methods, performance declined as the representation grew beyond five components; ICA was selected because it targets statistically independent components.Source: §2.5 · pp. 9–10, Figure 4 · p. 9
  • H&E-inferred expression outperformed direct prediction from the same embeddings and added information beyond clinical variables alone.The representation was explainable and biologically aligned with established response correlates; a broad, heterogeneous gene list reduced to five components carried the signal, and no selection strategy consistently outperformed random selection.Source: §2.5 · p. 10

2.6 Robustness of NEO predictions to tumor sampling bias

  • NEO reached the W/T ≤0.15 stability threshold with one slide (W/T = 0.05), versus 2.5 slides for ITR and 5.8 for muTILs.The imaging-only variant required 1.4 slides, still fewer than any other biomarker; NEO was the most stable with or without its clinical inputs.Source: §2.6 · p. 10, Figure 5 · p. 11
  • For imaging-only NEO, median prediction change was 3.4% with 500 patches and 0.8% with 5,000, rising to 11% at 50 patches.Change remained below 5% at all subset sizes of 500 patches or more; 500 patches represented approximately 6 mm2 of tissue.Source: §2.6 · p. 10, Figure 5 · p. 11
  • Across 247 patients with multiple slides, NEO predictions had ICC(1,1) = 0.934 (95% CI, 0.909–0.954), and four example slides showed similar predictions across biopsy cores.Patients had a median of 4 slides each, with up to 15.Source: §2.6 · p. 10, Figure 5 · p. 11

3 Discussion

NEO estimates pCR from diagnostic biopsy H&E and clinical variables through transcriptome-wide inference, and was reported to discriminate responders across molecular subtypes and add information beyond clinical variables. The discussion connects this approach to sampling robustness and established response biology while noting limits on treatment-decision use and the need for prospective evaluation.

3 Discussion

  • Across nine independent cohorts, NEO discriminated responders within three molecular subtypes and added information beyond baseline clinical variables.The discussion also reports that NEO outperformed computational TIL biomarkers and pathologist-assessed Ki-67.Source: §3 · p. 12
  • NEO uses five interpretable MORPHEUS-ICs; its two most-used components contribute nearly monotonically to predicted pCR probability.The five components together carry just under two-thirds the group importance of the eight clinical variables, and reported correlates match established response biology.Source: §3 · p. 12
  • NEO reached a stable estimate from one slide, while computational TIL biomarkers required 2.5 to 5.8 slides.Illustrative cores from one slide gave similar predictions, and the imaging-only variant changed little with patch subsampling down to approximately 6 mm^2.Source: §3 · p. 12
  • MORPHEUS recovered immune-infiltration and stromal-remodeling signals across nearly every cancer type, with proliferation and estrogen-response signals varying by tissue context.Board-certified breast pathologists reviewed morphological features corresponding to inferred programs; the authors suggest these findings may help prioritize targets and disease settings for future pathology AI.Source: §3 · pp. 12–13
  • Transcriptome-informed prediction compresses morphology into biological features before endpoint learning, a strategy the authors say may help when response-labeled cohorts are modest.The authors expect its advantage over direct pathology prediction to narrow as training cohorts grow enough to identify relevant morphology from response labels alone.Source: §3 · p. 13
  • The observational evaluation validates NEO as a pCR predictor but cannot, on its own, inform specific treatment decisions.Other stated limitations include re-implemented TIL benchmarks, no genomic-assay comparison, random patch subsampling rather than genuinely small specimens, and exploratory secondary endpoints.Source: §3 · p. 13

4 Methods

The methods assemble pre-treatment biopsy slides, clinical variables, and pCR outcomes across development and independent evaluation cohorts. They use MORPHEUS to infer expression from H&E, then train and assess pCR models with cohort-aware validation and complementary robustness and biological analyses.

4.1 Neoadjuvant cohorts and data

  • The study used five development cohorts (1,080 patients) and nine evaluation cohorts (1,412); pCR was taken from recorded status or inferred from post-treatment pathological staging.Inputs were restricted to pre-treatment biopsy slides and baseline clinical variables, excluding resection specimens and pathological stages.Source: §4.1 · p. 14

4.2 NEO overview

  • NEO combines biopsy WSIs and baseline clinical variables, using Falcon patch embeddings from frozen foundation models with no NAT-cohort fine-tuning.Tissue was segmented into 224×224-pixel patches at 0.5 µm per pixel; Falcon variants were pretrained on large unlabeled H&E corpora.Source: §4.2 · p. 14, §4.3 · p. 14

4.4 Gene expression prediction dataset

  • MORPHEUS aggregates patch embeddings to predict 14,773 genes jointly, using paired TCGA expression and histopathology for development and CPTAC for evaluation.The TCGA dataset included 8,742 patients and 10,520 WSIs; the two Falcon variants had separate expression-prediction models.Source: §4.4 · pp. 14–15, §4.5 · p. 15

4.6 Use of H&E-inferred expression

  • For pCR prediction, 2,427 curated genes were reduced to five ICA components, which were combined with clinical variables in EBM and tabular ResNet models.The final score averages calibrated predictions across folds, model classes, and Falcon variants; ResNets also used auxiliary pathological-stage targets during training, while EBMs predicted pCR alone.Source: §4.6 · p. 15, §4.8 · p. 16

4.7 Multiple-source cross-validation

  • Multiple-source cross-validation held out one entire source cohort per fold, while keeping TransNEO and Yale in every training fold.Three cohorts—Omica, NYU, and Gundersen—were each held out in turn, limiting validation shortcuts from cohort artifacts or differing pCR prevalence.Source: §4.7 · p. 15

4.9 Model evaluation

  • Performance was assessed with cohort-level AUROCs pooled by random-effects models, with a minimum of 30 positive–negative pairs per cohort-endpoint estimate.The analyses also used bootstrap confidence intervals, Hartung–Knapp tests for specified comparisons, and a one-sided Jonckheere–Terpstra test for the prespecified RCB trend.Source: §4.9 · p. 16, §4.16 · p. 18

4.10 Calibration assessment

  • Calibration was assessed using quantile bins and logistic recalibration; the score’s added association with pCR was tested using adjusted, cohort-specific Firth logistic regressions.The adjusted analysis required all nine covariates, leaving 1,031 of 1,412 patients across seven cohorts.Source: §4.10 · pp. 16–17, §4.11 · p. 17

4.12 Measuring fidelity of H&E-inferred expression

  • Expression fidelity and biological correlates were assessed using gene-level correlations, pathway enrichment, and pathologist review of high- and low-scoring patches.Biopsy robustness was measured using repeated-slide analyses and random patch subsampling; the sampling comparison used an imaging-only variant alongside NEO and computational TIL biomarkers.Source: §4.12 · p. 17, §4.13 · p. 17, §4.14 · pp. 17–18

Data availability

Data access depends on cohort-specific custodians and agreements; some public cohorts are directly accessible.

Data availability

  • Most individual-level datasets are restricted by cohort data-use agreements; TCGA and CPTAC data are available through the NCI Genomic Data Commons and The Cancer Imaging Archive.The passage also identifies access routes for Yale, TransNEO, IMPRESS, Providence, and Nexus cohorts.Source: Data availability · p. 27

Code availability

The complete model codebase is proprietary and unavailable for public release, though access may be considered for non-commercial academic research.

Code availability

  • The complete AI model codebase cannot be publicly released; academic researchers may request access for non-commercial research subject to institutional approval and agreements.Code used to generate main figures and aggregate results may be made available upon reasonable request.Source: Code availability · p. 27

Competing interests

Several authors report financial and intellectual-property interests related to Ataraxis AI.

Competing interests

  • Twelve authors hold Ataraxis AI equity, and six are inventors on a corresponding US patent application.The remaining authors declare no competing interests.Source: Competing interests · p. 27

5 Extended Data

Extended-data analyses report NEO’s overall and subtype discrimination, assess secondary endpoints and compare NEO with benchmark biomarkers and alternative models. MORPHEUS gene-level expression fidelity is also summarized across cancer types, with the TCGA validation estimates subject to possible optimism from model selection.

Extended Data Fig. 1

  • Overall pooled AUROC was 0.792 (95% CI, 0.732–0.852); subtype pooled AUROCs were 0.699 for TNBC, 0.669 for HER2+, and 0.762 for HR+/HER2−.All-patient estimates included nine cohorts; subtype estimates included six to eight cohorts.Source: Figure 1 · p. 28, §5 · p. 28

Extended Data Figs. 2–3

  • The ablations compare NEO with five histopathological biomarkers, a clinical-only baseline, and direct pathology-based pCR models, but these captions provide no comparative outcomes.Subtype analyses compare NEO with cTILs, LTR, ITR, and muTILs on identical patients and slides.Source: Figure 2 · p. 29, Figure 3 · p. 30

Extended Data Fig. 4

  • NEO’s median score decreased from 0.486 in RCB-0 to 0.139 in RCB-III; pooled AUROC was 0.736 for nodal pCR and 0.775 for overall pCR.The nodal and overall pCR pooled estimates used the remaining 536 patients across six cohorts.Source: Figure 4 · p. 31

Extended Data Fig. 5

  • MORPHEUS fidelity plots show per-gene Pearson-correlation distributions across cancer types for TCGA validation and CPTAC evaluation; TCGA estimates may be optimistic because that subset informed model selection.Only cancer types with at least 20 evaluated cases are shown; histogram widths share a common gene-count scale.Source: Figure 5 · p. 32

A.1.1 Neoadjuvant cohort characteristics

The neoadjuvant development and evaluation sets were assembled by applying the same eligibility criteria to patients from five and nine cohorts, respectively. The cohort tables summarize patient and treatment characteristics, with treatment groups describing recorded components rather than specific protocols or sequences.

A.1.1 Neoadjuvant cohort characteristics

  • Sequential eligibility filtering retained 1,080 of 13,240 development-cohort patients and 1,412 of 6,457 evaluation-cohort patients; most exclusions occurred at the first step.Eligible patients had received NAT, were non-metastatic at diagnosis and female, had assessable response, and had a pre-treatment biopsy WSI.Source: Figure 1 · p. 33
  • General breast cancer cohorts included many adjuvant-treated patients, placing them outside the population addressed by this neoadjuvant analysis.The eligibility criteria excluded patients not treated with neoadjuvant therapy.Source: Figure 1 · p. 33
  • The cohort tables report age as median [IQR] and categorical characteristics as n (%); treatment groups denote recorded components, not specific protocols or treatment sequences.Treatment percentages use molecular-subtype totals, while other percentages use each table column’s total patient count.Source: Table 1 · p. 34, Table 2 · p. 35

A.1.2 TCGA and CPTAC cohorts

MORPHEUS expression fidelity was assessed in TCGA and independent CPTAC cohorts, with each cohort’s composition summarized by cancer project or type. None of the breast cancer neoadjuvant cohorts overlapped with TCGA or CPTAC.

A.1.2 TCGA and CPTAC cohorts

  • TCGA project training cohorts ranged from 28 to 835 patients, and 13 projects with at least 250 patients supplied 72% of the 6,994 training patients.BRCA alone contributed 12% of training patients; validation cohorts ranged from 6 to 212 patients.Source: §A.1.2 · p. 36
  • Nine TCGA projects with fewer than 20 validation patients were excluded from per-type fidelity summaries and pathway analyses.These projects were ACC, CHOL, DLBC, KICH, MESO, OV, SKCM, UCS, and UVM.Source: §A.1.2 · p. 36
  • CPTAC cancer-type cohorts ranged from 61 patients for OV to 236 for UCEC; CCRCC contributed 314 RNA-seq samples from 222 patients.CCRCC accounted for 40 of the 41 patients with multiple primary-tumor samples.Source: §A.1.2 · p. 36
  • No patient from the 14 neoadjuvant therapy cohorts appeared in TCGA or CPTAC.TCGA and CPTAC were used for MORPHEUS expression-fidelity work.Source: §A.1.2 · p. 36

A.1.3 Expression preprocessing

Expression preprocessing retained genes by prevalence, transformed each dataset separately, and standardized training targets. Expression fidelity was assessed within cohorts because transformations differed across datasets.

A.1.3 Expression preprocessing

  • 14,773 of 19,962 protein-coding genes were retained after requiring more than 0.5 TPM in at least 35% of TCGA training samples.The variance-stabilizing transformation was fitted separately to TCGA training, TCGA validation, and CPTAC count matrices to accommodate protocol and preprocessing differences.Source: §A.1.3 · p. 36
  • CPTAC composition is reported by cancer type for patients, primary-tumor RNA-seq samples, and whole-slide images; fidelity is evaluated per RNA-seq sample.Forty-one patients contribute multiple samples, and every cohort exceeds the 20-case threshold for individual display in Extended Data Fig. 5.Source: Table 4 · p. 37
  • MORPHEUS training targets were standardized using TCGA training-subset gene means and standard deviations, and predictions remain on that subset’s transformed scale.Because transformations were fitted separately across datasets, fidelity was assessed by within-cohort correlation rather than absolute agreement.Source: §A.1.3 · p. 37

A.2 Model development

The figure outlines pathology, tabular, and fused model branches for pCR prediction, alongside MORPHEUS expression regression and NEO’s calibrated ensemble.

A.2 Model development

  • The multi-modal direct pCR model combines pathology and tabular branches through gated-attention fusion, with prediction heads on each branch and the fused representation.The branches produce same-width slide and tabular embeddings before fusion; tabular inputs vary by model and include clinical variables, five MORPHEUS-ICs, or both.Source: Figure 2 · p. 37
  • The pathology branch tiles each whole-slide image into 224 × 224 patches, encodes them with the frozen Falcon encoder, and aggregates embeddings by gated attention.The attention-weighted average is projected to the shared embedding width.Source: Figure 2 · p. 37
  • MORPHEUS follows the pathology branch but replaces the classification head with a linear expression-regression head.Source: Figure 2 · p. 37
  • NEO combines MORPHEUS-derived features with clinical variables, calibrates per-fold tabular ResNet and EBM outputs, and averages the calibrated outputs.Source: Figure 2 · p. 37

A.2.1 Clinical-input encoding

Clinical variables are handled differently by the tabular ResNet and EBM, including distinct categorical encoding and missing-age treatment.

A.2.1 Clinical-input encoding

  • The tabular ResNet one-hot encodes categorical variables and adds a missing-value category for each variable.Source: §A.2.1 · p. 38
  • Missing age values are imputed with the development-cohort mean for the tabular ResNet, whereas the EBM uses built-in missing-value bins without age imputation.Source: §A.2.1 · p. 38
  • Clinical inputs were processed identically during training and evaluation.Source: §A.2.1 · p. 38

A.2.2 Training and inference algorithms

This section describes the training, inference, and evaluation procedures for NEO and direct pCR models, including feature aggregation, objectives, calibration, ensembling, benchmarks, ablations, interpretability, secondary outcomes, and slide-level stability. It also reports comparisons of model components, pathway enrichments, and limitations in reproducing external biomarker performance.

A.2.2 Training and inference algorithms

  • MSCV trained one model class per fold, selected configurations by mean validation AUROC, averaged retained-model predictions, and fitted a beta calibrator on each fold’s held-out cohort.NEO applies this procedure to each Falcon variant and model class, then averages calibrated predictions; evaluation cohorts are excluded from fitting, selection, and calibration.Source: §A.2.2 · p. 38, §A.2.11 · p. 42

A.2.3 Slide-level aggregation by gated attention

  • Gated attention aggregates patch embeddings into slide representations, while the direct multimodal model fuses pathology and clinical embeddings and reports pCR from the fused head.MORPHEUS uses the slide representation to predict 14,773 expression values; NEO averages inferred expression across slides before ICA and response prediction.Source: §A.2.2 · p. 38, §A.2.3 · p. 39, §A.2.4 · p. 39

A.2.7 Multitask learning formulation and training objective

  • Neural pCR models jointly trained on pCR, pathological T and N stage, and T- and N-stage downstaging through ten binary subtasks with cross-entropy and elastic-net regularization.Unavailable labels and indeterminate ordinal thresholds were masked; auxiliary staging and downstaging targets were used only during training.Source: §A.2.6 · p. 39, §A.2.7 · pp. 40–41

A.2.8 MORPHEUS training objective

  • MORPHEUS optimized expression prediction by regularized mean squared error, whereas EBMs used fixed settings and one fit per MSCV fold.MORPHEUS predicted 14,773 protein-coding genes and was tuned on TCGA before the pCR models were trained.Source: §A.2.5 · p. 39, §A.2.8 · p. 41, §A.2.9 · p. 41

A.2.10 Hyperparameter tuning

  • Neural pCR models used matched search ranges and 100 trials, with early stopping and selection by mean validation AUROC; MORPHEUS used a separate search budget and selection criterion.The first half of trials used random sampling and the remainder Bayesian search; EBMs were not tuned.Source: §A.2.10 · p. 41

A.2.13 Direct pCR prediction models

  • Direct pCR models bypassed MORPHEUS and ICA, using pathology alone or gated-attention fusion with clinical variables, and were trained on matched patients, partitions, and tuning budgets.Direct models average slide-level pCR probabilities, whereas transcriptome-informed models average inferred expression before ICA and response prediction; calibrated predictions were averaged across Falcon variants.Source: §A.2.13 · p. 42, §A.2.12 · p. 42
  • Pooled AUROC was 0.792 for the full NEO ensemble versus 0.785 for the multimodal neural-network pCR model alone.The direct-model ablation isolates NEO’s neural-network component; the full ensemble additionally includes the EBM.Source: §A.2.13 · p. 43

A.3.1 Pooling across cohorts

  • AUROCs and paired AUROC differences were pooled across cohorts with random-effects models, while paired comparisons evaluated both predictors on the same patient bootstrap resamples.Pooled confidence intervals and tests used the Hartung–Knapp adjustment; resamples containing only one outcome class were discarded.Source: §A.3.1 · p. 43, §A.3.2 · p. 43

A.4 Benchmark biomarkers

  • NEO was compared with four computational TIL biomarkers and Ki-67, but re-implemented LTR and ITR results differed from published values, making segmentation differences relevant to interpretation.The cTILs implementation reproduced published results; the comparisons used shared slides for computational biomarkers, while Ki-67 was assessed in a separate subset.Source: §A.4 · p. 43, §A.4.4 · p. 44, §A.4.5 · p. 44

A.5.1 Gene list selection and expression-only ablation

  • Gene-list ablations compared random, fidelity-ranked, pCR-ranked, and published signatures, while a descriptive analysis varied reduction method and component count.Ranked gene subsets showed higher inferred-expression correlations than same-size random lists, with differences narrowing as lists approached all 14,773 genes; this redundancy may help explain limited gains as list size increases.Source: §A.5.1 · p. 45, Figure 4 · p. 45, §A.5.2 · p. 46

A.6.1 Interpreting an EBM

  • EBM interpretation used feature contributions to predicted log-odds and aggregated related contributions for group importance; adjusted associations used Firth-penalized regressions, while paired score differences were pooled with random effects.The adjusted models treated unresolved equivocal HER2 status and unspecified nodal stage as missing, and dropped linearly dependent covariates within each cohort.Source: §A.3.3 · p. 43, §A.3.2 · p. 43, §A.6.1 · p. 46

A.6.2 Spatially resolved examples of the H&E-inferred expression

  • Patch-level expression predictions were compared with measurements from two spatial transcriptomic sections, neither of which was used to train, select, or calibrate MORPHEUS.The sections used Xenium panels with 280 and 5,101 targets; prediction-measurement agreement was assessed by Spearman correlation across unsmoothed patch-level values.Source: §A.6.2 · p. 46

A.6.3 Pathway analyses of the MORPHEUS-ICs

  • MORPHEUS-IC0 was enriched for interferon gamma response, while MORPHEUS-IC2 was enriched for proliferation and growth programs including E2F, G2M, MYC, MTORC1, and apoptosis.A positive normalized enrichment score indicates pathway genes are concentrated among a component’s positively weighted genes; a negative score indicates concentration among negatively weighted genes.Source: §A.6.3 · p. 47, Figure 5 · p. 47

A.6.4 Molecular correlates of NEO predictions

  • NEO-score molecular correlates were analyzed in TCGA-BRCA using differential expression, pathway and immune-signature analyses, and PAM50 subtype summaries.The analysis included MORPHEUS training patients, so stronger agreement between inferred and measured expression in those patients may overstate associations relative to an independent cohort.Source: §A.6.4 · p. 47, Figure 6 · p. 48

A.7.1 Residual cancer burden

  • Secondary analyses evaluated overall-pCR scores against residual cancer burden and nodal pCR without retraining, using outcome-eligible subsets of the evaluation cohorts.The RCB analysis was conditional on RCB availability; the nodal analysis used 536 patients across six cohorts, and RCB-0/I versus RCB-II/III was the favorable-unfavorable contrast.Source: §A.7.1 · p. 48, §A.7.2 · p. 49

A.8.1 Intraclass correlation

  • Slide consistency was assessed with an intraclass correlation coefficient, and sampling variance was modeled as decreasing with the number of slides averaged.The analysis defined the number of slides at which sampling variance reaches 15% of total variance.Source: §A.8.1 · p. 49, §A.8.2 · p. 49

Related papers

Loading 2610.03693v1…