PaperlayerView exact revision on arXiv

Source-linked AI summary

Sparse concept attribution for histomorphological hypothesis generation from whole-slide classifiers

Tristan Lazard, Kenza Bouzid, Julius Hense, Shruthi Bannur, Daniel Coelho de Castro, Daniel Shao, Rajesh Jena, Drew Williamson, Stephanie Hyland

arXiv:2609.02985v1q-bio.QMeess.IV

TL;DR

Relating histological morphology to clinical attributes remains difficult because image interpretation is manual. SCOPE combines pathology-specific vision-language models with sparse concept attribution, recovering substantial known morphology across seven benchmark tasks while dense attribution performs like random selection.

  • Problem

    Relating histological morphology to clinical attributes remains challenging, despite slide-level models that predict clinical and biological variables from H&E images.

    Source: §1 · p. 1

  • Method

    SCOPE interprets whole-slide classifiers with a broad pathology concept bank, pathology-specific vision-language models, and benchmarked natural-language cohort explanations.

    Source: §1 · pp. 2–3

  • Results

    Across seven tasks and six organs, sparse attribution recovers substantial known morphology, whereas dense attribution is indistinguishable from random concept selection.

    Source: §1 · p. 3

  • Takeaways & Limitations

    SCOPE can express whole-slide classifiers in histopathology language and propose morphological hypotheses for expert validation, while validation on known tasks remains a proxy for discovery.

    Source: §3 · pp. 13–14

    The benchmark’s consensus references can limit measured precision because automated explanations may be more specific than the reference descriptions.

    Source: §3 · p. 13

Abstract

from arXiv · show

Histology images contain rich morphological information and can provide insights into pathological processes. However, deriving hypotheses relating morphological phenotypes to clinical attributes is bottlenecked by a manual image interpretation step. Here, we demonstrate that this process can be automated through interpretable deep learning. We present SCOPE, a method to interpret slide-level classifiers by combining pathology-specific vision--language models with sparse concept attribution onto a generalist histomorphological concept bank. To measure whether such explanations recover known morphology, we introduce MorphoRecoveryBench, a benchmark of seven tasks with pathologist-curated reference descriptions. On this benchmark, dense concept attribution is indistinguishable from a random baseline, whereas sparse attribution recovers substantial known morphology; decomposing the pooled slide embedding reaches similar explanation correctness at a fraction of the computational cost. Post-hoc interpretation of whole-slide classifiers can thus generate morphological hypotheses at scale, for expert validation.

1 Introduction

SCOPE combines pathology vision–language models, sparse concept attribution, and whole-slide classifiers to generate interpretable morphological hypotheses. MorphoRecoveryBench evaluates whether these explanations recover known morphology, showing sparse attribution outperforms dense similarity-based attribution and pooled-slide decomposition preserves correctness at lower cost.

Motivation

  • Whole-slide classifiers can predict clinical and biological quantities from H&E images, but interpreting the learned morphology requires human-verifiable explanations.MIL, pathology foundation models, and multimodal datasets have enabled prediction of cancer grades, molecular biomarkers, and patient outcomes3–9.Source: §1 · pp. 1–2

Limitations of existing explanations

  • Heatmaps, prototypes, sparse autoencoder directions, and counterfactuals identify important regions or patterns but leave morphological naming and dataset-level interpretation to pathologists10–20.These approaches therefore require laborious, subjective post-hoc interpretation and do not directly express the responsible histological features in language.Source: §1 · p. 2

Method

  • SCOPE uses pathology vision–language models21–24 with sparse, non-negative decomposition to translate image features into language while addressing spurious activations from large concept vocabularies.Its pathology-specific concept bank composes standard clinical terms into a broad hypothesis space, while sparse attribution replaces dense similarity scores.Source: §1 · p. 2

Contributions

  • SCOPE decomposes pooled slide representations directly, achieving similar explanation correctness at a fraction of tile-level computation cost.The post-hoc method allows concept banks and sparsity levels to change without retraining the classifier.Source: §1 · p. 3

Benchmark

  • MorphoRecoveryBench comprises seven binary whole-slide classification tasks with pathologist-curated ground-truth explanations and metrics for measuring recovery of known morphological differences between cohorts.The benchmark intentionally includes cancer subtyping tasks in which histological morphology is definitional.Source: §1 · p. 3

Contributions

  • Across seven tasks and six organs, sparse attribution recovers substantial known morphology, whereas dense similarity-based attribution is indistinguishable from random concept selection.This result establishes sparsity as necessary for reliable explanation on MorphoRecoveryBench.Source: §1 · p. 3

2 Results

Across six results subsections, SCOPE uses sparse concept decomposition to recover known morphology, retain predictive performance, transfer across cohorts, and generate speculative hypotheses while exposing design limits and failure modes.

2.1 Overview of SCOPE and MorphoRecoveryBench

  • SCOPE combines a 989-term generalist histomorphological concept bank with sparse, non-negative decomposition of abMIL slide embeddings and evaluates explanations on seven tasks across six organs.MorphoRecoveryBench compares generated hypotheses with pathologist-curated ground truths using LLM-judge precision, recall, jF1, direct concept-ranking nDCG@10, and expert review.Source: §2.1 · pp. 3, 5, Figure 1 · p. 4

2.2 Sparse concept attribution recovers known class morphology.

  • jF1 was 0.50 for SCOPE versus 0.47 for tile-sparse, while tile-dense reached 0.11, indistinguishable from the random floor of 0.09.Sparse attribution improves recovery by suppressing correlated, likely spurious concepts; SCOPE and tile-sparse outperform dense attribution, with gains supported by recall.Source: §2.2 · pp. 5, 7, Figure 2 · p. 6
  • SCOPE decomposes the attention-pooled slide embedding directly, propagates sparse weights into additive class log-odds contributions, and reduces decomposition time from 66.5 seconds to 5.5 milliseconds per slide.This post-hoc approach preserves explanation quality while allowing the frozen classifier to be re-explained with different concept banks or sparsity settings without retraining.Source: Figure 2 · p. 6, §2.2 · p. 7

2.3 Interpretable models retain predictive performance and transfer to an external cohort

  • Interpretable models retained near-black-box predictive performance, with CBM-SCOPE averaging AUROC 0.93 versus 0.96 for the black-box upper bound, while transfer remained task dependent.Without retraining, tile-sparse reached AUROCs of 0.86, 0.70, and 0.97 on CPTAC BRCA PAM50, COAD-MSI, and lung subtyping; explanation transfer was strongest for lung subtyping and weaker for BRCA and COAD-MSI.Source: §2.3 · pp. 8–9

2.4 What drives explanation quality?

  • Explanation quality depended on backbone, sparsity, and concept coverage: PathGen-CLIP-L reached macro jF1 0.50 versus 0.27 for MUSK and 0.22 for CONCH.jF1 peaked around 25 concepts per slide, explanation quality saturated at 989 concepts, and targeted augmentation increased jF1 by 5 points and recall by 8 points on average.Source: Figure 4 · p. 10, §2.4 · p. 11

2.5 Taxonomy of explanation errors.

  • The main explanation errors were confabulation and granularity mismatches between specific, hierarchy-free bank concepts and coarse literature-derived ground truths.The LLM templater and judge mitigated granularity differences by merging near-duplicates and matching specific terms to broader statements.Source: §2.5 · pp. 11–12

2.6 SCOPE generates morphological hypotheses across 26 molecular classification based tasks.

  • Across 26 molecular tasks, SCOPE generated 34 class-level morphological hypotheses without task-specific curation or expert input, shifting work toward selecting and validating hypotheses.For HNSC HPV status, the classifier achieved 0.835 test AUROC and the hypothesis recovered broad basaloid and lymphoid components, while more specific features remained unvalidated.Source: §2.6 · p. 12

3 Discussion

SCOPE converts trained whole-slide classifiers into named, visually grounded hypotheses about morphological differences between patient cohorts. Validation on known cancer and molecular groups demonstrates feasibility, while benchmark, representation, and architectural limitations define its role within broader discovery workflows.

  • The approach re-discovers known cancer subtypes and morphologically distinct molecular groups without prior task-specific classification specifications.It combines pre-trained vision-language models with a generalist pathology concept bank to produce natural-language explanations of discriminative morphology.Source: §3 · p. 12
  • SCOPE’s best method achieves average Judge-F1 0.50 despite classifiers reaching AUROC 0.93 across tasks, showing that explanation remains challenging even with strong discrimination.These results support feasibility but indicate that classifier performance alone does not guarantee correct morphological explanations.Source: §3 · p. 13
  • At scale, SCOPE generates 34 candidate morphologies across 8 organs for 26 tasks, shifting expert effort toward selecting and validating hypotheses.The generated concepts remain hypotheses until tested outside the producing model, for example through prospective-cohort grading and association with class labels.Source: §3 · p. 13
  • Sparse decomposition is crucial for discovery with an overcomplete concept bank, but concept coverage remains incomplete and explanation quality depends on the vision-language backbone.Explanation correctness nearly doubles with PathGen-CLIP-L relative to two state-of-the-art CLIP competitors despite similar classification accuracy, possibly because its captions contain richer morphological detail.Source: §3 · p. 13
  • MorphoRecoveryBench quantifies explanation quality but is limited by reference granularity and dataset support, which impose distinct ceilings on precision and recall.Reporting precision and recall separately preserves these ceilings and leaves their trade-off to the discovery pipeline, while the benchmark can support more elaborate evaluation criteria.Source: §3 · p. 13
  • MIL-based SCOPE cannot represent slide-level spatial distribution, limiting hypotheses about tissue architecture, co-localization, and spatial organisation beyond the tile scale.Spatially aware classifiers and explainability methods are required to extend SCOPE to such hypotheses, so it should complement rather than replace broader discovery systems.Source: §3 · pp. 13–14

4 Methods

The methods establish an interpretable framework for slide classification and explanation, spanning morphological concept representation, model pipelines, profile aggregation, and benchmark construction. They also define evaluation procedures for explanation quality and concept relevance, including automated validation, random-baseline calibration, and reproducible statistical assessment.

4.1 Describing tiles and slides with concepts

  • Slides are represented by comparable K-dimensional profiles over a fixed morphological concept bank, enabling task-independent concept attribution.The default bank contains K = 989 concepts embedded with PathGen-CLIP-L text features.Source: §4.1 · pp. 14, 16
  • The generalist overcomplete bank mines Quilt-1M captions, removes diagnostic or non-morphological terms, deduplicates concepts, and frequency-filters candidates.The medium bank retains the 989 most frequent concepts, roughly the top 10% of nominal groups.Source: §4.1 · pp. 14, 16, Figure 5 · p. 15

4.2 Classifier models and interpretable features

  • Four pipelines vary classifier architecture and attribution granularity, producing slide profiles whose concept contributions additively decompose class predictions.Tile-level methods use SI-MIL with dense or sparse attribution, whereas SCOPE decomposes an attention-pooled slide embedding and CBM-SCOPE drops the residual.Source: §4.2 · pp. 16–18
  • Sparse attribution reconstructs aligned tile or slide embeddings with non-negative concept combinations, suppressing spurious dense activations while preserving exact additivity through a residual term.Tile and concept embeddings are centered by modality-specific means before SpLiCE decomposition to address the CLIP modality gap.Source: §4.2 · pp. 17–18

4.3 Training and inference

  • Training uses frozen tile and concept features with downstream MIL optimization, capped 2,000-tile slide bags, balanced sampling, and patient-disjoint test evaluation.SI-MIL combines cross-entropy, weighted prediction alignment, and a differentiable top-k selection of 20 informative tiles.Source: §4.3 · pp. 18–19

4.4 From profiles to class explanations

  • Class explanations average slide profiles, separate positive and negative contributions, and use an LLM templater to convert sufficiently large concept scores into concise prose.Only concepts with absolute contribution exceeding 10^-4 are passed to the templater, which coalesces semantically close concepts.Source: §4.4 · p. 19

4.5 The MorphoRecoveryBench benchmark

  • MorphoRecoveryBench contains seven TCGA binary tasks across six organs, with literature-derived and pathologist-validated ground truths covering 11 informative class-directions.The benchmark includes five histological subtyping and two molecular tasks; PANDA prostate detection is development-only, while CPTAC cohorts provide external validation.Source: §4.5 · p. 20

4.6 Evaluation framework

  • Evaluation reports AUROC for classification and uses GPT-5.2 claim matching against ground-truth morphology to compute precision, recall, and F1, with metrics validated against expert annotation.Generated claims are labeled CORRECT or INCORRECT, while reference patterns are labeled MATCHED or MISSING.Source: §4.6 · pp. 20–21
  • Generated explanations are evaluated by decomposing proposed claims and ground-truth features to compute clinically tolerant precision, recall, and F1.Paraphrases, biological equivalents, and well-known clinical correlates count as matches rather than requiring exact wording.Source: §4.6 · p. 21
  • Concept-ranking evaluation labels each bank concept as POS, NEG, or NONE and reports nDCG@10 alongside judge-based jF1.An LLM determines whether each concept matches the positive or negative class description, and nDCG measures the relevance and ordering of top-ranked concepts.Source: §4.6 · p. 22
  • A pathologist validates the automated judge by applying the same claim-level protocol to a subset of generated explanations, yielding human precision, recall, and F1.The evaluated explanations come from one arbitrarily selected experimental run across all tasks and models.Source: §4.6 · p. 22
  • The random-concept floor samples size-matched concepts with random positive weights, repeats the process over 20 draws per class, and passes profiles through the same templating, judging, and ranking pipeline.This baseline estimates explanation quality attainable without learned attribution; results are reported as mean and standard deviation.Source: §4.6 · p. 22
  • Metrics are reported across five training seeds or aggregate units with mean and standard deviation or standard error, while paired comparisons use two-sided Wilcoxon tests at a 0.05 significance level.Pearson and Spearman correlations assess judge–expert agreement and augmentation effects, and LLM-stage sampling variance is separately estimated through repeated reruns.Source: §4.6 · pp. 22–23

Extended Data

The Extended Data evaluates held-out classification, explanation metrics, calibration, stochasticity, concept-bank and sparsity sensitivity, ground-truth augmentation, and transfer to an external cohort. Together, these analyses probe robustness and operating conditions of SCOPE beyond the main benchmark.

Held-out classification

  • Held-out AUROC is reported per task for interpretable models and AbMIL (Virchow2), a black-box attention-MIL comparator using a different backbone.Results are summarized as mean±std over five seeds.Source: Table 1 · p. 29

Metric agreement

  • Human checklist metrics and LLM-judge metrics are correlated over all 47 graded explanations from tile-level pipelines and SCOPE.The analysis pools explanations across the evaluated pipelines.Source: Table 2 · p. 29

Margin calibration

  • CBM-SCOPE margin calibration relates the residual-free reconstruction to the full SCOPE margin, while AUROC remains comparatively stable after residual removal but balanced accuracy can change.The figure includes seven-task fits, cross-task Pearson correlations, and performance changes at the uncalibrated default decision boundary.Source: Figure 1 · p. 30

Robustness analyses

  • Concept-bank-size sweeps, ground-truth concept augmentation, and templating/judging repeats assess how explanation quality depends on representation coverage and evaluation stochasticity.The augmentation analysis reports per-class metrics, macro averages, and missing canonical concepts, while the stochasticity analysis separates judge and templater variability.Source: Table 3 · p. 29, Table 4 · p. 31, Table 6 · p. 33

Sparsity sensitivity

  • Judge F1 peaks at the default L0 = 25 active-concept budget and declines in sparser or denser regimes, while nDCG@10 is roughly flat up to default before declining.Per-class optima span 2–100 concepts per slide, whereas the default shared budget is L0 = 25.Source: Table 5 · p. 31, Figure 2 · p. 32

External transfer

  • Transfer experiments report full SCOPE, residual-free CBM-SCOPE, and tile-sparse AUROCs together with templated-profile judge precision, recall, and F1 on external CPTAC and CPTAC-trained settings.External evaluation applies TCGA-trained models to the full CPTAC cohort; CPTAC-trained models use a held-out 15% test split.Source: Table 7 · p. 33

Supplementary Information · S1 CBM-SCOPE margin calibration

CBM-SCOPE retains most black-box classification accuracy, but dropping the residual alters decision-margin scale and offset. Although AUROC changes little, uncalibrated balanced accuracy can shift substantially, necessitating calibration before use.

S1 CBM-SCOPE margin calibration

S2 Benchmark creation details

MorphoRecoveryBench ground truths combine literature-derived candidate features with pathologist curation into concise reference descriptions used for judging explanations and checklist-based recovery assessment. Table S1 records these explanations across evaluated task directions and identifies bilateral versus unilateral evaluation, excluding uninformative negative classes.

Consensus ground truth

  • Ground truths were built by collecting candidate morphological features from literature, then having a board-certified pathologist remove generic or ambiguous features and condense the remainder into one-to-three sentences.These reference sentences support both explanation judging and expert checklists marking features as recovered or missed.Source: S2 Benchmark creation details · p. 34

Reference scope

  • The benchmark’s reference descriptions cover seven task domains: breast, lung, kidney, liver, brain, breast PAM50 subtype, and colorectal microsatellite instability.The listed task-specific literature references informed consensus construction, with WHO Classification of Tumours books preferred when sufficiently clear descriptions were available.Source: S2 Benchmark creation details · pp. 34–35

Ground-truth task structure

  • Table S1 lists the complete ground-truth explanation for every evaluated class direction and labels each task as bilateral or unilateral.Unilateral tasks omit the uninformative negative class and provide no gathered ground-truth explanation for it.Source: S2 Benchmark creation details · p. 35, Table S1 · p. 35

Example reference descriptions

  • Examples of curated explanations include biliary or hepatocellular architecture in liver tumors, fried-egg and chicken-wire patterns in oligodendroglioma, and fibrillary microcystic glial features in astrocytoma.Additional unilateral descriptions specify high-grade basal-like breast morphology and immune-, mucin-, plasma-cell-, and medullary features of MSI-high colorectal cancer.Source: S2 Benchmark creation details · p. 36

S3 Human evaluation details

The expert checklist subset contains 47 explanations graded by one pathologist, spanning tile-level pipelines and SCOPE across the evaluated tasks. Full method-level means and pooled human–LLM judge correlations are reported in Table S2 and Extended Data Table 2.

Expert checklist subset

  • The evaluation comprised 47 explanations: 36 from tile-level pipelines across eight datasets and 11 from SCOPE on seven TCGA tasks.The SCOPE explanations used a single seed and were paired with matching-seed judge scores.Source: S3 Human evaluation details · p. 37

Method-level evaluation

  • Table S2 reports human precision, recall, and F1 alongside corresponding LLM-judge scores for the three tile-level pipelines and SCOPE.Human metrics were computed from pathologist annotations on the same explanations evaluated by the automated judge.Source: Table S2 · p. 37, S3 Human evaluation details · p. 37

S4 Judge and templating stochasticity

Stochastic templating and judging contribute substantially to apparent variability in SCOPE explanations. The judge accounts for 39% and the templater for 61% of their combined variance, with phrasing contributing the larger share.

Noise quantification

  • Templating and judging alone produce a Judge-F1 standard deviation of σtotal = 0.09 on fixed SCOPE profiles.The template-then-judge pipeline was rerun 20 times while holding the SCOPE model fixed; SpLiCE decomposition made the profiles deterministic.Source: S4 Judge and templating stochasticity · p. 37
  • The 0.09 templating-and-judge standard deviation is comparable to Table 2’s 0.08 across-seed standard deviation, indicating that much nominal seed variability is LLM sampling noise.Each Table 2 per-seed cell combines one stochastic templating call with one stochastic judge call.Source: S4 Judge and templating stochasticity · p. 37

Variance decomposition

  • Freezing the templater and re-judging each class 10 times isolates judge variability at σjudge = 0.05.The templater’s contribution is then inferred by quadrature from the combined standard deviation.Source: S4 Judge and templating stochasticity · p. 37
  • The templater contributes 61% and the judge 39% of combined templating-plus-judge variance, making phrasing the larger but both stages modest contributors.These variance shares were estimated from repeated re-judging and templating experiments.Source: S4 Judge and templating stochasticity · p. 37

S5 Qualitative explanation panels

The appendix makes SCOPE’s qualitative grading transparent by pairing model explanations with matched, hallucinated, and missed concepts and repeating judge precision, recall, and F1. Across diverse histologic classes, the panels show variable explanation quality, with F1 scores from 0.33 to 0.71 and explanations that often describe morphology overlapping the ground truth.

Panel grading

  • Panel annotations deterministically expose grading decisions: matched concepts share tints, hallucinations are marked red, missed concepts grey, and each header reports precision, recall, and F1.Matched model words are linked to corresponding ground-truth phrases, while false positives count against precision and false negatives against recall.Source: S5 Qualitative explanation panels · pp. 37–38

Qualitative panels

  • For BRCA lobular, the reported precision, recall, and F1 are 0.60, 0.67, and 0.63, while CRC MSI-high has precision 0.57, recall 0.50, and F1 0.53.The BRCA lobular explanation describes dense sclerotic/collagenous fibrosis with infiltrative plasmacytoid/signet-ring–like cells. The CRC MSI-high explanation describes mucinous epithelial proliferation, abundant blue mucin, glandular/intestinal-type architecture, and an associated inflammatory component.Source: S5 Qualitative explanation panels · p. 38
  • 0.71 F1 was achieved for LGG oligodendroglioma, with P 0.62 and R 0.83, alongside descriptions of blue-cell sheets, perivascular pseudorosettes, halos, salt-and-pepper chromatin, and calcifications.These features overlap the ground truth’s monotonous nuclei, clear cytoplasm, microcalcifications, and chicken-wire capillary pattern.Source: S5 Qualitative explanation panels · p. 39
  • 0.33 F1 was obtained for RCC papillary/chromophobe, with P 0.29 and R 0.40, while the explanation emphasized papillary architecture, oncocytic cytoplasm, and fibrovascular cores despite ground-truth differences in vascular networks.The ground truth allows papillary architecture and eosinophilic cytoplasm but notes that the small-vessel network may be absent.Source: S5 Qualitative explanation panels · p. 40
  • 0.60 F1 with perfect recall was reported for RCC clear-cell, whose explanation identified clear cells, nests or trabeculae, thin septa, and a delicate capillary network matching the ground truth.The model also added possible xanthomatous differentiation and partial encapsulation, which are not stated in the supplied ground truth.Source: S5 Qualitative explanation panels · p. 40

S6 Qualitative TCGA-to-CPTAC explanation transfer

Qualitative paired runs show that TCGA-trained explanations transfer to CPTAC with task-dependent precision, recall, and F1, while preserving some shared morphological concepts. Transfer is strongest for NSCLC squamous carcinoma and weakest for CRC MSI-high.

Cross-cohort comparison

  • Across paired TCGA-to-CPTAC runs, explanations are compared using matched, unmatched, and missed concepts alongside independently judged precision, recall, and F1.Each panel uses the same TCGA-trained model evaluated on its internal TCGA cohort and external CPTAC cohort.Source: S6 Qualitative TCGA-to-CPTAC… · p. 40

BRCA basal-like

  • BRCA basal-like declined to CPTAC P 0.36 · R 0.30 · F1 0.33 from TCGA P 0.56 · R 0.60 · F1 0.58, despite shared necrosis and keratinizing features.The CPTAC explanation emphasized debris, fat necrosis, microorganisms, stromal features, and epithelial atypia, whereas the TCGA explanation included lymphoid and heterologous elements.Source: S6 Qualitative TCGA-to-CPTAC… · p. 40

CRC MSI-high

  • CRC MSI-high changed from TCGA P 0.43 · R 0.38 · F1 0.40 to CPTAC P 0.60 · R 0.14 · F1 0.23, reflecting low recall despite shared mucinous morphology.Both explanations identified abundant extracellular mucin and mucinous epithelium, while CPTAC additionally emphasized signet-ring change, papillary growth, sebaceous differentiation, and chronic inflammation.Source: S6 Qualitative TCGA-to-CPTAC… · p. 41

NSCLC adenocarcinoma

  • NSCLC adenocarcinoma achieved CPTAC P 0.44 · R 0.60 · F1 0.51 versus TCGA P 0.43 · R 0.80 · F1 0.56, with partially preserved glandular morphology.Both cohorts captured irregular glands, alveolar patterns, cribriform and papillary/micropapillary architecture, mucinous differentiation, and cytologic atypia.Source: S6 Qualitative TCGA-to-CPTAC… · p. 41

NSCLC squamous

  • In NSCLC squamous, transfer from TCGA to CPTAC yielded P 0.40, R 1.00, and F1 0.57, compared with TCGA P 0.36, R 0.67, and F1 0.47.Both explanations describe squamous differentiation, keratinization, intercellular bridges/desmosomes, basaloid areas or components, peripheral palisading, and comedo-type necrosis.Source: S6 Qualitative TCGA-to-CPTAC… · p. 41

S7 Evaluation of concept scoring

Concept scoring achieves high predictive accuracy but produces substantial co-prediction noise, which sparsification reduces while improving downstream explanation quality. Simulation results show that both concept accuracy and sparsity are crucial for recovering the ground-truth concept.

S7.1 Measuring the quality of concept scoring

  • The evaluation scored 200 kidney-relevant concepts on positive and control patches, using target concept AUROC and the number of concepts reaching AUROC ≥0.65 as metrics.Target concept AUROC measures identification accuracy, whereas co-predicted concepts quantify predictive noise in the concept bank.Source: S7.1 Measuring the quality of concept… · p. 43
  • For “clear cytoplasm,” target concepts achieved AUROCs above 0.88, but semantically related and unexpected concepts were also highly predictive.These activations indicate systematic model- or data-specific spurious concept responses.Source: Figure S1 · p. 42, S7.1 Measuring the quality of concept… · p. 43
  • SpLiCE reduced co-predicted concepts from 83.0 to 1.9 on average, while target concept AUROC decreased from 0.731 to 0.582.The sparse representation enabled stronger global explanation quality despite lower target prediction performance.Source: Figure S1 · p. 42, S7.1 Measuring the quality of concept… · p. 43

S7.2 Connecting concept scoring quality to explanation outcomes

  • When AUROC was 0.75 with at least 15 spurious concepts, ground-truth explanations usually failed to rank first, with MRR ≤0.3.Increasing spurious concepts rapidly reduced explanation quality, especially when the ground-truth concept score was imperfect.Source: Figure S1 · p. 42, S7.2 Connecting concept scoring… · p. 44

S8 Supplementary atlas of generated morphological hypotheses

The supplementary atlas presents generated morphological hypotheses and descriptions across molecular-classification tasks and multiple organs, linking molecular or histological attributes to tissue phenotypes. Examples span cellular morphology, architecture, stroma, vasculature, and inflammatory features, while remaining unvalidated hypotheses.

Atlas scope

  • The atlas comprises 34 class-level morphological hypotheses across 26 tasks and 8 organs, including all positive-class hypotheses.Of 35 candidate TCGA molecular classification tasks, 9 were excluded because their held-out classifiers did not exceed test AUROC 0.6.Source: S8 Supplementary atlas of generated… · p. 45

Interpretation and limitations

Breast atlas examples

  • Breast hypotheses linked molecular classes to pleomorphic, necrotic, inflammatory, fibrotic, ductal, and squamous morphological patterns.Examples included HRD-positive and TP53-mutant pleomorphic tumors, PIK3CA-mutant sclerotic fibrosis, and PR-positive versus PR-negative architectural differences.Source: S8 Supplementary atlas of generated… · p. 46

Colorectal atlas examples

  • Colorectal hypotheses associated APC, BRAF, consensus, hypermutation, MSI, and TP53 classes with intestinal, mucinous, desmoplastic, inflammatory, and necrotic patterns.Reported features included villi and Paneth cells, mucin-rich cribriform or micropapillary glands, collagen-rich stroma, intraepithelial lymphocytes, and comedo-type debris.Source: S8 Supplementary atlas of generated… · p. 47

Head-and-neck atlas examples

  • Head-and-neck hypotheses distinguished HPV-positive basaloid lesions with lymphoid infiltration from HPV-negative keratinizing squamous lesions with dyskeratosis and koilocytic change.Both hypotheses had test AUROC 0.835.Source: S8 Supplementary atlas of generated… · p. 48

Kidney

  • Kidney hypotheses distinguish BAP1- and PBRM1-mutant classes through clear or sebaceous cells, papillary or tubular architecture, adipocytic and xanthomatous change, and vascular or fibrous features.BAP1-mutant examples include prominent sebaceous differentiation and papillary projections on fibrovascular stalks, whereas PBRM1-mutant examples emphasize tubular structures, cystic dilatation, dense fibrous tissue, and mature fat.Source: S8 Supplementary atlas of generated… · pp. 48–49

Brain

  • Brain hypotheses describe oligodendroglioma with perivascular pseudorosettes and perinuclear halos, versus astrocytoma with astrocytic differentiation, neuropil, microglia, and variable atypia.The oligodendroglioma description includes salt-and-pepper chromatin, solid sheets, minimal inflammation, and possible microvascular proliferation; astrocytoma includes focal myxoid or reticular degeneration and occasional atypical mitoses.Source: S8 Supplementary atlas of generated… · p. 49

Liver

  • Liver hypotheses differentiate CTNNB1-mutant lesions by trabecular hepatic architecture, sinusoids, bile, steatosis, and eosinophilic cytoplasm from TP53-mutant lesions with pleomorphism, inflammation, and giant cells.The CTNNB1 description emphasizes Kupffer cells and lobular–trabecular growth, while TP53 includes sinusoidal vasculature, fibrovascular stroma, macronucleoli, osteoclast-type giant cells, and hobnail- or glomus-like morphology.Source: S8 Supplementary atlas of generated… · p. 49

Lung

  • Lung hypotheses distinguish EGFR-mutant papillary or micropapillary epithelial tumors, KRAS-mutant alveolar-pattern lesions, and TP53-mutant high-grade pleomorphic infiltrates.EGFR-mutant lesions show fibrovascular cores and ductal–glandular architecture; KRAS-mutant lesions show enlarged air spaces, macrophages, inflammation, and ground-glass appearance; TP53-mutant lesions include monster cells, necrosis, acantholysis, and pigment-laden macrophages.Source: S8 Supplementary atlas of generated… · p. 50

Stomach

  • Stomach hypotheses associate ARID1A-mutant and hypermutated classes with distinct inflammatory or gland-forming phenotypes, while Lauren subtypes emphasize signet-ring or intestinal-type glandular morphology.ARID1A-mutant lesions are predominantly lymphoplasmacytic and lichenoid with ulceration, whereas hypermutated lesions show irregular or cribriform mucin-producing glands, pleomorphism, necrosis, and lymphoid inflammation.Source: S8 Supplementary atlas of generated… · pp. 50–51

Related papers

Loading 2609.02985v1…