Source-linked AI summary
TIRA: Tumor Immune Representation Adaptation for Zero-Shot Cross-Cancer MSI and TMB Prediction
Dasari Naga Raju
TL;DR
Histopathological MSI and TMB prediction remains challenging when models transfer across morphologically distinct cancers, despite recurring immune-associated spatial patterns. TIRA uses source-derived topology to guide frozen foundation-model representations without target-domain adaptation. Across transfer settings, topology-informed adaptation improves robustness, while attention analysis links guidance to morphological evidence around tumor–immune interfaces.
Problem
Histology-based MSI and TMB prediction must generalize across morphologically distinct cancers, while foundation-model predictors do not explicitly use recurring spatial immune patterns.
Source: §1 · pp. 1–2Method
TIRA uses source-derived spatial immune topology to adapt frozen pathology foundation-model representations and guide tile attention while pooling morphology for joint prediction.
Source: §1 · p. 2, §6 · p. 11, §5 · p. 11Results
Across three foundation models and multiple transfer settings, TIRA improved robustness under cancer-type shift, and attention analysis implicated tumor–immune interfaces as selected morphological evidence.
Source: §5 · p. 11, §6 · p. 12Takeaways & Limitations
Spatial immune topology provides a source-derived biological constraint for improving foundation-model transfer across cancer types.
Source: §6 · p. 12The tissue classifier is trained on colorectal histology, so organ-specific morphology in gastric and endometrial cohorts may affect tissue-class probabilities and downstream topology estimates.
Source: §5 · p. 11
Abstract
from arXiv · showhide
Microsatellite instability-high (MSI-H) and high tumor mutational burden (TMB-H) are clinically relevant biomarkers, yet their histopathological prediction remains challenging when models are transferred across morphologically distinct cancer types. Immune-associated spatial patterns can persist across cancers despite these morphological differences, but foundation-model-based predictors trained on a single cancer do not explicitly use this information, limiting cross-cancer generalization. To address this limitation, we propose TIRA (Tumor Immune Representation Adaptation), a target-free framework that refines frozen foundation-model representations using spatial immune topology, without requiring target-domain data during model development or test-time adaptation. TIRA uses a topology-supervised biology representation to condition tile-level attention while pooling only morphological features for joint MSI and TMB prediction. We train TIRA on TCGA-COAD+READ and evaluate it zero-shot on CPTAC-COAD, TCGA-STAD, TCGA-UCEC, and CPTAC-UCEC, covering cross-site, cross-cancer, and combined cross-cancer-site distribution shifts under UNI2, CONCH, and Virchow2. With UNI2, TIRA improved zero-shot AUROC on TCGA-STAD from 0.633 to 0.766 for MSI and from 0.651 to 0.772 for TMB. Source-derived spatial immune topology improved the cross-cancer robustness of frozen pathology foundation-model representations.
1 Introduction
MSI and TMB prediction from histology faces a cross-cancer generalization challenge because biomarker-relevant information must transfer across distinct morphologies. TIRA uses source-derived spatial immune topology to guide morphology-based prediction without target-domain adaptation.
1 Introduction
- Recurring immune-associated histological patterns motivate using spatial immune organization as a prior for joint MSI and TMB prediction across morphologically distinct cancers.MSI and TMB are related but distinct biologically, and their clinical assessment relies on molecular tests, motivating prediction from routine H&E images.Source: §1 · pp. 1–2
- TIRA uses source-derived spatial immune topology to refine frozen pathology foundation-model representations for prediction on unseen cancer types.Its biology-guided MIL strategy conditions tile attention on topology-informed representations while pooling morphological features for joint biomarker prediction.Source: §1 · p. 2
2.1 Overview
- TIRA learns a topology-supervised biology representation in Stage 1, then uses it to guide tile attention while pooling only morphology in Stage 2.Stage 1 uses no MSI/TMB supervision, and the projector is frozen before Stage 2.Source: §2.1 · pp. 2–3, Figure 1 · p. 3
2.2 Source–Reference Direction Suppression
- Source–reference direction suppression estimates a direction from pooled source-subset embeddings, removes each tile’s component along it, and applies the frozen direction unchanged to targets.The source cohorts are paired as cancer cohort C and READ reference.Source: §2.2 · p. 3
2.3 Spatial Immune Topology
- A colorectal-trained tissue classifier supplies tile-class probabilities and coordinates for 18 candidate descriptors of immune organization and spatial distribution.The candidates cover tissue abundance, lymphocyte confidence, local organization, multiscale TIL density, immune mixing, stromal infiltration, and TLS-related features.Source: §2.3 · pp. 3–4
2.4 Biological Screening of Descriptors
- In 227 TCGA-COAD patients, descriptors with mean Spearman correlation above 0.10 across five immune genes are retained, yielding 12 topology targets without MSI or TMB labels.The descriptor screen uses matched bulk RNA-seq; the selected identities remain fixed across foundation models.Source: §2.4 · p. 4, Table 1 · p. 4
2.5 Stage 1: Topology-Supervised Projector
- Stage 1 trains a projector on slide-level mean residual embeddings to predict standardized retained descriptors and align the projected source-subset distributions with MMD.The projector is frozen after training, which uses no MSI or TMB labels.Source: §2.5 · pp. 4–5
2.6 Stage 2: Biology-Guided Attention MIL
- Stage 2 conditions tile attention jointly on morphology and topology-informed biology, but pools only morphological features for separate MSI and TMB classification heads.The topology-informed representation guides tile selection rather than serving as the predictive representation.Source: §2.6 · pp. 5–6
3 Experimental Setup
The study trains on TCGA-COAD+READ and evaluates zero-shot transfer across cohorts, cancer types, and combined cancer-site shifts. It compares frozen foundation-model representations and MIL baselines under a shared evaluation setup.
3.1 Cohorts and Labels
- Training uses TCGA-COAD+READ, with zero-shot evaluation on CPTAC-COAD, TCGA-STAD, TCGA-UCEC, and CPTAC-UCEC across cross-site, cross-cancer, and combined shifts.Reciprocal transfer also uses TCGA-STAD+READ and TCGA-UCEC+READ as alternative sources, while the 12-descriptor topology set remains fixed from the initial screening.Source: §3.1 · p. 6
3.2 Foundation Models, Baselines, and Evaluation
- The evaluation compares frozen UNI2, CONCH, and Virchow2 representations with five MIL baselines using shared patient partitions, frozen features, and an 8,000-tile-per-bag input limit.Patient-level AUROC is primary; confidence intervals and paired AUROC differences use 10,000 bootstrap resamples, while source results are cross-validated Stage 2 performance rather than fully nested out-of-fold pipeline evaluation.Source: §3.2 · p. 6
4 Results
Across zero-shot cross-cancer targets, TIRA improved matched ABMIL performance across foundation models, but gains depended on encoder, target cohort, and source–target pairing. Ablations and attention analyses implicate topology supervision and biology-guided selection, while calibration gains were not uniform.
4.1 Three-Foundation-Model Benchmark
- Across all three feature spaces, TIRA improved over matched ABMIL on TCGA-STAD and TCGA-UCEC; under UNI2, STAD MSI AUROC rose from 0.633 to 0.766 and TMB AUROC from 0.651 to 0.772.On TCGA-UCEC under UNI2, MSI increased from 0.515 to 0.595 and TMB from 0.528 to 0.587; improvement magnitude varied by foundation model and target cohort.Source: Table 2 · p. 7, Table 3 · p. 7, §4.1 · pp. 7–8
4.2 Reciprocal Cross-Cancer Transfer
- Reciprocal transfer was source–target dependent: UCEC+READ improved both biomarkers on TCGA-STAD but reduced both on TCGA-COAD, while CPTAC-UCEC showed a different pattern.Across several source configurations TIRA improved cross-cancer transfer, but the magnitude and direction of change varied.Source: Table 4 · p. 8, §4.2 · p. 8
4.3 Ablation Study
- Removing topology supervision caused the largest cross-cancer performance reduction; removing biology-guided attention also reduced performance, while direction suppression helped cancer-type shift more than same-cancer site variation.Removing direction suppression improved CPTAC-COAD MSI but reduced performance on TCGA-STAD and TCGA-UCEC; the full model combined these mechanisms.Source: Table 5 · p. 8, §4.3 · pp. 8–9
- READ as the suppression reference yielded higher AUROC than matched random-COAD and no suppression across STAD/UCEC endpoints, but the comparison does not establish READ as uniquely optimal.The reference-choice analysis used only one matched random-COAD reference.Source: Table 6 · p. 9, §4.3 · p. 9
4.4 Attention–Immune Alignment
- TIRA increased tumor–lymphocyte interface attention enrichment from 3.34 to 6.09 on TCGA-STAD and from 3.19 to 5.03 on TCGA-UCEC, with redistribution varying by cohort and biomarker subgroup.Interface tiles are tumor-classified tiles with at least one lymphocyte-classified neighbor among their 30 nearest spatial neighbors; representative maps encode attention as log10 weights and outline interface tiles in cyan.Source: §4.4 · pp. 9–10, Figure 2 · p. 9, Table 7 · p. 10
4.5 Calibration Analysis
- On TCGA-STAD, TIRA reduced both ECE and Brier score for MSI and TMB; on TCGA-UCEC, it improved TMB calibration while ABMIL retained lower MSI ECE and Brier score.Calibration used raw zero-shot probabilities without post-hoc target recalibration, and discrimination and calibration did not change uniformly under target shift.Source: Table 8 · p. 10, §4.5 · pp. 10–11
5 Discussion
TIRA improved cross-cancer transfer across three foundation-model representations, while gains varied by encoder and target cohort. The discussion connects topology-guided attention to morphological evidence selection and notes important validation limits.
5 Discussion
- TIRA improved cross-cancer transfer across all three foundation-model representations, with improvement magnitude varying by encoder and target cohort.Direction suppression reduces source-associated variation, MMD aligns projected source distributions, and topology supervision constrains adaptation using spatial immune descriptors screened for RNA concordance.Source: §5 · p. 11
- Reference-choice analysis showed that the effect of direction suppression depends on the cohort used to estimate the suppression direction.Source: §5 · p. 11
- TIRA assigns more attention to tumor–lymphocyte interfaces, while pooling only morphological features, and interface enrichment varies across cancer types and biomarker groups.The passage reports this pattern for TCGA-STAD and TCGA-UCEC.Source: §5 · p. 11
- Reciprocal transfer was source–target dependent: UCEC+READ improved both biomarkers on TCGA-STAD but reduced performance on TCGA-COAD, and the experiments did not isolate why.The paper cautions that TIRA improves robustness across diverse settings rather than producing cancer-invariant representations.Source: §5 · p. 11
- The tissue classifier was trained on colorectal histology, while spatial validation, reference configurations, descriptor sources, and prospective clinical utility remain limited or untested.The paired AUROC-difference confidence interval for UCEC TMB included zero in the primary UNI2 comparison.Source: §5 · p. 11
6 Conclusion
TIRA uses source-derived spatial immune topology to adapt frozen foundation-model representations without target-domain adaptation. The paper concludes that this approach improves robustness under cancer-type shift and guides selection of morphological evidence.
6 Conclusion
- TIRA uses source-derived spatial immune topology to adapt frozen foundation-model representations without target-domain adaptation.Source: §6 · p. 11
- Across three foundation models and multiple transfer settings, TIRA improved the robustness of MSI and TMB prediction under cancer-type shift.Source: §6 · pp. 11–12
- Component ablations and attention analysis indicate that topology-informed representations guide morphological evidence selection, particularly around tumor–immune interfaces.The paper concludes that spatial immune topology provides a source-derived biological constraint for foundation-model transfer across cancer types.Source: §6 · p. 12
1 Spatial Immune Topology Descriptor Definitions
The descriptors are computed per slide from foundation-model-specific tissue-class probabilities and tile coordinates. Twelve screened candidates are retained for topology supervision, with additional candidates defined separately.
1 Spatial Immune Topology Descriptor Definitions
- Descriptors use foundation-model-specific tissue-class probabilities and slide-normalized tile coordinates; 12 of 18 screened candidates are retained for topology supervision.The six other candidates are defined separately.Source: §1 · p. 15
1.1 Lymphocyte and Stromal Abundance (d1–d4)
- The lymphocyte and stromal abundance descriptors include mean lymphocyte confidence.Source: §1.1 · p. 15
1.2 Local Immune Neighbourhood Score (d5)
- The local immune-neighbourhood score uses a small ε constant for numerical stability.Source: §1.2 · p. 16
1.3 Peripheral and Core Lymphocyte Scores (d6–d7)
- Peripheral and core lymphocyte scores use each tile’s normalized radial distance from the slide centroid.Source: §1.3 · p. 16
1.4 Multiscale TIL Density (d8–d10)
- Multiscale TIL density is evaluated using neighbourhood sizes k = 10, 30, and 100.Source: §1.4 · p. 16
1.5 Immune Mixing Entropy (d11)
- Immune mixing entropy measures local intermixing of lymphocyte- and non-lymphocyte-labelled tissue.Source: §1.5 · p. 16
1.6 Stromal TIL Infiltration (d12)
- Stromal TIL infiltration uses a normalized spatial-distance threshold of 0.05 and is evaluated only with more than five stromal tiles and at least one lymphocyte tile.Source: §1 · p. 15, §1.6 · p. 17
1.7 Additional Candidate Descriptors Used in Screening
- The six additional screened candidates include local immune heterogeneity, intra-tumoral TIL density, and DBSCAN-based TLS count and fraction.Source: §1.7 · p. 17
2 Biological Validation: CONCH and Virchow2
The fixed UNI2-derived descriptor set showed biological concordance in CONCH and Virchow2 representations. Immune-related descriptors retained positive concordance with selected immune genes across both models.
2 Biological Validation: CONCH and Virchow2
- Immune mixing, lymphocyte-confidence measures, and multiscale TIL density showed positive concordance with selected immune genes across CONCH and Virchow2.Source: §2 · pp. 17–18
- A descriptor passed the screen when its mean correlation, ρ̄d, was greater than 0.10.Source: §2 · p. 18
- 12 of 18 CONCH descriptors and 11 of 18 Virchow2 descriptors exceeded mean Spearman ρ > 0.10; Stromal TIL infiltration fell below threshold for Virchow2 (ρ̄ = 0.048).The same 12 UNI2-derived descriptors were retained across foundation models.Source: §2 · p. 17, Table S1 · p. 18, Table S2 · p. 18
3 Zero-Shot Probability Calibration
Zero-shot calibration differed between TCGA-STAD and TCGA-UCEC. TIRA was generally closer to perfect calibration on STAD, while UCEC showed mixed comparisons and substantial departures from the diagonal.
3 Zero-Shot Probability Calibration
- The calibration assessment used the same patient-level predictions as the zero-shot evaluation on TCGA-STAD and TCGA-UCEC.Source: §3 · p. 19
- Calibration curves used five quantile-based bins, plotting mean predicted probability against observed event frequency without post-hoc target-cohort recalibration.Source: §3 · p. 19
- On TCGA-STAD, TIRA was generally closer to the perfect-calibration diagonal than matched ABMIL for both MSI and TMB; on UCEC, ABMIL calibrated MSI better while TIRA modestly improved TMB.Both UCEC panels still showed substantial departures from the diagonal.Source: Figure S1 · p. 19, §3 · pp. 19–20
4 Decision-Curve Analysis
Decision-curve analysis found a broader positive-net-benefit range for TIRA than ABMIL on TCGA-STAD, but no consistent advantage on TCGA-UCEC. The comparison spans threshold probabilities from 0.01 to 0.50.
4 Decision-Curve Analysis
- The analysis compared TIRA and matched ABMIL with treat-all and treat-none strategies across threshold probabilities from 0.01 to 0.50.Source: §4 · p. 20, Figure S2 · p. 20
- On TCGA-STAD, TIRA maintained positive net benefit across a broader range of moderate-to-high thresholds than ABMIL for both MSI and TMB.Source: Figure S2 · p. 20, §4 · p. 21
- On TCGA-UCEC, net benefit was more limited and threshold-dependent, with no consistent advantage across the evaluated range.Source: Figure S2 · p. 20, §4 · p. 21
5 Additional Classification Metric
The section reports balanced accuracy for MSI and TMB prediction across three foundation models and evaluation cohorts, applying source-derived decision thresholds unchanged to external cohorts.
5 Additional Classification Metric
- Balanced accuracy is reported for both MSI and TMB prediction across all three foundation models.Source: §5 · p. 21
- Source-derived decision thresholds are applied unchanged when evaluating external cohorts.Source: §5 · p. 21
- Tables S3 and S4 present balanced accuracy across foundation models and evaluation cohorts for MSI and TMB, respectively.Source: Table S3 · p. 21, Table S4 · p. 21