Conserved Immune Topology Improves Pathology Foundation Model Generalization for Cross-Cancer MSI-H Prediction

Dasari Naga Raju

cs.CVSep 4, 2026

Conserved Immune Topology Improves Pathology Foundation Model Generalization for Cross-Cancer MSI-H Prediction

Authors: Dasari Naga Raju

Organizations: Independent Researcher, Andhra Pradesh, India

Abstract

Pathology foundation models integrated with multiple instance learning achieve competitive accuracy within single-cancer cohorts, yet cross-cancer generalization remains unresolved due to organ-specific histological and architectural differences. In this paper, we propose Conserved Immune Topology (CIT), a lightweight spatial representation for cross-cancer MSI-H prediction that augments foundation-model embeddings with biologically motivated immune descriptors. CIT uses unsupervised clustering to identify immune-associated tiles, then encodes tertiary lymphoid structures, peritumoral immune reactions, multi-scale tumor-infiltrating lymphocyte density, and immune-tumor mixing from frozen foundation-model embeddings and tile coordinates without requiring annotations or target-domain data. The proposed method was evaluated under cross-site and cross-cancer settings using CPTAC-COAD and TCGA-STAD cohorts, which introduce scanner variability, distribution shifts, and organ-specific architectural variations. Zero-shot cross-cancer transfer with CIT increased TransMIL AUC from 0.6627 to 0.7161, an absolute gain of 0.0534 (p=0.003), with consistent improvements across all three MIL aggregators. These results suggest that spatial immune topology provides potentially an organ-invariant representation for MSI-H prediction, supporting cross-cancer generalization of pathology foundation models.

Figures & tables

Figure 1: Overview of the proposed framework. WSI patches are encoded via UNI2-h and concatenated with a CIT descriptor.

Dataset	Patients	MSI-H	MSS	MSI-H (%)
TCGA-COAD	301	54	247	17.9
TCGA-STAD	308	54	254	17.5
CPTAC-COAD	105	24	81	22.9

Table 1: Patient counts and MSI-H/MSS class distribution for study cohorts.

Model	MSI-H AUC
CONCH + ABMIL	$0.9145\pm 0.0499$
CONCH + CLAM-SB	$0.9181\pm 0.0563$
CONCH + TransMIL	$0.9198\pm 0.0615$
UNI2-h + ABMIL	$0.9337\pm 0.0166$
UNI2-h + ABMIL + CIT	$0.9348\pm 0.0197$
UNI2-h + CLAM-SB	$0.9469\pm 0.0233$

Table 2: Internal TCGA-COAD performance (

301

patients,

54

MSI-H /

247

MSS; MSI-H AUC, mean

\pm

SD over five folds).

Model	Baseline	CIT
ABMIL	$\mathbf{0.8089}$	0.8086
CLAM-SB	0.7989	$\mathbf{0.8344}$
TransMIL	0.7870	$\mathbf{0.8277}$

Table 3: Cross-site generalization on the independent CPTAC-COAD cohort. CIT improves CLAM-SB and TransMIL despite institution and scanner shift, while ABMIL is unchanged.

Model	Baseline	CIT	$\Delta$ AUC (98.33% CI)	DeLong $p$
ABMIL	0.5681	$\mathbf{0.6313}$	$+0.0632$ ( $+0.0280$ to $+0.0984$ )	$<0.001$
CLAM-SB	0.6242	$\mathbf{0.6414}$	$+0.0172$ ( $-0.0025$ to $+0.0369$ )	$0.037$
TransMIL	0.6627	$\mathbf{0.7161}$	$+0.0534$ ( $+0.0111$ to $+0.0957$ )	$0.003$

Table 4: Zero-shot cross-cancer evaluation on TCGA-STAD (

308

patients;

54

MSI-H /

254

MSS), models trained on TCGA-COAD.

\Delta

AUC with 98.33% DeLong CIs (Bonferroni-adjusted,

\alpha=0.0167

Figure 2: ROC curves for TransMIL with and without CIT on TCGA-STAD. CIT improves the true positive rate across most operating points.

Model	Baseline	CIT	$\Delta$
ABMIL	$0.7066$	$\mathbf{0.7228}$	$+0.016$
CLAM-SB	$0.7490$	$\mathbf{0.7547}$	$+0.006$
TransMIL	$0.7470$	$\mathbf{0.7878}$	$+0.041$

Table 5: Reverse-direction zero-shot transfer, models trained on TCGA-STAD and tested on TCGA-COAD (

301

patients;

54

MSI-H /

247

MSS). CIT improves MSI-H AUC across all aggregators, with the largest gain for TransMIL (

+0.041

Configuration	COAD AUC	STAD AUC	$\Delta$ STAD
Baseline (no CIT)	$0.9398\pm 0.0274$	$0.6627$	—
$-$ G1 (TLS)	$0.9559\pm 0.0115$	$0.7115$	$-0.0046$
$-$ G2 (margin)	$0.9554\pm 0.0117$	$0.7085$	$-0.0076$
$-$ G3 (TIL density)	$0.9562\pm 0.0114$	$0.7134$	$-0.0027$
$-$ G4 (mixing)	$0.9547\pm 0.0119$	$0.7050$	$-0.0111$
Full CIT (all 4 groups)	$\mathbf{0.9567\pm 0.0112}$	$\mathbf{0.7161}$	—

Table 6: Leave-one-group-out ablation (UNI2-h

+

TransMIL). Internal TCGA-COAD AUC (mean

\pm

SD) and zero-shot TCGA-STAD AUC with

\Delta

STAD relative to full CIT.

Explore similar work

TIRA: Tumor Immune Representation Adaptation for Zero-Shot Cross-Cancer MSI and TMB Prediction
Oct 7, 2026Dasari Naga RajuClinical Prediction Domain Generalization
Independent Researcher, Andhra Pradesh, India
Biological Spatial Priors Regularize Foundation Model Representations for Cross-Site MSI Generalization in Colorectal Cancer
May 4, 2026Dasari Naga RajuSpatial Generalization Pathology Foundation Models
Predicting Immune Biomarkers with MultiModal Mixture-of-Expert Pathology Foundation Models Empowers Precision Oncology
Jun 16, 2026Tianyu Liu, Ziqing Wang, Zhaokang Liang +12Pathology Foundation Models Clinical Prediction
Program of Computational Biology and Bioinforamtics, Yale University, USA · Broad Institute of MIT and Harvard, USA · These authors contributed equally to this work. +12

Oct 7, 2026J/K move · Enter open · S save

TIRA: Tumor Immune Representation Adaptation for Zero-Shot Cross-Cancer MSI and TMB Prediction

Dasari Naga Raju

Independent Researcher, Andhra Pradesh, India

Clinical Prediction Domain Generalization Vision Foundation Model Adaptation Topological Representation Learning

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

TIRA: Tumor Immune Representation Adaptation for Zero-Shot Cross-Cancer MSI and TMB Prediction

Biological Spatial Priors Regularize Foundation Model Representations for Cross-Site MSI Generalization in Colorectal Cancer

Predicting Immune Biomarkers with MultiModal Mixture-of-Expert Pathology Foundation Models Empowers Precision Oncology