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

Dasari Naga Raju

cs.CVOct 7, 2026

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

Authors: Dasari Naga Raju

Organizations: Independent Researcher, Andhra Pradesh, India

Abstract

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.

Figures & tables

Figure 1: Overview of TIRA . Stage 1 learns a topology-supervised biology representation from source data, and Stage 2 uses this representation to condition tile attention while pooling only morphological features for joint MSI/TMB prediction.

	RNA immune-gene correlation
Descriptor	CD8A	CD3E	FOXP3	PRF1	PDCD1	Mean $\rho$	Use
Lymphocyte abundance	0.161	0.188	0.064	0.013	0.164	0.118	$\checkmark$
Tumor abundance	-0.168	-0.200	-0.180	-0.174	-0.110	-0.166	$\times$
Stromal abundance	0.171	0.146	0.192	0.146	0.124	0.156	$\checkmark$
Mean lymphocyte conf.	0.173	0.201	0.073	0.027	0.175	0.130	$\checkmark$
Lymphocyte variability	0.143	0.177	0.053	-0.005	0.151	0.104	$\checkmark$

Table 1: Source-cohort RNA-concordance screening of 18 spatial immune descriptors in TCGA-COAD (

n=227

). Columns report Spearman

\rho

with five immune genes and their mean for UNI2.

		Source CV	Cross-site	Cross-cancer		Cross-cancer + site
FM	Method	COAD+READ ( $n$ =443, MSI-H=64)	CPTAC-COAD ( $n$ =75, MSI-H=15)	TCGA-STAD ( $n$ =308, MSI-H=54)	TCGA-UCEC ( $n$ =297, MSI-H=105)	CPTAC-UCEC ( $n$ =95, MSI-H=25)
UNI2	ABMIL	0.8778 (0.8244–0.9244)	0.8022 (0.6628–0.9184)	0.6333 (0.5458–0.7210)	0.5153 (0.4437–0.5870)	0.4314 (0.2868–0.5822)
	CLAM-SB	0.8715 (0.8136–0.9232)	0.7089 (0.5675–0.8364)	0.6490 (0.5628–0.7326)	0.5061 (0.4358–0.5771)	0.4829 (0.3414–0.6260)
	TransMIL	0.8861 (0.8369–0.9289)	0.7889 (0.6487–0.9100)	0.6395 (0.5489–0.7281)	0.5215 (0.4486–0.5937)	0.5069 (0.3721–0.6401)
	CasNet-FM	0.8543 (0.7871–0.9151)	0.7489 (0.6094–0.8712)	0.6492 (0.5632–0.7340)	0.5116 (0.4427–0.5823)	0.4543 (0.3161–0.5948)
	ILRA	0.8361 (0.7675–0.8975)	0.4489 (0.2804–0.6195)	0.6457 (0.5550–0.7324)	0.4887 (0.4187–0.5605)	0.5023 (0.3673–0.6370)

Table 2: Patient-level MSI prediction AUROC (95% bootstrap CI) for models trained on TCGA-COAD+READ and evaluated zero-shot on external cohorts spanning cross-site, cross-cancer, and combined shift.

		Source CV	Cross-site	Cross-cancer		Cross-cancer + site
FM	Method	COAD+READ ( $n$ =428, TMB-H=62)	CPTAC-COAD ( $n$ =75, TMB-H=15)	TCGA-STAD ( $n$ =308, TMB-H=56)	TCGA-UCEC ( $n$ =213, TMB-H=36)	CPTAC-UCEC ( $n$ =95, TMB-H=32)
UNI2	ABMIL	0.9180 (0.8721–0.9552)	0.8244 (0.6862–0.9385)	0.6509 (0.5658–0.7352)	0.5279 (0.4275–0.6275)	0.4807 (0.3485–0.6097)
	CLAM-SB	0.9040 (0.8604–0.9414)	0.7900 (0.6413–0.9176)	0.6503 (0.5623–0.7365)	0.5367 (0.4389–0.6347)	0.5069 (0.3750–0.6371)
	TransMIL	0.8832 (0.8329–0.9279)	0.8300 (0.7078–0.9289)	0.6451 (0.5558–0.7334)	0.5323 (0.4282–0.6347)	0.5184 (0.3883–0.6462)
	CasNet-FM	0.9149 (0.8636–0.9577)	0.8200 (0.6902–0.9272)	0.6698 (0.5865–0.7509)	0.5430 (0.4436–0.6419)	0.4742 (0.3421–0.6057)
	ILRA	0.8623 (0.8032–0.9138)	0.5189 (0.3561–0.6799)	0.6787 (0.5955–0.7580)	0.4735 (0.3725–0.5741)	0.4747 (0.3486–0.6035)

Table 3: Patient-level TMB prediction AUROC (95% bootstrap CI) for models trained on TCGA-COAD+READ and evaluated zero-shot on external cohorts spanning cross-site, cross-cancer, and combined shift.

Source	Target	Shift	ABMIL	TIRA	$\Delta$ AUROC [95% CI]	ABMIL	TIRA	$\Delta$ AUROC [95% CI]
			MSI			TMB
COAD+READ	TCGA-STAD	Cross-cancer	0.633	0.766	+0.133 [0.076, 0.193]	0.651	0.772	+0.121 [0.063, 0.182]
COAD+READ	TCGA-UCEC	Cross-cancer	0.515	0.595	+0.080 [0.029, 0.131]	0.528	0.587	+0.059 [ $-$ 0.026, 0.143]
COAD+READ	CPTAC-UCEC	Cross-cancer + site	0.431	0.561	+0.130 [ $-$ 0.032, 0.297]	0.481	0.532	+0.051 [ $-$ 0.081, 0.185]
STAD+READ	TCGA-COAD	Cross-cancer	0.786	0.859	+0.073 [0.015, 0.137]	0.805	0.889	+0.084 [0.028, 0.145]
STAD+READ	TCGA-UCEC	Cross-cancer	0.519	0.678	+0.159 [0.094, 0.226]	0.550	0.646	+0.096 [ $-$ 0.014, 0.204]

Table 4: Reciprocal robustness analysis under UNI2: patient-level AUROC for matched ABMIL and TIRA with paired 95% CIs. Descriptor identities are fixed from the primary TCGA-COAD RNA-concordance screening across all configurations.

	Source CV		Cross-site		Cross-cancer				Cross-cancer + site		Gen. Drop $\downarrow$		CC Drop from Full
Model	COAD+READ		CPTAC-COAD		TCGA-STAD		TCGA-UCEC		CPTAC-UCEC		Source $\rightarrow$ CC		Mean STAD/UCEC
	MSI	TMB	MSI	TMB	MSI	TMB	MSI	TMB	MSI	TMB	MSI	TMB	MSI	TMB
ABMIL	0.878	0.918	0.802	0.824	0.633	0.651	0.515	0.528	0.431	0.481	0.304	0.329	0.106	0.090
TIRA Full	0.893	0.922	0.793	0.843	0.766	0.772	0.595	0.587	0.561	0.532	0.212	0.242	–	–
$-$ Biology-Guided Attention	0.876	0.913	0.782	0.802	0.665	0.673	0.526	0.525	0.533	0.554	0.281	0.314	0.085	0.081
$-$ Direction Suppression	0.891	0.918	0.823	0.841	0.682	0.690	0.520	0.540	0.498	0.523	0.290	0.303	0.080	0.065

Table 5: Component ablation of TIRA (UNI2, COAD+READ source): patient-level AUROC across source and zero-shot cohorts, with CC Drop from Full denoting the reduction in mean STAD/UCEC AUROC relative to full TIRA.

Dataset	Task	No Supp.	Random	READ	$\Delta$ READ–Random
TCGA-STAD	MSI	0.682	0.712	0.766	+0.054 [+0.010,+0.100]
TCGA-STAD	TMB	0.690	0.724	0.772	+0.048 [+0.000,+0.090]
TCGA-UCEC	MSI	0.520	0.553	0.595	+0.042 [+0.001,+0.055]
TCGA-UCEC	TMB	0.540	0.521	0.587	+0.066 [+0.001,+0.145]

Table 6: Reference-choice ablation under UNI2. READ is compared with matched random-COAD and no-suppression conditions;

\Delta

and 95% CIs report READ versus random-COAD using 10,000 paired bootstrap resamples.

Figure 2: Representative UNI2 attention maps for MSI-H and TMB-H cases from TCGA-STAD and TCGA-UCEC. Tiles are colored by $\log_{10}$ attention weight; cyan outlines denote tumor–lymphocyte interface tiles.

Cohort	Group	N	ABMIL	TIRA	$\boldsymbol{\Delta}$	95% CI
TCGA-STAD	Overall	308	3.34	6.09	+2.75	[+1.94, +3.72]
TCGA-STAD	MSI-H	54	2.03	2.39	+0.36	[-0.21, +0.82]
TCGA-STAD	TMB-H	56	2.02	2.35	+0.34	[-0.21, +0.78]
TCGA-UCEC	Overall	297	3.19	5.03	+1.84	[+0.84, +3.54]
TCGA-UCEC	MSI-H	105	3.34	6.07	+2.73	[+0.43, +6.96]
TCGA-UCEC	TMB-H	36	4.00	4.74	+0.74	[-0.001, +1.46]

Table 7: Tumor–lymphocyte interface attention enrichment on UNI2 zero-shot targets; 10,000 paired bootstrap resamples.

Cohort	Task	Model	$n$	ECE	Brier
TCGA-STAD	MSI	ABMIL	308	0.179	0.184
	MSI	TIRA	308	0.063	0.134
	TMB	ABMIL	308	0.213	0.199
	TMB	TIRA	308	0.079	0.140
TCGA-UCEC	MSI	ABMIL	297	0.265	0.303
	MSI	TIRA	297	0.281	0.318

Table 8: Raw probability calibration for TIRA and ABMIL on UNI2 zero-shot target cohorts.

Descriptor	CD8A	CD3E	FOXP3	PRF1	PDCD1	Mean $\rho$	Pass
Lymphocyte abundance	0.184	0.234	0.136	0.113	0.249	0.183	$\checkmark$
Tumor abundance	-0.049	-0.065	-0.098	-0.024	0.004	-0.047	$\times$
Stromal abundance	0.063	0.053	0.119	0.049	0.028	0.062	$\times$
Mean lymphocyte conf.	0.188	0.240	0.148	0.119	0.257	0.190	$\checkmark$
Lymphocyte variability	0.183	0.232	0.122	0.098	0.241	0.175	$\checkmark$
Local immune nbhd.	0.098	0.129	0.106	0.024	0.082	0.088	$\times$

Table S1: RNA-concordance for CONCH in TCGA-COAD (

n=227

Descriptor	CD8A	CD3E	FOXP3	PRF1	PDCD1	Mean $\rho$	Pass
Lymphocyte abundance	0.181	0.185	0.062	0.065	0.186	0.136	$\checkmark$
Tumor abundance	-0.101	-0.132	-0.168	-0.055	-0.038	-0.099	$\times$
Stromal abundance	0.015	-0.010	0.050	-0.008	0.027	0.015	$\times$
Mean lymphocyte conf.	0.205	0.202	0.074	0.076	0.205	0.153	$\checkmark$
Lymphocyte variability	0.160	0.170	0.040	0.036	0.168	0.115	$\checkmark$
Local immune nbhd.	0.181	0.208	0.157	0.068	0.123	0.147	$\checkmark$

Table S2: RNA-concordance for Virchow2 in TCGA-COAD (

n=227

Figure S1: Zero-shot calibration of ABMIL and TIRA under UNI2 on TCGA-STAD and TCGA-UCEC. Curves use five quantile-based bins; the dashed diagonal denotes perfect calibration.

Figure S2: Decision-curve analysis of ABMIL and TIRA under UNI2 on TCGA-STAD and TCGA-UCEC for MSI and TMB. Net benefit is shown across threshold probabilities from 0.01 to 0.50, with treat-all and treat-none strategies included as references.

FM	Model	COAD+READ	CPTAC-COAD	TCGA-STAD	TCGA-UCEC	CPTAC-UCEC
UNI2	ABMIL	0.823	0.500	0.566	0.496	0.477
	CLAM-SB	0.833	0.492	0.582	0.513	0.511
	TransMIL	0.814	0.608	0.566	0.514	0.543
	CasNet-FM	0.844	0.617	0.544	0.510	0.489
	ILRA	0.798	0.467	0.598	0.512	0.501
	TIRA (Ours)	0.841	0.742	0.647	0.531	0.540

Table S3: Balanced accuracy for MSI prediction across foundation models and evaluation cohorts.

FM	Model	COAD+READ	CPTAC-COAD	TCGA-STAD	TCGA-UCEC	CPTAC-UCEC
UNI2	ABMIL	0.866	0.625	0.557	0.473	0.501
	CLAM-SB	0.833	0.742	0.571	0.531	0.485
	TransMIL	0.815	0.708	0.565	0.543	0.494
	CasNet-FM	0.881	0.558	0.574	0.468	0.478
	ILRA	0.813	0.550	0.635	0.457	0.492
	TIRA (Ours)	0.876	0.734	0.640	0.538	0.499

Table S4: Balanced accuracy for TMB prediction across foundation models and evaluation cohorts.

Explore similar work

Biological Spatial Priors Regularize Foundation Model Representations for Cross-Site MSI Generalization in Colorectal Cancer
May 4, 2026Dasari Naga RajuDigital Pathology Multiple Instance Learning
Probing, Fusion, and Trustworthiness: A Systematic Evaluation of Foundation Model Representations for Multimodal Cancer Analysis
Jun 15, 2026Jingyu Hu, Giuseppe Tripodi, Reed Naidoo +2Pathology Foundation Models Multimodal Clinical Data
The Alan Turing Institute, London, United Kingdom · University of Manchester, Manchester, United Kingdom · The Institute of Cancer Research, London, United Kingdom +1
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 Biomarker
Program of Computational Biology and Bioinforamtics, Yale University, USA · Broad Institute of MIT and Harvard, USA · These authors contributed equally to this work. +12

May 4, 2026, eess.IVJ/K move · Enter open · S save

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

Dasari Naga Raju

Digital Pathology Multiple Instance Learning Geospatial Foundation Models Colorectal Cancer Spatial Priors

Open paper AskPDF arXiv

Predicting microsatellite instability (MSI) status from routine hematoxylin and eosin (H&E) whole slide images (WSIs) offers a practical alternative to molecular testing, but models trained at one institution tend to generalize poorly to slides acquired at a different site. Foundation model representations, despite their generality, still encode site-specific texture alongside the conserved biological morphology underlying MSI. We investigate whether tile-level spatial priors derived from known MSI histology can guide these representations toward more site-invariant features. We introduce a biologically motivated spatial prior based on peripheral distance encoding, reflecting the Crohn's-like peripheral lymphocytic reaction at the tumor invasive margin, and evaluate a secondary local immune neighborhood encoding reflecting the lymphocyte-to-tumor ratio in each tile's immediate spatial neighborhood. Both priors are injected into a TransMIL aggregator before self-attention, allowing the transformer to integrate spatial biological context with UNI2-h or Virchow2 features across all attention layers. We evaluate six foundation model and MIL aggregator combinations as a reference, then assess the effect of each spatial prior. Training on TCGA-COAD (137 slides) and evaluating externally on TCGA-READ (50 slides) without retraining, peripheral distance encoding achieves MSI AUC 0.959 +/- 0.012 on COAD and MSS specificity 1.000 on READ, compared to 0.957 and 0.939 for the strongest reference configuration. Local immune neighborhood encoding achieves comparable internal AUC but lower cross-site specificity, suggesting margin proximity encodes a more site-invariant biological signal than local immune density. Results suggest biologically grounded spatial priors act as regularizers that reduce reliance on site-specific imaging patterns.

Abstract

Explore similar work

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

Probing, Fusion, and Trustworthiness: A Systematic Evaluation of Foundation Model Representations for Multimodal Cancer Analysis

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