MoSPR: Histology-to-Gene Expression Prediction with Morpho-Spatial Macrostates and Low-Rank Molecular Programs

Dongmyung Shin; Geongyu Lee; Yesung Cho; Park Jong Bae

cs.AISep 28, 2026

MoSPR: Histology-to-Gene Expression Prediction with Morpho-Spatial Macrostates and Low-Rank Molecular Programs

Authors: Dongmyung Shin, Geongyu Lee, Yesung Cho, Park Jong Bae

Organizations: OmixAI Co. Ltd. · Oncocross Co. Ltd. · Kyunghee University

Abstract

Predicting molecular profiles from histopathology remains challenging because whole-slide images contain spatially organized, heterogeneous tissue patterns, while gene expression comprises thousands of correlated targets. We introduce MoSPR (Morpho-Spatial Program Regression), a linear framework that couples an adjacency-informed histology representation with a low-rank molecular basis. MoSPR clusters frozen patch embeddings into morphology microstates, aggregates their spatial adjacencies across the training cohort, and groups microstates with similar adjacency patterns into shared macrostates. Each slide is then represented by global morphology and macrostate-specific deviations, which are linearly mapped to coefficients of a training-derived low-rank gene-expression basis. Across three cancer cohorts from The Cancer Genome Atlas, MoSPR achieves the highest mean gene-expression prediction scores among all evaluated methods. Without pathway-level supervision, pathway scores derived from its predicted expression profiles rank first in eight of nine comparisons across three pathway collections. Ablation studies on the breast cancer cohort show complementary gains from adjacency-derived macrostate representation and low-rank molecular prediction. Moreover, with half of the training data on this cohort, MoSPR exceeds the full-data gene-prediction score of the strongest competing baseline. Finally, its linear formulation enables exact decomposition of each predicted expression profile into global and macrostate-specific molecular contributions, providing an interpretable link between spatially coherent macrostate regions and their associated molecular programs. Our code is available at https://github.com/Radisen-Panthera/MoSPR.

Figures & tables

	BRCA		KIRC		LUAD
Method	PCC	SCC	PCC	SCC	PCC	SCC
Max ( Wang et al., 2018 )	0.294	0.283	0.236	0.244	0.258	0.278
Mean ( Wang et al., 2018 )	0.336	0.338	0.285	0.293	0.292	0.315
ABMIL ( Ilse et al., 2018 )	0.370	0.363	0.293	0.299	0.308	0.324
HE2RNA ( Schmauch et al., 2020 )	0.300	0.293	0.256	0.273	0.276	0.297
AbReg ( Graziani et al., 2022 )	0.355	0.349	0.289	0.293	0.296	0.311

Table 1: Cross-cancer gene-expression prediction. Entries report mean gene-wise PCC and SCC over four cross-validation folds. Best and second-best results are shown in bold and underlined, respectively.

	BRCA			KIRC			LUAD
Method	Hallmark	GO-BP	KEGG	Hallmark	GO-BP	KEGG	Hallmark	GO-BP	KEGG
Max ( Wang et al., 2018 )	0.356	0.337	0.321	0.281	0.284	0.309	0.420	0.379	0.388
Mean ( Wang et al., 2018 )	0.469	0.448	0.415	0.367	0.379	0.415	0.498	0.446	0.452
ABMIL ( Ilse et al., 2018 )	0.500	0.476	0.446	0.378	0.388	0.418	0.519	0.465	0.474
HE2RNA ( Schmauch et al., 2020 )	0.333	0.329	0.300	0.299	0.321	0.340	0.388	0.349	0.365
AbReg ( Graziani et al., 2022 )	0.481	0.453	0.425	0.397	0.389	0.420	0.507	0.443	0.462

Table 2: Gene-derived pathway reconstruction across three TCGA cohorts. Entries report the median pathway-wise SCC within each collection, averaged over four cross-validation folds. Pathway scores are derived solely from predicted gene-expression profiles and are not used for training. Best and second-best results are shown in bold and underlined, respectively.

Variant	Histology Input	Molecular Output	Gene SCC	Hallmark SCC
Global Direct	$M$	Direct	0.348	0.458
Global Low-Rank	$M$	$U_{q}$	0.380	0.481
Shuffled-State Direct	$[M\mid S_{\mathrm{shuffle}}]$	Direct	0.360	0.499
Shuffled-State Low-Rank	$[M\mid S_{\mathrm{shuffle}}]$	$U_{q}$	0.379	0.499
Spatial Direct	$[M\mid S]$	Direct	0.395	0.552
Spatial Low-Rank (MoSPR)	$[M\mid S]$	$U_{q}$	0.411	0.551

Table 3: Component ablation on TCGA-BRCA. Best and second-best results are shown in bold and underlined, respectively.

Figure 2: Training-set scaling on TCGA-BRCA.

Figure 3: Representative macrostates and their associated Hallmark pathway enrichments.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Cohort	Patients	Slides	Genes	Fold	Patients	Slides	Patients	Slides	Patients	Slides
	Cohort Summary				Train		Validation		Test
TCGA-BRCA	1,037	1,467	14,042	0	519	734	259	366	259	367
				1	519	733	259	367	259	367
				2	518	733	259	367	260	367
				3	518	734	260	367	259	366
TCGA-KIRC	342	681	14,295	0	170	340	86	170	86	171

Appendix

Table S1: Dataset summary and fold-specific sample counts used for four-fold cross-validation.

Best-predicted
	BRCA		KIRC		LUAD
Rank	Gene	SCC	Gene	SCC	Gene	SCC
1	PODN	0.781	ACAA2	0.728	TPX2	0.750
2	HTRA1	0.762	LINC01507	0.719	PRR11	0.742
3	ZCCHC24	0.759	EMX2OS	0.715	CENPA	0.736
4	COL14A1	0.759	LDB2	0.715	CCNA2	0.736

Appendix

Table S2: Per-gene MoSPR performance on TCGA-BRCA, TCGA-KIRC, and TCGA-LUAD. Shown are the 10 best- and worst-predicted genes ranked by mean SCC across four cross-validation folds.

Method	# Params	GFLOPs
Max	7.47M	5.14
Mean	7.47M	5.14
ABMIL	7.86M	12.84
HE2RNA	15.97M	22.40
AbReg	7.66M	149.86
tRNAformer	5.22M	1.86

Appendix

Table S3: Model size and inference complexity on TCGA-BRCA after frozen CONCH feature extraction. Parameter counts exclude the shared CONCH encoder and buffers such as standardization statistics. The MoSPR count includes the microstate centroids and the PCA basis, which are fixed during training and required at inference. GFLOPs are computed using the cohort median of

9{,}609

patches for full-bag methods; HE2RNA, tRNAformer, and SEQUOIA VIS use their architecture-specific clustered inputs of 100, 49, and 100 super-patches, respectively.

Cohort	Model	Full-Rank	Low-Rank	Post-Hoc
TCGA-BRCA	ABMIL	0.363	0.354	0.380
	tRNAformer	0.332	0.327	0.338
	MoSPR (ours)	0.411
TCGA-KIRC	ABMIL	0.299	0.322	0.314
	tRNAformer	0.319	0.308	0.324
	MoSPR (ours)	0.348

Appendix

Table S4: Low-rank output controls for ABMIL and tRNAformer. Values report mean gene SCC over four cross-validation folds. Low-Rank fits each baseline in its own validation-selected

q

-dimensional PCA target space, whereas Post-Hoc projects the full-rank predictions onto the same subspace (

U_{q}^{\top}U_{q}

) without retraining. The rank

q

is selected independently for each baseline and fold; MoSPR is shown for reference.

Cohort	Patients	Mean ARI	Median ARI
TCGA-BRCA	1,037	0.769	0.801
TCGA-KIRC	342	0.653	0.646
TCGA-LUAD	482	0.786	0.798

Appendix

Table S5: Macrostate partition stability under patient-level bootstrap resampling. Microstates were fixed, and 300 patient-level bootstrap resamples were generated for each cancer cohort. For each resample, the global microstate adjacency matrix was reconstructed and re-clustered into

K=8

macrostates. Partition stability was quantified using the Adjusted Rand Index (ARI; Hubert and Arabie, 1985 ) between the bootstrap-derived macrostate partition and the reference partition obtained from the full cohort.

Figure S1: Training-set scaling on TCGA-KIRC. Gene-expression prediction performance (left) and Hallmark pathway reconstruction performance (right) are shown across increasing fractions of the training data.

Figure S2: Training-set scaling on TCGA-LUAD. Gene-expression prediction performance (left) and Hallmark pathway reconstruction performance (right) are shown across increasing fractions of the training data.

Explore similar work

HistoGPA: A Context-Conditioned Gene-Prior Attention Framework for Histology-Based Spatial Gene Expression Prediction
Jul 27, 2026Ziang Liu, Xinhai Chen, Yigui Feng +3Spatial Transcriptomics Digital Pathology
Laboratory of Digitizing Software for Frontier Equipment, National University of Defense Technology · National Key Laboratory of Parallel and Distributed Computing, National University of Defense Technology · College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, China
Data-Efficient Multimodal Alignment for Histopathology-based Molecular Prediction
Jun 29, 2026Dominik Winter, Dominik Vonficht, Loïc Le Bescond +6Digital Pathology Biomarker
AstraZeneca Computational Pathology and Biomarkers, Munich, Germany · AstraZeneca, Early Oncology and Translation Medicine, Cambridge, UK
CHRep: Cross-modal Histology Representation and Post-hoc Calibration for Spatial Gene Expression Prediction
Apr 23, 2026Changfan Wang, Xinran Wang, Donghai Liu +4Spatial Transcriptomics Digital Pathology
Beijing University of Posts and Telecommunications, Beijing 100876, China, and also with the Beijing Key Laboratory of Network System and Network Culture, Beijing 100876, China · Peking University Third Hospital, Beijing 100191, China

Jul 27, 2026, cs.CVJ/K move · Enter open · S save

HistoGPA: A Context-Conditioned Gene-Prior Attention Framework for Histology-Based Spatial Gene Expression Prediction

Ziang Liu, Xinhai Chen, Yigui Feng, Shuai Li+2

Laboratory of Digitizing Software for Frontier Equipment, National University of Defense Technology · National Key Laboratory of Parallel and Distributed Computing, National University of Defense Technology · College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, China

Spatial Transcriptomics Digital Pathology Tissue

Open paper AskPDF arXiv

Predicting spatial gene expression from routine hematoxylin and eosin (H&E) images provides a practical complement to experimental spatial transcriptomics. Existing approaches focus on local or multi-scale visual features and often treat pretrained gene representations as fixed priors, although the interpretation of local morphology and the relevance of gene priors depend on tissue context. We propose HistoGPA, a context-conditioned gene-prior attention framework that uses a shared slide-level representation in two parallel pathways: one modulates local morphological features, whereas the other conditions pretrained gene embeddings and retrieves gene-prior information through cross-attention. This design enables each spatial location to retrieve context-adapted gene-prior information using its local morphology, position, and slide context. Across ten cancer types in HEST-1k, HistoGPA achieves the highest macro-averaged gene-wise Pearson correlation coefficient among the compared methods under the same evaluation protocol for both the top-50 and top-1,500 highly variable gene sets. Additional analyses show that HistoGPA better recovers the spatial expression patterns of cancer-associated genes and yields greater agreement between clusters derived independently from predicted and ground-truth expression profiles. Together, these findings motivate a context-dependent view of histology-to-expression prediction, in which local morphological representations and gene priors are jointly adapted to the broader tissue context.

Abstract

Explore similar work

HistoGPA: A Context-Conditioned Gene-Prior Attention Framework for Histology-Based Spatial Gene Expression Prediction

Data-Efficient Multimodal Alignment for Histopathology-based Molecular Prediction

CHRep: Cross-modal Histology Representation and Post-hoc Calibration for Spatial Gene Expression Prediction